Merge pull request 'Add structured learnings updates' (#35) from structured-learnings into main

Reviewed-on: #35
Reviewed-by: SG Command <sgcommand@warky.dev>

.gitea/workflows/ci.yml

@@ -0,0 +1,44 @@
+name: CI
+
+on:
+  push:
+    branches: ['**']
+    tags-ignore: ['v*']
+  pull_request:
+    branches: ['**']
+
+jobs:
+  build-and-test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Go
+        uses: actions/setup-go@v5
+        with:
+          go-version: '1.26'
+
+      - name: Cache Go modules
+        uses: actions/cache@v4
+        with:
+          path: |
+            ~/.cache/go-build
+            ~/go/pkg/mod
+          key: ${{ runner.os }}-go-${{ hashFiles('**/go.sum') }}
+          restore-keys: |
+            ${{ runner.os }}-go-
+
+      - name: Download dependencies
+        run: go mod download
+
+      - name: Tidy modules
+        run: go mod tidy
+
+      - name: Run tests
+        run: go test ./...
+
+      - name: Build amcs-server
+        run: go build -o /dev/null ./cmd/amcs-server
+
+      - name: Build amcs-cli
+        run: go build -o /dev/null ./cmd/amcs-cli

.gitea/workflows/release.yml

@@ -0,0 +1,122 @@
+name: Release
+
+on:
+  push:
+    tags:
+      - 'v*.*.*'
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: 'Tag to release (e.g. v1.2.3)'
+        required: true
+
+env:
+  GITEA_SERVER: https://git.warky.dev
+  GITEA_REPO: wdevs/amcs
+
+jobs:
+  release:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Set up Go
+        uses: actions/setup-go@v5
+        with:
+          go-version: '1.26'
+
+      - name: Cache Go modules
+        uses: actions/cache@v4
+        with:
+          path: |
+            ~/.cache/go-build
+            ~/go/pkg/mod
+          key: ${{ runner.os }}-go-${{ hashFiles('**/go.sum') }}
+          restore-keys: |
+            ${{ runner.os }}-go-
+
+      - name: Download dependencies
+        run: go mod download
+
+      - name: Set up Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 'lts/*'
+
+      - name: Install pnpm
+        run: npm install -g pnpm
+
+      - name: Build UI
+        run: |
+          cd ui
+          pnpm install --frozen-lockfile
+          pnpm run build
+
+      - name: Set build vars
+        id: vars
+        run: |
+          TAG="${{ github.event.inputs.tag || github.ref_name }}"
+          echo "VERSION=${TAG}" >> $GITHUB_OUTPUT
+          echo "COMMIT=$(git rev-parse --short HEAD)" >> $GITHUB_OUTPUT
+          echo "BUILD_DATE=$(date -u +%Y-%m-%dT%H:%M:%SZ)" >> $GITHUB_OUTPUT
+
+      - name: Build release binaries
+        run: |
+          VERSION="${{ steps.vars.outputs.VERSION }}"
+          COMMIT="${{ steps.vars.outputs.COMMIT }}"
+          BUILD_DATE="${{ steps.vars.outputs.BUILD_DATE }}"
+          LDFLAGS="-s -w -X git.warky.dev/wdevs/amcs/internal/buildinfo.Version=${VERSION} -X git.warky.dev/wdevs/amcs/internal/buildinfo.TagName=${VERSION} -X git.warky.dev/wdevs/amcs/internal/buildinfo.Commit=${COMMIT} -X git.warky.dev/wdevs/amcs/internal/buildinfo.BuildDate=${BUILD_DATE}"
+          mkdir -p dist
+          for BINARY in amcs-server amcs-cli; do
+            CMD="./cmd/${BINARY}"
+            for PLATFORM in linux/amd64 linux/arm64 darwin/amd64 darwin/arm64 windows/amd64; do
+              OS="${PLATFORM%/*}"
+              ARCH="${PLATFORM#*/}"
+              EXT=""
+              [ "$OS" = "windows" ] && EXT=".exe"
+              OUTPUT="dist/${BINARY}-${OS}-${ARCH}${EXT}"
+              echo "Building ${OUTPUT}..."
+              GOOS=$OS GOARCH=$ARCH go build -ldflags "${LDFLAGS}" -o "${OUTPUT}" "${CMD}"
+            done
+          done
+          cd dist && sha256sum * > checksums.txt && cd ..
+
+      - name: Create Gitea Release
+        id: create_release
+        run: |
+          export VERSION="${{ steps.vars.outputs.VERSION }}"
+          BODY=$(python3 <<'PY'
+          import json, subprocess, os
+          version = os.environ['VERSION']
+          commit = subprocess.check_output(['git', 'rev-parse', '--short', 'HEAD'], text=True).strip()
+          body = f"## {version}\n\nBuilt from commit {commit}.\n\nSee `checksums.txt` to verify downloads."
+          print(json.dumps({
+              'tag_name': version,
+              'name': version,
+              'body': body,
+              'draft': False,
+              'prerelease': False,
+          }))
+          PY
+          )
+          RESPONSE=$(curl -fsS -X POST "${{ env.GITEA_SERVER }}/api/v1/repos/${{ env.GITEA_REPO }}/releases" \
+            -H "Authorization: token ${{ secrets.GITHUB_TOKEN }}" \
+            -H "Content-Type: application/json" \
+            -d "$BODY")
+          RELEASE_ID=$(printf '%s' "$RESPONSE" | python3 -c 'import sys,json; print(json.load(sys.stdin)["id"])')
+          echo "RELEASE_ID=${RELEASE_ID}" >> $GITHUB_OUTPUT
+
+      - name: Upload release assets
+        run: |
+          RELEASE_ID="${{ steps.create_release.outputs.RELEASE_ID }}"
+          for f in dist/*; do
+            name=$(basename "$f")
+            echo "Uploading ${name}..."
+            curl -fsS -X POST \
+              "${{ env.GITEA_SERVER }}/api/v1/repos/${{ env.GITEA_REPO }}/releases/${RELEASE_ID}/assets?name=${name}" \
+              -H "Authorization: token ${{ secrets.GITHUB_TOKEN }}" \
+              -H "Content-Type: application/octet-stream" \
+              --data-binary @"${f}"
+          done

.gitignore

@@ -31,3 +31,7 @@ cmd/amcs-server/__debug_*
 bin/
 .cache/
 OB1/
+ui/node_modules/
+ui/.svelte-kit/
+internal/app/ui/dist/
+.codex

Dockerfile

@@ -1,3 +1,14 @@
+FROM node:22-bookworm AS ui-builder
+
+RUN npm install -g pnpm
+WORKDIR /src/ui
+
+COPY ui/package.json ui/pnpm-lock.yaml ./
+RUN pnpm install --frozen-lockfile
+
+COPY ui/ ./
+RUN pnpm run build
+
 FROM golang:1.26.1-bookworm AS builder
 
 WORKDIR /src
@@ -6,6 +17,7 @@ COPY go.mod go.sum ./
 RUN go mod download
 
 COPY . .
+COPY --from=ui-builder /src/internal/app/ui/dist ./internal/app/ui/dist
 
 RUN set -eu; \
     VERSION_TAG="$(git describe --tags --exact-match 2>/dev/null || echo dev)"; \
@@ -17,7 +29,14 @@ RUN set -eu; \
         -X git.warky.dev/wdevs/amcs/internal/buildinfo.TagName=${VERSION_TAG} \
         -X git.warky.dev/wdevs/amcs/internal/buildinfo.Commit=${COMMIT_SHA} \
         -X git.warky.dev/wdevs/amcs/internal/buildinfo.BuildDate=${BUILD_DATE}" \
-      -o /out/amcs-server ./cmd/amcs-server
+      -o /out/amcs-server ./cmd/amcs-server; \
+    CGO_ENABLED=0 GOOS=linux go build -trimpath \
+      -ldflags="-s -w \
+        -X git.warky.dev/wdevs/amcs/internal/buildinfo.Version=${VERSION_TAG} \
+        -X git.warky.dev/wdevs/amcs/internal/buildinfo.TagName=${VERSION_TAG} \
+        -X git.warky.dev/wdevs/amcs/internal/buildinfo.Commit=${COMMIT_SHA} \
+        -X git.warky.dev/wdevs/amcs/internal/buildinfo.BuildDate=${BUILD_DATE}" \
+      -o /out/amcs-migrate-config ./cmd/amcs-migrate-config
 
 FROM debian:bookworm-slim
 
@@ -29,6 +48,7 @@ RUN apt-get update \
 WORKDIR /app
 
 COPY --from=builder /out/amcs-server /app/amcs-server
+COPY --from=builder /out/amcs-migrate-config /app/amcs-migrate-config
 COPY --chown=appuser:appuser configs /app/configs
 
 USER appuser

Makefile

@@ -3,25 +3,43 @@ GO_CACHE_DIR := $(CURDIR)/.cache/go-build
 SERVER_BIN := $(BIN_DIR)/amcs-server
 CMD_SERVER := ./cmd/amcs-server
 BUILDINFO_PKG := git.warky.dev/wdevs/amcs/internal/buildinfo
+UI_DIR := $(CURDIR)/ui
 PATCH_INCREMENT ?= 1
 VERSION_TAG ?= $(shell git describe --tags --exact-match 2>/dev/null || echo dev)
 COMMIT_SHA ?= $(shell git rev-parse --short HEAD 2>/dev/null || echo unknown)
 BUILD_DATE ?= $(shell date -u +%Y-%m-%dT%H:%M:%SZ)
+RELSPEC ?= $(shell command -v relspec 2>/dev/null || echo $(HOME)/go/bin/relspec)
+SCHEMA_FILES := $(sort $(wildcard schema/*.dbml))
+MERGE_TARGET_TMP := $(CURDIR)/.cache/schema.merge-target.dbml
+GENERATED_SCHEMA_MIGRATION := migrations/020_generated_schema.sql
+PNPM ?= pnpm
 LDFLAGS := -s -w \
 	-X $(BUILDINFO_PKG).Version=$(VERSION_TAG) \
 	-X $(BUILDINFO_PKG).TagName=$(VERSION_TAG) \
 	-X $(BUILDINFO_PKG).Commit=$(COMMIT_SHA) \
 	-X $(BUILDINFO_PKG).BuildDate=$(BUILD_DATE)
 
-.PHONY: all build clean migrate release-version test
+.PHONY: all build clean migrate release-version test generate-migrations check-schema-drift build-cli ui-install ui-build ui-dev ui-check
 
 all: build
 
-build:
+build: ui-build
 	@mkdir -p $(BIN_DIR)
 	go build -ldflags "$(LDFLAGS)" -o $(SERVER_BIN) $(CMD_SERVER)
 
-test:
+ui-install:
+	cd $(UI_DIR) && $(PNPM) install --frozen-lockfile
+
+ui-build: ui-install
+	cd $(UI_DIR) && $(PNPM) run build
+
+ui-dev: ui-install
+	cd $(UI_DIR) && $(PNPM) run dev
+
+ui-check: ui-install
+	cd $(UI_DIR) && $(PNPM) run check
+
+test: ui-check
 	@mkdir -p $(GO_CACHE_DIR)
 	GOCACHE=$(GO_CACHE_DIR) go test ./...
 
@@ -43,6 +61,7 @@ release-version:
 		exit 1; \
 	fi; \
 	git tag -a "$$next_tag" -m "Release $$next_tag"; \
+	git push origin "$$next_tag"; \
 	echo "$$next_tag"
 
 migrate:
@@ -50,3 +69,31 @@ migrate:
 
 clean:
 	rm -rf $(BIN_DIR)
+
+generate-migrations:
+	@test -n "$(SCHEMA_FILES)" || (echo "No DBML schema files found in schema/" >&2; exit 1)
+	@command -v $(RELSPEC) >/dev/null 2>&1 || (echo "relspec not found; install git.warky.dev/wdevs/relspecgo/cmd/relspec@latest" >&2; exit 1)
+	@mkdir -p $(dir $(MERGE_TARGET_TMP))
+	@: > $(MERGE_TARGET_TMP)
+	@schema_list=$$(printf '%s\n' $(SCHEMA_FILES) | paste -sd, -); \
+	$(RELSPEC) merge --target dbml --target-path $(MERGE_TARGET_TMP) --source dbml --from-list "$$schema_list" --output pgsql --output-path $(GENERATED_SCHEMA_MIGRATION)
+
+check-schema-drift:
+	@test -f $(GENERATED_SCHEMA_MIGRATION) || (echo "$(GENERATED_SCHEMA_MIGRATION) is missing; run make generate-migrations" >&2; exit 1)
+	@command -v $(RELSPEC) >/dev/null 2>&1 || (echo "relspec not found; install git.warky.dev/wdevs/relspecgo/cmd/relspec@latest" >&2; exit 1)
+	@mkdir -p $(dir $(MERGE_TARGET_TMP))
+	@tmpfile=$$(mktemp); \
+	: > $(MERGE_TARGET_TMP); \
+	schema_list=$$(printf '%s\n' $(SCHEMA_FILES) | paste -sd, -); \
+	$(RELSPEC) merge --target dbml --target-path $(MERGE_TARGET_TMP) --source dbml --from-list "$$schema_list" --output pgsql --output-path $$tmpfile; \
+	if ! cmp -s $$tmpfile $(GENERATED_SCHEMA_MIGRATION); then \
+		echo "Schema drift detected between schema/*.dbml and $(GENERATED_SCHEMA_MIGRATION)" >&2; \
+		diff -u $(GENERATED_SCHEMA_MIGRATION) $$tmpfile || true; \
+		rm -f $$tmpfile; \
+		exit 1; \
+	fi; \
+	rm -f $$tmpfile
+
+build-cli:
+	@mkdir -p $(BIN_DIR)
+	go build -o $(BIN_DIR)/amcs-cli ./cmd/amcs-cli

README.md

@@ -1,24 +1,18 @@
-# Avalon Memory Crystal Server  (amcs)
+# AMCS Directory
 
-![Avalon Memory Crystal](assets/avelonmemorycrystal.jpg)
+This is the AMCS (Advanced Module Control System) directory.
 
-A Go MCP server for capturing and retrieving thoughts, memory, and project context. Exposes tools over Streamable HTTP, backed by Postgres with pgvector for semantic search.
+## Purpose
 
-## What it does
+The AMCS directory is used to store configuration and code for the Advanced Module Control System, which handles...
 
-- **Capture** thoughts with automatic embedding and metadata extraction
+## Structure
-- **Search** thoughts semantically via vector similarity
-- **Organise** thoughts into projects and retrieve full project context
-- **Summarise** and recall memory across topics and time windows
-- **Link** related thoughts and traverse relationships
 
-## Stack
+- `configs/` - Configuration files
+- `scripts/` - Scripts for managing the system
+- `assets/` - Asset files
 
-- Go — MCP server over Streamable HTTP
+## Next Steps
-- Postgres + pgvector — storage and vector search
-- LiteLLM — primary hosted AI provider (embeddings + metadata extraction)
-- OpenRouter — default upstream behind LiteLLM
-- Ollama — supported local or self-hosted OpenAI-compatible provider
 
 ## Tools
 
@@ -46,21 +40,55 @@ A Go MCP server for capturing and retrieving thoughts, memory, and project conte
 | `load_file` | Retrieve a stored file by ID; returns metadata, base64 content, and an embedded MCP binary resource |
 | `list_files` | Browse stored files by thought, project, or kind |
 | `backfill_embeddings` | Generate missing embeddings for stored thoughts |
-| `reparse_thought_metadata` | Re-extract and normalize metadata for stored thoughts |
+| `reparse_thought_metadata` | Re-extract metadata from thought content |
-| `retry_failed_metadata` | Retry metadata extraction for thoughts still pending or failed |
+| `retry_failed_metadata` | Retry pending/failed metadata extraction |
-| `add_skill` | Store a reusable agent skill (behavioural instruction or capability prompt) |
+| `add_maintenance_task` | Create a recurring or one-time home maintenance task |
+| `log_maintenance` | Log completed maintenance; updates next due date |
+| `get_upcoming_maintenance` | List maintenance tasks due within the next N days |
+| `search_maintenance_history` | Search the maintenance log by task name, category, or date range |
+| `save_chat_history` | Save chat messages with optional title, summary, channel, agent, and project |
+| `get_chat_history` | Fetch chat history by UUID or session_id |
+| `list_chat_histories` | List chat histories; filter by project, channel, agent_id, session_id, or days |
+| `delete_chat_history` | Delete a chat history by id |
+| `add_skill` | Store an agent skill (instruction or capability prompt) |
 | `remove_skill` | Delete an agent skill by id |
 | `list_skills` | List all agent skills, optionally filtered by tag |
-| `add_guardrail` | Store a reusable agent guardrail (constraint or safety rule) |
+| `add_guardrail` | Store an agent guardrail (constraint or safety rule) |
 | `remove_guardrail` | Delete an agent guardrail by id |
 | `list_guardrails` | List all agent guardrails, optionally filtered by tag or severity |
-| `add_project_skill` | Link an agent skill to a project; pass `project` explicitly if your client does not preserve MCP sessions |
+| `add_project_skill` | Link a skill to a project; pass `project` if client is stateless |
-| `remove_project_skill` | Unlink an agent skill from a project; pass `project` explicitly if your client does not preserve MCP sessions |
+| `remove_project_skill` | Unlink a skill from a project; pass `project` if client is stateless |
-| `list_project_skills` | List all skills linked to a project; pass `project` explicitly if your client does not preserve MCP sessions |
+| `list_project_skills` | Skills for a project; pass `project` if client is stateless |
-| `add_project_guardrail` | Link an agent guardrail to a project; pass `project` explicitly if your client does not preserve MCP sessions |
+| `add_project_guardrail` | Link a guardrail to a project; pass `project` if client is stateless |
-| `remove_project_guardrail` | Unlink an agent guardrail from a project; pass `project` explicitly if your client does not preserve MCP sessions |
+| `remove_project_guardrail` | Unlink a guardrail from a project; pass `project` if client is stateless |
-| `list_project_guardrails` | List all guardrails linked to a project; pass `project` explicitly if your client does not preserve MCP sessions |
+| `list_project_guardrails` | Guardrails for a project; pass `project` if client is stateless |
-| `get_version_info` | Return the server build version information, including version, tag name, commit, and build date |
+| `get_version_info` | Build version, commit, and date |
+| `describe_tools` | List all available MCP tools with names, descriptions, categories, and model-authored usage notes; call this at the start of a session to orient yourself |
+| `annotate_tool` | Persist your own usage notes for a specific tool; notes are returned by `describe_tools` in future sessions |
+
+## Self-Documenting Tools
+
+AMCS includes a built-in tool directory that models can read and annotate.
+
+**`describe_tools`** returns every registered tool with its name, description, category, and any model-written notes. Call it with no arguments to get the full list, or filter by category:
+
+```json
+{ "category": "thoughts" }
+```
+
+Available categories: `system`, `thoughts`, `projects`, `files`, `admin`, `maintenance`, `skills`, `chat`, `meta`.
+
+**`annotate_tool`** lets a model write persistent usage notes against a tool name. Notes survive across sessions and are returned by `describe_tools`:
+
+```json
+{ "tool_name": "capture_thought", "notes": "Always pass project explicitly — session state is not reliable in this client." }
+```
+
+Pass an empty string to clear notes. The intended workflow is:
+
+1. At the start of a session, call `describe_tools` to discover tools and read accumulated notes.
+2. As you learn something non-obvious about a tool — a gotcha, a workflow pattern, a required field ordering — call `annotate_tool` to record it.
+3. Future sessions receive the annotation automatically via `describe_tools`.
 
 ## MCP Error Contract
 
@@ -210,12 +238,25 @@ Link existing skills and guardrails to a project so they are automatically avail
 Config is YAML-driven. Copy `configs/config.example.yaml` and set:
 
 - `database.url` — Postgres connection string
-- `auth.mode` — `api_keys` or `oauth_client_credentials`
+- `auth.keys` — static API keys for MCP access via `x-brain-key` or `Authorization: Bearer <key>`
-- `auth.keys` — API keys for MCP access via `x-brain-key` or `Authorization: Bearer <key>` when `auth.mode=api_keys`
+- `auth.oauth.clients` — optional OAuth client credentials registry
-- `auth.oauth.clients` — client registry when `auth.mode=oauth_client_credentials`
+- `ai.providers` — named provider definitions (`litellm`, `ollama`, `openrouter`)
+- `ai.embeddings.primary` / `ai.metadata.primary` — primary role targets (`provider` + `model`)
+- `ai.embeddings.fallbacks` / `ai.metadata.fallbacks` — sequential fallback targets
 - `mcp.version` is build-generated and should not be set in config
 
-**OAuth Client Credentials flow** (`auth.mode=oauth_client_credentials`):
+Config schema is versioned. Current schema version is `2`.
+
+Use the migration helper to rewrite legacy configs in-place:
+
+```bash
+go run ./cmd/amcs-migrate-config --config ./configs/dev.yaml
+```
+
+Use `--dry-run` to print migrated YAML without writing.
+Server startup migrates older config formats in memory only and does not write files.
+
+**OAuth Client Credentials flow**:
 
 1. Obtain a token — `POST /oauth/token` (public, no auth required):
    ```
@@ -233,10 +274,11 @@ Config is YAML-driven. Copy `configs/config.example.yaml` and set:
    ```
 
 Alternatively, pass `client_id` and `client_secret` as body parameters instead of `Authorization: Basic`. Direct `Authorization: Basic` credential validation on the MCP endpoint is also supported as a fallback (no token required).
-- `ai.litellm.base_url` and `ai.litellm.api_key` — LiteLLM proxy
+- `AMCS_LITELLM_BASE_URL` / `AMCS_LITELLM_API_KEY` override all configured LiteLLM providers
-- `ai.ollama.base_url` and `ai.ollama.api_key` — Ollama local or remote server
+- `AMCS_OLLAMA_BASE_URL` / `AMCS_OLLAMA_API_KEY` override all configured Ollama providers
+- `AMCS_OPENROUTER_API_KEY` overrides all configured OpenRouter providers
 
-See `llm/plan.md` for full architecture and implementation plan.
+See `llm/plan.md` for an audited high-level status summary of the original implementation plan, and `llm/todo.md` for the audited backfill/fallback follow-up status.
 
 ## Backfill
 
@@ -499,13 +541,90 @@ Recommended Apache settings:
 - `ProxyTimeout 600` and `ProxyPass ... timeout=600` give Apache enough time to wait for the Go backend.
 - If another proxy or load balancer sits in front of Apache, align its size and timeout settings too.
 
+## CLI
+
+`amcs-cli` is a pre-built CLI client for the AMCS MCP server. Download it from https://git.warky.dev/wdevs/amcs/releases
+
+The primary purpose is to give agents and MCP clients a ready-made bridge to the AMCS server so they do not need to implement their own HTTP MCP client. Configure it once and any stdio-based MCP client can use AMCS immediately.
+
+### Commands
+
+| Command | Purpose |
+|---|---|
+| `amcs-cli tools` | List all tools available on the remote server |
+| `amcs-cli call <tool>` | Call a tool by name with `--arg key=value` flags |
+| `amcs-cli stdio` | Start a stdio MCP bridge backed by the remote server |
+
+`stdio` is the main integration point. It connects to the remote HTTP MCP server, discovers all its tools, and re-exposes them over stdio. Register it as a stdio MCP server in your agent config and it proxies every tool call through to AMCS.
+
+### Configuration
+
+Config file: `~/.config/amcs/config.yaml`
+
+```yaml
+server: https://your-amcs-server
+token: your-bearer-token
+```
+
+Env vars override the config file: `AMCS_SERVER` (preferred), `AMCS_URL` (legacy alias), and `AMCS_TOKEN`. Flags `--server` and `--token` override env vars.
+
+### stdio MCP client setup
+
+#### Claude Code
+
+```bash
+claude mcp add --transport stdio amcs amcs-cli stdio
+```
+
+With inline credentials (no config file):
+
+```bash
+claude mcp add --transport stdio amcs amcs-cli stdio \
+  --env AMCS_SERVER=https://your-amcs-server \
+  --env AMCS_TOKEN=your-bearer-token
+```
+
+#### Output format
+
+`call` outputs JSON by default. Pass `--output yaml` for YAML.
+
 ## Development
 
 Run the SQL migrations against a local database with:
 
 `DATABASE_URL=postgres://... make migrate`
 
-LLM integration instructions are served at `/llm`.
+### Backend + embedded UI build
+
+The web UI now lives in the top-level `ui/` module and is embedded into the Go binary at build time with `go:embed`.
+
+**Use `pnpm` for all UI work in this repo.**
+
+- `make build` — runs the real UI build first, then compiles the Go server
+- `make test` — runs `svelte-check` for the frontend and `go test ./...` for the backend
+- `make ui-install` — installs frontend dependencies with `pnpm install --frozen-lockfile`
+- `make ui-build` — builds only the frontend bundle
+- `make ui-dev` — starts the Vite dev server with hot reload on `http://localhost:5173`
+- `make ui-check` — runs the frontend type and Svelte checks
+
+### Local UI workflow
+
+For the normal production-style local flow:
+
+1. Start the backend: `./scripts/run-local.sh configs/dev.yaml`
+2. Open `http://localhost:8080`
+
+For frontend iteration with hot reload and no Go rebuilds:
+
+1. Start the backend once: `go run ./cmd/amcs-server --config configs/dev.yaml`
+2. In another shell start the UI dev server: `make ui-dev`
+3. Open `http://localhost:5173`
+
+The Vite dev server proxies backend routes such as `/api/status`, `/llm`, `/healthz`, `/readyz`, `/files`, `/mcp`, and the OAuth endpoints back to the Go server on `http://127.0.0.1:8080` by default. Override that target with `AMCS_UI_BACKEND` if needed.
+
+The root page (`/`) is now the Svelte frontend. It preserves the existing landing-page content and status information by fetching data from `GET /api/status`.
+
+LLM integration instructions are still served at `/llm`.
 
 ## Containers
 
@@ -530,29 +649,50 @@ Notes:
 - Database migrations `001` through `005` run automatically when the Postgres volume is created for the first time.
 - `migrations/006_rls_and_grants.sql` is intentionally skipped during container bootstrap because it contains deployment-specific grants for a role named `amcs_user`.
 
+### Run config migration with Compose
+
+The container image now includes `/app/amcs-migrate-config`.
+
+Dry-run (prints migrated YAML, does not write files):
+
+```bash
+docker compose --profile tools run --rm migrate-config --config /app/configs/dev.yaml --dry-run
+```
+
+Apply migration in-place (writes file + creates backup):
+
+```bash
+docker compose --profile tools run --rm migrate-config --config /app/configs/dev.yaml
+```
+
 ## Ollama
 
-Set `ai.provider: "ollama"` to use a local or self-hosted Ollama server through its OpenAI-compatible API.
+Set your role targets to an Ollama provider to use a local or self-hosted Ollama server through its OpenAI-compatible API.
 
 Example:
 
 ```yaml
 ai:
-  provider: "ollama"
+  providers:
-  embeddings:
+    local:
-    model: "nomic-embed-text"
+      type: "ollama"
-    dimensions: 768
-  metadata:
-    model: "llama3.2"
-    temperature: 0.1
-  ollama:
       base_url: "http://localhost:11434/v1"
       api_key: "ollama"
       request_headers: {}
+  embeddings:
+    dimensions: 768
+    primary:
+      provider: "local"
+      model: "nomic-embed-text"
+  metadata:
+    temperature: 0.1
+    primary:
+      provider: "local"
+      model: "llama3.2"
 ```
 
 Notes:
 
-- For remote Ollama servers, point `ai.ollama.base_url` at the remote `/v1` endpoint.
+- For remote Ollama servers, point `ai.providers.<name>.base_url` at the remote `/v1` endpoint.
 - The client always sends Bearer auth; Ollama ignores it locally, so `api_key: "ollama"` is a safe default.
 - `ai.embeddings.dimensions` must match the embedding model you actually use, or startup will fail the database vector-dimension check.

changelog.md

@@ -0,0 +1,90 @@
+# Changelog
+
+## 2026-04-21
+
+### 2026-04-21 21h - Config Schema v2 Introduced
+
+- Refactored configuration to schema version `2` with named AI providers and role-based model chains.
+- Added support for per-role primary and fallback targets for embeddings and metadata.
+- Added optional background role overrides for backfill and metadata retry workers.
+
+### 2026-04-21 21h - Automatic v1 -> v2 Migration
+
+- Added config migration framework with explicit schema versioning.
+- Implemented `v1 -> v2` migration to transform legacy provider blocks into named providers + role chains.
+- Loader now auto-migrates older config files, rewrites migrated YAML, and creates timestamped backups.
+
+### 2026-04-21 21h - AI Registry and Role Runners
+
+- Added `ai.Registry` to build provider clients from named provider config entries.
+- Added `EmbeddingRunner` and `MetadataRunner` with sequential fallback execution.
+- Added target health tracking with cooldowns for transient/permanent/empty-response failures.
+
+### 2026-04-21 21h - App and Tool Wiring Updates
+
+- Rewired app startup to use provider registry + role runners for foreground and background flows.
+- Updated capture, search, summarize, context, recall, backfill, metadata retry, and reparse paths to use new runners.
+- Preserved environment override behavior for provider credentials/endpoints across matching provider types.
+
+### 2026-04-21 21h - Migrate Config CLI Added
+
+- Added `cmd/amcs-migrate-config` CLI to migrate config files to the current schema version.
+- Supports dry-run output and in-place write mode with automatic backup file creation.
+
+### 2026-04-21 21h - Tests and Documentation Updated
+
+- Added focused tests for config migration, AI registry behavior, and runner fallback behavior.
+- Updated `configs/config.example.yaml` to the new v2 schema.
+- Updated README configuration sections and migration guidance to reflect v2 and `amcs-migrate-config` usage.
+
+### 2026-04-21 21h - Uncommitted File Change List
+
+- Modified: `.gitignore`
+- Modified: `README.md`
+- Modified: `configs/config.example.yaml`
+- Modified: `internal/ai/compat/client.go`
+- Modified: `internal/ai/compat/client_test.go`
+- Modified: `internal/app/app.go`
+- Modified: `internal/config/config.go`
+- Modified: `internal/config/loader.go`
+- Modified: `internal/config/loader_test.go`
+- Modified: `internal/config/validate.go`
+- Modified: `internal/config/validate_test.go`
+- Modified: `internal/mcpserver/server.go`
+- Modified: `internal/mcpserver/streamable_integration_test.go`
+- Modified: `internal/tools/backfill.go`
+- Modified: `internal/tools/capture.go`
+- Modified: `internal/tools/context.go`
+- Modified: `internal/tools/enrichment_retry.go`
+- Modified: `internal/tools/links.go`
+- Modified: `internal/tools/metadata_retry.go`
+- Modified: `internal/tools/recall.go`
+- Modified: `internal/tools/reparse_metadata.go`
+- Modified: `internal/tools/retrieval.go`
+- Modified: `internal/tools/search.go`
+- Modified: `internal/tools/summarize.go`
+- Modified: `internal/tools/update.go`
+- Deleted: `internal/ai/factory.go`
+- Deleted: `internal/ai/factory_test.go`
+- Deleted: `internal/ai/litellm/client.go`
+- Deleted: `internal/ai/ollama/client.go`
+- Deleted: `internal/ai/openrouter/client.go`
+- Deleted: `internal/ai/provider.go`
+- New: `changelog.md`
+- New: `cmd/amcs-migrate-config/main.go`
+- New: `internal/ai/registry.go`
+- New: `internal/ai/registry_test.go`
+- New: `internal/ai/runner.go`
+- New: `internal/ai/runner_test.go`
+- New: `internal/config/migrate.go`
+- New: `internal/config/migrate_test.go`
+
+### 2026-04-21 21h - Docker Support for Config Migration CLI
+- Added `amcs-migrate-config` binary to the Docker image build output.
+- Added `migrate-config` service in `docker-compose.yml` under the `tools` profile.
+- Documented compose-based migration commands (dry-run and in-place apply) in the README.
+
+### 2026-04-21 21h - Startup Migration Write Disabled
+- Changed config loading to migrate legacy schemas in memory only during startup.
+- Removed automatic file rewrite and backup creation from the startup config loader.
+- Added loader log hint to use `amcs-migrate-config` when persistent conversion is needed.

cmd/amcs-cli/cmd/call.go

@@ -0,0 +1,98 @@
+package cmd
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"strconv"
+	"strings"
+
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+	"github.com/spf13/cobra"
+	"gopkg.in/yaml.v3"
+)
+
+var argFlags []string
+
+var callCmd = &cobra.Command{
+	Use:   "call <tool>",
+	Short: "Call a remote AMCS tool",
+	Args:  cobra.ExactArgs(1),
+	RunE: func(cmd *cobra.Command, args []string) error {
+		toolName := args[0]
+		toolArgs, err := parseArgs(argFlags)
+		if err != nil {
+			return err
+		}
+
+		session, err := connectRemote(cmd.Context())
+		if err != nil {
+			return err
+		}
+		defer func() { _ = session.Close() }()
+
+		res, err := session.CallTool(cmd.Context(), &mcp.CallToolParams{Name: toolName, Arguments: toolArgs})
+		if err != nil {
+			return fmt.Errorf("call tool %q: %w", toolName, err)
+		}
+		return printOutput(res)
+	},
+}
+
+func init() {
+	callCmd.Flags().StringArrayVar(&argFlags, "arg", nil, "Tool argument in key=value format (repeatable)")
+	rootCmd.AddCommand(callCmd)
+}
+
+func parseArgs(items []string) (map[string]any, error) {
+	result := make(map[string]any, len(items))
+	for _, item := range items {
+		key, value, ok := strings.Cut(item, "=")
+		if !ok || strings.TrimSpace(key) == "" {
+			return nil, fmt.Errorf("invalid --arg %q: want key=value", item)
+		}
+		result[key] = parseScalar(value)
+	}
+	return result, nil
+}
+
+func parseScalar(s string) any {
+	if s == "true" || s == "false" {
+		b, _ := strconv.ParseBool(s)
+		return b
+	}
+	if i, err := strconv.ParseInt(s, 10, 64); err == nil {
+		return i
+	}
+	if f, err := strconv.ParseFloat(s, 64); err == nil && strings.ContainsAny(s, ".eE") {
+		return f
+	}
+	var v any
+	if err := json.Unmarshal([]byte(s), &v); err == nil {
+		switch v.(type) {
+		case map[string]any, []any, float64, bool, nil:
+			return v
+		}
+	}
+	return s
+}
+
+func printOutput(v any) error {
+	switch outputFlag {
+	case "yaml":
+		data, err := yaml.Marshal(v)
+		if err != nil {
+			return fmt.Errorf("marshal yaml: %w", err)
+		}
+		_, err = os.Stdout.Write(data)
+		return err
+	default:
+		data, err := json.MarshalIndent(v, "", "  ")
+		if err != nil {
+			return fmt.Errorf("marshal json: %w", err)
+		}
+		data = append(data, '\n')
+		_, err = os.Stdout.Write(data)
+		return err
+	}
+}

cmd/amcs-cli/cmd/config.go

@@ -0,0 +1,60 @@
+package cmd
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"gopkg.in/yaml.v3"
+)
+
+type Config struct {
+	Server string `yaml:"server"`
+	Token  string `yaml:"token"`
+}
+
+func defaultConfigPath() (string, error) {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return "", fmt.Errorf("resolve home dir: %w", err)
+	}
+	return filepath.Join(home, ".config", "amcs", "config.yaml"), nil
+}
+
+func resolveConfigPath(path string) (string, error) {
+	if strings.TrimSpace(path) != "" {
+		return path, nil
+	}
+	return defaultConfigPath()
+}
+
+func loadConfigFile(path string) (Config, error) {
+	var cfg Config
+	data, err := os.ReadFile(path)
+	if err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return cfg, nil
+		}
+		return cfg, fmt.Errorf("read config: %w", err)
+	}
+	if err := yaml.Unmarshal(data, &cfg); err != nil {
+		return cfg, fmt.Errorf("parse config: %w", err)
+	}
+	return cfg, nil
+}
+
+func saveConfigFile(path string, cfg Config) error {
+	if err := os.MkdirAll(filepath.Dir(path), 0o700); err != nil {
+		return fmt.Errorf("create config dir: %w", err)
+	}
+	data, err := yaml.Marshal(cfg)
+	if err != nil {
+		return fmt.Errorf("marshal config: %w", err)
+	}
+	if err := os.WriteFile(path, data, 0o600); err != nil {
+		return fmt.Errorf("write config: %w", err)
+	}
+	return nil
+}

cmd/amcs-cli/cmd/root.go

@@ -0,0 +1,151 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"os"
+	"strings"
+
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+	"github.com/spf13/cobra"
+)
+
+var (
+	cfgFile    string
+	serverFlag string
+	tokenFlag  string
+	outputFlag string
+	verbose    bool
+	cfg        Config
+)
+
+const cliUserAgent = "amcs-cli/0.0.1"
+
+var rootCmd = &cobra.Command{
+	Use:           "amcs-cli",
+	Short:         "CLI for connecting to a remote AMCS MCP server",
+	SilenceUsage:  true,
+	SilenceErrors: true,
+	PersistentPreRunE: func(cmd *cobra.Command, _ []string) error {
+		return loadConfig()
+	},
+}
+
+func Execute() {
+	if err := rootCmd.Execute(); err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+}
+
+func init() {
+	rootCmd.PersistentFlags().StringVar(&cfgFile, "config", "", "Path to config file")
+	rootCmd.PersistentFlags().StringVar(&serverFlag, "server", "", "AMCS server URL")
+	rootCmd.PersistentFlags().StringVar(&tokenFlag, "token", "", "AMCS bearer token")
+	rootCmd.PersistentFlags().StringVar(&outputFlag, "output", "json", "Output format: json or yaml")
+	rootCmd.PersistentFlags().BoolVar(&verbose, "verbose", false, "Enable verbose logging to stderr")
+}
+
+func loadConfig() error {
+	path, err := resolveConfigPath(cfgFile)
+	if err != nil {
+		return err
+	}
+	loaded, err := loadConfigFile(path)
+	if err != nil {
+		return err
+	}
+	cfg = loaded
+	if v := strings.TrimSpace(os.Getenv("AMCS_SERVER")); v != "" {
+		cfg.Server = v
+	}
+	if v := strings.TrimSpace(os.Getenv("AMCS_URL")); v != "" {
+		cfg.Server = v
+	}
+	if v := strings.TrimSpace(os.Getenv("AMCS_TOKEN")); v != "" {
+		cfg.Token = v
+	}
+	if v := strings.TrimSpace(serverFlag); v != "" {
+		cfg.Server = v
+	}
+	if v := strings.TrimSpace(tokenFlag); v != "" {
+		cfg.Token = v
+	}
+	outputFlag = strings.ToLower(strings.TrimSpace(outputFlag))
+	if outputFlag != "json" && outputFlag != "yaml" {
+		return fmt.Errorf("invalid --output %q: must be json or yaml", outputFlag)
+	}
+	return nil
+}
+
+func requireServer() error {
+	if strings.TrimSpace(cfg.Server) == "" {
+		return fmt.Errorf("server URL is required; set --server, AMCS_SERVER, AMCS_URL, or config server")
+	}
+	return nil
+}
+
+func endpointURL() string {
+	base := strings.TrimRight(strings.TrimSpace(cfg.Server), "/")
+	if strings.HasSuffix(base, "/mcp") {
+		return base
+	}
+	return base + "/mcp"
+}
+
+func newHTTPClient() *http.Client {
+	return &http.Client{
+		Timeout: 0,
+		Transport: &bearerTransport{
+			base:  http.DefaultTransport,
+			token: cfg.Token,
+		},
+	}
+}
+
+type bearerTransport struct {
+	base  http.RoundTripper
+	token string
+}
+
+func (t *bearerTransport) RoundTrip(req *http.Request) (*http.Response, error) {
+	base := t.base
+	if base == nil {
+		base = http.DefaultTransport
+	}
+	clone := req.Clone(req.Context())
+	if strings.TrimSpace(clone.Header.Get("User-Agent")) == "" {
+		clone.Header.Set("User-Agent", cliUserAgent)
+	}
+	if strings.TrimSpace(t.token) != "" {
+		clone.Header.Set("Authorization", "Bearer "+t.token)
+	}
+	return base.RoundTrip(clone)
+}
+
+func connectRemote(ctx context.Context) (*mcp.ClientSession, error) {
+	if err := requireServer(); err != nil {
+		return nil, err
+	}
+	verboseLogf("connecting to %s", endpointURL())
+	client := mcp.NewClient(&mcp.Implementation{Name: "amcs-cli", Version: "0.0.1"}, nil)
+	transport := &mcp.StreamableClientTransport{
+		Endpoint:             endpointURL(),
+		HTTPClient:           newHTTPClient(),
+		DisableStandaloneSSE: true,
+	}
+	session, err := client.Connect(ctx, transport, nil)
+	if err != nil {
+		return nil, fmt.Errorf("connect to AMCS server: %w", err)
+	}
+	verboseLogf("connected to %s", endpointURL())
+	return session, nil
+}
+
+func verboseLogf(format string, args ...any) {
+	if !verbose {
+		return
+	}
+	_, _ = fmt.Fprintf(os.Stderr, "[amcs-cli] "+format+"\n", args...)
+}

cmd/amcs-cli/cmd/root_test.go

@@ -0,0 +1,35 @@
+package cmd
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestBearerTransportFormatsBearerToken(t *testing.T) {
+	const want = "Bearer X"
+	const wantUA = "amcs-cli/0.0.1"
+
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if got := r.Header.Get("Authorization"); got != want {
+			t.Fatalf("Authorization header = %q, want %q", got, want)
+		}
+		if got := r.Header.Get("User-Agent"); got != wantUA {
+			t.Fatalf("User-Agent header = %q, want %q", got, wantUA)
+		}
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer ts.Close()
+
+	client := &http.Client{Transport: &bearerTransport{token: "X"}}
+	req, err := http.NewRequest(http.MethodGet, ts.URL, nil)
+	if err != nil {
+		t.Fatalf("NewRequest() error = %v", err)
+	}
+
+	res, err := client.Do(req)
+	if err != nil {
+		t.Fatalf("client.Do() error = %v", err)
+	}
+	_ = res.Body.Close()
+}

cmd/amcs-cli/cmd/sse.go

@@ -0,0 +1,89 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+	"github.com/spf13/cobra"
+)
+
+var sseCmd = &cobra.Command{
+	Use:   "sse",
+	Short: "Run a stdio MCP bridge backed by a remote AMCS server using SSE transport (widely supported by hosted MCP clients)",
+	RunE: func(cmd *cobra.Command, _ []string) error {
+		ctx := cmd.Context()
+
+		if err := requireServer(); err != nil {
+			return err
+		}
+
+		client := mcp.NewClient(&mcp.Implementation{Name: "amcs-cli", Version: "0.0.1"}, nil)
+		transport := &mcp.SSEClientTransport{
+			Endpoint:   sseEndpointURL(),
+			HTTPClient: newHTTPClient(),
+		}
+
+		verboseLogf("connecting to SSE endpoint %s", sseEndpointURL())
+		remote, err := client.Connect(ctx, transport, nil)
+		if err != nil {
+			return fmt.Errorf("connect to AMCS SSE endpoint: %w", err)
+		}
+		defer func() { _ = remote.Close() }()
+		verboseLogf("connected to SSE endpoint %s", sseEndpointURL())
+
+		tools, err := remote.ListTools(ctx, nil)
+		if err != nil {
+			return fmt.Errorf("load remote tools: %w", err)
+		}
+
+		server := mcp.NewServer(&mcp.Implementation{
+			Name:    "amcs-cli",
+			Title:   "AMCS CLI Bridge (SSE)",
+			Version: "0.0.1",
+		}, nil)
+
+		for _, tool := range tools.Tools {
+			remoteTool := tool
+			server.AddTool(&mcp.Tool{
+				Name:         remoteTool.Name,
+				Description:  remoteTool.Description,
+				InputSchema:  remoteTool.InputSchema,
+				OutputSchema: remoteTool.OutputSchema,
+				Annotations:  remoteTool.Annotations,
+			}, func(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+				return remote.CallTool(ctx, &mcp.CallToolParams{
+					Name:      req.Params.Name,
+					Arguments: req.Params.Arguments,
+				})
+			})
+		}
+
+		session, err := server.Connect(ctx, &mcp.StdioTransport{}, nil)
+		if err != nil {
+			return fmt.Errorf("start stdio bridge: %w", err)
+		}
+		defer func() { _ = session.Close() }()
+		verboseLogf("sse stdio bridge ready")
+		verboseLogf("waiting for MCP commands on stdin")
+
+		<-ctx.Done()
+		return nil
+	},
+}
+
+func sseEndpointURL() string {
+	base := strings.TrimRight(strings.TrimSpace(cfg.Server), "/")
+	if strings.HasSuffix(base, "/mcp") {
+		base = strings.TrimSuffix(base, "/mcp")
+	}
+	if strings.HasSuffix(base, "/sse") {
+		return base
+	}
+	return base + "/sse"
+}
+
+func init() {
+	rootCmd.AddCommand(sseCmd)
+}

cmd/amcs-cli/cmd/stdio.go

@@ -0,0 +1,64 @@
+package cmd
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+	"github.com/spf13/cobra"
+)
+
+var stdioCmd = &cobra.Command{
+	Use:   "stdio",
+	Short: "Run a stdio MCP bridge backed by a remote AMCS server",
+	RunE: func(cmd *cobra.Command, _ []string) error {
+		ctx := cmd.Context()
+		remote, err := connectRemote(ctx)
+		if err != nil {
+			return err
+		}
+		defer func() { _ = remote.Close() }()
+
+		tools, err := remote.ListTools(ctx, nil)
+		if err != nil {
+			return fmt.Errorf("load remote tools: %w", err)
+		}
+
+		server := mcp.NewServer(&mcp.Implementation{
+			Name:    "amcs-cli",
+			Title:   "AMCS CLI Bridge",
+			Version: "0.0.1",
+		}, nil)
+
+		for _, tool := range tools.Tools {
+			remoteTool := tool
+			server.AddTool(&mcp.Tool{
+				Name:         remoteTool.Name,
+				Description:  remoteTool.Description,
+				InputSchema:  remoteTool.InputSchema,
+				OutputSchema: remoteTool.OutputSchema,
+				Annotations:  remoteTool.Annotations,
+			}, func(ctx context.Context, req *mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+				return remote.CallTool(ctx, &mcp.CallToolParams{
+					Name:      req.Params.Name,
+					Arguments: req.Params.Arguments,
+				})
+			})
+		}
+
+		session, err := server.Connect(ctx, &mcp.StdioTransport{}, nil)
+		if err != nil {
+			return fmt.Errorf("start stdio bridge: %w", err)
+		}
+		defer func() { _ = session.Close() }()
+		verboseLogf("stdio bridge connected to remote AMCS and ready")
+		verboseLogf("waiting for MCP commands on stdin")
+
+		<-ctx.Done()
+		return nil
+	},
+}
+
+func init() {
+	rootCmd.AddCommand(stdioCmd)
+}

cmd/amcs-cli/cmd/tools.go

@@ -0,0 +1,38 @@
+package cmd
+
+import (
+	"fmt"
+	"os"
+	"strings"
+	"text/tabwriter"
+
+	"github.com/spf13/cobra"
+)
+
+var toolsCmd = &cobra.Command{
+	Use:   "tools",
+	Short: "List tools available on the remote AMCS server",
+	RunE: func(cmd *cobra.Command, _ []string) error {
+		session, err := connectRemote(cmd.Context())
+		if err != nil {
+			return err
+		}
+		defer func() { _ = session.Close() }()
+
+		res, err := session.ListTools(cmd.Context(), nil)
+		if err != nil {
+			return fmt.Errorf("list tools: %w", err)
+		}
+
+		w := tabwriter.NewWriter(os.Stdout, 0, 0, 2, ' ', 0)
+		fmt.Fprintln(w, "NAME\tDESCRIPTION")
+		for _, tool := range res.Tools {
+			fmt.Fprintf(w, "%s\t%s\n", tool.Name, strings.TrimSpace(tool.Description))
+		}
+		return w.Flush()
+	},
+}
+
+func init() {
+	rootCmd.AddCommand(toolsCmd)
+}

cmd/amcs-cli/main.go

@@ -0,0 +1,7 @@
+package main
+
+import "git.warky.dev/wdevs/amcs/cmd/amcs-cli/cmd"
+
+func main() {
+	cmd.Execute()
+}

cmd/amcs-migrate-config/main.go

@@ -0,0 +1,105 @@
+package main
+
+import (
+	"flag"
+	"fmt"
+	"log"
+	"os"
+	"time"
+
+	"gopkg.in/yaml.v3"
+
+	"git.warky.dev/wdevs/amcs/internal/config"
+)
+
+func main() {
+	var (
+		configPath string
+		dryRun     bool
+		toVersion  int
+	)
+	flag.StringVar(&configPath, "config", "", "Path to the YAML config file (default: $AMCS_CONFIG or ./configs/dev.yaml)")
+	flag.BoolVar(&dryRun, "dry-run", false, "Print the migrated config to stdout instead of writing it back")
+	flag.IntVar(&toVersion, "to-version", config.CurrentConfigVersion, "Stop migrating after reaching this version")
+	flag.Parse()
+
+	if toVersion <= 0 || toVersion > config.CurrentConfigVersion {
+		log.Fatalf("invalid -to-version %d (must be between 1 and %d)", toVersion, config.CurrentConfigVersion)
+	}
+
+	path := config.ResolvePath(configPath)
+	original, err := os.ReadFile(path)
+	if err != nil {
+		log.Fatalf("read config %q: %v", path, err)
+	}
+
+	raw := map[string]any{}
+	if err := yaml.Unmarshal(original, &raw); err != nil {
+		log.Fatalf("decode config %q: %v", path, err)
+	}
+	if raw == nil {
+		raw = map[string]any{}
+	}
+
+	applied, err := migrateUpTo(raw, toVersion)
+	if err != nil {
+		log.Fatalf("migrate: %v", err)
+	}
+
+	if len(applied) == 0 {
+		fmt.Fprintf(os.Stderr, "%s already at version %d; nothing to do\n", path, currentVersion(raw))
+		return
+	}
+
+	out, err := yaml.Marshal(raw)
+	if err != nil {
+		log.Fatalf("marshal migrated config: %v", err)
+	}
+
+	for _, step := range applied {
+		fmt.Fprintf(os.Stderr, "applied migration v%d -> v%d: %s\n", step.From, step.To, step.Describe)
+	}
+
+	if dryRun {
+		_, _ = os.Stdout.Write(out)
+		return
+	}
+
+	backup := fmt.Sprintf("%s.bak.%d", path, time.Now().Unix())
+	if err := os.WriteFile(backup, original, 0o600); err != nil {
+		log.Fatalf("write backup %q: %v", backup, err)
+	}
+	if err := os.WriteFile(path, out, 0o600); err != nil {
+		log.Fatalf("write migrated config %q: %v", path, err)
+	}
+	fmt.Fprintf(os.Stderr, "wrote migrated config to %s (backup: %s)\n", path, backup)
+}
+
+// migrateUpTo runs the migration ladder but stops at the requested version.
+func migrateUpTo(raw map[string]any, target int) ([]config.ConfigMigration, error) {
+	if currentVersion(raw) >= target {
+		return nil, nil
+	}
+	if target == config.CurrentConfigVersion {
+		return config.Migrate(raw)
+	}
+	// Partial migrations are rare; for now reject anything other than the
+	// current version target since the migration ladder is short.
+	return nil, fmt.Errorf("partial migration to v%d is not supported (use -to-version=%d)", target, config.CurrentConfigVersion)
+}
+
+func currentVersion(raw map[string]any) int {
+	v, ok := raw["version"]
+	if !ok {
+		return 1
+	}
+	switch n := v.(type) {
+	case int:
+		return n
+	case int64:
+		return int(n)
+	case float64:
+		return int(n)
+	}
+	return 1
+}

configs/config.example.yaml

@@ -1,3 +1,5 @@
+version: 2
+
 server:
   host: "0.0.0.0"
   port: 8080
@@ -9,6 +11,7 @@ server:
 
 mcp:
   path: "/mcp"
+  sse_path: "/sse"
   server_name: "amcs"
   transport: "streamable_http"
   session_timeout: "10m"
@@ -26,7 +29,7 @@ auth:
       - id: "oauth-client"
         client_id: ""
         client_secret: ""
-        description: "used when auth.mode=oauth_client_credentials"
+        description: "optional OAuth client credentials"
 
 database:
   url: "postgres://postgres:postgres@localhost:5432/amcs?sslmode=disable"
@@ -36,33 +39,58 @@ database:
   max_conn_idle_time: "10m"
 
 ai:
-  provider: "litellm"
+  providers:
-  embeddings:
+    default:
-    model: "openai/text-embedding-3-small"
+      type: "litellm"
-    dimensions: 1536
-  metadata:
-    model: "gpt-4o-mini"
-    fallback_models: []
-    temperature: 0.1
-    log_conversations: false
-  litellm:
       base_url: "http://localhost:4000/v1"
       api_key: "replace-me"
-    use_responses_api: false
       request_headers: {}
-    embedding_model: "openrouter/openai/text-embedding-3-small"
+
-    metadata_model: "gpt-4o-mini"
+    ollama_local:
-    fallback_metadata_models: []
+      type: "ollama"
-  ollama:
       base_url: "http://localhost:11434/v1"
       api_key: "ollama"
       request_headers: {}
+
     openrouter:
+      type: "openrouter"
       base_url: "https://openrouter.ai/api/v1"
-    api_key: ""
+      api_key: "replace-me"
       app_name: "amcs"
       site_url: ""
-    extra_headers: {}
+      request_headers: {}
+
+  embeddings:
+    dimensions: 1536
+    primary:
+      provider: "default"
+      model: "openai/text-embedding-3-small"
+    fallbacks:
+      - provider: "ollama_local"
+        model: "nomic-embed-text"
+
+  metadata:
+    temperature: 0.1
+    log_conversations: false
+    timeout: "10s"
+    primary:
+      provider: "default"
+      model: "gpt-4o-mini"
+    fallbacks:
+      - provider: "openrouter"
+        model: "openai/gpt-4.1-mini"
+
+  # Optional overrides for background jobs (backfill_embeddings,
+  # retry_failed_metadata, reparse_thought_metadata).
+  background:
+    embeddings:
+      primary:
+        provider: "default"
+        model: "openai/text-embedding-3-small"
+    metadata:
+      primary:
+        provider: "default"
+        model: "gpt-4o-mini"
 
 capture:
   source: "mcp"

configs/dev.yaml

@@ -9,6 +9,7 @@ server:
 
 mcp:
   path: "/mcp"
+  sse_path: "/sse"
   server_name: "amcs"
   transport: "streamable_http"
   session_timeout: "10m"

configs/docker.yaml

@@ -9,6 +9,7 @@ server:
 
 mcp:
   path: "/mcp"
+  sse_path: "/sse"
   server_name: "amcs"
   transport: "streamable_http"
   session_timeout: "10m"

docker-compose.yml

@@ -36,6 +36,18 @@ services:
     ports:
       - "8080:8080"
 
+  migrate-config:
+    build:
+      context: .
+    profiles: ["tools"]
+    restart: "no"
+    volumes:
+      - ./configs:/app/configs
+    environment:
+      AMCS_CONFIG: /app/configs/docker.yaml
+    entrypoint: ["/app/amcs-migrate-config"]
+    command: ["--config", "/app/configs/docker.yaml", "--dry-run"]
+
 volumes:
   postgres_data:
 

go.mod

@@ -8,11 +8,13 @@ require (
 	github.com/jackc/pgx/v5 v5.9.1
 	github.com/modelcontextprotocol/go-sdk v1.4.1
 	github.com/pgvector/pgvector-go v0.3.0
+	github.com/spf13/cobra v1.10.2
 	golang.org/x/sync v0.17.0
 	gopkg.in/yaml.v3 v3.0.1
 )
 
 require (
+	github.com/inconshreveable/mousetrap v1.1.0 // indirect
 	github.com/jackc/pgpassfile v1.0.0 // indirect
 	github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761 // indirect
 	github.com/jackc/puddle/v2 v2.2.2 // indirect
@@ -20,6 +22,7 @@ require (
 	github.com/rogpeppe/go-internal v1.14.1 // indirect
 	github.com/segmentio/asm v1.1.3 // indirect
 	github.com/segmentio/encoding v0.5.4 // indirect
+	github.com/spf13/pflag v1.0.9 // indirect
 	github.com/x448/float16 v0.8.4 // indirect
 	github.com/yosida95/uritemplate/v3 v3.0.2 // indirect
 	golang.org/x/oauth2 v0.34.0 // indirect

go.sum

@@ -1,5 +1,6 @@
 entgo.io/ent v0.14.3 h1:wokAV/kIlH9TeklJWGGS7AYJdVckr0DloWjIcO9iIIQ=
 entgo.io/ent v0.14.3/go.mod h1:aDPE/OziPEu8+OWbzy4UlvWmD2/kbRuWfK2A40hcxJM=
+github.com/cpuguy83/go-md2man/v2 v2.0.6/go.mod h1:oOW0eioCTA6cOiMLiUPZOpcVxMig6NIQQ7OS05n1F4g=
 github.com/creack/pty v1.1.9/go.mod h1:oKZEueFk5CKHvIhNR5MUki03XCEU+Q6VDXinZuGJ33E=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
@@ -16,6 +17,8 @@ github.com/google/jsonschema-go v0.4.2 h1:tmrUohrwoLZZS/P3x7ex0WAVknEkBZM46iALbc
 github.com/google/jsonschema-go v0.4.2/go.mod h1:r5quNTdLOYEz95Ru18zA0ydNbBuYoo9tgaYcxEYhJVE=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8=
+github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw=
 github.com/jackc/pgpassfile v1.0.0 h1:/6Hmqy13Ss2zCq62VdNG8tM1wchn8zjSGOBJ6icpsIM=
 github.com/jackc/pgpassfile v1.0.0/go.mod h1:CEx0iS5ambNFdcRtxPj5JhEz+xB6uRky5eyVu/W2HEg=
 github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761 h1:iCEnooe7UlwOQYpKFhBabPMi4aNAfoODPEFNiAnClxo=
@@ -44,10 +47,15 @@ github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZb
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=
 github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7so1lCWt35ZSgc=
+github.com/russross/blackfriday/v2 v2.1.0/go.mod h1:+Rmxgy9KzJVeS9/2gXHxylqXiyQDYRxCVz55jmeOWTM=
 github.com/segmentio/asm v1.1.3 h1:WM03sfUOENvvKexOLp+pCqgb/WDjsi7EK8gIsICtzhc=
 github.com/segmentio/asm v1.1.3/go.mod h1:Ld3L4ZXGNcSLRg4JBsZ3//1+f/TjYl0Mzen/DQy1EJg=
 github.com/segmentio/encoding v0.5.4 h1:OW1VRern8Nw6ITAtwSZ7Idrl3MXCFwXHPgqESYfvNt0=
 github.com/segmentio/encoding v0.5.4/go.mod h1:HS1ZKa3kSN32ZHVZ7ZLPLXWvOVIiZtyJnO1gPH1sKt0=
+github.com/spf13/cobra v1.10.2 h1:DMTTonx5m65Ic0GOoRY2c16WCbHxOOw6xxezuLaBpcU=
+github.com/spf13/cobra v1.10.2/go.mod h1:7C1pvHqHw5A4vrJfjNwvOdzYu0Gml16OCs2GRiTUUS4=
+github.com/spf13/pflag v1.0.9 h1:9exaQaMOCwffKiiiYk6/BndUBv+iRViNW+4lEMi0PvY=
+github.com/spf13/pflag v1.0.9/go.mod h1:McXfInJRrz4CZXVZOBLb0bTZqETkiAhM9Iw0y3An2Bg=
 github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME=
 github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI=
 github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg=
@@ -73,6 +81,7 @@ github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=
 github.com/x448/float16 v0.8.4/go.mod h1:14CWIYCyZA/cWjXOioeEpHeN/83MdbZDRQHoFcYsOfg=
 github.com/yosida95/uritemplate/v3 v3.0.2 h1:Ed3Oyj9yrmi9087+NczuL5BwkIc4wvTb5zIM+UJPGz4=
 github.com/yosida95/uritemplate/v3 v3.0.2/go.mod h1:ILOh0sOhIJR3+L/8afwt/kE++YT040gmv5BQTMR2HP4=
+go.yaml.in/yaml/v3 v3.0.4/go.mod h1:DhzuOOF2ATzADvBadXxruRBLzYTpT36CKvDb3+aBEFg=
 golang.org/x/crypto v0.36.0 h1:AnAEvhDddvBdpY+uR+MyHmuZzzNqXSe/GvuDeob5L34=
 golang.org/x/crypto v0.36.0/go.mod h1:Y4J0ReaxCR1IMaabaSMugxJES1EpwhBHhv2bDHklZvc=
 golang.org/x/oauth2 v0.34.0 h1:hqK/t4AKgbqWkdkcAeI8XLmbK+4m4G5YeQRrmiotGlw=

internal/ai/compat/client.go

@@ -14,7 +14,6 @@ import (
 	"regexp"
 	"slices"
 	"strings"
-	"sync"
 	"time"
 
 	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
@@ -36,38 +35,41 @@ Rules:
 - If unsure, prefer "observation".
 - Do not include any text outside the JSON object.`
 
+// Client is a low-level OpenAI-compatible HTTP client. It knows nothing about
+// role chains, fallbacks, or health — those concerns belong to ai.Runner. Each
+// method takes the model name per-call so a single Client instance can service
+// many different models on the same base URL.
 type Client struct {
 	name       string
 	baseURL    string
 	apiKey     string
-	embeddingModel         string
-	metadataModel          string
-	fallbackMetadataModels []string
-	temperature            float64
 	headers    map[string]string
 	httpClient *http.Client
 	log        *slog.Logger
-	dimensions             int
-	logConversations       bool
-	modelHealthMu          sync.Mutex
-	modelHealth            map[string]modelHealthState
 }
 
 type Config struct {
 	Name       string
 	BaseURL    string
 	APIKey     string
-	EmbeddingModel         string
-	MetadataModel          string
-	FallbackMetadataModels []string
-	Temperature            float64
 	Headers    map[string]string
 	HTTPClient *http.Client
 	Log        *slog.Logger
-	Dimensions             int
+}
+
+// MetadataOptions control a single ExtractMetadataWith call.
+type MetadataOptions struct {
+	Model            string
+	Temperature      float64
 	LogConversations bool
 }
 
+// SummarizeOptions control a single SummarizeWith call.
+type SummarizeOptions struct {
+	Model       string
+	Temperature float64
+}
+
 type embeddingsRequest struct {
 	Input string `json:"input"`
 	Model string `json:"model"`
@@ -127,65 +129,38 @@ type providerError struct {
 
 const maxMetadataAttempts = 3
 
-const (
+// ErrEmptyResponse and ErrNoJSONObject are sentinel errors callers can inspect
-	emptyResponseCircuitThreshold = 3
+// to classify metadata failures (e.g. bump empty-response health counters).
-	emptyResponseCircuitTTL       = 5 * time.Minute
-	permanentModelFailureTTL      = 24 * time.Hour
-)
-
 var (
-	errMetadataEmptyResponse = errors.New("metadata empty response")
+	ErrEmptyResponse = errors.New("metadata empty response")
-	errMetadataNoJSONObject  = errors.New("metadata response contains no JSON object")
+	ErrNoJSONObject  = errors.New("metadata response contains no JSON object")
 )
 
-type modelHealthState struct {
-	consecutiveEmpty int
-	unhealthyUntil   time.Time
-}
-
 func New(cfg Config) *Client {
-	fallbacks := make([]string, 0, len(cfg.FallbackMetadataModels))
-	seen := make(map[string]struct{}, len(cfg.FallbackMetadataModels))
-	for _, model := range cfg.FallbackMetadataModels {
-		model = strings.TrimSpace(model)
-		if model == "" {
-			continue
-		}
-		if _, ok := seen[model]; ok {
-			continue
-		}
-		seen[model] = struct{}{}
-		fallbacks = append(fallbacks, model)
-	}
-
 	return &Client{
 		name:       cfg.Name,
 		baseURL:    cfg.BaseURL,
 		apiKey:     cfg.APIKey,
-		embeddingModel:         cfg.EmbeddingModel,
-		metadataModel:          cfg.MetadataModel,
-		fallbackMetadataModels: fallbacks,
-		temperature:            cfg.Temperature,
 		headers:    cfg.Headers,
 		httpClient: cfg.HTTPClient,
 		log:        cfg.Log,
-		dimensions:             cfg.Dimensions,
-		logConversations:       cfg.LogConversations,
-		modelHealth:            make(map[string]modelHealthState),
 	}
 }
 
-func (c *Client) Embed(ctx context.Context, input string) ([]float32, error) {
+func (c *Client) Name() string { return c.name }
+
+// EmbedWith generates an embedding for the given input using model.
+func (c *Client) EmbedWith(ctx context.Context, model, input string) ([]float32, error) {
 	input = strings.TrimSpace(input)
 	if input == "" {
 		return nil, fmt.Errorf("%s embed: input must not be empty", c.name)
 	}
+	if strings.TrimSpace(model) == "" {
+		return nil, fmt.Errorf("%s embed: model is required", c.name)
+	}
 
 	var resp embeddingsResponse
-	err := c.doJSON(ctx, "/embeddings", embeddingsRequest{
+	err := c.doJSON(ctx, "/embeddings", embeddingsRequest{Input: input, Model: model}, &resp)
-		Input: input,
-		Model: c.embeddingModel,
-	}, &resp)
 	if err != nil {
 		return nil, err
 	}
@@ -195,133 +170,26 @@ func (c *Client) Embed(ctx context.Context, input string) ([]float32, error) {
 	if len(resp.Data) == 0 {
 		return nil, fmt.Errorf("%s embed: no embedding returned", c.name)
 	}
-	if c.dimensions > 0 && len(resp.Data[0].Embedding) != c.dimensions {
-		return nil, fmt.Errorf("%s embed: expected %d dimensions, got %d", c.name, c.dimensions, len(resp.Data[0].Embedding))
-	}
-
 	return resp.Data[0].Embedding, nil
 }
 
-func (c *Client) ExtractMetadata(ctx context.Context, input string) (thoughttypes.ThoughtMetadata, error) {
+// ExtractMetadataWith extracts structured metadata for input using opts.Model.
+// Returns compat.ErrEmptyResponse / ErrNoJSONObject wrapped when the model
+// produces unusable output so callers can classify the failure.
+func (c *Client) ExtractMetadataWith(ctx context.Context, opts MetadataOptions, input string) (thoughttypes.ThoughtMetadata, error) {
 	input = strings.TrimSpace(input)
 	if input == "" {
 		return thoughttypes.ThoughtMetadata{}, fmt.Errorf("%s extract metadata: input must not be empty", c.name)
 	}
-
+	if strings.TrimSpace(opts.Model) == "" {
-	start := time.Now()
+		return thoughttypes.ThoughtMetadata{}, fmt.Errorf("%s extract metadata: model is required", c.name)
-	if c.log != nil {
-		c.log.Info("metadata client started",
-			slog.String("provider", c.name),
-			slog.String("model", c.metadataModel),
-		)
-	}
-
-	logCompletion := func(model string, err error) {
-		if c.log == nil {
-			return
-		}
-
-		attrs := []any{
-			slog.String("provider", c.name),
-			slog.String("model", model),
-			slog.String("duration", formatLogDuration(time.Since(start))),
-		}
-		if err != nil {
-			attrs = append(attrs, slog.String("error", err.Error()))
-			c.log.Error("metadata client completed", attrs...)
-			return
-		}
-
-		c.log.Info("metadata client completed", attrs...)
-	}
-
-	result, err := c.extractMetadataWithModel(ctx, input, c.metadataModel)
-	if errors.Is(err, errMetadataEmptyResponse) {
-		c.noteEmptyResponse(c.metadataModel)
-	}
-	if isPermanentModelError(err) {
-		c.notePermanentModelFailure(c.metadataModel, err)
-	}
-	if err == nil {
-		c.noteModelSuccess(c.metadataModel)
-		logCompletion(c.metadataModel, nil)
-		return result, nil
-	}
-
-	for _, fallbackModel := range c.fallbackMetadataModels {
-		if ctx.Err() != nil {
-			break
-		}
-		if fallbackModel == "" || fallbackModel == c.metadataModel {
-			continue
-		}
-		if c.shouldBypassModel(fallbackModel) {
-			continue
-		}
-		if c.log != nil {
-			c.log.Warn("metadata extraction failed, trying fallback model",
-				slog.String("provider", c.name),
-				slog.String("primary_model", c.metadataModel),
-				slog.String("fallback_model", fallbackModel),
-				slog.String("error", err.Error()),
-			)
-		}
-		fallbackResult, fallbackErr := c.extractMetadataWithModel(ctx, input, fallbackModel)
-		if errors.Is(fallbackErr, errMetadataEmptyResponse) {
-			c.noteEmptyResponse(fallbackModel)
-		}
-		if isPermanentModelError(fallbackErr) {
-			c.notePermanentModelFailure(fallbackModel, fallbackErr)
-		}
-		if fallbackErr == nil {
-			c.noteModelSuccess(fallbackModel)
-			logCompletion(fallbackModel, nil)
-			return fallbackResult, nil
-		}
-		err = fallbackErr
-	}
-
-	if ctx.Err() != nil {
-		err = fmt.Errorf("%s metadata: %w", c.name, ctx.Err())
-		logCompletion(c.metadataModel, err)
-		return thoughttypes.ThoughtMetadata{}, err
-	}
-
-	heuristic := heuristicMetadataFromInput(input)
-	if c.log != nil {
-		c.log.Warn("metadata extraction failed for all models, using heuristic fallback",
-			slog.String("provider", c.name),
-			slog.String("error", err.Error()),
-		)
-	}
-	logCompletion(c.metadataModel, nil)
-	return heuristic, nil
-}
-
-func formatLogDuration(d time.Duration) string {
-	if d < 0 {
-		d = -d
-	}
-
-	totalMilliseconds := d.Milliseconds()
-	minutes := totalMilliseconds / 60000
-	seconds := (totalMilliseconds / 1000) % 60
-	milliseconds := totalMilliseconds % 1000
-	return fmt.Sprintf("%02d:%02d:%03d", minutes, seconds, milliseconds)
-}
-
-func (c *Client) extractMetadataWithModel(ctx context.Context, input, model string) (thoughttypes.ThoughtMetadata, error) {
-	if c.shouldBypassModel(model) {
-		return thoughttypes.ThoughtMetadata{}, fmt.Errorf("%s metadata: model %q temporarily bypassed after repeated empty responses", c.name, model)
 	}
 
 	stream := true
 	req := chatCompletionsRequest{
-		Model:       model,
+		Model:          opts.Model,
-		Temperature: c.temperature,
+		Temperature:    opts.Temperature,
-		ResponseFormat: &responseType{
+		ResponseFormat: &responseType{Type: "json_object"},
-			Type: "json_object",
-		},
 		Stream:         &stream,
 		Messages: []chatMessage{
 			{Role: "system", Content: metadataSystemPrompt},
@@ -329,7 +197,7 @@ func (c *Client) extractMetadataWithModel(ctx context.Context, input, model stri
 		},
 	}
 
-	metadata, err := c.extractMetadataWithRequest(ctx, req, input, model)
+	metadata, err := c.extractMetadataWithRequest(ctx, req, input, opts)
 	if err == nil || !shouldRetryWithoutJSONMode(err) {
 		return metadata, err
 	}
@@ -337,23 +205,22 @@ func (c *Client) extractMetadataWithModel(ctx context.Context, input, model stri
 	if c.log != nil {
 		c.log.Warn("metadata json mode failed, retrying without response_format",
 			slog.String("provider", c.name),
-			slog.String("model", model),
+			slog.String("model", opts.Model),
 			slog.String("error", err.Error()),
 		)
 	}
 
 	req.ResponseFormat = nil
-	return c.extractMetadataWithRequest(ctx, req, input, model)
+	return c.extractMetadataWithRequest(ctx, req, input, opts)
 }
 
-func (c *Client) extractMetadataWithRequest(ctx context.Context, req chatCompletionsRequest, input, model string) (thoughttypes.ThoughtMetadata, error) {
+func (c *Client) extractMetadataWithRequest(ctx context.Context, req chatCompletionsRequest, input string, opts MetadataOptions) (thoughttypes.ThoughtMetadata, error) {
-
 	var lastErr error
 	for attempt := 1; attempt <= maxMetadataAttempts; attempt++ {
-		if c.logConversations && c.log != nil {
+		if opts.LogConversations && c.log != nil {
 			c.log.Info("metadata conversation request",
 				slog.String("provider", c.name),
-				slog.String("model", model),
+				slog.String("model", opts.Model),
 				slog.Int("attempt", attempt),
 				slog.String("system", metadataSystemPrompt),
 				slog.String("input", input),
@@ -373,10 +240,10 @@ func (c *Client) extractMetadataWithRequest(ctx context.Context, req chatComplet
 
 		rawResponse := extractChoiceText(resp.Choices[0].Message, resp.Choices[0].Text)
 
-		if c.logConversations && c.log != nil {
+		if opts.LogConversations && c.log != nil {
 			c.log.Info("metadata conversation response",
 				slog.String("provider", c.name),
-				slog.String("model", model),
+				slog.String("model", opts.Model),
 				slog.Int("attempt", attempt),
 				slog.String("response", rawResponse),
 			)
@@ -387,13 +254,13 @@ func (c *Client) extractMetadataWithRequest(ctx context.Context, req chatComplet
 		metadataText = stripCodeFence(metadataText)
 		metadataText = extractJSONObject(metadataText)
 		if metadataText == "" {
-			lastErr = fmt.Errorf("%s metadata: %w", c.name, errMetadataNoJSONObject)
+			lastErr = fmt.Errorf("%s metadata: %w", c.name, ErrNoJSONObject)
 			if strings.TrimSpace(rawResponse) == "" && attempt < maxMetadataAttempts && ctx.Err() == nil {
-				lastErr = fmt.Errorf("%s metadata: %w", c.name, errMetadataEmptyResponse)
+				lastErr = fmt.Errorf("%s metadata: %w", c.name, ErrEmptyResponse)
 				if c.log != nil {
 					c.log.Warn("metadata response empty, waiting and retrying",
 						slog.String("provider", c.name),
-						slog.String("model", model),
+						slog.String("model", opts.Model),
 						slog.Int("attempt", attempt+1),
 					)
 				}
@@ -403,7 +270,7 @@ func (c *Client) extractMetadataWithRequest(ctx context.Context, req chatComplet
 				continue
 			}
 			if strings.TrimSpace(rawResponse) == "" {
-				lastErr = fmt.Errorf("%s metadata: %w", c.name, errMetadataEmptyResponse)
+				lastErr = fmt.Errorf("%s metadata: %w", c.name, ErrEmptyResponse)
 			}
 			return thoughttypes.ThoughtMetadata{}, lastErr
 		}
@@ -420,13 +287,17 @@ func (c *Client) extractMetadataWithRequest(ctx context.Context, req chatComplet
 	if lastErr != nil {
 		return thoughttypes.ThoughtMetadata{}, lastErr
 	}
-	return thoughttypes.ThoughtMetadata{}, fmt.Errorf("%s metadata: %w", c.name, errMetadataNoJSONObject)
+	return thoughttypes.ThoughtMetadata{}, fmt.Errorf("%s metadata: %w", c.name, ErrNoJSONObject)
 }
 
-func (c *Client) Summarize(ctx context.Context, systemPrompt, userPrompt string) (string, error) {
+// SummarizeWith runs a chat-completion summarisation using opts.Model.
+func (c *Client) SummarizeWith(ctx context.Context, opts SummarizeOptions, systemPrompt, userPrompt string) (string, error) {
+	if strings.TrimSpace(opts.Model) == "" {
+		return "", fmt.Errorf("%s summarize: model is required", c.name)
+	}
 	req := chatCompletionsRequest{
-		Model:       c.metadataModel,
+		Model:       opts.Model,
-		Temperature: 0.2,
+		Temperature: opts.Temperature,
 		Messages: []chatMessage{
 			{Role: "system", Content: systemPrompt},
 			{Role: "user", Content: userPrompt},
@@ -447,12 +318,49 @@ func (c *Client) Summarize(ctx context.Context, systemPrompt, userPrompt string)
 	return extractChoiceText(resp.Choices[0].Message, resp.Choices[0].Text), nil
 }
 
-func (c *Client) Name() string {
+// IsPermanentModelError reports whether err indicates the model itself is
-	return c.name
+// invalid or missing (vs. a transient outage). Runners use this to mark a
+// target unhealthy for longer.
+func IsPermanentModelError(err error) bool {
+	if err == nil {
+		return false
+	}
+	lower := strings.ToLower(err.Error())
+	for _, marker := range []string{
+		"invalid model name",
+		"model_not_found",
+		"model not found",
+		"unknown model",
+		"no such model",
+		"does not exist",
+	} {
+		if strings.Contains(lower, marker) {
+			return true
+		}
+	}
+	return false
 }
 
-func (c *Client) EmbeddingModel() string {
+// HeuristicMetadataFromInput produces best-effort metadata from the note text
-	return c.embeddingModel
+// when every model in the chain has failed. Exported so ai.Runner can use it.
+func HeuristicMetadataFromInput(input string) thoughttypes.ThoughtMetadata {
+	text := strings.TrimSpace(input)
+	lower := strings.ToLower(text)
+
+	metadata := thoughttypes.ThoughtMetadata{
+		People:         heuristicPeople(text),
+		ActionItems:    heuristicActionItems(text),
+		DatesMentioned: heuristicDates(text),
+		Topics:         heuristicTopics(lower),
+		Type:           heuristicType(lower),
+	}
+	if len(metadata.Topics) == 0 {
+		metadata.Topics = []string{"uncategorized"}
+	}
+	if metadata.Type == "" {
+		metadata.Type = "observation"
+	}
+	return metadata
 }
 
 func (c *Client) doJSON(ctx context.Context, path string, requestBody any, dest any) error {
@@ -724,8 +632,6 @@ func isRetryableChatResponseError(err error) bool {
 	return strings.Contains(lower, "read response") || strings.Contains(lower, "read stream response")
 }
 
-// extractJSONObject finds the first complete {...} block in s.
-// It handles models that prepend prose to a JSON response despite json_object mode.
 func extractJSONObject(s string) string {
 	for start := 0; start < len(s); start++ {
 		if s[start] != '{' {
@@ -768,10 +674,6 @@ func extractJSONObject(s string) string {
 	return ""
 }
 
-// stripThinkingBlocks removes <think>...</think> and <thinking>...</thinking>
-// blocks produced by reasoning models (DeepSeek R1, QwQ, etc.) so that the
-// remaining text can be parsed as JSON without interference from thinking content
-// that may itself contain braces.
 func stripThinkingBlocks(s string) string {
 	for _, tag := range []string{"think", "thinking"} {
 		open := "<" + tag + ">"
@@ -857,7 +759,6 @@ func extractTextFromAny(value any) string {
 		}
 		return strings.Join(parts, "\n")
 	case map[string]any:
-		// Common provider shapes for chat content parts.
 		for _, key := range []string{"text", "output_text", "content", "value"} {
 			if nested, ok := typed[key]; ok {
 				if text := strings.TrimSpace(extractTextFromAny(nested)); text != "" {
@@ -875,28 +776,6 @@ var (
 	wordPattern      = regexp.MustCompile(`[a-zA-Z][a-zA-Z0-9_/-]{2,}`)
 )
 
-func heuristicMetadataFromInput(input string) thoughttypes.ThoughtMetadata {
-	text := strings.TrimSpace(input)
-	lower := strings.ToLower(text)
-
-	metadata := thoughttypes.ThoughtMetadata{
-		People:         heuristicPeople(text),
-		ActionItems:    heuristicActionItems(text),
-		DatesMentioned: heuristicDates(text),
-		Topics:         heuristicTopics(lower),
-		Type:           heuristicType(lower),
-		Source:         "",
-	}
-
-	if len(metadata.Topics) == 0 {
-		metadata.Topics = []string{"uncategorized"}
-	}
-	if metadata.Type == "" {
-		metadata.Type = "observation"
-	}
-	return metadata
-}
-
 func heuristicType(lower string) string {
 	switch {
 	case strings.Contains(lower, "preferred name"), strings.Contains(lower, "personal profile"), strings.Contains(lower, "wife:"), strings.Contains(lower, "daughter:"), strings.Contains(lower, "born:"):
@@ -1055,7 +934,7 @@ func shouldRetryWithoutJSONMode(err error) bool {
 	if err == nil {
 		return false
 	}
-	if errors.Is(err, errMetadataEmptyResponse) || errors.Is(err, errMetadataNoJSONObject) {
+	if errors.Is(err, ErrEmptyResponse) || errors.Is(err, ErrNoJSONObject) {
 		return true
 	}
 
@@ -1063,27 +942,6 @@ func shouldRetryWithoutJSONMode(err error) bool {
 	return strings.Contains(lower, "parse json")
 }
 
-func isPermanentModelError(err error) bool {
-	if err == nil {
-		return false
-	}
-
-	lower := strings.ToLower(err.Error())
-	for _, marker := range []string{
-		"invalid model name",
-		"model_not_found",
-		"model not found",
-		"unknown model",
-		"no such model",
-		"does not exist",
-	} {
-		if strings.Contains(lower, marker) {
-			return true
-		}
-	}
-	return false
-}
-
 func sleepRetry(ctx context.Context, attempt int, log *slog.Logger, provider string) error {
 	delay := time.Duration(attempt*attempt) * 200 * time.Millisecond
 	if log != nil {
@@ -1110,59 +968,3 @@ func sleepMetadataRetry(ctx context.Context, attempt int) error {
 		return nil
 	}
 }
-
-func (c *Client) shouldBypassModel(model string) bool {
-	c.modelHealthMu.Lock()
-	defer c.modelHealthMu.Unlock()
-
-	state, ok := c.modelHealth[model]
-	if !ok {
-		return false
-	}
-	return !state.unhealthyUntil.IsZero() && time.Now().Before(state.unhealthyUntil)
-}
-
-func (c *Client) noteEmptyResponse(model string) {
-	c.modelHealthMu.Lock()
-	defer c.modelHealthMu.Unlock()
-
-	state := c.modelHealth[model]
-	state.consecutiveEmpty++
-	if state.consecutiveEmpty >= emptyResponseCircuitThreshold {
-		state.unhealthyUntil = time.Now().Add(emptyResponseCircuitTTL)
-		if c.log != nil {
-			c.log.Warn("metadata model marked temporarily unhealthy after repeated empty responses",
-				slog.String("provider", c.name),
-				slog.String("model", model),
-				slog.Time("until", state.unhealthyUntil),
-			)
-		}
-	}
-	c.modelHealth[model] = state
-}
-
-func (c *Client) noteModelSuccess(model string) {
-	c.modelHealthMu.Lock()
-	defer c.modelHealthMu.Unlock()
-
-	delete(c.modelHealth, model)
-}
-
-func (c *Client) notePermanentModelFailure(model string, err error) {
-	c.modelHealthMu.Lock()
-	defer c.modelHealthMu.Unlock()
-
-	state := c.modelHealth[model]
-	state.consecutiveEmpty = emptyResponseCircuitThreshold
-	state.unhealthyUntil = time.Now().Add(permanentModelFailureTTL)
-	c.modelHealth[model] = state
-
-	if c.log != nil {
-		c.log.Warn("metadata model marked unhealthy after permanent failure",
-			slog.String("provider", c.name),
-			slog.String("model", model),
-			slog.String("error", err.Error()),
-			slog.Time("until", state.unhealthyUntil),
-		)
-	}
-}

internal/ai/compat/client_test.go

@@ -11,6 +11,17 @@ import (
 	"testing"
 )
 
+func newTestClient(t *testing.T, url string) *Client {
+	t.Helper()
+	return New(Config{
+		Name:       "litellm",
+		BaseURL:    url,
+		APIKey:     "test-key",
+		HTTPClient: http.DefaultClient,
+		Log:        slog.New(slog.NewTextHandler(io.Discard, nil)),
+	})
+}
+
 func TestExtractMetadataFromStreamingResponse(t *testing.T) {
 	t.Parallel()
 
@@ -26,6 +37,9 @@ func TestExtractMetadataFromStreamingResponse(t *testing.T) {
 		if req.Stream == nil || !*req.Stream {
 			t.Fatalf("stream flag = %v, want true", req.Stream)
 		}
+		if req.Model != "qwen3.5:latest" {
+			t.Fatalf("model = %q, want qwen3.5:latest", req.Model)
+		}
 
 		w.Header().Set("Content-Type", "text/event-stream")
 		_, _ = io.WriteString(w, "data: {\"choices\":[{\"delta\":{\"content\":\"{\\\"people\\\":[],\"}}]}\n\n")
@@ -35,20 +49,13 @@ func TestExtractMetadataFromStreamingResponse(t *testing.T) {
 	}))
 	defer server.Close()
 
-	client := New(Config{
+	client := newTestClient(t, server.URL)
-		Name:           "litellm",
+	metadata, err := client.ExtractMetadataWith(context.Background(), MetadataOptions{
-		BaseURL:        server.URL,
+		Model:       "qwen3.5:latest",
-		APIKey:         "test-key",
-		MetadataModel:  "qwen3.5:latest",
 		Temperature: 0.1,
-		HTTPClient:     server.Client(),
+	}, "Project idea: Build an Android companion app.")
-		Log:            slog.New(slog.NewTextHandler(io.Discard, nil)),
-		EmbeddingModel: "unused",
-	})
-
-	metadata, err := client.ExtractMetadata(context.Background(), "Project idea: Build an Android companion app.")
 	if err != nil {
-		t.Fatalf("ExtractMetadata() error = %v", err)
+		t.Fatalf("ExtractMetadataWith() error = %v", err)
 	}
 
 	if metadata.Type != "idea" {
@@ -94,20 +101,13 @@ func TestExtractMetadataRetriesWithoutJSONMode(t *testing.T) {
 	}))
 	defer server.Close()
 
-	client := New(Config{
+	client := newTestClient(t, server.URL)
-		Name:           "litellm",
+	metadata, err := client.ExtractMetadataWith(context.Background(), MetadataOptions{
-		BaseURL:        server.URL,
+		Model:       "qwen3.5:latest",
-		APIKey:         "test-key",
-		MetadataModel:  "qwen3.5:latest",
 		Temperature: 0.1,
-		HTTPClient:     server.Client(),
+	}, "Project idea: Build an Android companion app.")
-		Log:            slog.New(slog.NewTextHandler(io.Discard, nil)),
-		EmbeddingModel: "unused",
-	})
-
-	metadata, err := client.ExtractMetadata(context.Background(), "Project idea: Build an Android companion app.")
 	if err != nil {
-		t.Fatalf("ExtractMetadata() error = %v", err)
+		t.Fatalf("ExtractMetadataWith() error = %v", err)
 	}
 
 	if metadata.Type != "idea" {
@@ -127,71 +127,33 @@ func TestExtractMetadataRetriesWithoutJSONMode(t *testing.T) {
 	}
 }
 
-func TestExtractMetadataBypassesInvalidFallbackModelAfterFirstFailure(t *testing.T) {
+func TestIsPermanentModelError(t *testing.T) {
 	t.Parallel()
 
-	var mu sync.Mutex
+	cases := []struct {
-	primaryCalls := 0
+		name string
-	invalidFallbackCalls := 0
+		err  error
-
+		want bool
-	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	}{
-		defer func() {
+		{"nil", nil, false},
-			_ = r.Body.Close()
+		{"invalid model", errMsg("Invalid model name passed in model=qwen3"), true},
-		}()
+		{"model not found", errMsg("model_not_found"), true},
-
+		{"no such model", errMsg("no such model"), true},
-		var req chatCompletionsRequest
+		{"transient", errMsg("connection refused"), false},
-		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-			t.Fatalf("decode request: %v", err)
 	}
 
-		switch req.Model {
+	for _, tc := range cases {
-		case "empty-primary":
+		tc := tc
-			_, _ = io.WriteString(w, `{"choices":[{"message":{"role":"assistant","content":""}}]}`)
+		t.Run(tc.name, func(t *testing.T) {
-		case "qwen3.5:latest":
+			if got := IsPermanentModelError(tc.err); got != tc.want {
-			mu.Lock()
+				t.Fatalf("IsPermanentModelError(%v) = %v, want %v", tc.err, got, tc.want)
-			primaryCalls++
-			mu.Unlock()
-			_, _ = io.WriteString(w, `{"choices":[{"message":{"role":"assistant","content":"{\"people\":[],\"action_items\":[],\"dates_mentioned\":[],\"topics\":[\"metadata\"],\"type\":\"observation\",\"source\":\"primary\"}"}}]}`)
-		case "qwen3":
-			mu.Lock()
-			invalidFallbackCalls++
-			mu.Unlock()
-			w.WriteHeader(http.StatusBadRequest)
-			_, _ = io.WriteString(w, "{\"error\":{\"message\":\"{'error': '/chat/completions: Invalid model name passed in model=qwen3. Call `/v1/models` to view available models for your key.'}\"}}")
-		default:
-			t.Fatalf("unexpected model %q", req.Model)
 			}
-	}))
-	defer server.Close()
-
-	client := New(Config{
-		Name:                   "litellm",
-		BaseURL:                server.URL,
-		APIKey:                 "test-key",
-		MetadataModel:          "empty-primary",
-		FallbackMetadataModels: []string{"qwen3", "qwen3.5:latest"},
-		Temperature:            0.1,
-		HTTPClient:             server.Client(),
-		Log:                    slog.New(slog.NewTextHandler(io.Discard, nil)),
-		EmbeddingModel:         "unused",
 		})
-
-	for i := 0; i < 2; i++ {
-		metadata, err := client.ExtractMetadata(context.Background(), "A short note about metadata.")
-		if err != nil {
-			t.Fatalf("ExtractMetadata() error = %v", err)
-		}
-		if metadata.Source != "primary" {
-			t.Fatalf("metadata source = %q, want primary", metadata.Source)
-		}
-	}
-
-	mu.Lock()
-	defer mu.Unlock()
-	if invalidFallbackCalls != 1 {
-		t.Fatalf("invalid fallback calls = %d, want 1", invalidFallbackCalls)
-	}
-	if primaryCalls != 2 {
-		t.Fatalf("valid fallback calls = %d, want 2", primaryCalls)
 	}
 }
+
+type stringError string
+
+func (s stringError) Error() string { return string(s) }
+
+func errMsg(s string) error { return stringError(s) }

internal/ai/factory.go

@@ -1,25 +0,0 @@
-package ai
-
-import (
-	"fmt"
-	"log/slog"
-	"net/http"
-
-	"git.warky.dev/wdevs/amcs/internal/ai/litellm"
-	"git.warky.dev/wdevs/amcs/internal/ai/ollama"
-	"git.warky.dev/wdevs/amcs/internal/ai/openrouter"
-	"git.warky.dev/wdevs/amcs/internal/config"
-)
-
-func NewProvider(cfg config.AIConfig, httpClient *http.Client, log *slog.Logger) (Provider, error) {
-	switch cfg.Provider {
-	case "litellm":
-		return litellm.New(cfg, httpClient, log)
-	case "ollama":
-		return ollama.New(cfg, httpClient, log)
-	case "openrouter":
-		return openrouter.New(cfg, httpClient, log)
-	default:
-		return nil, fmt.Errorf("unsupported ai.provider: %s", cfg.Provider)
-	}
-}

internal/ai/factory_test.go

@@ -1,33 +0,0 @@
-package ai
-
-import (
-	"io"
-	"log/slog"
-	"net/http"
-	"testing"
-
-	"git.warky.dev/wdevs/amcs/internal/config"
-)
-
-func TestNewProviderSupportsOllama(t *testing.T) {
-	provider, err := NewProvider(config.AIConfig{
-		Provider: "ollama",
-		Embeddings: config.AIEmbeddingConfig{
-			Model:      "nomic-embed-text",
-			Dimensions: 768,
-		},
-		Metadata: config.AIMetadataConfig{
-			Model: "llama3.2",
-		},
-		Ollama: config.OllamaConfig{
-			BaseURL: "http://localhost:11434/v1",
-			APIKey:  "ollama",
-		},
-	}, &http.Client{}, slog.New(slog.NewTextHandler(io.Discard, nil)))
-	if err != nil {
-		t.Fatalf("NewProvider() error = %v", err)
-	}
-	if provider.Name() != "ollama" {
-		t.Fatalf("provider name = %q, want ollama", provider.Name())
-	}
-}

internal/ai/litellm/client.go

@@ -1,30 +0,0 @@
-package litellm
-
-import (
-	"log/slog"
-	"net/http"
-
-	"git.warky.dev/wdevs/amcs/internal/ai/compat"
-	"git.warky.dev/wdevs/amcs/internal/config"
-)
-
-func New(cfg config.AIConfig, httpClient *http.Client, log *slog.Logger) (*compat.Client, error) {
-	fallbacks := cfg.LiteLLM.EffectiveFallbackMetadataModels()
-	if len(fallbacks) == 0 {
-		fallbacks = cfg.Metadata.EffectiveFallbackModels()
-	}
-	return compat.New(compat.Config{
-		Name:                   "litellm",
-		BaseURL:                cfg.LiteLLM.BaseURL,
-		APIKey:                 cfg.LiteLLM.APIKey,
-		EmbeddingModel:         cfg.LiteLLM.EmbeddingModel,
-		MetadataModel:          cfg.LiteLLM.MetadataModel,
-		FallbackMetadataModels: fallbacks,
-		Temperature:            cfg.Metadata.Temperature,
-		Headers:                cfg.LiteLLM.RequestHeaders,
-		HTTPClient:             httpClient,
-		Log:                    log,
-		Dimensions:             cfg.Embeddings.Dimensions,
-		LogConversations:       cfg.Metadata.LogConversations,
-	}), nil
-}

internal/ai/ollama/client.go

@@ -1,26 +0,0 @@
-package ollama
-
-import (
-	"log/slog"
-	"net/http"
-
-	"git.warky.dev/wdevs/amcs/internal/ai/compat"
-	"git.warky.dev/wdevs/amcs/internal/config"
-)
-
-func New(cfg config.AIConfig, httpClient *http.Client, log *slog.Logger) (*compat.Client, error) {
-	return compat.New(compat.Config{
-		Name:                   "ollama",
-		BaseURL:                cfg.Ollama.BaseURL,
-		APIKey:                 cfg.Ollama.APIKey,
-		EmbeddingModel:         cfg.Embeddings.Model,
-		MetadataModel:          cfg.Metadata.Model,
-		FallbackMetadataModels: cfg.Metadata.EffectiveFallbackModels(),
-		Temperature:            cfg.Metadata.Temperature,
-		Headers:                cfg.Ollama.RequestHeaders,
-		HTTPClient:             httpClient,
-		Log:                    log,
-		Dimensions:             cfg.Embeddings.Dimensions,
-		LogConversations:       cfg.Metadata.LogConversations,
-	}), nil
-}

internal/ai/openrouter/client.go

@@ -1,37 +0,0 @@
-package openrouter
-
-import (
-	"log/slog"
-	"net/http"
-
-	"git.warky.dev/wdevs/amcs/internal/ai/compat"
-	"git.warky.dev/wdevs/amcs/internal/config"
-)
-
-func New(cfg config.AIConfig, httpClient *http.Client, log *slog.Logger) (*compat.Client, error) {
-	headers := make(map[string]string, len(cfg.OpenRouter.ExtraHeaders)+2)
-	for key, value := range cfg.OpenRouter.ExtraHeaders {
-		headers[key] = value
-	}
-	if cfg.OpenRouter.SiteURL != "" {
-		headers["HTTP-Referer"] = cfg.OpenRouter.SiteURL
-	}
-	if cfg.OpenRouter.AppName != "" {
-		headers["X-Title"] = cfg.OpenRouter.AppName
-	}
-
-	return compat.New(compat.Config{
-		Name:                   "openrouter",
-		BaseURL:                cfg.OpenRouter.BaseURL,
-		APIKey:                 cfg.OpenRouter.APIKey,
-		EmbeddingModel:         cfg.Embeddings.Model,
-		MetadataModel:          cfg.Metadata.Model,
-		FallbackMetadataModels: cfg.Metadata.EffectiveFallbackModels(),
-		Temperature:            cfg.Metadata.Temperature,
-		Headers:                headers,
-		HTTPClient:             httpClient,
-		Log:                    log,
-		Dimensions:             cfg.Embeddings.Dimensions,
-		LogConversations:       cfg.Metadata.LogConversations,
-	}), nil
-}

internal/ai/provider.go

@@ -1,15 +0,0 @@
-package ai
-
-import (
-	"context"
-
-	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
-)
-
-type Provider interface {
-	Embed(ctx context.Context, input string) ([]float32, error)
-	ExtractMetadata(ctx context.Context, input string) (thoughttypes.ThoughtMetadata, error)
-	Summarize(ctx context.Context, systemPrompt, userPrompt string) (string, error)
-	Name() string
-	EmbeddingModel() string
-}

internal/ai/registry.go

@@ -0,0 +1,96 @@
+package ai
+
+import (
+	"fmt"
+	"log/slog"
+	"net/http"
+	"strings"
+
+	"git.warky.dev/wdevs/amcs/internal/ai/compat"
+	"git.warky.dev/wdevs/amcs/internal/config"
+)
+
+// Registry holds one compat.Client per named provider. Runners look up clients
+// by provider name when walking a role chain.
+type Registry struct {
+	clients map[string]*compat.Client
+}
+
+// NewRegistry builds a Registry from the configured providers. Each provider
+// type maps onto a compat.Client with type-specific header plumbing (e.g.
+// openrouter's HTTP-Referer / X-Title).
+func NewRegistry(providers map[string]config.ProviderConfig, httpClient *http.Client, log *slog.Logger) (*Registry, error) {
+	if httpClient == nil {
+		return nil, fmt.Errorf("ai registry: http client is required")
+	}
+	if len(providers) == 0 {
+		return nil, fmt.Errorf("ai registry: no providers configured")
+	}
+
+	clients := make(map[string]*compat.Client, len(providers))
+	for name, p := range providers {
+		headers, err := providerHeaders(p)
+		if err != nil {
+			return nil, fmt.Errorf("ai registry: provider %q: %w", name, err)
+		}
+		clients[name] = compat.New(compat.Config{
+			Name:       name,
+			BaseURL:    p.BaseURL,
+			APIKey:     p.APIKey,
+			Headers:    headers,
+			HTTPClient: httpClient,
+			Log:        log,
+		})
+	}
+	return &Registry{clients: clients}, nil
+}
+
+// Client returns the compat.Client registered under name.
+func (r *Registry) Client(name string) (*compat.Client, error) {
+	c, ok := r.clients[name]
+	if !ok {
+		return nil, fmt.Errorf("ai registry: provider %q is not configured", name)
+	}
+	return c, nil
+}
+
+// Names returns the registered provider names.
+func (r *Registry) Names() []string {
+	names := make([]string, 0, len(r.clients))
+	for name := range r.clients {
+		names = append(names, name)
+	}
+	return names
+}
+
+func providerHeaders(p config.ProviderConfig) (map[string]string, error) {
+	switch p.Type {
+	case "litellm", "ollama":
+		return cloneHeaders(p.RequestHeaders), nil
+	case "openrouter":
+		headers := cloneHeaders(p.RequestHeaders)
+		if headers == nil {
+			headers = map[string]string{}
+		}
+		if s := strings.TrimSpace(p.SiteURL); s != "" {
+			headers["HTTP-Referer"] = s
+		}
+		if s := strings.TrimSpace(p.AppName); s != "" {
+			headers["X-Title"] = s
+		}
+		return headers, nil
+	default:
+		return nil, fmt.Errorf("unsupported provider type %q", p.Type)
+	}
+}
+
+func cloneHeaders(in map[string]string) map[string]string {
+	if len(in) == 0 {
+		return nil
+	}
+	out := make(map[string]string, len(in))
+	for k, v := range in {
+		out[k] = v
+	}
+	return out
+}

internal/ai/registry_test.go

@@ -0,0 +1,80 @@
+package ai
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.warky.dev/wdevs/amcs/internal/ai/compat"
+	"git.warky.dev/wdevs/amcs/internal/config"
+)
+
+func TestNewRegistryOpenRouterHeaders(t *testing.T) {
+	var (
+		gotReferer string
+		gotTitle   string
+		gotCustom  string
+	)
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotReferer = r.Header.Get("HTTP-Referer")
+		gotTitle = r.Header.Get("X-Title")
+		gotCustom = r.Header.Get("X-Custom")
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"choices": []map[string]any{{"message": map[string]any{"role": "assistant", "content": "ok"}}},
+		})
+	}))
+	defer srv.Close()
+
+	providers := map[string]config.ProviderConfig{
+		"router": {
+			Type:    "openrouter",
+			BaseURL: srv.URL,
+			APIKey:  "secret",
+			RequestHeaders: map[string]string{
+				"X-Custom": "value",
+			},
+			AppName: "amcs",
+			SiteURL: "https://example.com",
+		},
+	}
+
+	reg, err := NewRegistry(providers, srv.Client(), nil)
+	if err != nil {
+		t.Fatalf("NewRegistry() error = %v", err)
+	}
+
+	client, err := reg.Client("router")
+	if err != nil {
+		t.Fatalf("Client(router) error = %v", err)
+	}
+
+	if _, err := client.SummarizeWith(context.Background(), compat.SummarizeOptions{Model: "gpt-4.1-mini"}, "system", "user"); err != nil {
+		t.Fatalf("SummarizeWith() error = %v", err)
+	}
+	if gotReferer != "https://example.com" {
+		t.Fatalf("HTTP-Referer = %q, want https://example.com", gotReferer)
+	}
+	if gotTitle != "amcs" {
+		t.Fatalf("X-Title = %q, want amcs", gotTitle)
+	}
+	if gotCustom != "value" {
+		t.Fatalf("X-Custom = %q, want value", gotCustom)
+	}
+}
+
+func TestNewRegistryRejectsUnsupportedProviderType(t *testing.T) {
+	providers := map[string]config.ProviderConfig{
+		"bad": {
+			Type:    "unknown",
+			BaseURL: "http://localhost:4000/v1",
+			APIKey:  "secret",
+		},
+	}
+
+	_, err := NewRegistry(providers, &http.Client{}, nil)
+	if err == nil {
+		t.Fatal("NewRegistry() error = nil, want unsupported provider type error")
+	}
+}

internal/ai/runner.go

@@ -0,0 +1,367 @@
+package ai
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"log/slog"
+	"sync"
+	"time"
+
+	"git.warky.dev/wdevs/amcs/internal/ai/compat"
+	"git.warky.dev/wdevs/amcs/internal/config"
+	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
+)
+
+// Health TTLs per failure class. These are short enough that a healed target
+// gets retried without manual intervention, but long enough to avoid hammering
+// a broken provider every call.
+const (
+	transientCooldown        = 30 * time.Second
+	permanentCooldown        = 10 * time.Minute
+	emptyResponseThreshold   = 3
+	emptyResponseCooldown    = 2 * time.Minute
+	dimensionMismatchWarning = "embedding dimension mismatch"
+)
+
+// EmbedResult carries the vector plus the (provider, model) that produced it —
+// callers store the actual model so later searches against that row use the
+// matching query embedding.
+type EmbedResult struct {
+	Vector   []float32
+	Provider string
+	Model    string
+}
+
+// EmbeddingRunner executes the embeddings role chain with sequential fallback.
+type EmbeddingRunner struct {
+	registry   *Registry
+	chain      []config.RoleTarget
+	dimensions int
+	health     *healthTracker
+	log        *slog.Logger
+}
+
+// MetadataRunner executes the metadata role chain with sequential fallback and
+// a heuristic fallthrough when every target is unhealthy or fails.
+type MetadataRunner struct {
+	registry *Registry
+	chain    []config.RoleTarget
+	opts     metadataRunOpts
+	health   *healthTracker
+	log      *slog.Logger
+}
+
+type metadataRunOpts struct {
+	temperature      float64
+	logConversations bool
+}
+
+// NewEmbeddingRunner builds a runner for the embeddings role. chain must be
+// non-empty and every target must be registered.
+func NewEmbeddingRunner(registry *Registry, chain []config.RoleTarget, dimensions int, log *slog.Logger) (*EmbeddingRunner, error) {
+	if registry == nil {
+		return nil, fmt.Errorf("embedding runner: registry is required")
+	}
+	if len(chain) == 0 {
+		return nil, fmt.Errorf("embedding runner: chain is empty")
+	}
+	if dimensions <= 0 {
+		return nil, fmt.Errorf("embedding runner: dimensions must be > 0")
+	}
+	for i, t := range chain {
+		if _, err := registry.Client(t.Provider); err != nil {
+			return nil, fmt.Errorf("embedding runner: chain[%d]: %w", i, err)
+		}
+	}
+	return &EmbeddingRunner{
+		registry:   registry,
+		chain:      chain,
+		dimensions: dimensions,
+		health:     newHealthTracker(),
+		log:        log,
+	}, nil
+}
+
+// NewMetadataRunner builds a runner for the metadata role.
+func NewMetadataRunner(registry *Registry, chain []config.RoleTarget, temperature float64, logConversations bool, log *slog.Logger) (*MetadataRunner, error) {
+	if registry == nil {
+		return nil, fmt.Errorf("metadata runner: registry is required")
+	}
+	if len(chain) == 0 {
+		return nil, fmt.Errorf("metadata runner: chain is empty")
+	}
+	for i, t := range chain {
+		if _, err := registry.Client(t.Provider); err != nil {
+			return nil, fmt.Errorf("metadata runner: chain[%d]: %w", i, err)
+		}
+	}
+	return &MetadataRunner{
+		registry: registry,
+		chain:    chain,
+		opts: metadataRunOpts{
+			temperature:      temperature,
+			logConversations: logConversations,
+		},
+		health: newHealthTracker(),
+		log:    log,
+	}, nil
+}
+
+// PrimaryProvider returns the first provider in the chain.
+func (r *EmbeddingRunner) PrimaryProvider() string { return r.chain[0].Provider }
+
+// PrimaryModel returns the first model in the chain — the one used as the
+// storage key for search matching.
+func (r *EmbeddingRunner) PrimaryModel() string { return r.chain[0].Model }
+
+// Dimensions returns the required vector dimension.
+func (r *EmbeddingRunner) Dimensions() int { return r.dimensions }
+
+// Embed walks the chain and returns the first successful embedding. The
+// returned EmbedResult names the actual (provider, model) that produced the
+// vector — callers use that when recording the row.
+func (r *EmbeddingRunner) Embed(ctx context.Context, input string) (EmbedResult, error) {
+	var errs []error
+	for _, target := range r.chain {
+		if r.health.skip(target) {
+			continue
+		}
+		client, err := r.registry.Client(target.Provider)
+		if err != nil {
+			errs = append(errs, err)
+			continue
+		}
+		vec, err := client.EmbedWith(ctx, target.Model, input)
+		if err != nil {
+			if ctx.Err() != nil {
+				return EmbedResult{}, ctx.Err()
+			}
+			r.classify(target, err)
+			r.logFailure("embed", target, err)
+			errs = append(errs, fmt.Errorf("%s/%s: %w", target.Provider, target.Model, err))
+			continue
+		}
+		if len(vec) != r.dimensions {
+			dimErr := fmt.Errorf("%s: expected %d, got %d", dimensionMismatchWarning, r.dimensions, len(vec))
+			r.health.markTransient(target)
+			r.logFailure("embed", target, dimErr)
+			errs = append(errs, fmt.Errorf("%s/%s: %w", target.Provider, target.Model, dimErr))
+			continue
+		}
+		r.health.markHealthy(target)
+		return EmbedResult{Vector: vec, Provider: target.Provider, Model: target.Model}, nil
+	}
+	return EmbedResult{}, fmt.Errorf("all embedding targets failed: %w", errors.Join(errs...))
+}
+
+// EmbedPrimary embeds using only the primary target — used for search queries
+// so the query vector matches rows stored under the primary model. Falls back
+// to returning the error without walking the chain.
+func (r *EmbeddingRunner) EmbedPrimary(ctx context.Context, input string) ([]float32, error) {
+	target := r.chain[0]
+	client, err := r.registry.Client(target.Provider)
+	if err != nil {
+		return nil, err
+	}
+	vec, err := client.EmbedWith(ctx, target.Model, input)
+	if err != nil {
+		r.classify(target, err)
+		return nil, err
+	}
+	if len(vec) != r.dimensions {
+		return nil, fmt.Errorf("%s: expected %d, got %d", dimensionMismatchWarning, r.dimensions, len(vec))
+	}
+	r.health.markHealthy(target)
+	return vec, nil
+}
+
+// PrimaryProvider / PrimaryModel for metadata mirror the embedding runner.
+func (r *MetadataRunner) PrimaryProvider() string { return r.chain[0].Provider }
+func (r *MetadataRunner) PrimaryModel() string    { return r.chain[0].Model }
+
+// ExtractMetadata walks the chain sequentially. If every target fails or is
+// unhealthy, it returns a heuristic metadata so capture never hard-fails.
+func (r *MetadataRunner) ExtractMetadata(ctx context.Context, input string) (thoughttypes.ThoughtMetadata, error) {
+	var errs []error
+	for _, target := range r.chain {
+		if r.health.skip(target) {
+			continue
+		}
+		client, err := r.registry.Client(target.Provider)
+		if err != nil {
+			errs = append(errs, err)
+			continue
+		}
+		md, err := client.ExtractMetadataWith(ctx, compat.MetadataOptions{
+			Model:            target.Model,
+			Temperature:      r.opts.temperature,
+			LogConversations: r.opts.logConversations,
+		}, input)
+		if err != nil {
+			if ctx.Err() != nil {
+				return thoughttypes.ThoughtMetadata{}, ctx.Err()
+			}
+			r.classify(target, err)
+			r.logFailure("metadata", target, err)
+			errs = append(errs, fmt.Errorf("%s/%s: %w", target.Provider, target.Model, err))
+			continue
+		}
+		r.health.markHealthy(target)
+		return md, nil
+	}
+	if r.log != nil {
+		r.log.Warn("metadata chain exhausted, using heuristic fallback",
+			slog.Int("targets", len(r.chain)),
+			slog.String("error", errors.Join(errs...).Error()),
+		)
+	}
+	return compat.HeuristicMetadataFromInput(input), nil
+}
+
+// Summarize walks the chain; unlike metadata, there is no heuristic fallback —
+// returns the joined error when everything fails.
+func (r *MetadataRunner) Summarize(ctx context.Context, systemPrompt, userPrompt string) (string, error) {
+	var errs []error
+	for _, target := range r.chain {
+		if r.health.skip(target) {
+			continue
+		}
+		client, err := r.registry.Client(target.Provider)
+		if err != nil {
+			errs = append(errs, err)
+			continue
+		}
+		out, err := client.SummarizeWith(ctx, compat.SummarizeOptions{
+			Model:       target.Model,
+			Temperature: r.opts.temperature,
+		}, systemPrompt, userPrompt)
+		if err != nil {
+			if ctx.Err() != nil {
+				return "", ctx.Err()
+			}
+			r.classify(target, err)
+			r.logFailure("summarize", target, err)
+			errs = append(errs, fmt.Errorf("%s/%s: %w", target.Provider, target.Model, err))
+			continue
+		}
+		r.health.markHealthy(target)
+		return out, nil
+	}
+	return "", fmt.Errorf("all summarize targets failed: %w", errors.Join(errs...))
+}
+
+func (r *EmbeddingRunner) classify(target config.RoleTarget, err error) {
+	switch {
+	case compat.IsPermanentModelError(err):
+		r.health.markPermanent(target)
+	default:
+		r.health.markTransient(target)
+	}
+}
+
+func (r *MetadataRunner) classify(target config.RoleTarget, err error) {
+	switch {
+	case compat.IsPermanentModelError(err):
+		r.health.markPermanent(target)
+	case errors.Is(err, compat.ErrEmptyResponse):
+		r.health.markEmpty(target)
+	default:
+		r.health.markTransient(target)
+	}
+}
+
+func (r *EmbeddingRunner) logFailure(role string, target config.RoleTarget, err error) {
+	if r.log == nil {
+		return
+	}
+	r.log.Warn("ai target failed",
+		slog.String("role", role),
+		slog.String("provider", target.Provider),
+		slog.String("model", target.Model),
+		slog.String("error", err.Error()),
+	)
+}
+
+func (r *MetadataRunner) logFailure(role string, target config.RoleTarget, err error) {
+	if r.log == nil {
+		return
+	}
+	r.log.Warn("ai target failed",
+		slog.String("role", role),
+		slog.String("provider", target.Provider),
+		slog.String("model", target.Model),
+		slog.String("error", err.Error()),
+	)
+}
+
+// healthTracker records per-(provider, model) failure state. skip returns true
+// when a target is still inside its cooldown window; the caller then tries the
+// next target in the chain.
+type healthTracker struct {
+	mu     sync.Mutex
+	states map[config.RoleTarget]*healthState
+}
+
+type healthState struct {
+	unhealthyUntil time.Time
+	emptyCount     int
+}
+
+func newHealthTracker() *healthTracker {
+	return &healthTracker{states: map[config.RoleTarget]*healthState{}}
+}
+
+func (h *healthTracker) skip(target config.RoleTarget) bool {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	s, ok := h.states[target]
+	if !ok {
+		return false
+	}
+	return time.Now().Before(s.unhealthyUntil)
+}
+
+func (h *healthTracker) markTransient(target config.RoleTarget) {
+	h.setCooldown(target, transientCooldown)
+}
+
+func (h *healthTracker) markPermanent(target config.RoleTarget) {
+	h.setCooldown(target, permanentCooldown)
+}
+
+func (h *healthTracker) markEmpty(target config.RoleTarget) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	s := h.states[target]
+	if s == nil {
+		s = &healthState{}
+		h.states[target] = s
+	}
+	s.emptyCount++
+	if s.emptyCount >= emptyResponseThreshold {
+		s.unhealthyUntil = time.Now().Add(emptyResponseCooldown)
+		s.emptyCount = 0
+	}
+}
+
+func (h *healthTracker) markHealthy(target config.RoleTarget) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	if s, ok := h.states[target]; ok {
+		s.unhealthyUntil = time.Time{}
+		s.emptyCount = 0
+	}
+}
+
+func (h *healthTracker) setCooldown(target config.RoleTarget, d time.Duration) {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	s := h.states[target]
+	if s == nil {
+		s = &healthState{}
+		h.states[target] = s
+	}
+	s.unhealthyUntil = time.Now().Add(d)
+	s.emptyCount = 0
+}

internal/ai/runner_test.go

@@ -0,0 +1,139 @@
+package ai
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"sync"
+	"testing"
+
+	"git.warky.dev/wdevs/amcs/internal/config"
+)
+
+func TestEmbeddingRunnerFallsBackAndSkipsUnhealthyPrimary(t *testing.T) {
+	var (
+		mu           sync.Mutex
+		primaryCalls int
+	)
+
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/embeddings" {
+			http.NotFound(w, r)
+			return
+		}
+		var req struct {
+			Model string `json:"model"`
+		}
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			http.Error(w, err.Error(), http.StatusBadRequest)
+			return
+		}
+		switch req.Model {
+		case "embed-primary":
+			mu.Lock()
+			primaryCalls++
+			mu.Unlock()
+			http.Error(w, "upstream down", http.StatusBadGateway)
+		case "embed-fallback":
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"data": []map[string]any{{"embedding": []float32{0.1, 0.2, 0.3}}},
+			})
+		default:
+			http.Error(w, "unknown model", http.StatusBadRequest)
+		}
+	}))
+	defer srv.Close()
+
+	reg, err := NewRegistry(map[string]config.ProviderConfig{
+		"p1": {Type: "litellm", BaseURL: srv.URL, APIKey: "k1"},
+		"p2": {Type: "litellm", BaseURL: srv.URL, APIKey: "k2"},
+	}, srv.Client(), nil)
+	if err != nil {
+		t.Fatalf("NewRegistry() error = %v", err)
+	}
+
+	runner, err := NewEmbeddingRunner(reg, []config.RoleTarget{
+		{Provider: "p1", Model: "embed-primary"},
+		{Provider: "p2", Model: "embed-fallback"},
+	}, 3, nil)
+	if err != nil {
+		t.Fatalf("NewEmbeddingRunner() error = %v", err)
+	}
+
+	res, err := runner.Embed(context.Background(), "hello")
+	if err != nil {
+		t.Fatalf("Embed() first call error = %v", err)
+	}
+	if res.Provider != "p2" || res.Model != "embed-fallback" {
+		t.Fatalf("Embed() first call target = %s/%s, want p2/embed-fallback", res.Provider, res.Model)
+	}
+
+	res, err = runner.Embed(context.Background(), "hello again")
+	if err != nil {
+		t.Fatalf("Embed() second call error = %v", err)
+	}
+	if res.Provider != "p2" || res.Model != "embed-fallback" {
+		t.Fatalf("Embed() second call target = %s/%s, want p2/embed-fallback", res.Provider, res.Model)
+	}
+
+	mu.Lock()
+	calls := primaryCalls
+	mu.Unlock()
+	if calls != 3 {
+		t.Fatalf("primary calls = %d, want 3 (first request retries 3x; second call should skip unhealthy primary)", calls)
+	}
+}
+
+func TestMetadataRunnerSummarizeFallsBack(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/chat/completions" {
+			http.NotFound(w, r)
+			return
+		}
+		var req struct {
+			Model string `json:"model"`
+		}
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			http.Error(w, err.Error(), http.StatusBadRequest)
+			return
+		}
+		switch req.Model {
+		case "sum-primary":
+			http.Error(w, "provider error", http.StatusBadGateway)
+		case "sum-fallback":
+			_ = json.NewEncoder(w).Encode(map[string]any{
+				"choices": []map[string]any{{
+					"message": map[string]any{"role": "assistant", "content": "fallback summary"},
+				}},
+			})
+		default:
+			http.Error(w, "unknown model", http.StatusBadRequest)
+		}
+	}))
+	defer srv.Close()
+
+	reg, err := NewRegistry(map[string]config.ProviderConfig{
+		"p1": {Type: "litellm", BaseURL: srv.URL, APIKey: "k1"},
+		"p2": {Type: "litellm", BaseURL: srv.URL, APIKey: "k2"},
+	}, srv.Client(), nil)
+	if err != nil {
+		t.Fatalf("NewRegistry() error = %v", err)
+	}
+
+	runner, err := NewMetadataRunner(reg, []config.RoleTarget{
+		{Provider: "p1", Model: "sum-primary"},
+		{Provider: "p2", Model: "sum-fallback"},
+	}, 0.1, false, nil)
+	if err != nil {
+		t.Fatalf("NewMetadataRunner() error = %v", err)
+	}
+
+	summary, err := runner.Summarize(context.Background(), "system", "user")
+	if err != nil {
+		t.Fatalf("Summarize() error = %v", err)
+	}
+	if summary != "fallback summary" {
+		t.Fatalf("summary = %q, want %q", summary, "fallback summary")
+	}
+}

internal/app/app.go

@@ -34,7 +34,7 @@ func Run(ctx context.Context, configPath string) error {
 
 	logger.Info("loaded configuration",
 		slog.String("path", loadedFrom),
-		slog.String("provider", cfg.AI.Provider),
+		slog.Int("config_version", cfg.Version),
 		slog.String("version", info.Version),
 		slog.String("tag_name", info.TagName),
 		slog.String("build_date", info.BuildDate),
@@ -52,11 +52,37 @@ func Run(ctx context.Context, configPath string) error {
 	}
 
 	httpClient := &http.Client{Timeout: 30 * time.Second}
-	provider, err := ai.NewProvider(cfg.AI, httpClient, logger)
+	registry, err := ai.NewRegistry(cfg.AI.Providers, httpClient, logger)
 	if err != nil {
 		return err
 	}
 
+	foregroundEmbeddings, err := ai.NewEmbeddingRunner(registry, cfg.AI.Embeddings.Chain(), cfg.AI.Embeddings.Dimensions, logger)
+	if err != nil {
+		return err
+	}
+	foregroundMetadata, err := ai.NewMetadataRunner(registry, cfg.AI.Metadata.Chain(), cfg.AI.Metadata.Temperature, cfg.AI.Metadata.LogConversations, logger)
+	if err != nil {
+		return err
+	}
+
+	backgroundEmbeddings := foregroundEmbeddings
+	backgroundMetadata := foregroundMetadata
+	if cfg.AI.Background != nil {
+		if cfg.AI.Background.Embeddings != nil {
+			backgroundEmbeddings, err = ai.NewEmbeddingRunner(registry, cfg.AI.Background.Embeddings.AsTargets(), cfg.AI.Embeddings.Dimensions, logger)
+			if err != nil {
+				return err
+			}
+		}
+		if cfg.AI.Background.Metadata != nil {
+			backgroundMetadata, err = ai.NewMetadataRunner(registry, cfg.AI.Background.Metadata.AsTargets(), cfg.AI.Metadata.Temperature, cfg.AI.Metadata.LogConversations, logger)
+			if err != nil {
+				return err
+			}
+		}
+	}
+
 	var keyring *auth.Keyring
 	var oauthRegistry *auth.OAuthRegistry
 	var tokenStore *auth.TokenStore
@@ -77,12 +103,13 @@ func Run(ctx context.Context, configPath string) error {
 	dynClients := auth.NewDynamicClientStore()
 	activeProjects := session.NewActiveProjects()
 
-	logger.Info("database connection verified",
+	logger.Info("ai providers initialised",
-		slog.String("provider", provider.Name()),
+		slog.String("embedding_primary", foregroundEmbeddings.PrimaryProvider()+"/"+foregroundEmbeddings.PrimaryModel()),
+		slog.String("metadata_primary", foregroundMetadata.PrimaryProvider()+"/"+foregroundMetadata.PrimaryModel()),
 	)
 
 	if cfg.Backfill.Enabled && cfg.Backfill.RunOnStartup {
-		go runBackfillPass(ctx, db, provider, cfg.Backfill, logger)
+		go runBackfillPass(ctx, db, backgroundEmbeddings, cfg.Backfill, logger)
 	}
 
 	if cfg.Backfill.Enabled && cfg.Backfill.Interval > 0 {
@@ -94,14 +121,14 @@ func Run(ctx context.Context, configPath string) error {
 				case <-ctx.Done():
 					return
 				case <-ticker.C:
-					runBackfillPass(ctx, db, provider, cfg.Backfill, logger)
+					runBackfillPass(ctx, db, backgroundEmbeddings, cfg.Backfill, logger)
 				}
 			}
 		}()
 	}
 
 	if cfg.MetadataRetry.Enabled && cfg.MetadataRetry.RunOnStartup {
-		go runMetadataRetryPass(ctx, db, provider, cfg, activeProjects, logger)
+		go runMetadataRetryPass(ctx, db, backgroundMetadata, cfg, activeProjects, logger)
 	}
 
 	if cfg.MetadataRetry.Enabled && cfg.MetadataRetry.Interval > 0 {
@@ -113,13 +140,13 @@ func Run(ctx context.Context, configPath string) error {
 				case <-ctx.Done():
 					return
 				case <-ticker.C:
-					runMetadataRetryPass(ctx, db, provider, cfg, activeProjects, logger)
+					runMetadataRetryPass(ctx, db, backgroundMetadata, cfg, activeProjects, logger)
 				}
 			}
 		}()
 	}
 
-	handler, err := routes(logger, cfg, info, db, provider, keyring, oauthRegistry, tokenStore, authCodes, dynClients, activeProjects)
+	handler, err := routes(logger, cfg, info, db, foregroundEmbeddings, foregroundMetadata, backgroundEmbeddings, backgroundMetadata, keyring, oauthRegistry, tokenStore, authCodes, dynClients, activeProjects)
 	if err != nil {
 		return err
 	}
@@ -156,48 +183,53 @@ func Run(ctx context.Context, configPath string) error {
 	}
 }
 
-func routes(logger *slog.Logger, cfg *config.Config, info buildinfo.Info, db *store.DB, provider ai.Provider, keyring *auth.Keyring, oauthRegistry *auth.OAuthRegistry, tokenStore *auth.TokenStore, authCodes *auth.AuthCodeStore, dynClients *auth.DynamicClientStore, activeProjects *session.ActiveProjects) (http.Handler, error) {
+func routes(logger *slog.Logger, cfg *config.Config, info buildinfo.Info, db *store.DB, embeddings *ai.EmbeddingRunner, metadata *ai.MetadataRunner, bgEmbeddings *ai.EmbeddingRunner, bgMetadata *ai.MetadataRunner, keyring *auth.Keyring, oauthRegistry *auth.OAuthRegistry, tokenStore *auth.TokenStore, authCodes *auth.AuthCodeStore, dynClients *auth.DynamicClientStore, activeProjects *session.ActiveProjects) (http.Handler, error) {
 	mux := http.NewServeMux()
-	authMiddleware := auth.Middleware(cfg.Auth, keyring, oauthRegistry, tokenStore, logger)
+	accessTracker := auth.NewAccessTracker()
+	oauthEnabled := oauthRegistry != nil && tokenStore != nil
+	authMiddleware := auth.Middleware(cfg.Auth, keyring, oauthRegistry, tokenStore, accessTracker, logger)
 	filesTool := tools.NewFilesTool(db, activeProjects)
-	metadataRetryer := tools.NewMetadataRetryer(context.Background(), db, provider, cfg.Capture, cfg.AI.Metadata.Timeout, activeProjects, logger)
+	enrichmentRetryer := tools.NewEnrichmentRetryer(context.Background(), db, bgMetadata, cfg.Capture, cfg.AI.Metadata.Timeout, activeProjects, logger)
+	backfillTool := tools.NewBackfillTool(db, bgEmbeddings, activeProjects, logger)
 
 	toolSet := mcpserver.ToolSet{
-		Capture:       tools.NewCaptureTool(db, provider, cfg.Capture, cfg.AI.Metadata.Timeout, activeProjects, metadataRetryer, logger),
+		Capture:       tools.NewCaptureTool(db, embeddings, cfg.Capture, activeProjects, enrichmentRetryer, backfillTool),
-		Search:        tools.NewSearchTool(db, provider, cfg.Search, activeProjects),
+		Search:        tools.NewSearchTool(db, embeddings, cfg.Search, activeProjects),
 		List:          tools.NewListTool(db, cfg.Search, activeProjects),
 		Stats:         tools.NewStatsTool(db),
 		Get:           tools.NewGetTool(db),
-		Update:        tools.NewUpdateTool(db, provider, cfg.Capture, logger),
+		Update:        tools.NewUpdateTool(db, embeddings, metadata, cfg.Capture, logger),
 		Delete:        tools.NewDeleteTool(db),
 		Archive:       tools.NewArchiveTool(db),
 		Projects:      tools.NewProjectsTool(db, activeProjects),
 		Version:       tools.NewVersionTool(cfg.MCP.ServerName, info),
-		Context:       tools.NewContextTool(db, provider, cfg.Search, activeProjects),
+		Learnings:     tools.NewLearningsTool(db, activeProjects, cfg.Search),
-		Recall:        tools.NewRecallTool(db, provider, cfg.Search, activeProjects),
+		Context:       tools.NewContextTool(db, embeddings, cfg.Search, activeProjects),
-		Summarize:     tools.NewSummarizeTool(db, provider, cfg.Search, activeProjects),
+		Recall:        tools.NewRecallTool(db, embeddings, cfg.Search, activeProjects),
-		Links:         tools.NewLinksTool(db, provider, cfg.Search),
+		Summarize:     tools.NewSummarizeTool(db, embeddings, metadata, cfg.Search, activeProjects),
+		Links:         tools.NewLinksTool(db, embeddings, cfg.Search),
 		Files:         filesTool,
-		Backfill:      tools.NewBackfillTool(db, provider, activeProjects, logger),
+		Backfill:      backfillTool,
-		Reparse:       tools.NewReparseMetadataTool(db, provider, cfg.Capture, activeProjects, logger),
+		Reparse:       tools.NewReparseMetadataTool(db, bgMetadata, cfg.Capture, activeProjects, logger),
-		RetryMetadata: tools.NewRetryMetadataTool(metadataRetryer),
+		RetryMetadata: tools.NewRetryEnrichmentTool(enrichmentRetryer),
-		Household:     tools.NewHouseholdTool(db),
 		Maintenance:   tools.NewMaintenanceTool(db),
-		Calendar:      tools.NewCalendarTool(db),
-		Meals:         tools.NewMealsTool(db),
-		CRM:           tools.NewCRMTool(db),
 		Skills:        tools.NewSkillsTool(db, activeProjects),
 		ChatHistory:   tools.NewChatHistoryTool(db, activeProjects),
+		Describe:      tools.NewDescribeTool(db, mcpserver.BuildToolCatalog()),
 	}
 
-	mcpHandler, err := mcpserver.New(cfg.MCP, logger, toolSet, activeProjects.Clear)
+	mcpHandlers, err := mcpserver.NewHandlers(cfg.MCP, logger, toolSet, activeProjects.Clear)
 	if err != nil {
 		return nil, fmt.Errorf("build mcp handler: %w", err)
 	}
-	mux.Handle(cfg.MCP.Path, authMiddleware(mcpHandler))
+	mux.Handle(cfg.MCP.Path, authMiddleware(mcpHandlers.StreamableHTTP))
+	if mcpHandlers.SSE != nil {
+		mux.Handle(cfg.MCP.SSEPath, authMiddleware(mcpHandlers.SSE))
+		logger.Info("SSE transport enabled", slog.String("sse_path", cfg.MCP.SSEPath))
+	}
 	mux.Handle("/files", authMiddleware(fileHandler(filesTool)))
 	mux.Handle("/files/{id}", authMiddleware(fileHandler(filesTool)))
-	if oauthRegistry != nil && tokenStore != nil {
+	if oauthEnabled {
 		mux.HandleFunc("/.well-known/oauth-authorization-server", oauthMetadataHandler())
 		mux.HandleFunc("/oauth-authorization-server", oauthMetadataHandler())
 		mux.HandleFunc("/oauth/register", oauthRegisterHandler(dynClients, logger))
@@ -207,7 +239,9 @@ func routes(logger *slog.Logger, cfg *config.Config, info buildinfo.Info, db *st
 	}
 	mux.HandleFunc("/favicon.ico", serveFavicon)
 	mux.HandleFunc("/images/project.jpg", serveHomeImage)
+	mux.HandleFunc("/images/icon.png", serveIcon)
 	mux.HandleFunc("/llm", serveLLMInstructions)
+	mux.HandleFunc("/api/status", statusAPIHandler(info, accessTracker, oauthEnabled))
 
 	mux.HandleFunc("/healthz", func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(http.StatusOK)
@@ -225,59 +259,7 @@ func routes(logger *slog.Logger, cfg *config.Config, info buildinfo.Info, db *st
 		_, _ = w.Write([]byte("ready"))
 	})
 
-	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+	mux.HandleFunc("/", homeHandler(info, accessTracker, oauthEnabled))
-		if r.URL.Path != "/" {
-			http.NotFound(w, r)
-			return
-		}
-
-		if r.Method != http.MethodGet && r.Method != http.MethodHead {
-			w.Header().Set("Allow", "GET, HEAD")
-			http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
-			return
-		}
-
-		const homePage = `<!doctype html>
-<html lang="en">
-<head>
-  <meta charset="utf-8">
-  <meta name="viewport" content="width=device-width, initial-scale=1">
-  <title>AMCS</title>
-  <style>
-    body { margin: 0; font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif; background: #f5f7fb; color: #172033; }
-    main { max-width: 860px; margin: 48px auto; background: #fff; border-radius: 12px; box-shadow: 0 10px 28px rgba(23, 32, 51, 0.12); overflow: hidden; }
-    .content { padding: 28px; }
-    h1 { margin: 0 0 12px 0; font-size: 2rem; }
-    p { margin: 0; line-height: 1.5; color: #334155; }
-    .actions { margin-top: 18px; }
-    .link { display: inline-block; padding: 10px 14px; border-radius: 8px; background: #172033; color: #fff; text-decoration: none; font-weight: 600; }
-    .link:hover { background: #0f172a; }
-    img { display: block; width: 100%; height: auto; }
-  </style>
-</head>
-<body>
-  <main>
-    <img src="/images/project.jpg" alt="Avelon Memory Crystal project image">
-    <div class="content">
-      <h1>Avelon Memory Crystal Server (AMCS)</h1>
-      <p>AMCS is a memory server that captures, links, and retrieves structured project thoughts for AI assistants using semantic search, summaries, and MCP tools.</p>
-      <div class="actions">
-        <a class="link" href="/llm">LLM Instructions</a>
-        <a class="link" href="/oauth-authorization-server">OAuth Authorization Server</a>
-        <a class="link" href="/healthz">Health Check</a>
-      </div>
-    </div>
-  </main>
-</body>
-</html>`
-		w.Header().Set("Content-Type", "text/html; charset=utf-8")
-		w.WriteHeader(http.StatusOK)
-		if r.Method == http.MethodHead {
-			return
-		}
-
-		_, _ = w.Write([]byte(homePage))
-	})
 
 	return observability.Chain(
 		mux,
@@ -288,8 +270,8 @@ func routes(logger *slog.Logger, cfg *config.Config, info buildinfo.Info, db *st
 	), nil
 }
 
-func runMetadataRetryPass(ctx context.Context, db *store.DB, provider ai.Provider, cfg *config.Config, activeProjects *session.ActiveProjects, logger *slog.Logger) {
+func runMetadataRetryPass(ctx context.Context, db *store.DB, metadataRunner *ai.MetadataRunner, cfg *config.Config, activeProjects *session.ActiveProjects, logger *slog.Logger) {
-	retryer := tools.NewMetadataRetryer(ctx, db, provider, cfg.Capture, cfg.AI.Metadata.Timeout, activeProjects, logger)
+	retryer := tools.NewMetadataRetryer(ctx, db, metadataRunner, cfg.Capture, cfg.AI.Metadata.Timeout, activeProjects, logger)
 	_, out, err := retryer.Handle(ctx, nil, tools.RetryMetadataInput{
 		Limit:           cfg.MetadataRetry.MaxPerRun,
 		IncludeArchived: cfg.MetadataRetry.IncludeArchived,
@@ -307,8 +289,8 @@ func runMetadataRetryPass(ctx context.Context, db *store.DB, provider ai.Provide
 	)
 }
 
-func runBackfillPass(ctx context.Context, db *store.DB, provider ai.Provider, cfg config.BackfillConfig, logger *slog.Logger) {
+func runBackfillPass(ctx context.Context, db *store.DB, embeddings *ai.EmbeddingRunner, cfg config.BackfillConfig, logger *slog.Logger) {
-	backfiller := tools.NewBackfillTool(db, provider, nil, logger)
+	backfiller := tools.NewBackfillTool(db, embeddings, nil, logger)
 	_, out, err := backfiller.Handle(ctx, nil, tools.BackfillInput{
 		Limit:           cfg.MaxPerRun,
 		IncludeArchived: cfg.IncludeArchived,
@@ -342,3 +324,26 @@ func serveHomeImage(w http.ResponseWriter, r *http.Request) {
 
 	_, _ = w.Write(homeImage)
 }
+
+func serveIcon(w http.ResponseWriter, r *http.Request) {
+	if iconImage == nil {
+		http.NotFound(w, r)
+		return
+	}
+
+	if r.Method != http.MethodGet && r.Method != http.MethodHead {
+		w.Header().Set("Allow", "GET, HEAD")
+		http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
+		return
+	}
+
+	w.Header().Set("Content-Type", "image/png")
+	w.Header().Set("Cache-Control", "public, max-age=31536000, immutable")
+	w.WriteHeader(http.StatusOK)
+
+	if r.Method == http.MethodHead {
+		return
+	}
+
+	_, _ = w.Write(iconImage)
+}

internal/app/oauth.go

@@ -14,6 +14,7 @@ import (
 	"time"
 
 	"git.warky.dev/wdevs/amcs/internal/auth"
+	"git.warky.dev/wdevs/amcs/internal/requestip"
 )
 
 // --- JSON types ---
@@ -261,7 +262,7 @@ func handleClientCredentials(w http.ResponseWriter, r *http.Request, oauthRegist
 	}
 	keyID, ok := oauthRegistry.Lookup(clientID, clientSecret)
 	if !ok {
-		log.Warn("oauth token: invalid client credentials", slog.String("remote_addr", r.RemoteAddr))
+		log.Warn("oauth token: invalid client credentials", slog.String("remote_addr", requestip.FromRequest(r)))
 		w.Header().Set("WWW-Authenticate", `Basic realm="oauth"`)
 		writeTokenError(w, "invalid_client", http.StatusUnauthorized)
 		return
@@ -290,7 +291,7 @@ func handleAuthorizationCode(w http.ResponseWriter, r *http.Request, authCodes *
 		return
 	}
 	if !verifyPKCE(codeVerifier, entry.CodeChallenge, entry.CodeChallengeMethod) {
-		log.Warn("oauth token: PKCE verification failed", slog.String("remote_addr", r.RemoteAddr))
+		log.Warn("oauth token: PKCE verification failed", slog.String("remote_addr", requestip.FromRequest(r)))
 		writeTokenError(w, "invalid_grant", http.StatusBadRequest)
 		return
 	}

internal/app/static_assets.go

@@ -12,6 +12,7 @@ var (
 
 	faviconICO = mustReadStaticFile("favicon.ico")
 	homeImage  = mustReadStaticFile("avelonmemorycrystal.jpg")
+	iconImage  = tryReadStaticFile("icon.png")
 )
 
 func mustReadStaticFile(name string) []byte {
@@ -22,3 +23,11 @@ func mustReadStaticFile(name string) []byte {
 
 	return data
 }
+
+func tryReadStaticFile(name string) []byte {
+	data, err := fs.ReadFile(staticFiles, "static/"+name)
+	if err != nil {
+		return nil
+	}
+	return data
+}

internal/app/status.go

@@ -0,0 +1,140 @@
+package app
+
+import (
+	"bytes"
+	"encoding/json"
+	"io/fs"
+	"net/http"
+	"path"
+	"strings"
+	"time"
+
+	"git.warky.dev/wdevs/amcs/internal/auth"
+	"git.warky.dev/wdevs/amcs/internal/buildinfo"
+)
+
+const connectedWindow = 10 * time.Minute
+
+type statusAPIResponse struct {
+	Title           string                `json:"title"`
+	Description     string                `json:"description"`
+	Version         string                `json:"version"`
+	BuildDate       string                `json:"build_date"`
+	Commit          string                `json:"commit"`
+	ConnectedCount  int                   `json:"connected_count"`
+	TotalKnown      int                   `json:"total_known"`
+	ConnectedWindow string                `json:"connected_window"`
+	Entries         []auth.AccessSnapshot `json:"entries"`
+	OAuthEnabled    bool                  `json:"oauth_enabled"`
+}
+
+func statusSnapshot(info buildinfo.Info, tracker *auth.AccessTracker, oauthEnabled bool, now time.Time) statusAPIResponse {
+	entries := tracker.Snapshot()
+	return statusAPIResponse{
+		Title:           "Avelon Memory Crystal Server (AMCS)",
+		Description:     "AMCS is a memory server that captures, links, and retrieves structured project thoughts for AI assistants using semantic search, summaries, and MCP tools.",
+		Version:         fallback(info.Version, "dev"),
+		BuildDate:       fallback(info.BuildDate, "unknown"),
+		Commit:          fallback(info.Commit, "unknown"),
+		ConnectedCount:  tracker.ConnectedCount(now, connectedWindow),
+		TotalKnown:      len(entries),
+		ConnectedWindow: "last 10 minutes",
+		Entries:         entries,
+		OAuthEnabled:    oauthEnabled,
+	}
+}
+
+func fallback(value, defaultValue string) string {
+	if strings.TrimSpace(value) == "" {
+		return defaultValue
+	}
+	return value
+}
+
+func statusAPIHandler(info buildinfo.Info, tracker *auth.AccessTracker, oauthEnabled bool) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/api/status" {
+			http.NotFound(w, r)
+			return
+		}
+		if r.Method != http.MethodGet && r.Method != http.MethodHead {
+			w.Header().Set("Allow", "GET, HEAD")
+			http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
+			return
+		}
+
+		w.Header().Set("Content-Type", "application/json; charset=utf-8")
+		w.WriteHeader(http.StatusOK)
+		if r.Method == http.MethodHead {
+			return
+		}
+
+		_ = json.NewEncoder(w).Encode(statusSnapshot(info, tracker, oauthEnabled, time.Now()))
+	}
+}
+
+func homeHandler(_ buildinfo.Info, _ *auth.AccessTracker, _ bool) http.HandlerFunc {
+	return func(w http.ResponseWriter, r *http.Request) {
+		if r.Method != http.MethodGet && r.Method != http.MethodHead {
+			w.Header().Set("Allow", "GET, HEAD")
+			http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
+			return
+		}
+
+		requestPath := strings.TrimPrefix(path.Clean(r.URL.Path), "/")
+		if requestPath == "." {
+			requestPath = ""
+		}
+
+		if requestPath != "" {
+			if serveUIAsset(w, r, requestPath) {
+				return
+			}
+			http.NotFound(w, r)
+			return
+		}
+
+		serveUIIndex(w, r)
+	}
+}
+
+func serveUIAsset(w http.ResponseWriter, r *http.Request, name string) bool {
+	if uiDistFS == nil {
+		return false
+	}
+	if strings.Contains(name, "..") {
+		return false
+	}
+	file, err := uiDistFS.Open(name)
+	if err != nil {
+		return false
+	}
+	defer file.Close()
+
+	info, err := file.Stat()
+	if err != nil || info.IsDir() {
+		return false
+	}
+
+	data, err := fs.ReadFile(uiDistFS, name)
+	if err != nil {
+		return false
+	}
+
+	http.ServeContent(w, r, info.Name(), info.ModTime(), bytes.NewReader(data))
+	return true
+}
+
+func serveUIIndex(w http.ResponseWriter, r *http.Request) {
+	if indexHTML == nil {
+		http.Error(w, "ui assets not built", http.StatusServiceUnavailable)
+		return
+	}
+
+	w.Header().Set("Content-Type", "text/html; charset=utf-8")
+	w.WriteHeader(http.StatusOK)
+	if r.Method == http.MethodHead {
+		return
+	}
+	_, _ = w.Write(indexHTML)
+}

internal/app/status_test.go

@@ -0,0 +1,133 @@
+package app
+
+import (
+	"encoding/json"
+	"io"
+	"log/slog"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"git.warky.dev/wdevs/amcs/internal/auth"
+	"git.warky.dev/wdevs/amcs/internal/buildinfo"
+	"git.warky.dev/wdevs/amcs/internal/config"
+)
+
+func TestStatusSnapshotHidesOAuthLinkWhenDisabled(t *testing.T) {
+	tracker := auth.NewAccessTracker()
+	snapshot := statusSnapshot(buildinfo.Info{Version: "v1.2.3", BuildDate: "2026-04-04", Commit: "abc123"}, tracker, false, time.Date(2026, 4, 4, 12, 0, 0, 0, time.UTC))
+
+	if snapshot.OAuthEnabled {
+		t.Fatal("OAuthEnabled = true, want false")
+	}
+	if snapshot.ConnectedCount != 0 {
+		t.Fatalf("ConnectedCount = %d, want 0", snapshot.ConnectedCount)
+	}
+	if snapshot.Title == "" {
+		t.Fatal("Title = empty, want non-empty")
+	}
+}
+
+func TestStatusSnapshotShowsTrackedAccess(t *testing.T) {
+	tracker := auth.NewAccessTracker()
+	now := time.Date(2026, 4, 4, 12, 0, 0, 0, time.UTC)
+	tracker.Record("client-a", "/files", "127.0.0.1:1234", "tester", now)
+
+	snapshot := statusSnapshot(buildinfo.Info{Version: "v1.2.3"}, tracker, true, now)
+
+	if !snapshot.OAuthEnabled {
+		t.Fatal("OAuthEnabled = false, want true")
+	}
+	if snapshot.ConnectedCount != 1 {
+		t.Fatalf("ConnectedCount = %d, want 1", snapshot.ConnectedCount)
+	}
+	if len(snapshot.Entries) != 1 {
+		t.Fatalf("len(Entries) = %d, want 1", len(snapshot.Entries))
+	}
+	if snapshot.Entries[0].KeyID != "client-a" || snapshot.Entries[0].LastPath != "/files" {
+		t.Fatalf("entry = %+v, want keyID client-a and path /files", snapshot.Entries[0])
+	}
+}
+
+func TestStatusAPIHandlerReturnsJSON(t *testing.T) {
+	handler := statusAPIHandler(buildinfo.Info{Version: "v1"}, auth.NewAccessTracker(), true)
+	req := httptest.NewRequest(http.MethodGet, "/api/status", nil)
+	rec := httptest.NewRecorder()
+
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusOK)
+	}
+	if got := rec.Header().Get("Content-Type"); !strings.Contains(got, "application/json") {
+		t.Fatalf("content-type = %q, want application/json", got)
+	}
+
+	var payload statusAPIResponse
+	if err := json.Unmarshal(rec.Body.Bytes(), &payload); err != nil {
+		t.Fatalf("json.Unmarshal() error = %v", err)
+	}
+	if payload.Version != "v1" {
+		t.Fatalf("version = %q, want %q", payload.Version, "v1")
+	}
+}
+
+func TestHomeHandlerAllowsHead(t *testing.T) {
+	handler := homeHandler(buildinfo.Info{Version: "v1"}, auth.NewAccessTracker(), false)
+	req := httptest.NewRequest(http.MethodHead, "/", nil)
+	rec := httptest.NewRecorder()
+
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusOK)
+	}
+	if body := rec.Body.String(); body != "" {
+		t.Fatalf("body = %q, want empty for HEAD", body)
+	}
+}
+
+func TestHomeHandlerServesIndex(t *testing.T) {
+	handler := homeHandler(buildinfo.Info{Version: "v1"}, auth.NewAccessTracker(), false)
+	req := httptest.NewRequest(http.MethodGet, "/", nil)
+	rec := httptest.NewRecorder()
+
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusOK)
+	}
+	if !strings.Contains(rec.Body.String(), "<div id=\"app\"></div>") {
+		t.Fatalf("body = %q, want embedded UI index", rec.Body.String())
+	}
+}
+
+func TestMiddlewareRecordsAuthenticatedAccess(t *testing.T) {
+	keyring, err := auth.NewKeyring([]config.APIKey{{ID: "client-a", Value: "secret"}})
+	if err != nil {
+		t.Fatalf("NewKeyring() error = %v", err)
+	}
+	tracker := auth.NewAccessTracker()
+	logger := slog.New(slog.NewTextHandler(io.Discard, nil))
+	handler := auth.Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, tracker, logger)(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+
+	req := httptest.NewRequest(http.MethodGet, "/files", nil)
+	req.Header.Set("x-brain-key", "secret")
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNoContent {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusNoContent)
+	}
+	snap := tracker.Snapshot()
+	if len(snap) != 1 {
+		t.Fatalf("len(snapshot) = %d, want 1", len(snap))
+	}
+	if snap[0].KeyID != "client-a" || snap[0].LastPath != "/files" {
+		t.Fatalf("snapshot[0] = %+v, want keyID client-a and path /files", snap[0])
+	}
+}

internal/app/ui_assets.go

@@ -0,0 +1,22 @@
+package app
+
+import (
+	"embed"
+	"io/fs"
+)
+
+var (
+	//go:embed ui/dist
+	uiFiles embed.FS
+	uiDistFS fs.FS
+	indexHTML []byte
+)
+
+func init() {
+	dist, err := fs.Sub(uiFiles, "ui/dist")
+	if err != nil {
+		return
+	}
+	uiDistFS = dist
+	indexHTML, _ = fs.ReadFile(uiDistFS, "index.html")
+}

internal/auth/access_tracker.go

@@ -0,0 +1,81 @@
+package auth
+
+import (
+	"sort"
+	"sync"
+	"time"
+)
+
+type AccessSnapshot struct {
+	KeyID          string    `json:"key_id"`
+	LastPath       string    `json:"last_path"`
+	RemoteAddr     string    `json:"remote_addr"`
+	UserAgent      string    `json:"user_agent"`
+	RequestCount   int       `json:"request_count"`
+	LastAccessedAt time.Time `json:"last_accessed_at"`
+}
+
+type AccessTracker struct {
+	mu      sync.RWMutex
+	entries map[string]AccessSnapshot
+}
+
+func NewAccessTracker() *AccessTracker {
+	return &AccessTracker{entries: make(map[string]AccessSnapshot)}
+}
+
+func (t *AccessTracker) Record(keyID, path, remoteAddr, userAgent string, now time.Time) {
+	if t == nil || keyID == "" {
+		return
+	}
+
+	t.mu.Lock()
+	defer t.mu.Unlock()
+
+	entry := t.entries[keyID]
+	entry.KeyID = keyID
+	entry.LastPath = path
+	entry.RemoteAddr = remoteAddr
+	entry.UserAgent = userAgent
+	entry.LastAccessedAt = now.UTC()
+	entry.RequestCount++
+	t.entries[keyID] = entry
+}
+
+func (t *AccessTracker) Snapshot() []AccessSnapshot {
+	if t == nil {
+		return nil
+	}
+
+	t.mu.RLock()
+	defer t.mu.RUnlock()
+
+	items := make([]AccessSnapshot, 0, len(t.entries))
+	for _, entry := range t.entries {
+		items = append(items, entry)
+	}
+
+	sort.Slice(items, func(i, j int) bool {
+		return items[i].LastAccessedAt.After(items[j].LastAccessedAt)
+	})
+
+	return items
+}
+
+func (t *AccessTracker) ConnectedCount(now time.Time, window time.Duration) int {
+	if t == nil {
+		return 0
+	}
+
+	cutoff := now.UTC().Add(-window)
+	t.mu.RLock()
+	defer t.mu.RUnlock()
+
+	count := 0
+	for _, entry := range t.entries {
+		if !entry.LastAccessedAt.Before(cutoff) {
+			count++
+		}
+	}
+	return count
+}

internal/auth/access_tracker_test.go

@@ -0,0 +1,45 @@
+package auth
+
+import (
+	"testing"
+	"time"
+)
+
+func TestAccessTrackerRecordAndSnapshot(t *testing.T) {
+	tracker := NewAccessTracker()
+	older := time.Date(2026, 4, 4, 10, 0, 0, 0, time.UTC)
+	newer := older.Add(2 * time.Minute)
+
+	tracker.Record("client-a", "/files", "10.0.0.1:1234", "agent-a", older)
+	tracker.Record("client-b", "/mcp", "10.0.0.2:1234", "agent-b", newer)
+	tracker.Record("client-a", "/files/1", "10.0.0.1:1234", "agent-a2", newer.Add(30*time.Second))
+
+	snap := tracker.Snapshot()
+	if len(snap) != 2 {
+		t.Fatalf("len(snapshot) = %d, want 2", len(snap))
+	}
+	if snap[0].KeyID != "client-a" {
+		t.Fatalf("snapshot[0].KeyID = %q, want client-a", snap[0].KeyID)
+	}
+	if snap[0].RequestCount != 2 {
+		t.Fatalf("snapshot[0].RequestCount = %d, want 2", snap[0].RequestCount)
+	}
+	if snap[0].LastPath != "/files/1" {
+		t.Fatalf("snapshot[0].LastPath = %q, want /files/1", snap[0].LastPath)
+	}
+	if snap[0].UserAgent != "agent-a2" {
+		t.Fatalf("snapshot[0].UserAgent = %q, want agent-a2", snap[0].UserAgent)
+	}
+}
+
+func TestAccessTrackerConnectedCount(t *testing.T) {
+	tracker := NewAccessTracker()
+	now := time.Date(2026, 4, 4, 12, 0, 0, 0, time.UTC)
+
+	tracker.Record("recent", "/mcp", "", "", now.Add(-2*time.Minute))
+	tracker.Record("stale", "/mcp", "", "", now.Add(-11*time.Minute))
+
+	if got := tracker.ConnectedCount(now, 10*time.Minute); got != 1 {
+		t.Fatalf("ConnectedCount() = %d, want 1", got)
+	}
+}

internal/auth/keyring_test.go

@@ -39,7 +39,7 @@ func TestMiddlewareAllowsHeaderAuthAndSetsContext(t *testing.T) {
 		t.Fatalf("NewKeyring() error = %v", err)
 	}
 
-	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		keyID, ok := KeyIDFromContext(r.Context())
 		if !ok || keyID != "client-a" {
 			t.Fatalf("KeyIDFromContext() = (%q, %v), want (client-a, true)", keyID, ok)
@@ -63,7 +63,7 @@ func TestMiddlewareAllowsBearerAuthAndSetsContext(t *testing.T) {
 		t.Fatalf("NewKeyring() error = %v", err)
 	}
 
-	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		keyID, ok := KeyIDFromContext(r.Context())
 		if !ok || keyID != "client-a" {
 			t.Fatalf("KeyIDFromContext() = (%q, %v), want (client-a, true)", keyID, ok)
@@ -90,7 +90,7 @@ func TestMiddlewarePrefersExplicitHeaderOverBearerAuth(t *testing.T) {
 		t.Fatalf("NewKeyring() error = %v", err)
 	}
 
-	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		keyID, ok := KeyIDFromContext(r.Context())
 		if !ok || keyID != "client-a" {
 			t.Fatalf("KeyIDFromContext() = (%q, %v), want (client-a, true)", keyID, ok)
@@ -119,7 +119,7 @@ func TestMiddlewareAllowsQueryParamWhenEnabled(t *testing.T) {
 		HeaderName:      "x-brain-key",
 		QueryParam:      "key",
 		AllowQueryParam: true,
-	}, keyring, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	}, keyring, nil, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(http.StatusNoContent)
 	}))
 
@@ -138,7 +138,7 @@ func TestMiddlewareRejectsMissingOrInvalidKey(t *testing.T) {
 		t.Fatalf("NewKeyring() error = %v", err)
 	}
 
-	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		t.Fatal("next handler should not be called")
 	}))
 
@@ -157,3 +157,34 @@ func TestMiddlewareRejectsMissingOrInvalidKey(t *testing.T) {
 		t.Fatalf("invalid key status = %d, want %d", rec.Code, http.StatusUnauthorized)
 	}
 }
+
+func TestMiddlewareRecordsForwardedRemoteAddr(t *testing.T) {
+	keyring, err := NewKeyring([]config.APIKey{{ID: "client-a", Value: "secret"}})
+	if err != nil {
+		t.Fatalf("NewKeyring() error = %v", err)
+	}
+	tracker := NewAccessTracker()
+
+	handler := Middleware(config.AuthConfig{HeaderName: "x-brain-key"}, keyring, nil, nil, tracker, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+
+	req := httptest.NewRequest(http.MethodGet, "/mcp", nil)
+	req.RemoteAddr = "10.0.0.5:2222"
+	req.Header.Set("x-brain-key", "secret")
+	req.Header.Set("X-Real-IP", "203.0.113.99")
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNoContent {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusNoContent)
+	}
+
+	snap := tracker.Snapshot()
+	if len(snap) != 1 {
+		t.Fatalf("len(snapshot) = %d, want 1", len(snap))
+	}
+	if snap[0].RemoteAddr != "203.0.113.99" {
+		t.Fatalf("snapshot remote_addr = %q, want %q", snap[0].RemoteAddr, "203.0.113.99")
+	}
+}

internal/auth/middleware.go

@@ -6,30 +6,39 @@ import (
 	"log/slog"
 	"net/http"
 	"strings"
+	"time"
 
 	"git.warky.dev/wdevs/amcs/internal/config"
+	"git.warky.dev/wdevs/amcs/internal/requestip"
 )
 
 type contextKey string
 
 const keyIDContextKey contextKey = "auth.key_id"
 
-func Middleware(cfg config.AuthConfig, keyring *Keyring, oauthRegistry *OAuthRegistry, tokenStore *TokenStore, log *slog.Logger) func(http.Handler) http.Handler {
+func Middleware(cfg config.AuthConfig, keyring *Keyring, oauthRegistry *OAuthRegistry, tokenStore *TokenStore, tracker *AccessTracker, log *slog.Logger) func(http.Handler) http.Handler {
 	headerName := cfg.HeaderName
 	if headerName == "" {
 		headerName = "x-brain-key"
 	}
+	recordAccess := func(r *http.Request, keyID string) {
+		if tracker != nil {
+			tracker.Record(keyID, r.URL.Path, requestip.FromRequest(r), r.UserAgent(), time.Now())
+		}
+	}
 	return func(next http.Handler) http.Handler {
 		return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			remoteAddr := requestip.FromRequest(r)
 			// 1. Custom header → keyring only.
 			if keyring != nil {
 				if token := strings.TrimSpace(r.Header.Get(headerName)); token != "" {
 					keyID, ok := keyring.Lookup(token)
 					if !ok {
-						log.Warn("authentication failed", slog.String("remote_addr", r.RemoteAddr))
+						log.Warn("authentication failed", slog.String("remote_addr", remoteAddr))
 						http.Error(w, "invalid API key", http.StatusUnauthorized)
 						return
 					}
+					recordAccess(r, keyID)
 					next.ServeHTTP(w, r.WithContext(context.WithValue(r.Context(), keyIDContextKey, keyID)))
 					return
 				}
@@ -39,17 +48,19 @@ func Middleware(cfg config.AuthConfig, keyring *Keyring, oauthRegistry *OAuthReg
 			if bearer := extractBearer(r); bearer != "" {
 				if tokenStore != nil {
 					if keyID, ok := tokenStore.Lookup(bearer); ok {
+						recordAccess(r, keyID)
 						next.ServeHTTP(w, r.WithContext(context.WithValue(r.Context(), keyIDContextKey, keyID)))
 						return
 					}
 				}
 				if keyring != nil {
 					if keyID, ok := keyring.Lookup(bearer); ok {
+						recordAccess(r, keyID)
 						next.ServeHTTP(w, r.WithContext(context.WithValue(r.Context(), keyIDContextKey, keyID)))
 						return
 					}
 				}
-				log.Warn("bearer token rejected", slog.String("remote_addr", r.RemoteAddr))
+				log.Warn("bearer token rejected", slog.String("remote_addr", remoteAddr))
 				http.Error(w, "invalid token or API key", http.StatusUnauthorized)
 				return
 			}
@@ -62,10 +73,11 @@ func Middleware(cfg config.AuthConfig, keyring *Keyring, oauthRegistry *OAuthReg
 				}
 				keyID, ok := oauthRegistry.Lookup(clientID, clientSecret)
 				if !ok {
-					log.Warn("oauth client authentication failed", slog.String("remote_addr", r.RemoteAddr))
+					log.Warn("oauth client authentication failed", slog.String("remote_addr", remoteAddr))
 					http.Error(w, "invalid OAuth client credentials", http.StatusUnauthorized)
 					return
 				}
+				recordAccess(r, keyID)
 				next.ServeHTTP(w, r.WithContext(context.WithValue(r.Context(), keyIDContextKey, keyID)))
 				return
 			}
@@ -75,10 +87,11 @@ func Middleware(cfg config.AuthConfig, keyring *Keyring, oauthRegistry *OAuthReg
 				if token := strings.TrimSpace(r.URL.Query().Get(cfg.QueryParam)); token != "" {
 					keyID, ok := keyring.Lookup(token)
 					if !ok {
-						log.Warn("authentication failed", slog.String("remote_addr", r.RemoteAddr))
+						log.Warn("authentication failed", slog.String("remote_addr", remoteAddr))
 						http.Error(w, "invalid API key", http.StatusUnauthorized)
 						return
 					}
+					recordAccess(r, keyID)
 					next.ServeHTTP(w, r.WithContext(context.WithValue(r.Context(), keyIDContextKey, keyID)))
 					return
 				}

internal/auth/oauth_registry_test.go

@@ -42,7 +42,7 @@ func TestMiddlewareAllowsOAuthBasicAuthAndSetsContext(t *testing.T) {
 		t.Fatalf("NewOAuthRegistry() error = %v", err)
 	}
 
-	handler := Middleware(config.AuthConfig{}, nil, oauthRegistry, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	handler := Middleware(config.AuthConfig{}, nil, oauthRegistry, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		keyID, ok := KeyIDFromContext(r.Context())
 		if !ok || keyID != "oauth-client" {
 			t.Fatalf("KeyIDFromContext() = (%q, %v), want (oauth-client, true)", keyID, ok)
@@ -70,7 +70,7 @@ func TestMiddlewareRejectsOAuthMissingOrInvalidCredentials(t *testing.T) {
 		t.Fatalf("NewOAuthRegistry() error = %v", err)
 	}
 
-	handler := Middleware(config.AuthConfig{}, nil, oauthRegistry, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+	handler := Middleware(config.AuthConfig{}, nil, oauthRegistry, nil, nil, testLogger())(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		t.Fatal("next handler should not be called")
 	}))
 

internal/config/config.go

@@ -8,6 +8,7 @@ const (
 )
 
 type Config struct {
+	Version       int                 `yaml:"version"`
 	Server        ServerConfig        `yaml:"server"`
 	MCP           MCPConfig           `yaml:"mcp"`
 	Auth          AuthConfig          `yaml:"auth"`
@@ -32,10 +33,13 @@ type ServerConfig struct {
 
 type MCPConfig struct {
 	Path           string        `yaml:"path"`
+	SSEPath        string        `yaml:"sse_path"`
 	ServerName     string        `yaml:"server_name"`
 	Version        string        `yaml:"version"`
 	Transport      string        `yaml:"transport"`
 	SessionTimeout time.Duration `yaml:"session_timeout"`
+	PublicURL      string        `yaml:"public_url"`
+	Instructions   string        `yaml:"-"`
 }
 
 type AuthConfig struct {
@@ -71,52 +75,82 @@ type DatabaseConfig struct {
 	MaxConnIdleTime time.Duration `yaml:"max_conn_idle_time"`
 }
 
+// AIConfig (v2): named providers + per-role chains.
 type AIConfig struct {
+	Providers  map[string]ProviderConfig `yaml:"providers"`
+	Embeddings EmbeddingsRoleConfig      `yaml:"embeddings"`
+	Metadata   MetadataRoleConfig        `yaml:"metadata"`
+	Background *BackgroundRolesConfig    `yaml:"background,omitempty"`
+}
+
+type ProviderConfig struct {
+	Type           string            `yaml:"type"`
+	BaseURL        string            `yaml:"base_url"`
+	APIKey         string            `yaml:"api_key"`
+	RequestHeaders map[string]string `yaml:"request_headers,omitempty"`
+	AppName        string            `yaml:"app_name,omitempty"`
+	SiteURL        string            `yaml:"site_url,omitempty"`
+}
+
+type RoleTarget struct {
 	Provider string `yaml:"provider"`
-	Embeddings AIEmbeddingConfig  `yaml:"embeddings"`
+	Model    string `yaml:"model"`
-	Metadata   AIMetadataConfig   `yaml:"metadata"`
-	LiteLLM    LiteLLMConfig      `yaml:"litellm"`
-	Ollama     OllamaConfig       `yaml:"ollama"`
-	OpenRouter OpenRouterAIConfig `yaml:"openrouter"`
 }
 
-type AIEmbeddingConfig struct {
+type RoleChain struct {
-	Model      string `yaml:"model"`
+	Primary   RoleTarget   `yaml:"primary"`
+	Fallbacks []RoleTarget `yaml:"fallbacks,omitempty"`
+}
+
+type EmbeddingsRoleConfig struct {
 	Dimensions int          `yaml:"dimensions"`
+	Primary    RoleTarget   `yaml:"primary"`
+	Fallbacks  []RoleTarget `yaml:"fallbacks,omitempty"`
 }
 
-type AIMetadataConfig struct {
+type MetadataRoleConfig struct {
-	Model            string        `yaml:"model"`
-	FallbackModels   []string      `yaml:"fallback_models"`
-	FallbackModel    string        `yaml:"fallback_model"` // legacy single fallback
 	Temperature      float64       `yaml:"temperature"`
 	LogConversations bool          `yaml:"log_conversations"`
 	Timeout          time.Duration `yaml:"timeout"`
+	Primary          RoleTarget    `yaml:"primary"`
+	Fallbacks        []RoleTarget  `yaml:"fallbacks,omitempty"`
 }
 
-type LiteLLMConfig struct {
+// BackgroundRolesConfig overrides the foreground chains for background workers
-	BaseURL                string            `yaml:"base_url"`
+// (backfill_embeddings, metadata_retry, reparse_metadata). Either field may be
-	APIKey                 string            `yaml:"api_key"`
+// nil to inherit the foreground role unchanged.
-	UseResponsesAPI        bool              `yaml:"use_responses_api"`
+type BackgroundRolesConfig struct {
-	RequestHeaders         map[string]string `yaml:"request_headers"`
+	Embeddings *RoleChain `yaml:"embeddings,omitempty"`
-	EmbeddingModel         string            `yaml:"embedding_model"`
+	Metadata   *RoleChain `yaml:"metadata,omitempty"`
-	MetadataModel          string            `yaml:"metadata_model"`
-	FallbackMetadataModels []string          `yaml:"fallback_metadata_models"`
-	FallbackMetadataModel  string            `yaml:"fallback_metadata_model"` // legacy single fallback
 }
 
-type OllamaConfig struct {
+// Chain returns primary followed by fallbacks (deduped, blanks dropped).
-	BaseURL        string            `yaml:"base_url"`
+func (e EmbeddingsRoleConfig) Chain() []RoleTarget {
-	APIKey         string            `yaml:"api_key"`
+	return dedupeTargets(append([]RoleTarget{e.Primary}, e.Fallbacks...))
-	RequestHeaders map[string]string `yaml:"request_headers"`
 }
 
-type OpenRouterAIConfig struct {
+func (m MetadataRoleConfig) Chain() []RoleTarget {
-	BaseURL      string            `yaml:"base_url"`
+	return dedupeTargets(append([]RoleTarget{m.Primary}, m.Fallbacks...))
-	APIKey       string            `yaml:"api_key"`
+}
-	AppName      string            `yaml:"app_name"`
+
-	SiteURL      string            `yaml:"site_url"`
+func (c RoleChain) AsTargets() []RoleTarget {
-	ExtraHeaders map[string]string `yaml:"extra_headers"`
+	return dedupeTargets(append([]RoleTarget{c.Primary}, c.Fallbacks...))
+}
+
+func dedupeTargets(in []RoleTarget) []RoleTarget {
+	out := make([]RoleTarget, 0, len(in))
+	seen := make(map[RoleTarget]struct{}, len(in))
+	for _, t := range in {
+		if t.Provider == "" || t.Model == "" {
+			continue
+		}
+		if _, ok := seen[t]; ok {
+			continue
+		}
+		seen[t] = struct{}{}
+		out = append(out, t)
+	}
+	return out
 }
 
 type CaptureConfig struct {
@@ -161,45 +195,3 @@ type MetadataRetryConfig struct {
 	MaxPerRun       int           `yaml:"max_per_run"`
 	IncludeArchived bool          `yaml:"include_archived"`
 }
-
-func (c AIMetadataConfig) EffectiveFallbackModels() []string {
-	models := make([]string, 0, len(c.FallbackModels)+1)
-	for _, model := range c.FallbackModels {
-		if model != "" {
-			models = append(models, model)
-		}
-	}
-	if c.FallbackModel != "" {
-		models = append(models, c.FallbackModel)
-	}
-	return dedupeNonEmpty(models)
-}
-
-func (c LiteLLMConfig) EffectiveFallbackMetadataModels() []string {
-	models := make([]string, 0, len(c.FallbackMetadataModels)+1)
-	for _, model := range c.FallbackMetadataModels {
-		if model != "" {
-			models = append(models, model)
-		}
-	}
-	if c.FallbackMetadataModel != "" {
-		models = append(models, c.FallbackMetadataModel)
-	}
-	return dedupeNonEmpty(models)
-}
-
-func dedupeNonEmpty(values []string) []string {
-	seen := make(map[string]struct{}, len(values))
-	out := make([]string, 0, len(values))
-	for _, value := range values {
-		if value == "" {
-			continue
-		}
-		if _, ok := seen[value]; ok {
-			continue
-		}
-		seen[value] = struct{}{}
-		out = append(out, value)
-	}
-	return out
-}

internal/config/loader.go

@@ -2,6 +2,7 @@ package config
 
 import (
 	"fmt"
+	"log/slog"
 	"os"
 	"strconv"
 	"strings"
@@ -12,6 +13,12 @@ import (
 )
 
 func Load(explicitPath string) (*Config, string, error) {
+	return LoadWithLogger(explicitPath, nil)
+}
+
+// LoadWithLogger is Load with a logger surface for migration notices. Passing
+// nil is fine — migration events will simply not be logged.
+func LoadWithLogger(explicitPath string, log *slog.Logger) (*Config, string, error) {
 	path := ResolvePath(explicitPath)
 
 	data, err := os.ReadFile(path)
@@ -19,10 +26,38 @@ func Load(explicitPath string) (*Config, string, error) {
 		return nil, path, fmt.Errorf("read config %q: %w", path, err)
 	}
 
-	cfg := defaultConfig()
+	raw := map[string]any{}
-	if err := yaml.Unmarshal(data, &cfg); err != nil {
+	if err := yaml.Unmarshal(data, &raw); err != nil {
 		return nil, path, fmt.Errorf("decode config %q: %w", path, err)
 	}
+	if raw == nil {
+		raw = map[string]any{}
+	}
+
+	applied, err := Migrate(raw)
+	if err != nil {
+		return nil, path, fmt.Errorf("migrate config %q: %w", path, err)
+	}
+
+	if len(applied) > 0 {
+		if log != nil {
+			for _, step := range applied {
+				log.Warn("config migrated in memory",
+					slog.String("path", path),
+					slog.Int("from_version", step.From),
+					slog.Int("to_version", step.To),
+					slog.String("describe", step.Describe),
+					slog.String("hint", "persist with amcs-migrate-config"),
+				)
+			}
+		}
+	}
+
+	cfg, err := decodeTyped(raw)
+	if err != nil {
+		return nil, path, fmt.Errorf("decode migrated config %q: %w", path, err)
+	}
+	cfg.Version = CurrentConfigVersion
 
 	applyEnvOverrides(&cfg)
 	if err := cfg.Validate(); err != nil {
@@ -32,6 +67,18 @@ func Load(explicitPath string) (*Config, string, error) {
 	return &cfg, path, nil
 }
 
+func decodeTyped(raw map[string]any) (Config, error) {
+	out, err := yaml.Marshal(raw)
+	if err != nil {
+		return Config{}, fmt.Errorf("re-marshal migrated config: %w", err)
+	}
+	cfg := defaultConfig()
+	if err := yaml.Unmarshal(out, &cfg); err != nil {
+		return Config{}, err
+	}
+	return cfg, nil
+}
+
 func ResolvePath(explicitPath string) string {
 	if path := strings.TrimSpace(explicitPath); path != "" {
 		if path != ".yaml" && path != ".yml" {
@@ -49,6 +96,7 @@ func ResolvePath(explicitPath string) string {
 func defaultConfig() Config {
 	info := buildinfo.Current()
 	return Config{
+		Version: CurrentConfigVersion,
 		Server: ServerConfig{
 			Host:         "0.0.0.0",
 			Port:         8080,
@@ -58,6 +106,7 @@ func defaultConfig() Config {
 		},
 		MCP: MCPConfig{
 			Path:           "/mcp",
+			SSEPath:        "/sse",
 			ServerName:     "amcs",
 			Version:        info.Version,
 			Transport:      "streamable_http",
@@ -68,20 +117,14 @@ func defaultConfig() Config {
 			QueryParam: "key",
 		},
 		AI: AIConfig{
-			Provider: "litellm",
+			Providers: map[string]ProviderConfig{},
-			Embeddings: AIEmbeddingConfig{
+			Embeddings: EmbeddingsRoleConfig{
-				Model:      "openai/text-embedding-3-small",
 				Dimensions: 1536,
 			},
-			Metadata: AIMetadataConfig{
+			Metadata: MetadataRoleConfig{
-				Model:       "gpt-4o-mini",
 				Temperature: 0.1,
 				Timeout:     10 * time.Second,
 			},
-			Ollama: OllamaConfig{
-				BaseURL: "http://localhost:11434/v1",
-				APIKey:  "ollama",
-			},
 		},
 		Capture: CaptureConfig{
 			Source: DefaultSource,
@@ -117,11 +160,13 @@ func defaultConfig() Config {
 
 func applyEnvOverrides(cfg *Config) {
 	overrideString(&cfg.Database.URL, "AMCS_DATABASE_URL")
-	overrideString(&cfg.AI.LiteLLM.BaseURL, "AMCS_LITELLM_BASE_URL")
+	overrideString(&cfg.MCP.PublicURL, "AMCS_PUBLIC_URL")
-	overrideString(&cfg.AI.LiteLLM.APIKey, "AMCS_LITELLM_API_KEY")
+
-	overrideString(&cfg.AI.Ollama.BaseURL, "AMCS_OLLAMA_BASE_URL")
+	overrideProviderField(cfg, "AMCS_LITELLM_BASE_URL", "litellm", func(p *ProviderConfig, v string) { p.BaseURL = v })
-	overrideString(&cfg.AI.Ollama.APIKey, "AMCS_OLLAMA_API_KEY")
+	overrideProviderField(cfg, "AMCS_LITELLM_API_KEY", "litellm", func(p *ProviderConfig, v string) { p.APIKey = v })
-	overrideString(&cfg.AI.OpenRouter.APIKey, "AMCS_OPENROUTER_API_KEY")
+	overrideProviderField(cfg, "AMCS_OLLAMA_BASE_URL", "ollama", func(p *ProviderConfig, v string) { p.BaseURL = v })
+	overrideProviderField(cfg, "AMCS_OLLAMA_API_KEY", "ollama", func(p *ProviderConfig, v string) { p.APIKey = v })
+	overrideProviderField(cfg, "AMCS_OPENROUTER_API_KEY", "openrouter", func(p *ProviderConfig, v string) { p.APIKey = v })
 
 	if value, ok := os.LookupEnv("AMCS_SERVER_PORT"); ok {
 		if port, err := strconv.Atoi(strings.TrimSpace(value)); err == nil {
@@ -130,6 +175,24 @@ func applyEnvOverrides(cfg *Config) {
 	}
 }
 
+// overrideProviderField applies an env var to every configured provider of the
+// given type. This preserves the v1 behaviour where e.g. AMCS_LITELLM_API_KEY
+// rewrote the single litellm block — in v2 it rewrites every litellm provider.
+func overrideProviderField(cfg *Config, envKey, providerType string, apply func(*ProviderConfig, string)) {
+	value, ok := os.LookupEnv(envKey)
+	if !ok {
+		return
+	}
+	value = strings.TrimSpace(value)
+	for name, p := range cfg.AI.Providers {
+		if p.Type != providerType {
+			continue
+		}
+		apply(&p, value)
+		cfg.AI.Providers[name] = p
+	}
+}
+
 func overrideString(target *string, envKey string) {
 	if value, ok := os.LookupEnv(envKey); ok {
 		*target = strings.TrimSpace(value)

internal/config/loader_test.go

@@ -3,6 +3,7 @@ package config
 import (
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"
 	"time"
 )
@@ -31,9 +32,8 @@ func TestResolvePathIgnoresBareYAMLExtension(t *testing.T) {
 	}
 }
 
-func TestLoadAppliesEnvOverrides(t *testing.T) {
+const v2ConfigYAML = `
-	configPath := filepath.Join(t.TempDir(), "test.yaml")
+version: 2
-	if err := os.WriteFile(configPath, []byte(`
 server:
   port: 8080
 mcp:
@@ -46,18 +46,30 @@ auth:
 database:
   url: "postgres://from-file"
 ai:
-  provider: "litellm"
+  providers:
-  embeddings:
+    default:
-    dimensions: 1536
+      type: "litellm"
-  litellm:
       base_url: "http://localhost:4000/v1"
       api_key: "file-key"
+  embeddings:
+    dimensions: 1536
+    primary:
+      provider: "default"
+      model: "text-embed"
+  metadata:
+    primary:
+      provider: "default"
+      model: "gpt-4"
 search:
   default_limit: 10
   max_limit: 50
 logging:
   level: "info"
-`), 0o600); err != nil {
+`
+
+func TestLoadAppliesEnvOverrides(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "test.yaml")
+	if err := os.WriteFile(configPath, []byte(v2ConfigYAML), 0o600); err != nil {
 		t.Fatalf("write config: %v", err)
 	}
 
@@ -76,8 +88,8 @@ logging:
 	if cfg.Database.URL != "postgres://from-env" {
 		t.Fatalf("database url = %q, want env override", cfg.Database.URL)
 	}
-	if cfg.AI.LiteLLM.APIKey != "env-key" {
+	if cfg.AI.Providers["default"].APIKey != "env-key" {
-		t.Fatalf("litellm api key = %q, want env override", cfg.AI.LiteLLM.APIKey)
+		t.Fatalf("litellm api key = %q, want env override", cfg.AI.Providers["default"].APIKey)
 	}
 	if cfg.Server.Port != 9090 {
 		t.Fatalf("server port = %d, want 9090", cfg.Server.Port)
@@ -90,10 +102,12 @@ logging:
 func TestLoadAppliesOllamaEnvOverrides(t *testing.T) {
 	configPath := filepath.Join(t.TempDir(), "test.yaml")
 	if err := os.WriteFile(configPath, []byte(`
+version: 2
 server:
   port: 8080
 mcp:
   path: "/mcp"
+  session_timeout: "10m"
 auth:
   keys:
     - id: "test"
@@ -101,15 +115,20 @@ auth:
 database:
   url: "postgres://from-file"
 ai:
-  provider: "ollama"
+  providers:
-  embeddings:
+    local:
-    model: "nomic-embed-text"
+      type: "ollama"
-    dimensions: 768
-  metadata:
-    model: "llama3.2"
-  ollama:
       base_url: "http://localhost:11434/v1"
       api_key: "ollama"
+  embeddings:
+    dimensions: 768
+    primary:
+      provider: "local"
+      model: "nomic-embed-text"
+  metadata:
+    primary:
+      provider: "local"
+      model: "llama3.2"
 search:
   default_limit: 10
   max_limit: 50
@@ -127,10 +146,85 @@ logging:
 		t.Fatalf("Load() error = %v", err)
 	}
 
-	if cfg.AI.Ollama.BaseURL != "https://ollama.example.com/v1" {
+	p := cfg.AI.Providers["local"]
-		t.Fatalf("ollama base url = %q, want env override", cfg.AI.Ollama.BaseURL)
+	if p.BaseURL != "https://ollama.example.com/v1" {
+		t.Fatalf("ollama base url = %q, want env override", p.BaseURL)
 	}
-	if cfg.AI.Ollama.APIKey != "remote-key" {
+	if p.APIKey != "remote-key" {
-		t.Fatalf("ollama api key = %q, want env override", cfg.AI.Ollama.APIKey)
+		t.Fatalf("ollama api key = %q, want env override", p.APIKey)
+	}
+}
+
+func TestLoadMigratesV1Config(t *testing.T) {
+	configPath := filepath.Join(t.TempDir(), "v1.yaml")
+	v1 := `
+server:
+  port: 8080
+mcp:
+  path: "/mcp"
+  session_timeout: "10m"
+auth:
+  keys:
+    - id: "test"
+      value: "secret"
+database:
+  url: "postgres://from-file"
+ai:
+  provider: "litellm"
+  embeddings:
+    model: "text-embed"
+    dimensions: 1536
+  metadata:
+    model: "gpt-4"
+    temperature: 0.2
+    fallback_models: ["gpt-3.5"]
+  litellm:
+    base_url: "http://localhost:4000/v1"
+    api_key: "file-key"
+search:
+  default_limit: 10
+  max_limit: 50
+logging:
+  level: "info"
+`
+	if err := os.WriteFile(configPath, []byte(v1), 0o600); err != nil {
+		t.Fatalf("write config: %v", err)
+	}
+
+	cfg, _, err := Load(configPath)
+	if err != nil {
+		t.Fatalf("Load() error = %v", err)
+	}
+
+	if cfg.Version != CurrentConfigVersion {
+		t.Fatalf("version = %d, want %d", cfg.Version, CurrentConfigVersion)
+	}
+	if p, ok := cfg.AI.Providers["default"]; !ok || p.Type != "litellm" || p.APIKey != "file-key" {
+		t.Fatalf("providers[default] = %+v, want litellm/file-key", p)
+	}
+	if cfg.AI.Embeddings.Primary.Model != "text-embed" || cfg.AI.Embeddings.Primary.Provider != "default" {
+		t.Fatalf("embeddings.primary = %+v, want default/text-embed", cfg.AI.Embeddings.Primary)
+	}
+	if cfg.AI.Metadata.Primary.Model != "gpt-4" || cfg.AI.Metadata.Primary.Provider != "default" {
+		t.Fatalf("metadata.primary = %+v, want default/gpt-4", cfg.AI.Metadata.Primary)
+	}
+	if len(cfg.AI.Metadata.Fallbacks) != 1 || cfg.AI.Metadata.Fallbacks[0].Model != "gpt-3.5" {
+		t.Fatalf("metadata.fallbacks = %+v, want [default/gpt-3.5]", cfg.AI.Metadata.Fallbacks)
+	}
+
+	entries, err := filepath.Glob(configPath + ".bak.*")
+	if err != nil {
+		t.Fatalf("glob backups: %v", err)
+	}
+	if len(entries) != 0 {
+		t.Fatalf("backup files = %d, want 0 (load should not rewrite config)", len(entries))
+	}
+
+	originalOnDisk, err := os.ReadFile(configPath)
+	if err != nil {
+		t.Fatalf("read original config: %v", err)
+	}
+	if !strings.Contains(string(originalOnDisk), "provider: \"litellm\"") {
+		t.Fatalf("expected source config to remain unchanged on disk")
 	}
 }

internal/config/migrate.go

@@ -0,0 +1,341 @@
+package config
+
+import (
+	"fmt"
+	"sort"
+)
+
+// CurrentConfigVersion is the schema version this binary expects. Files at a
+// lower version are migrated automatically when loaded.
+const CurrentConfigVersion = 2
+
+// ConfigMigration upgrades a raw YAML map by one version.
+type ConfigMigration struct {
+	From, To int
+	Describe string
+	Apply    func(map[string]any) error
+}
+
+// migrations is the ordered ladder of upgrades. Add new entries at the end.
+var migrations = []ConfigMigration{
+	{From: 1, To: 2, Describe: "named providers + role chains", Apply: migrateV1toV2},
+}
+
+// Migrate brings raw up to CurrentConfigVersion in place. Returns the list of
+// migrations that were applied (may be empty if already current).
+func Migrate(raw map[string]any) ([]ConfigMigration, error) {
+	if raw == nil {
+		return nil, fmt.Errorf("migrate: raw config is nil")
+	}
+
+	version := readVersion(raw)
+	if version > CurrentConfigVersion {
+		return nil, fmt.Errorf("migrate: config version %d is newer than supported version %d", version, CurrentConfigVersion)
+	}
+
+	applied := make([]ConfigMigration, 0)
+	for {
+		if version >= CurrentConfigVersion {
+			break
+		}
+		step, ok := findMigration(version)
+		if !ok {
+			return nil, fmt.Errorf("migrate: no migration registered from version %d", version)
+		}
+		if err := step.Apply(raw); err != nil {
+			return nil, fmt.Errorf("migrate v%d->v%d: %w", step.From, step.To, err)
+		}
+		raw["version"] = step.To
+		version = step.To
+		applied = append(applied, step)
+	}
+	return applied, nil
+}
+
+func findMigration(from int) (ConfigMigration, bool) {
+	for _, m := range migrations {
+		if m.From == from {
+			return m, true
+		}
+	}
+	return ConfigMigration{}, false
+}
+
+// readVersion returns the version from raw. Files without a version field are
+// treated as version 1 (the original schema).
+func readVersion(raw map[string]any) int {
+	v, ok := raw["version"]
+	if !ok {
+		return 1
+	}
+	switch n := v.(type) {
+	case int:
+		return n
+	case int64:
+		return int(n)
+	case float64:
+		return int(n)
+	}
+	return 1
+}
+
+// migrateV1toV2 lifts the single-provider config into the named-providers +
+// role-chains layout. The pre-v2 config implicitly used one provider for both
+// embeddings and metadata; we materialise that as a provider named "default".
+func migrateV1toV2(raw map[string]any) error {
+	aiRaw := mapValue(raw, "ai")
+	if aiRaw == nil {
+		aiRaw = map[string]any{}
+	}
+
+	providerType := stringValue(aiRaw, "provider")
+	if providerType == "" {
+		providerType = "litellm"
+	}
+
+	providers, embeddingModel, metadataModel, fallbackModels := buildV1Provider(aiRaw, providerType)
+
+	embeddingsOld := mapValue(aiRaw, "embeddings")
+	dimensions := intValue(embeddingsOld, "dimensions")
+	if dimensions <= 0 {
+		dimensions = 1536
+	}
+	if embeddingModel == "" {
+		embeddingModel = stringValue(embeddingsOld, "model")
+	}
+
+	metadataOld := mapValue(aiRaw, "metadata")
+	if metadataModel == "" {
+		metadataModel = stringValue(metadataOld, "model")
+	}
+	temperature := floatValue(metadataOld, "temperature")
+	logConversations := boolValue(metadataOld, "log_conversations")
+	timeoutStr := stringValue(metadataOld, "timeout")
+
+	if list := stringListValue(metadataOld, "fallback_models"); len(list) > 0 {
+		fallbackModels = append(fallbackModels, list...)
+	}
+	if v := stringValue(metadataOld, "fallback_model"); v != "" {
+		fallbackModels = append(fallbackModels, v)
+	}
+
+	embeddings := map[string]any{
+		"dimensions": dimensions,
+		"primary":    map[string]any{"provider": "default", "model": embeddingModel},
+	}
+
+	metadata := map[string]any{
+		"temperature":       temperature,
+		"log_conversations": logConversations,
+		"primary":           map[string]any{"provider": "default", "model": metadataModel},
+	}
+	if timeoutStr != "" {
+		metadata["timeout"] = timeoutStr
+	}
+	if fallbacks := chainTargets("default", fallbackModels); len(fallbacks) > 0 {
+		metadata["fallbacks"] = fallbacks
+	}
+
+	raw["ai"] = map[string]any{
+		"providers":  providers,
+		"embeddings": embeddings,
+		"metadata":   metadata,
+	}
+	return nil
+}
+
+func buildV1Provider(aiRaw map[string]any, providerType string) (map[string]any, string, string, []string) {
+	providers := map[string]any{}
+	defaultEntry := map[string]any{"type": providerType}
+	embedModel := ""
+	metaModel := ""
+	var fallbacks []string
+
+	switch providerType {
+	case "litellm":
+		block := mapValue(aiRaw, "litellm")
+		copyKeys(defaultEntry, block, "base_url", "api_key")
+		copyHeaders(defaultEntry, block, "request_headers")
+		embedModel = stringValue(block, "embedding_model")
+		metaModel = stringValue(block, "metadata_model")
+		if list := stringListValue(block, "fallback_metadata_models"); len(list) > 0 {
+			fallbacks = append(fallbacks, list...)
+		}
+		if v := stringValue(block, "fallback_metadata_model"); v != "" {
+			fallbacks = append(fallbacks, v)
+		}
+	case "ollama":
+		block := mapValue(aiRaw, "ollama")
+		copyKeys(defaultEntry, block, "base_url", "api_key")
+		copyHeaders(defaultEntry, block, "request_headers")
+	case "openrouter":
+		block := mapValue(aiRaw, "openrouter")
+		copyKeys(defaultEntry, block, "base_url", "api_key", "app_name", "site_url")
+		copyHeaders(defaultEntry, block, "extra_headers")
+		// rename: extra_headers → request_headers
+		if hdr, ok := defaultEntry["extra_headers"]; ok {
+			defaultEntry["request_headers"] = hdr
+			delete(defaultEntry, "extra_headers")
+		}
+	}
+
+	providers["default"] = defaultEntry
+	return providers, embedModel, metaModel, fallbacks
+}
+
+func chainTargets(provider string, models []string) []any {
+	out := make([]any, 0, len(models))
+	seen := map[string]struct{}{}
+	for _, m := range models {
+		if m == "" {
+			continue
+		}
+		key := provider + "|" + m
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+		out = append(out, map[string]any{"provider": provider, "model": m})
+	}
+	return out
+}
+
+func mapValue(raw map[string]any, key string) map[string]any {
+	if raw == nil {
+		return nil
+	}
+	v, ok := raw[key]
+	if !ok {
+		return nil
+	}
+	switch m := v.(type) {
+	case map[string]any:
+		return m
+	case map[any]any:
+		return convertAnyMap(m)
+	}
+	return nil
+}
+
+func convertAnyMap(in map[any]any) map[string]any {
+	out := make(map[string]any, len(in))
+	keys := make([]string, 0, len(in))
+	for k, v := range in {
+		ks, ok := k.(string)
+		if !ok {
+			continue
+		}
+		keys = append(keys, ks)
+		out[ks] = v
+	}
+	sort.Strings(keys)
+	return out
+}
+
+func stringValue(raw map[string]any, key string) string {
+	if raw == nil {
+		return ""
+	}
+	v, ok := raw[key]
+	if !ok {
+		return ""
+	}
+	if s, ok := v.(string); ok {
+		return s
+	}
+	return ""
+}
+
+func intValue(raw map[string]any, key string) int {
+	if raw == nil {
+		return 0
+	}
+	switch n := raw[key].(type) {
+	case int:
+		return n
+	case int64:
+		return int(n)
+	case float64:
+		return int(n)
+	}
+	return 0
+}
+
+func floatValue(raw map[string]any, key string) float64 {
+	if raw == nil {
+		return 0
+	}
+	switch n := raw[key].(type) {
+	case float64:
+		return n
+	case int:
+		return float64(n)
+	case int64:
+		return float64(n)
+	}
+	return 0
+}
+
+func boolValue(raw map[string]any, key string) bool {
+	if raw == nil {
+		return false
+	}
+	if b, ok := raw[key].(bool); ok {
+		return b
+	}
+	return false
+}
+
+func stringListValue(raw map[string]any, key string) []string {
+	if raw == nil {
+		return nil
+	}
+	v, ok := raw[key]
+	if !ok {
+		return nil
+	}
+	list, ok := v.([]any)
+	if !ok {
+		return nil
+	}
+	out := make([]string, 0, len(list))
+	for _, item := range list {
+		if s, ok := item.(string); ok && s != "" {
+			out = append(out, s)
+		}
+	}
+	return out
+}
+
+func copyKeys(dst, src map[string]any, keys ...string) {
+	if src == nil {
+		return
+	}
+	for _, k := range keys {
+		if v, ok := src[k]; ok {
+			dst[k] = v
+		}
+	}
+}
+
+func copyHeaders(dst, src map[string]any, key string) {
+	if src == nil {
+		return
+	}
+	v, ok := src[key]
+	if !ok {
+		return
+	}
+	switch headers := v.(type) {
+	case map[string]any:
+		if len(headers) == 0 {
+			return
+		}
+		dst[key] = headers
+	case map[any]any:
+		if len(headers) == 0 {
+			return
+		}
+		dst[key] = convertAnyMap(headers)
+	}
+}

internal/config/migrate_test.go

@@ -0,0 +1,77 @@
+package config
+
+import "testing"
+
+func TestMigrateV1ToV2Litellm(t *testing.T) {
+	raw := map[string]any{
+		"ai": map[string]any{
+			"provider": "litellm",
+			"embeddings": map[string]any{
+				"model":      "text-embedding-3-small",
+				"dimensions": 1536,
+			},
+			"metadata": map[string]any{
+				"model":           "gpt-4o-mini",
+				"temperature":     0.2,
+				"fallback_models": []any{"gpt-4.1-mini"},
+			},
+			"litellm": map[string]any{
+				"base_url": "http://localhost:4000/v1",
+				"api_key":  "secret",
+			},
+		},
+	}
+
+	applied, err := Migrate(raw)
+	if err != nil {
+		t.Fatalf("Migrate() error = %v", err)
+	}
+	if len(applied) != 1 || applied[0].From != 1 || applied[0].To != 2 {
+		t.Fatalf("applied = %+v, want [v1->v2]", applied)
+	}
+	if got := readVersion(raw); got != CurrentConfigVersion {
+		t.Fatalf("version = %d, want %d", got, CurrentConfigVersion)
+	}
+
+	ai := mapValue(raw, "ai")
+	providers := mapValue(ai, "providers")
+	def := mapValue(providers, "default")
+	if got := stringValue(def, "type"); got != "litellm" {
+		t.Fatalf("providers.default.type = %q, want litellm", got)
+	}
+	if got := stringValue(def, "base_url"); got != "http://localhost:4000/v1" {
+		t.Fatalf("providers.default.base_url = %q", got)
+	}
+
+	emb := mapValue(ai, "embeddings")
+	embPrimary := mapValue(emb, "primary")
+	if stringValue(embPrimary, "provider") != "default" || stringValue(embPrimary, "model") != "text-embedding-3-small" {
+		t.Fatalf("embeddings.primary = %+v, want default/text-embedding-3-small", embPrimary)
+	}
+
+	meta := mapValue(ai, "metadata")
+	metaPrimary := mapValue(meta, "primary")
+	if stringValue(metaPrimary, "provider") != "default" || stringValue(metaPrimary, "model") != "gpt-4o-mini" {
+		t.Fatalf("metadata.primary = %+v, want default/gpt-4o-mini", metaPrimary)
+	}
+	fallbacks, ok := meta["fallbacks"].([]any)
+	if !ok || len(fallbacks) != 1 {
+		t.Fatalf("metadata.fallbacks = %#v, want len=1", meta["fallbacks"])
+	}
+	firstFallback, ok := fallbacks[0].(map[string]any)
+	if !ok {
+		t.Fatalf("metadata.fallbacks[0] type = %T, want map[string]any", fallbacks[0])
+	}
+	if stringValue(firstFallback, "provider") != "default" || stringValue(firstFallback, "model") != "gpt-4.1-mini" {
+		t.Fatalf("metadata fallback = %+v, want default/gpt-4.1-mini", firstFallback)
+	}
+}
+
+func TestMigrateRejectsNewerVersion(t *testing.T) {
+	raw := map[string]any{"version": CurrentConfigVersion + 1}
+
+	_, err := Migrate(raw)
+	if err == nil {
+		t.Fatal("Migrate() error = nil, want error for newer config version")
+	}
+}

internal/config/validate.go

@@ -33,42 +33,20 @@ func (c Config) Validate() error {
 	if strings.TrimSpace(c.MCP.Path) == "" {
 		return fmt.Errorf("invalid config: mcp.path is required")
 	}
+	if c.MCP.SSEPath != "" {
+		if strings.TrimSpace(c.MCP.SSEPath) == "" {
+			return fmt.Errorf("invalid config: mcp.sse_path must not be blank whitespace")
+		}
+		if c.MCP.SSEPath == c.MCP.Path {
+			return fmt.Errorf("invalid config: mcp.sse_path %q must differ from mcp.path", c.MCP.SSEPath)
+		}
+	}
 	if c.MCP.SessionTimeout <= 0 {
 		return fmt.Errorf("invalid config: mcp.session_timeout must be greater than zero")
 	}
 
-	switch c.AI.Provider {
+	if err := c.AI.validate(); err != nil {
-	case "litellm", "ollama", "openrouter":
+		return err
-	default:
-		return fmt.Errorf("invalid config: unsupported ai.provider %q", c.AI.Provider)
-	}
-
-	if c.AI.Embeddings.Dimensions <= 0 {
-		return fmt.Errorf("invalid config: ai.embeddings.dimensions must be greater than zero")
-	}
-
-	switch c.AI.Provider {
-	case "litellm":
-		if strings.TrimSpace(c.AI.LiteLLM.BaseURL) == "" {
-			return fmt.Errorf("invalid config: ai.litellm.base_url is required when ai.provider=litellm")
-		}
-		if strings.TrimSpace(c.AI.LiteLLM.APIKey) == "" {
-			return fmt.Errorf("invalid config: ai.litellm.api_key is required when ai.provider=litellm")
-		}
-	case "ollama":
-		if strings.TrimSpace(c.AI.Ollama.BaseURL) == "" {
-			return fmt.Errorf("invalid config: ai.ollama.base_url is required when ai.provider=ollama")
-		}
-		if strings.TrimSpace(c.AI.Ollama.APIKey) == "" {
-			return fmt.Errorf("invalid config: ai.ollama.api_key is required when ai.provider=ollama")
-		}
-	case "openrouter":
-		if strings.TrimSpace(c.AI.OpenRouter.BaseURL) == "" {
-			return fmt.Errorf("invalid config: ai.openrouter.base_url is required when ai.provider=openrouter")
-		}
-		if strings.TrimSpace(c.AI.OpenRouter.APIKey) == "" {
-			return fmt.Errorf("invalid config: ai.openrouter.api_key is required when ai.provider=openrouter")
-		}
 	}
 
 	if c.Server.Port <= 0 {
@@ -100,3 +78,61 @@ func (c Config) Validate() error {
 
 	return nil
 }
+
+func (a AIConfig) validate() error {
+	if len(a.Providers) == 0 {
+		return fmt.Errorf("invalid config: ai.providers must contain at least one entry")
+	}
+	for name, p := range a.Providers {
+		if strings.TrimSpace(name) == "" {
+			return fmt.Errorf("invalid config: ai.providers contains an entry with an empty name")
+		}
+		switch p.Type {
+		case "litellm", "ollama", "openrouter":
+		default:
+			return fmt.Errorf("invalid config: ai.providers.%s.type %q is not supported", name, p.Type)
+		}
+		if strings.TrimSpace(p.BaseURL) == "" {
+			return fmt.Errorf("invalid config: ai.providers.%s.base_url is required", name)
+		}
+		if strings.TrimSpace(p.APIKey) == "" {
+			return fmt.Errorf("invalid config: ai.providers.%s.api_key is required", name)
+		}
+	}
+
+	if a.Embeddings.Dimensions <= 0 {
+		return fmt.Errorf("invalid config: ai.embeddings.dimensions must be greater than zero")
+	}
+
+	if err := a.validateChain("ai.embeddings", a.Embeddings.Chain()); err != nil {
+		return err
+	}
+	if err := a.validateChain("ai.metadata", a.Metadata.Chain()); err != nil {
+		return err
+	}
+	if a.Background != nil {
+		if a.Background.Embeddings != nil {
+			if err := a.validateChain("ai.background.embeddings", a.Background.Embeddings.AsTargets()); err != nil {
+				return err
+			}
+		}
+		if a.Background.Metadata != nil {
+			if err := a.validateChain("ai.background.metadata", a.Background.Metadata.AsTargets()); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (a AIConfig) validateChain(prefix string, chain []RoleTarget) error {
+	if len(chain) == 0 {
+		return fmt.Errorf("invalid config: %s.primary must reference a configured provider and model", prefix)
+	}
+	for i, target := range chain {
+		if _, ok := a.Providers[target.Provider]; !ok {
+			return fmt.Errorf("invalid config: %s[%d] references unknown provider %q", prefix, i, target.Provider)
+		}
+	}
+	return nil
+}

internal/config/validate_test.go

@@ -7,6 +7,7 @@ import (
 
 func validConfig() Config {
 	return Config{
+		Version: CurrentConfigVersion,
 		Server:  ServerConfig{Port: 8080},
 		MCP:     MCPConfig{Path: "/mcp", SessionTimeout: 10 * time.Minute},
 		Auth: AuthConfig{
@@ -14,21 +15,15 @@ func validConfig() Config {
 		},
 		Database: DatabaseConfig{URL: "postgres://example"},
 		AI: AIConfig{
-			Provider: "litellm",
+			Providers: map[string]ProviderConfig{
-			Embeddings: AIEmbeddingConfig{
+				"default": {Type: "litellm", BaseURL: "http://localhost:4000/v1", APIKey: "key"},
+			},
+			Embeddings: EmbeddingsRoleConfig{
 				Dimensions: 1536,
+				Primary:    RoleTarget{Provider: "default", Model: "text-embed"},
 			},
-			LiteLLM: LiteLLMConfig{
+			Metadata: MetadataRoleConfig{
-				BaseURL: "http://localhost:4000/v1",
+				Primary: RoleTarget{Provider: "default", Model: "gpt-4"},
-				APIKey:  "key",
-			},
-			Ollama: OllamaConfig{
-				BaseURL: "http://localhost:11434/v1",
-				APIKey:  "ollama",
-			},
-			OpenRouter: OpenRouterAIConfig{
-				BaseURL: "https://openrouter.ai/api/v1",
-				APIKey:  "key",
 			},
 		},
 		Search:  SearchConfig{DefaultLimit: 10, MaxLimit: 50},
@@ -36,29 +31,44 @@ func validConfig() Config {
 	}
 }
 
-func TestValidateAcceptsSupportedProviders(t *testing.T) {
+func TestValidateAcceptsSupportedProviderTypes(t *testing.T) {
+	for _, providerType := range []string{"litellm", "ollama", "openrouter"} {
 		cfg := validConfig()
+		p := cfg.AI.Providers["default"]
+		p.Type = providerType
+		cfg.AI.Providers["default"] = p
 		if err := cfg.Validate(); err != nil {
-		t.Fatalf("Validate litellm error = %v", err)
+			t.Fatalf("Validate %s error = %v", providerType, err)
 		}
-
-	cfg.AI.Provider = "ollama"
-	if err := cfg.Validate(); err != nil {
-		t.Fatalf("Validate ollama error = %v", err)
-	}
-
-	cfg.AI.Provider = "openrouter"
-	if err := cfg.Validate(); err != nil {
-		t.Fatalf("Validate openrouter error = %v", err)
 	}
 }
 
-func TestValidateRejectsInvalidProvider(t *testing.T) {
+func TestValidateRejectsInvalidProviderType(t *testing.T) {
 	cfg := validConfig()
-	cfg.AI.Provider = "unknown"
+	p := cfg.AI.Providers["default"]
+	p.Type = "unknown"
+	cfg.AI.Providers["default"] = p
 
 	if err := cfg.Validate(); err == nil {
-		t.Fatal("Validate() error = nil, want error for unsupported provider")
+		t.Fatal("Validate() error = nil, want error for unsupported provider type")
+	}
+}
+
+func TestValidateRejectsChainWithUnknownProvider(t *testing.T) {
+	cfg := validConfig()
+	cfg.AI.Metadata.Primary = RoleTarget{Provider: "does-not-exist", Model: "x"}
+
+	if err := cfg.Validate(); err == nil {
+		t.Fatal("Validate() error = nil, want error for chain referencing unknown provider")
+	}
+}
+
+func TestValidateRejectsEmptyProviders(t *testing.T) {
+	cfg := validConfig()
+	cfg.AI.Providers = map[string]ProviderConfig{}
+
+	if err := cfg.Validate(); err == nil {
+		t.Fatal("Validate() error = nil, want error for empty providers")
 	}
 }
 

internal/mcpserver/schema.go

@@ -221,12 +221,19 @@ func formatLogDuration(d time.Duration) string {
 	return fmt.Sprintf("%02d:%02d:%03d", minutes, seconds, milliseconds)
 }
 
+func normalizeObjectSchema(schema *jsonschema.Schema) {
+	if schema != nil && schema.Type == "object" && schema.Properties == nil {
+		schema.Properties = map[string]*jsonschema.Schema{}
+	}
+}
+
 func setToolSchemas[In any, Out any](tool *mcp.Tool) error {
 	if tool.InputSchema == nil {
 		inputSchema, err := jsonschema.For[In](toolSchemaOptions)
 		if err != nil {
 			return fmt.Errorf("infer input schema: %w", err)
 		}
+		normalizeObjectSchema(inputSchema)
 		tool.InputSchema = inputSchema
 	}
 

internal/mcpserver/schema_test.go

@@ -13,6 +13,24 @@ import (
 	"git.warky.dev/wdevs/amcs/internal/tools"
 )
 
+func TestSetToolSchemasAddsEmptyPropertiesForNoArgInput(t *testing.T) {
+	type noArgInput struct{}
+	type anyOutput struct{}
+
+	tool := &mcp.Tool{Name: "no_args"}
+	if err := setToolSchemas[noArgInput, anyOutput](tool); err != nil {
+		t.Fatalf("set tool schemas: %v", err)
+	}
+
+	schema, ok := tool.InputSchema.(*jsonschema.Schema)
+	if !ok {
+		t.Fatalf("input schema type = %T, want *jsonschema.Schema", tool.InputSchema)
+	}
+	if schema.Properties == nil {
+		t.Fatal("input schema missing properties: strict MCP clients require properties:{} on object schemas")
+	}
+}
+
 func TestSetToolSchemasUsesStringUUIDsInListOutput(t *testing.T) {
 	tool := &mcp.Tool{Name: "list_thoughts"}
 

internal/mcpserver/server.go

@@ -3,11 +3,18 @@ package mcpserver
 import (
 	"log/slog"
 	"net/http"
+	"strings"
 
 	"github.com/modelcontextprotocol/go-sdk/mcp"
 
 	"git.warky.dev/wdevs/amcs/internal/config"
 	"git.warky.dev/wdevs/amcs/internal/tools"
+	amcsllm "git.warky.dev/wdevs/amcs/llm"
+)
+
+const (
+	serverTitle      = "Avalon Memory Crystal Server"
+	serverWebsiteURL = "https://git.warky.dev/wdevs/amcs"
 )
 
 type ToolSet struct {
@@ -28,37 +35,64 @@ type ToolSet struct {
 	Files         *tools.FilesTool
 	Backfill      *tools.BackfillTool
 	Reparse       *tools.ReparseMetadataTool
-	RetryMetadata *tools.RetryMetadataTool
+	RetryMetadata *tools.RetryEnrichmentTool
-	Household     *tools.HouseholdTool
 	Maintenance   *tools.MaintenanceTool
-	Calendar      *tools.CalendarTool
-	Meals         *tools.MealsTool
-	CRM           *tools.CRMTool
 	Skills        *tools.SkillsTool
 	ChatHistory   *tools.ChatHistoryTool
+	Describe      *tools.DescribeTool
+	Learnings     *tools.LearningsTool
 }
 
+// Handlers groups the HTTP handlers produced for an MCP server instance.
+type Handlers struct {
+	// StreamableHTTP is the primary MCP handler (always non-nil).
+	StreamableHTTP http.Handler
+	// SSE is the SSE transport handler; nil when SSEPath is empty.
+	// SSE is the de facto transport for MCP over the internet and is required by most hosted MCP clients.
+	SSE http.Handler
+}
+
+// New builds the StreamableHTTP MCP handler. It is a convenience wrapper
+// around NewHandlers for callers that only need the primary transport.
 func New(cfg config.MCPConfig, logger *slog.Logger, toolSet ToolSet, onSessionClosed func(string)) (http.Handler, error) {
+	h, err := NewHandlers(cfg, logger, toolSet, onSessionClosed)
+	if err != nil {
+		return nil, err
+	}
+	return h.StreamableHTTP, nil
+}
+
+// NewHandlers builds MCP HTTP handlers for both transports.
+// SSE is nil when cfg.SSEPath is empty.
+func NewHandlers(cfg config.MCPConfig, logger *slog.Logger, toolSet ToolSet, onSessionClosed func(string)) (Handlers, error) {
+	instructions := cfg.Instructions
+	if instructions == "" {
+		instructions = string(amcsllm.MemoryInstructions)
+	}
+
 	server := mcp.NewServer(&mcp.Implementation{
 		Name:       cfg.ServerName,
+		Title:      serverTitle,
 		Version:    cfg.Version,
-	}, nil)
+		WebsiteURL: serverWebsiteURL,
+		Icons:      buildServerIcons(cfg.PublicURL),
+	}, &mcp.ServerOptions{
+		Instructions: instructions,
+	})
 
 	for _, register := range []func(*mcp.Server, *slog.Logger, ToolSet) error{
 		registerSystemTools,
 		registerThoughtTools,
 		registerProjectTools,
+		registerLearningTools,
 		registerFileTools,
 		registerMaintenanceTools,
-		registerHouseholdTools,
-		registerCalendarTools,
-		registerMealTools,
-		registerCRMTools,
 		registerSkillTools,
 		registerChatHistoryTools,
+		registerDescribeTools,
 	} {
 		if err := register(server, logger, toolSet); err != nil {
-			return nil, err
+			return Handlers{}, err
 		}
 	}
 
@@ -70,15 +104,37 @@ func New(cfg config.MCPConfig, logger *slog.Logger, toolSet ToolSet, onSessionCl
 		opts.EventStore = newCleanupEventStore(mcp.NewMemoryEventStore(nil), onSessionClosed)
 	}
 
-	return mcp.NewStreamableHTTPHandler(func(*http.Request) *mcp.Server {
+	h := Handlers{
+		StreamableHTTP: mcp.NewStreamableHTTPHandler(func(*http.Request) *mcp.Server {
 			return server
-	}, opts), nil
+		}, opts),
+	}
+
+	if strings.TrimSpace(cfg.SSEPath) != "" {
+		h.SSE = mcp.NewSSEHandler(func(*http.Request) *mcp.Server {
+			return server
+		}, nil)
+	}
+
+	return h, nil
+}
+
+// buildServerIcons returns icon definitions referencing the server's own /images/icon.png endpoint.
+// Returns nil when publicURL is empty so the icons field is omitted from the MCP identity.
+func buildServerIcons(publicURL string) []mcp.Icon {
+	if strings.TrimSpace(publicURL) == "" {
+		return nil
+	}
+	base := strings.TrimRight(publicURL, "/")
+	return []mcp.Icon{
+		{Source: base + "/images/icon.png", MIMEType: "image/png"},
+	}
 }
 
 func registerSystemTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "get_version_info",
-		Description: "Return the server build version information, including version, tag name, commit, and build date.",
+		Description: "Build version, commit, and date.",
 	}, toolSet.Version.GetInfo); err != nil {
 		return err
 	}
@@ -88,13 +144,13 @@ func registerSystemTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSe
 func registerThoughtTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "capture_thought",
-		Description: "Store a thought with generated embeddings and extracted metadata.",
+		Description: "Store a thought; embeddings and metadata extracted async.",
 	}, toolSet.Capture.Handle); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "search_thoughts",
-		Description: "Search stored thoughts by semantic similarity.",
+		Description: "Semantic search; falls back to full-text if no embeddings.",
 	}, toolSet.Search.Handle); err != nil {
 		return err
 	}
@@ -106,7 +162,7 @@ func registerThoughtTools(server *mcp.Server, logger *slog.Logger, toolSet ToolS
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "thought_stats",
-		Description: "Get counts and top metadata buckets across stored thoughts.",
+		Description: "Counts and top metadata buckets for stored thoughts.",
 	}, toolSet.Stats.Handle); err != nil {
 		return err
 	}
@@ -130,19 +186,19 @@ func registerThoughtTools(server *mcp.Server, logger *slog.Logger, toolSet ToolS
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "archive_thought",
-		Description: "Archive a thought so it is hidden from default search and listing.",
+		Description: "Hide a thought from default search and listing.",
 	}, toolSet.Archive.Handle); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "summarize_thoughts",
-		Description: "Summarize a filtered or searched set of thoughts.",
+		Description: "LLM summary of a filtered set of thoughts.",
 	}, toolSet.Summarize.Handle); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "recall_context",
-		Description: "Recall semantically relevant and recent context.",
+		Description: "Semantic + recency context for prompt injection; falls back to full-text.",
 	}, toolSet.Recall.Handle); err != nil {
 		return err
 	}
@@ -154,7 +210,7 @@ func registerThoughtTools(server *mcp.Server, logger *slog.Logger, toolSet ToolS
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "related_thoughts",
-		Description: "Retrieve explicit links and semantic neighbors for a thought.",
+		Description: "Explicit links and semantic neighbours; falls back to full-text.",
 	}, toolSet.Links.Related); err != nil {
 		return err
 	}
@@ -176,25 +232,47 @@ func registerProjectTools(server *mcp.Server, logger *slog.Logger, toolSet ToolS
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "set_active_project",
-		Description: "Set the active project for the current MCP session. Requires a stateful MCP client that reuses the same session across calls.",
+		Description: "Set session's active project. Pass project per call if client is stateless.",
 	}, toolSet.Projects.SetActive); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "get_active_project",
-		Description: "Return the active project for the current MCP session. If your client does not preserve MCP sessions, pass project explicitly to project-scoped tools instead.",
+		Description: "Return session's active project. Pass project per call if client is stateless.",
 	}, toolSet.Projects.GetActive); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "get_project_context",
-		Description: "Get recent and semantic context for a project. Uses the explicit project when provided, otherwise the active MCP session project.",
+		Description: "Recent and semantic context for a project; falls back to full-text.",
 	}, toolSet.Context.Handle); err != nil {
 		return err
 	}
 	return nil
 }
 
+func registerLearningTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
+	if err := addTool(server, logger, &mcp.Tool{
+		Name:        "add_learning",
+		Description: "Create a curated learning record distinct from raw thoughts.",
+	}, toolSet.Learnings.Add); err != nil {
+		return err
+	}
+	if err := addTool(server, logger, &mcp.Tool{
+		Name:        "get_learning",
+		Description: "Retrieve a structured learning by id.",
+	}, toolSet.Learnings.Get); err != nil {
+		return err
+	}
+	if err := addTool(server, logger, &mcp.Tool{
+		Name:        "list_learnings",
+		Description: "List structured learnings with optional project, status, priority, tag, and text filters.",
+	}, toolSet.Learnings.List); err != nil {
+		return err
+	}
+	return nil
+}
+
 func registerFileTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
 	server.AddResourceTemplate(&mcp.ResourceTemplate{
 		Name:        "stored_file",
@@ -204,19 +282,19 @@ func registerFileTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet)
 
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "upload_file",
-		Description: "Stage a file and get an amcs://files/{id} resource URI. Provide content_path (absolute server-side path, no size limit) or content_base64 (≤10 MB). Optionally link immediately with thought_id/project, or omit them and pass the returned URI to save_file later.",
+		Description: "Stage a file; returns amcs://files/{id}. content_path for large/binary, content_base64 for ≤10 MB. Link now or pass URI to save_file.",
 	}, toolSet.Files.Upload); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "save_file",
-		Description: "Store a file and optionally link it to a thought. Supply either content_base64 (≤10 MB) or content_uri (amcs://files/{id} from a prior upload_file or POST /files call). For files larger than 10 MB, use upload_file with content_path first.",
+		Description: "Store and optionally link a file. content_base64 (≤10 MB) or content_uri from upload_file. >10 MB: use upload_file first.",
 	}, toolSet.Files.Save); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "load_file",
-		Description: "Load a previously stored file by id and return its metadata and base64 content.",
+		Description: "Fetch file metadata and content by id (UUID or amcs://files/{id}); includes embedded MCP resource.",
 	}, toolSet.Files.Load); err != nil {
 		return err
 	}
@@ -232,19 +310,19 @@ func registerFileTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet)
 func registerMaintenanceTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "backfill_embeddings",
-		Description: "Generate missing embeddings for stored thoughts using the active embedding model.",
+		Description: "Generate missing embeddings. Run after model switch or bulk import.",
 	}, toolSet.Backfill.Handle); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "reparse_thought_metadata",
-		Description: "Re-extract and normalize metadata for stored thoughts from their content.",
+		Description: "Re-extract metadata from thought content.",
 	}, toolSet.Reparse.Handle); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "retry_failed_metadata",
-		Description: "Retry metadata extraction for thoughts still marked pending or failed.",
+		Description: "Retry pending/failed metadata extraction.",
 	}, toolSet.RetryMetadata.Handle); err != nil {
 		return err
 	}
@@ -256,7 +334,7 @@ func registerMaintenanceTools(server *mcp.Server, logger *slog.Logger, toolSet T
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "log_maintenance",
-		Description: "Log completed maintenance work; automatically updates the task's next due date.",
+		Description: "Log completed maintenance; updates next due date.",
 	}, toolSet.Maintenance.LogWork); err != nil {
 		return err
 	}
@@ -275,176 +353,10 @@ func registerMaintenanceTools(server *mcp.Server, logger *slog.Logger, toolSet T
 	return nil
 }
 
-func registerHouseholdTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "add_household_item",
-		Description: "Store a household fact (paint color, appliance details, measurement, document, etc.).",
-	}, toolSet.Household.AddItem); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "search_household_items",
-		Description: "Search household items by name, category, or location.",
-	}, toolSet.Household.SearchItems); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "get_household_item",
-		Description: "Retrieve a household item by id.",
-	}, toolSet.Household.GetItem); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "add_vendor",
-		Description: "Add a service provider (plumber, electrician, landscaper, etc.).",
-	}, toolSet.Household.AddVendor); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "list_vendors",
-		Description: "List household service vendors, optionally filtered by service type.",
-	}, toolSet.Household.ListVendors); err != nil {
-		return err
-	}
-	return nil
-}
-
-func registerCalendarTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "add_family_member",
-		Description: "Add a family member to the household.",
-	}, toolSet.Calendar.AddMember); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "list_family_members",
-		Description: "List all family members.",
-	}, toolSet.Calendar.ListMembers); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "add_activity",
-		Description: "Schedule a one-time or recurring family activity.",
-	}, toolSet.Calendar.AddActivity); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "get_week_schedule",
-		Description: "Get all activities scheduled for a given week.",
-	}, toolSet.Calendar.GetWeekSchedule); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "search_activities",
-		Description: "Search activities by title, type, or family member.",
-	}, toolSet.Calendar.SearchActivities); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "add_important_date",
-		Description: "Track a birthday, anniversary, deadline, or other important date.",
-	}, toolSet.Calendar.AddImportantDate); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "get_upcoming_dates",
-		Description: "Get important dates coming up in the next N days.",
-	}, toolSet.Calendar.GetUpcomingDates); err != nil {
-		return err
-	}
-	return nil
-}
-
-func registerMealTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "add_recipe",
-		Description: "Save a recipe with ingredients and instructions.",
-	}, toolSet.Meals.AddRecipe); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "search_recipes",
-		Description: "Search recipes by name, cuisine, tags, or ingredient.",
-	}, toolSet.Meals.SearchRecipes); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "update_recipe",
-		Description: "Update an existing recipe.",
-	}, toolSet.Meals.UpdateRecipe); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "create_meal_plan",
-		Description: "Set the meal plan for a week; replaces any existing plan for that week.",
-	}, toolSet.Meals.CreateMealPlan); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "get_meal_plan",
-		Description: "Get the meal plan for a given week.",
-	}, toolSet.Meals.GetMealPlan); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "generate_shopping_list",
-		Description: "Auto-generate a shopping list from the meal plan for a given week.",
-	}, toolSet.Meals.GenerateShoppingList); err != nil {
-		return err
-	}
-	return nil
-}
-
-func registerCRMTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "add_professional_contact",
-		Description: "Add a professional contact to the CRM.",
-	}, toolSet.CRM.AddContact); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "search_contacts",
-		Description: "Search professional contacts by name, company, title, notes, or tags.",
-	}, toolSet.CRM.SearchContacts); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "log_interaction",
-		Description: "Log an interaction with a professional contact.",
-	}, toolSet.CRM.LogInteraction); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "get_contact_history",
-		Description: "Get full history (interactions and opportunities) for a contact.",
-	}, toolSet.CRM.GetHistory); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "create_opportunity",
-		Description: "Create a deal, project, or opportunity linked to a contact.",
-	}, toolSet.CRM.CreateOpportunity); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "get_follow_ups_due",
-		Description: "List contacts with a follow-up date due within the next N days.",
-	}, toolSet.CRM.GetFollowUpsDue); err != nil {
-		return err
-	}
-	if err := addTool(server, logger, &mcp.Tool{
-		Name:        "link_thought_to_contact",
-		Description: "Append a stored thought to a contact's notes.",
-	}, toolSet.CRM.LinkThought); err != nil {
-		return err
-	}
-	return nil
-}
-
 func registerSkillTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "add_skill",
-		Description: "Store a reusable agent skill (behavioural instruction or capability prompt).",
+		Description: "Store an agent skill (instruction or capability prompt).",
 	}, toolSet.Skills.AddSkill); err != nil {
 		return err
 	}
@@ -462,7 +374,7 @@ func registerSkillTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "add_guardrail",
-		Description: "Store a reusable agent guardrail (constraint or safety rule).",
+		Description: "Store an agent guardrail (constraint or safety rule).",
 	}, toolSet.Skills.AddGuardrail); err != nil {
 		return err
 	}
@@ -480,37 +392,37 @@ func registerSkillTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "add_project_skill",
-		Description: "Link an agent skill to a project. Pass project explicitly when your client does not preserve MCP sessions.",
+		Description: "Link a skill to a project. Pass project if client is stateless.",
 	}, toolSet.Skills.AddProjectSkill); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "remove_project_skill",
-		Description: "Unlink an agent skill from a project. Pass project explicitly when your client does not preserve MCP sessions.",
+		Description: "Unlink a skill from a project. Pass project if client is stateless.",
 	}, toolSet.Skills.RemoveProjectSkill); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "list_project_skills",
-		Description: "List all skills linked to a project. Call this at the start of a project session to load existing agent behaviour instructions before generating new ones. Pass project explicitly when your client does not preserve MCP sessions.",
+		Description: "Skills for a project. Load at session start; only add new if none returned. Pass project if stateless.",
 	}, toolSet.Skills.ListProjectSkills); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "add_project_guardrail",
-		Description: "Link an agent guardrail to a project. Pass project explicitly when your client does not preserve MCP sessions.",
+		Description: "Link a guardrail to a project. Pass project if client is stateless.",
 	}, toolSet.Skills.AddProjectGuardrail); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "remove_project_guardrail",
-		Description: "Unlink an agent guardrail from a project. Pass project explicitly when your client does not preserve MCP sessions.",
+		Description: "Unlink a guardrail from a project. Pass project if client is stateless.",
 	}, toolSet.Skills.RemoveProjectGuardrail); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "list_project_guardrails",
-		Description: "List all guardrails linked to a project. Call this at the start of a project session to load existing agent constraints before generating new ones. Pass project explicitly when your client does not preserve MCP sessions.",
+		Description: "Guardrails for a project. Load at session start; only add new if none returned. Pass project if stateless.",
 	}, toolSet.Skills.ListProjectGuardrails); err != nil {
 		return err
 	}
@@ -520,27 +432,119 @@ func registerSkillTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet
 func registerChatHistoryTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "save_chat_history",
-		Description: "Save a chat session's message history for later retrieval. Stores messages with optional title, summary, channel, agent, and project metadata.",
+		Description: "Save chat messages with optional title, summary, channel, agent, and project.",
 	}, toolSet.ChatHistory.SaveChatHistory); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "get_chat_history",
-		Description: "Retrieve a saved chat history by its UUID or session_id. Returns the full message list.",
+		Description: "Fetch chat history by UUID or session_id.",
 	}, toolSet.ChatHistory.GetChatHistory); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "list_chat_histories",
-		Description: "List saved chat histories with optional filters: project, channel, agent_id, session_id, or recent days.",
+		Description: "List chat histories; filter by project, channel, agent_id, session_id, or days.",
 	}, toolSet.ChatHistory.ListChatHistories); err != nil {
 		return err
 	}
 	if err := addTool(server, logger, &mcp.Tool{
 		Name:        "delete_chat_history",
-		Description: "Permanently delete a saved chat history by id.",
+		Description: "Delete a chat history by id.",
 	}, toolSet.ChatHistory.DeleteChatHistory); err != nil {
 		return err
 	}
 	return nil
 }
+
+func registerDescribeTools(server *mcp.Server, logger *slog.Logger, toolSet ToolSet) error {
+	if err := addTool(server, logger, &mcp.Tool{
+		Name:        "describe_tools",
+		Description: "Call first each session. All tools with categories and usage notes. Categories: system, thoughts, projects, files, admin, maintenance, skills, chat, meta.",
+	}, toolSet.Describe.Describe); err != nil {
+		return err
+	}
+	if err := addTool(server, logger, &mcp.Tool{
+		Name:        "annotate_tool",
+		Description: "Save usage notes for a tool; returned by describe_tools. Empty string clears.",
+	}, toolSet.Describe.Annotate); err != nil {
+		return err
+	}
+	return nil
+}
+
+// BuildToolCatalog returns the static catalog of all registered MCP tools.
+// Pass this to tools.NewDescribeTool when assembling the ToolSet.
+func BuildToolCatalog() []tools.ToolEntry {
+	return []tools.ToolEntry{
+		// system
+		{Name: "get_version_info", Description: "Return the server build version information, including version, tag name, commit, and build date.", Category: "system"},
+
+		// thoughts
+		{Name: "capture_thought", Description: "Store a thought with generated embeddings and extracted metadata. The thought is saved immediately even if metadata extraction times out; pending thoughts are retried in the background.", Category: "thoughts"},
+		{Name: "search_thoughts", Description: "Search stored thoughts by semantic similarity. Falls back to Postgres full-text search automatically when no embeddings exist for the active model.", Category: "thoughts"},
+		{Name: "list_thoughts", Description: "List recent thoughts with optional metadata filters.", Category: "thoughts"},
+		{Name: "thought_stats", Description: "Get counts and top metadata buckets across stored thoughts.", Category: "thoughts"},
+		{Name: "get_thought", Description: "Retrieve a full thought by id.", Category: "thoughts"},
+		{Name: "update_thought", Description: "Update thought content or merge metadata.", Category: "thoughts"},
+		{Name: "delete_thought", Description: "Hard-delete a thought by id.", Category: "thoughts"},
+		{Name: "archive_thought", Description: "Archive a thought so it is hidden from default search and listing.", Category: "thoughts"},
+		{Name: "summarize_thoughts", Description: "Produce an LLM prose summary of a filtered or searched set of thoughts.", Category: "thoughts"},
+		{Name: "recall_context", Description: "Recall semantically relevant and recent context for prompt injection. Combines vector similarity with recency. Falls back to full-text search when no embeddings exist.", Category: "thoughts"},
+		{Name: "link_thoughts", Description: "Create a typed relationship between two thoughts.", Category: "thoughts"},
+		{Name: "related_thoughts", Description: "Retrieve explicit links and semantic neighbours for a thought. Falls back to full-text search when no embeddings exist.", Category: "thoughts"},
+
+		// projects
+		{Name: "create_project", Description: "Create a named project container for thoughts.", Category: "projects"},
+		{Name: "list_projects", Description: "List projects and their current thought counts.", Category: "projects"},
+		{Name: "set_active_project", Description: "Set the active project for the current MCP session. Requires a stateful MCP client that reuses the same session across calls. If your client does not preserve sessions, pass project explicitly to each tool instead.", Category: "projects"},
+		{Name: "get_active_project", Description: "Return the active project for the current MCP session. If your client does not preserve MCP sessions, pass project explicitly to project-scoped tools instead of relying on this.", Category: "projects"},
+		{Name: "get_project_context", Description: "Get recent and semantic context for a project. Uses the explicit project when provided, otherwise the active MCP session project. Falls back to full-text search when no embeddings exist.", Category: "projects"},
+
+		// learnings
+		{Name: "add_learning", Description: "Create a curated learning record distinct from raw thoughts.", Category: "projects"},
+		{Name: "get_learning", Description: "Retrieve a structured learning by id.", Category: "projects"},
+		{Name: "list_learnings", Description: "List structured learnings with optional project, category, area, status, priority, tag, and text filters.", Category: "projects"},
+
+		// files
+		{Name: "upload_file", Description: "Stage a file and get an amcs://files/{id} resource URI. Use content_path (absolute server-side path, no size limit) for large or binary files, or content_base64 (≤10 MB) for small files. Pass thought_id/project to link immediately, or omit and pass the URI to save_file later.", Category: "files"},
+		{Name: "save_file", Description: "Store a file and optionally link it to a thought. Use content_base64 (≤10 MB) for small files, or content_uri (amcs://files/{id} from a prior upload_file) for previously staged files. For files larger than 10 MB, use upload_file with content_path first. If the goal is to retain the artifact, store the file directly instead of reading or summarising it first.", Category: "files"},
+		{Name: "load_file", Description: "Load a stored file by id. Returns metadata, base64 content, and an embedded MCP binary resource at amcs://files/{id}. Prefer the embedded resource when your client supports it. The id field accepts a bare UUID or full amcs://files/{id} URI.", Category: "files"},
+		{Name: "list_files", Description: "List stored files, optionally filtered by thought, project, or kind.", Category: "files"},
+
+		// admin
+		{Name: "backfill_embeddings", Description: "Generate missing embeddings for stored thoughts using the active embedding model. Run this after switching embedding models or importing thoughts that have no vectors.", Category: "admin"},
+		{Name: "reparse_thought_metadata", Description: "Re-extract and normalize metadata for stored thoughts from their content.", Category: "admin"},
+		{Name: "retry_failed_metadata", Description: "Retry metadata extraction for thoughts still marked pending or failed.", Category: "admin"},
+
+		// maintenance
+		{Name: "add_maintenance_task", Description: "Create a recurring or one-time home maintenance task.", Category: "maintenance"},
+		{Name: "log_maintenance", Description: "Log completed maintenance work; automatically updates the task's next due date.", Category: "maintenance"},
+		{Name: "get_upcoming_maintenance", Description: "List maintenance tasks due within the next N days.", Category: "maintenance"},
+		{Name: "search_maintenance_history", Description: "Search the maintenance log by task name, category, or date range.", Category: "maintenance"},
+
+		// skills
+		{Name: "add_skill", Description: "Store a reusable agent skill (behavioural instruction or capability prompt).", Category: "skills"},
+		{Name: "remove_skill", Description: "Delete an agent skill by id.", Category: "skills"},
+		{Name: "list_skills", Description: "List all agent skills, optionally filtered by tag.", Category: "skills"},
+		{Name: "add_guardrail", Description: "Store a reusable agent guardrail (constraint or safety rule).", Category: "skills"},
+		{Name: "remove_guardrail", Description: "Delete an agent guardrail by id.", Category: "skills"},
+		{Name: "list_guardrails", Description: "List all agent guardrails, optionally filtered by tag or severity.", Category: "skills"},
+		{Name: "add_project_skill", Description: "Link an agent skill to a project. Pass project explicitly when your client does not preserve MCP sessions.", Category: "skills"},
+		{Name: "remove_project_skill", Description: "Unlink an agent skill from a project. Pass project explicitly when your client does not preserve MCP sessions.", Category: "skills"},
+		{Name: "list_project_skills", Description: "List all skills linked to a project. Call this at the start of every project session to load agent behaviour instructions before generating new ones. Only create new skills if none are returned. Pass project explicitly when your client does not preserve MCP sessions.", Category: "skills"},
+		{Name: "add_project_guardrail", Description: "Link an agent guardrail to a project. Pass project explicitly when your client does not preserve MCP sessions.", Category: "skills"},
+		{Name: "remove_project_guardrail", Description: "Unlink an agent guardrail from a project. Pass project explicitly when your client does not preserve MCP sessions.", Category: "skills"},
+		{Name: "list_project_guardrails", Description: "List all guardrails linked to a project. Call this at the start of every project session to load agent constraints before generating new ones. Only create new guardrails if none are returned. Pass project explicitly when your client does not preserve MCP sessions.", Category: "skills"},
+
+		// chat
+		{Name: "save_chat_history", Description: "Save a chat session's message history for later retrieval. Stores messages with optional title, summary, channel, agent, and project metadata.", Category: "chat"},
+		{Name: "get_chat_history", Description: "Retrieve a saved chat history by its UUID or session_id. Returns the full message list.", Category: "chat"},
+		{Name: "list_chat_histories", Description: "List saved chat histories with optional filters: project, channel, agent_id, session_id, or recent days.", Category: "chat"},
+		{Name: "delete_chat_history", Description: "Permanently delete a saved chat history by id.", Category: "chat"},
+
+		// meta
+		{Name: "describe_tools", Description: "Call this first in every session. Returns all available MCP tools with names, descriptions, categories, and your accumulated usage notes. Filter by category to narrow results. Available categories: system, thoughts, projects, files, admin, household, maintenance, calendar, meals, crm, skills, chat, meta.", Category: "meta"},
+		{Name: "annotate_tool", Description: "Persist usage notes, gotchas, or workflow patterns for a specific tool. Notes survive across sessions and are returned by describe_tools. Call this whenever you discover something non-obvious about a tool's behaviour. Pass an empty string to clear notes.", Category: "meta"},
+	}
+}

internal/mcpserver/server_test.go

@@ -28,50 +28,38 @@ func TestNewListsAllRegisteredTools(t *testing.T) {
 	sort.Strings(got)
 
 	want := []string{
-		"add_activity",
-		"add_family_member",
 		"add_guardrail",
-		"add_household_item",
+		"add_learning",
-		"add_important_date",
 		"add_maintenance_task",
-		"add_professional_contact",
 		"add_project_guardrail",
 		"add_project_skill",
-		"add_recipe",
 		"add_skill",
-		"add_vendor",
+		"annotate_tool",
 		"archive_thought",
 		"backfill_embeddings",
 		"capture_thought",
-		"create_meal_plan",
-		"create_opportunity",
 		"create_project",
+		"delete_chat_history",
 		"delete_thought",
-		"generate_shopping_list",
+		"describe_tools",
 		"get_active_project",
-		"get_contact_history",
+		"get_chat_history",
-		"get_follow_ups_due",
+		"get_learning",
-		"get_household_item",
-		"get_meal_plan",
 		"get_project_context",
 		"get_thought",
-		"get_upcoming_dates",
 		"get_upcoming_maintenance",
 		"get_version_info",
-		"get_week_schedule",
-		"link_thought_to_contact",
 		"link_thoughts",
-		"list_family_members",
+		"list_chat_histories",
 		"list_files",
 		"list_guardrails",
+		"list_learnings",
 		"list_project_guardrails",
 		"list_project_skills",
 		"list_projects",
 		"list_skills",
 		"list_thoughts",
-		"list_vendors",
 		"load_file",
-		"log_interaction",
 		"log_maintenance",
 		"recall_context",
 		"related_thoughts",
@@ -81,17 +69,13 @@ func TestNewListsAllRegisteredTools(t *testing.T) {
 		"remove_skill",
 		"reparse_thought_metadata",
 		"retry_failed_metadata",
+		"save_chat_history",
 		"save_file",
-		"search_activities",
-		"search_contacts",
-		"search_household_items",
 		"search_maintenance_history",
-		"search_recipes",
 		"search_thoughts",
 		"set_active_project",
 		"summarize_thoughts",
 		"thought_stats",
-		"update_recipe",
 		"update_thought",
 		"upload_file",
 	}

internal/mcpserver/sse_test.go

@@ -0,0 +1,136 @@
+package mcpserver
+
+import (
+	"context"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+
+	"git.warky.dev/wdevs/amcs/internal/config"
+)
+
+func TestNewHandlers_SSEDisabledByDefault(t *testing.T) {
+	h, err := NewHandlers(config.MCPConfig{
+		ServerName:     "test",
+		Version:        "0.0.1",
+		SessionTimeout: time.Minute,
+	}, nil, streamableTestToolSet(), nil)
+	if err != nil {
+		t.Fatalf("NewHandlers() error = %v", err)
+	}
+	if h.StreamableHTTP == nil {
+		t.Fatal("StreamableHTTP handler is nil")
+	}
+	if h.SSE != nil {
+		t.Fatal("SSE handler should be nil when SSEPath is empty")
+	}
+}
+
+func TestNewHandlers_SSEEnabledWhenPathSet(t *testing.T) {
+	h, err := NewHandlers(config.MCPConfig{
+		ServerName:     "test",
+		Version:        "0.0.1",
+		SessionTimeout: time.Minute,
+		SSEPath:        "/sse",
+	}, nil, streamableTestToolSet(), nil)
+	if err != nil {
+		t.Fatalf("NewHandlers() error = %v", err)
+	}
+	if h.StreamableHTTP == nil {
+		t.Fatal("StreamableHTTP handler is nil")
+	}
+	if h.SSE == nil {
+		t.Fatal("SSE handler is nil when SSEPath is set")
+	}
+}
+
+func TestNew_BackwardCompatibility(t *testing.T) {
+	handler, err := New(config.MCPConfig{
+		ServerName:     "test",
+		Version:        "0.0.1",
+		SessionTimeout: time.Minute,
+	}, nil, streamableTestToolSet(), nil)
+	if err != nil {
+		t.Fatalf("New() error = %v", err)
+	}
+	if handler == nil {
+		t.Fatal("New() returned nil handler")
+	}
+}
+
+func TestSSEListTools(t *testing.T) {
+	h, err := NewHandlers(config.MCPConfig{
+		ServerName:     "test",
+		Version:        "0.0.1",
+		SessionTimeout: time.Minute,
+		SSEPath:        "/sse",
+	}, nil, streamableTestToolSet(), nil)
+	if err != nil {
+		t.Fatalf("NewHandlers() error = %v", err)
+	}
+
+	srv := httptest.NewServer(h.SSE)
+	t.Cleanup(srv.Close)
+
+	client := mcp.NewClient(&mcp.Implementation{Name: "client", Version: "0.0.1"}, nil)
+	cs, err := client.Connect(context.Background(), &mcp.SSEClientTransport{Endpoint: srv.URL}, nil)
+	if err != nil {
+		t.Fatalf("connect SSE client: %v", err)
+	}
+	t.Cleanup(func() { _ = cs.Close() })
+
+	result, err := cs.ListTools(context.Background(), nil)
+	if err != nil {
+		t.Fatalf("ListTools() error = %v", err)
+	}
+	if len(result.Tools) == 0 {
+		t.Fatal("ListTools() returned no tools")
+	}
+}
+
+func TestSSEAndStreamableShareTools(t *testing.T) {
+	h, err := NewHandlers(config.MCPConfig{
+		ServerName:     "test",
+		Version:        "0.0.1",
+		SessionTimeout: time.Minute,
+		SSEPath:        "/sse",
+	}, nil, streamableTestToolSet(), nil)
+	if err != nil {
+		t.Fatalf("NewHandlers() error = %v", err)
+	}
+
+	sseSrv := httptest.NewServer(h.SSE)
+	t.Cleanup(sseSrv.Close)
+
+	streamSrv := httptest.NewServer(h.StreamableHTTP)
+	t.Cleanup(streamSrv.Close)
+
+	sseClient := mcp.NewClient(&mcp.Implementation{Name: "sse-client", Version: "0.0.1"}, nil)
+	sseSession, err := sseClient.Connect(context.Background(), &mcp.SSEClientTransport{Endpoint: sseSrv.URL}, nil)
+	if err != nil {
+		t.Fatalf("connect SSE client: %v", err)
+	}
+	t.Cleanup(func() { _ = sseSession.Close() })
+
+	streamClient := mcp.NewClient(&mcp.Implementation{Name: "stream-client", Version: "0.0.1"}, nil)
+	streamSession, err := streamClient.Connect(context.Background(), &mcp.StreamableClientTransport{Endpoint: streamSrv.URL}, nil)
+	if err != nil {
+		t.Fatalf("connect StreamableHTTP client: %v", err)
+	}
+	t.Cleanup(func() { _ = streamSession.Close() })
+
+	sseTools, err := sseSession.ListTools(context.Background(), nil)
+	if err != nil {
+		t.Fatalf("SSE ListTools() error = %v", err)
+	}
+	streamTools, err := streamSession.ListTools(context.Background(), nil)
+	if err != nil {
+		t.Fatalf("StreamableHTTP ListTools() error = %v", err)
+	}
+
+	if len(sseTools.Tools) != len(streamTools.Tools) {
+		t.Fatalf("SSE tool count = %d, StreamableHTTP tool count = %d, want equal", len(sseTools.Tools), len(streamTools.Tools))
+	}
+}

internal/mcpserver/streamable_integration_test.go

@@ -126,12 +126,8 @@ func streamableTestToolSet() ToolSet {
 		Files:         new(tools.FilesTool),
 		Backfill:      new(tools.BackfillTool),
 		Reparse:       new(tools.ReparseMetadataTool),
-		RetryMetadata: new(tools.RetryMetadataTool),
+		RetryMetadata: new(tools.RetryEnrichmentTool),
-		Household:     new(tools.HouseholdTool),
 		Maintenance:   new(tools.MaintenanceTool),
-		Calendar:      new(tools.CalendarTool),
-		Meals:         new(tools.MealsTool),
-		CRM:           new(tools.CRMTool),
 		Skills:        new(tools.SkillsTool),
 	}
 }

internal/observability/http.go

@@ -1,19 +1,25 @@
 package observability
 
 import (
+	"bytes"
 	"context"
+	"encoding/json"
+	"io"
 	"log/slog"
-	"net"
 	"net/http"
 	"runtime/debug"
+	"strings"
 	"time"
 
 	"github.com/google/uuid"
+
+	"git.warky.dev/wdevs/amcs/internal/requestip"
 )
 
 type contextKey string
 
 const requestIDContextKey contextKey = "request_id"
+const mcpToolContextKey contextKey = "mcp_tool"
 
 func Chain(h http.Handler, middlewares ...func(http.Handler) http.Handler) http.Handler {
 	for i := len(middlewares) - 1; i >= 0; i-- {
@@ -57,18 +63,27 @@ func Recover(log *slog.Logger) func(http.Handler) http.Handler {
 func AccessLog(log *slog.Logger) func(http.Handler) http.Handler {
 	return func(next http.Handler) http.Handler {
 		return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			if tool := mcpToolFromRequest(r); tool != "" {
+				r = r.WithContext(context.WithValue(r.Context(), mcpToolContextKey, tool))
+			}
+
 			recorder := &statusRecorder{ResponseWriter: w, status: http.StatusOK}
 			started := time.Now()
 			next.ServeHTTP(recorder, r)
 
-			log.Info("http request",
+			attrs := []any{
 				slog.String("request_id", RequestIDFromContext(r.Context())),
 				slog.String("method", r.Method),
 				slog.String("path", r.URL.Path),
 				slog.Int("status", recorder.status),
 				slog.Duration("duration", time.Since(started)),
-				slog.String("remote_addr", stripPort(r.RemoteAddr)),
+				slog.String("remote_addr", requestip.FromRequest(r)),
-			)
+				slog.String("mcp_session_id", mcpSessionIDFromRequest(r)),
+			}
+			if tool, _ := r.Context().Value(mcpToolContextKey).(string); strings.TrimSpace(tool) != "" {
+				attrs = append(attrs, slog.String("tool", tool), slog.String("tool_call", tool))
+			}
+			log.Info("http request", attrs...)
 		})
 	}
 }
@@ -101,10 +116,67 @@ func (s *statusRecorder) WriteHeader(statusCode int) {
 	s.ResponseWriter.WriteHeader(statusCode)
 }
 
-func stripPort(remote string) string {
+func mcpToolFromRequest(r *http.Request) string {
-	host, _, err := net.SplitHostPort(remote)
+	if r == nil || r.Method != http.MethodPost || !strings.HasPrefix(r.URL.Path, "/mcp") || r.Body == nil {
-	if err != nil {
+		return ""
-		return remote
 	}
-	return host
+
+	raw, err := io.ReadAll(r.Body)
+	if err != nil {
+		return ""
+	}
+	r.Body = io.NopCloser(bytes.NewReader(raw))
+	if len(raw) == 0 {
+		return ""
+	}
+
+	// Support both single and batch JSON-RPC payloads.
+	if strings.HasPrefix(strings.TrimSpace(string(raw)), "[") {
+		var batch []rpcEnvelope
+		if err := json.Unmarshal(raw, &batch); err != nil {
+			return ""
+		}
+		for _, msg := range batch {
+			if tool := msg.toolName(); tool != "" {
+				return tool
+			}
+		}
+		return ""
+	}
+
+	var msg rpcEnvelope
+	if err := json.Unmarshal(raw, &msg); err != nil {
+		return ""
+	}
+	return msg.toolName()
+}
+
+func mcpSessionIDFromRequest(r *http.Request) string {
+	if r == nil {
+		return ""
+	}
+	if v := strings.TrimSpace(r.Header.Get("MCP-Session-Id")); v != "" {
+		return v
+	}
+	// Some clients/proxies may propagate the session in query params.
+	for _, key := range []string{"session_id", "sessionId", "mcp_session_id"} {
+		if v := strings.TrimSpace(r.URL.Query().Get(key)); v != "" {
+			return v
+		}
+	}
+	return ""
+}
+
+type rpcEnvelope struct {
+	Method string `json:"method"`
+	Params struct {
+		Name string `json:"name"`
+	} `json:"params"`
+}
+
+func (m rpcEnvelope) toolName() string {
+	if m.Method != "tools/call" {
+		return ""
+	}
+	return strings.TrimSpace(m.Params.Name)
 }

internal/observability/http_test.go

@@ -1,10 +1,13 @@
 package observability
 
 import (
+	"bytes"
+	"encoding/json"
 	"io"
 	"log/slog"
 	"net/http"
 	"net/http/httptest"
+	"strings"
 	"testing"
 	"time"
 )
@@ -57,3 +60,99 @@ func TestRecoverHandlesPanic(t *testing.T) {
 		t.Fatalf("status = %d, want %d", rec.Code, http.StatusInternalServerError)
 	}
 }
+
+func TestAccessLogUsesForwardedClientIP(t *testing.T) {
+	var buf bytes.Buffer
+	logger := slog.New(slog.NewTextHandler(&buf, nil))
+	handler := AccessLog(logger)(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+
+	req := httptest.NewRequest(http.MethodGet, "/mcp", nil)
+	req.RemoteAddr = "10.0.0.10:1234"
+	req.Header.Set("X-Real-IP", "203.0.113.7")
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNoContent {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusNoContent)
+	}
+	if !strings.Contains(buf.String(), "remote_addr=203.0.113.7") {
+		t.Fatalf("log output = %q, want remote_addr=203.0.113.7", buf.String())
+	}
+}
+
+func TestAccessLogIncludesMCPToolName(t *testing.T) {
+	var buf bytes.Buffer
+	logger := slog.New(slog.NewTextHandler(&buf, nil))
+	handler := AccessLog(logger)(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+
+	payload := map[string]any{
+		"jsonrpc": "2.0",
+		"id":      "1",
+		"method":  "tools/call",
+		"params": map[string]any{
+			"name":      "list_projects",
+			"arguments": map[string]any{},
+		},
+	}
+	body, err := json.Marshal(payload)
+	if err != nil {
+		t.Fatalf("json.Marshal() error = %v", err)
+	}
+
+	req := httptest.NewRequest(http.MethodPost, "/mcp", bytes.NewReader(body))
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNoContent {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusNoContent)
+	}
+	if !strings.Contains(buf.String(), "tool=list_projects") {
+		t.Fatalf("log output = %q, want tool=list_projects", buf.String())
+	}
+	if !strings.Contains(buf.String(), "tool_call=list_projects") {
+		t.Fatalf("log output = %q, want tool_call=list_projects", buf.String())
+	}
+}
+
+func TestAccessLogIncludesMCPSessionIDHeader(t *testing.T) {
+	var buf bytes.Buffer
+	logger := slog.New(slog.NewTextHandler(&buf, nil))
+	handler := AccessLog(logger)(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+
+	req := httptest.NewRequest(http.MethodGet, "/sse", nil)
+	req.Header.Set("MCP-Session-Id", "sess-123")
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNoContent {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusNoContent)
+	}
+	if !strings.Contains(buf.String(), "mcp_session_id=sess-123") {
+		t.Fatalf("log output = %q, want mcp_session_id=sess-123", buf.String())
+	}
+}
+
+func TestAccessLogIncludesMCPSessionIDQueryParam(t *testing.T) {
+	var buf bytes.Buffer
+	logger := slog.New(slog.NewTextHandler(&buf, nil))
+	handler := AccessLog(logger)(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusNoContent)
+	}))
+
+	req := httptest.NewRequest(http.MethodGet, "/sse?session_id=sess-q-1", nil)
+	rec := httptest.NewRecorder()
+	handler.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNoContent {
+		t.Fatalf("status = %d, want %d", rec.Code, http.StatusNoContent)
+	}
+	if !strings.Contains(buf.String(), "mcp_session_id=sess-q-1") {
+		t.Fatalf("log output = %q, want mcp_session_id=sess-q-1", buf.String())
+	}
+}

internal/requestip/requestip.go

@@ -0,0 +1,72 @@
+package requestip
+
+import (
+	"net"
+	"net/http"
+	"strings"
+)
+
+// FromRequest returns the best-effort client IP/host for a request, preferring
+// proxy headers before falling back to RemoteAddr.
+//
+// Header precedence:
+// 1) X-Real-IP
+// 2) X-Forwarded-For (first value)
+// 3) Forwarded (for=...)
+// 4) RemoteAddr (host part)
+func FromRequest(r *http.Request) string {
+	if r == nil {
+		return ""
+	}
+
+	if v := firstAddressToken(r.Header.Get("X-Real-IP")); v != "" {
+		return stripPort(v)
+	}
+	if v := firstAddressToken(r.Header.Get("X-Forwarded-For")); v != "" {
+		return stripPort(v)
+	}
+	if v := forwardedForValue(r.Header.Get("Forwarded")); v != "" {
+		return stripPort(v)
+	}
+	return stripPort(strings.TrimSpace(r.RemoteAddr))
+}
+
+func firstAddressToken(v string) string {
+	if v == "" {
+		return ""
+	}
+	part := strings.TrimSpace(strings.Split(v, ",")[0])
+	part = strings.Trim(part, `"`)
+	return strings.TrimSpace(part)
+}
+
+func forwardedForValue(v string) string {
+	for _, part := range strings.Split(v, ",") {
+		for _, kv := range strings.Split(part, ";") {
+			k, raw, ok := strings.Cut(strings.TrimSpace(kv), "=")
+			if !ok || !strings.EqualFold(strings.TrimSpace(k), "for") {
+				continue
+			}
+			candidate := strings.Trim(strings.TrimSpace(raw), `"`)
+			if candidate == "" {
+				continue
+			}
+			return candidate
+		}
+	}
+	return ""
+}
+
+func stripPort(addr string) string {
+	addr = strings.TrimSpace(addr)
+	if addr == "" {
+		return ""
+	}
+	// RFC 7239 quoted values may wrap IPv6 with brackets.
+	addr = strings.Trim(addr, "[]")
+	host, _, err := net.SplitHostPort(addr)
+	if err == nil {
+		return host
+	}
+	return addr
+}

internal/requestip/requestip_test.go

@@ -0,0 +1,37 @@
+package requestip
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"testing"
+)
+
+func TestFromRequestPrefersXRealIP(t *testing.T) {
+	req := httptest.NewRequest(http.MethodGet, "/", nil)
+	req.RemoteAddr = "10.0.0.10:5555"
+	req.Header.Set("X-Forwarded-For", "198.51.100.1")
+	req.Header.Set("X-Real-IP", "203.0.113.10")
+
+	if got := FromRequest(req); got != "203.0.113.10" {
+		t.Fatalf("FromRequest() = %q, want %q", got, "203.0.113.10")
+	}
+}
+
+func TestFromRequestUsesXForwardedForFirstValue(t *testing.T) {
+	req := httptest.NewRequest(http.MethodGet, "/", nil)
+	req.RemoteAddr = "10.0.0.10:5555"
+	req.Header.Set("X-Forwarded-For", "198.51.100.7, 10.1.1.2")
+
+	if got := FromRequest(req); got != "198.51.100.7" {
+		t.Fatalf("FromRequest() = %q, want %q", got, "198.51.100.7")
+	}
+}
+
+func TestFromRequestFallsBackToRemoteAddr(t *testing.T) {
+	req := httptest.NewRequest(http.MethodGet, "/", nil)
+	req.RemoteAddr = "192.0.2.5:1234"
+
+	if got := FromRequest(req); got != "192.0.2.5" {
+		t.Fatalf("FromRequest() = %q, want %q", got, "192.0.2.5")
+	}
+}

internal/store/learnings.go

@@ -0,0 +1,215 @@
+package store
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/google/uuid"
+	"github.com/jackc/pgx/v5"
+	"github.com/jackc/pgx/v5/pgtype"
+
+	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
+)
+
+func (db *DB) CreateLearning(ctx context.Context, learning thoughttypes.Learning) (thoughttypes.Learning, error) {
+	row := db.pool.QueryRow(ctx, `
+		insert into learnings (
+			summary, details, category, area, status, priority, confidence,
+			action_required, source_type, source_ref, project_id, related_thought_id,
+			related_skill_id, reviewed_by, reviewed_at, duplicate_of_learning_id,
+			supersedes_learning_id, tags
+		) values (
+			$1, $2, $3, $4, $5, $6, $7,
+			$8, $9, $10, $11, $12,
+			$13, $14, $15, $16,
+			$17, $18
+		)
+		returning id, created_at, updated_at
+	`,
+		strings.TrimSpace(learning.Summary),
+		strings.TrimSpace(learning.Details),
+		strings.TrimSpace(learning.Category),
+		strings.TrimSpace(learning.Area),
+		string(learning.Status),
+		string(learning.Priority),
+		string(learning.Confidence),
+		learning.ActionRequired,
+		nullableText(learning.SourceType),
+		nullableText(learning.SourceRef),
+		learning.ProjectID,
+		learning.RelatedThoughtID,
+		learning.RelatedSkillID,
+		nullableTextPtr(learning.ReviewedBy),
+		learning.ReviewedAt,
+		learning.DuplicateOfLearningID,
+		learning.SupersedesLearningID,
+		learning.Tags,
+	)
+
+	created := learning
+	if err := row.Scan(&created.ID, &created.CreatedAt, &created.UpdatedAt); err != nil {
+		return thoughttypes.Learning{}, fmt.Errorf("create learning: %w", err)
+	}
+	return created, nil
+}
+
+func (db *DB) GetLearning(ctx context.Context, id uuid.UUID) (thoughttypes.Learning, error) {
+	row := db.pool.QueryRow(ctx, `
+		select id, summary, details, category, area, status, priority, confidence,
+		       action_required, source_type, source_ref, project_id, related_thought_id,
+		       related_skill_id, reviewed_by, reviewed_at, duplicate_of_learning_id,
+		       supersedes_learning_id, tags, created_at, updated_at
+		from learnings
+		where id = $1
+	`, id)
+
+	learning, err := scanLearning(row)
+	if err != nil {
+		if err == pgx.ErrNoRows {
+			return thoughttypes.Learning{}, fmt.Errorf("learning not found: %s", id)
+		}
+		return thoughttypes.Learning{}, fmt.Errorf("get learning: %w", err)
+	}
+	return learning, nil
+}
+
+func (db *DB) ListLearnings(ctx context.Context, filter thoughttypes.LearningFilter) ([]thoughttypes.Learning, error) {
+	args := make([]any, 0, 8)
+	conditions := make([]string, 0, 8)
+
+	if filter.ProjectID != nil {
+		args = append(args, *filter.ProjectID)
+		conditions = append(conditions, fmt.Sprintf("project_id = $%d", len(args)))
+	}
+	if value := strings.TrimSpace(filter.Category); value != "" {
+		args = append(args, value)
+		conditions = append(conditions, fmt.Sprintf("category = $%d", len(args)))
+	}
+	if value := strings.TrimSpace(filter.Area); value != "" {
+		args = append(args, value)
+		conditions = append(conditions, fmt.Sprintf("area = $%d", len(args)))
+	}
+	if value := strings.TrimSpace(filter.Status); value != "" {
+		args = append(args, value)
+		conditions = append(conditions, fmt.Sprintf("status = $%d", len(args)))
+	}
+	if value := strings.TrimSpace(filter.Priority); value != "" {
+		args = append(args, value)
+		conditions = append(conditions, fmt.Sprintf("priority = $%d", len(args)))
+	}
+	if value := strings.TrimSpace(filter.Tag); value != "" {
+		args = append(args, value)
+		conditions = append(conditions, fmt.Sprintf("$%d = any(tags)", len(args)))
+	}
+	if value := strings.TrimSpace(filter.Query); value != "" {
+		args = append(args, value)
+		conditions = append(conditions, fmt.Sprintf("to_tsvector('simple', summary || ' ' || coalesce(details, '')) @@ websearch_to_tsquery('simple', $%d)", len(args)))
+	}
+
+	query := `
+		select id, summary, details, category, area, status, priority, confidence,
+		       action_required, source_type, source_ref, project_id, related_thought_id,
+		       related_skill_id, reviewed_by, reviewed_at, duplicate_of_learning_id,
+		       supersedes_learning_id, tags, created_at, updated_at
+		from learnings
+	`
+	if len(conditions) > 0 {
+		query += " where " + strings.Join(conditions, " and ")
+	}
+	query += " order by updated_at desc"
+	if filter.Limit > 0 {
+		args = append(args, filter.Limit)
+		query += fmt.Sprintf(" limit $%d", len(args))
+	}
+
+	rows, err := db.pool.Query(ctx, query, args...)
+	if err != nil {
+		return nil, fmt.Errorf("list learnings: %w", err)
+	}
+	defer rows.Close()
+
+	items := make([]thoughttypes.Learning, 0)
+	for rows.Next() {
+		item, err := scanLearning(rows)
+		if err != nil {
+			return nil, fmt.Errorf("scan learning: %w", err)
+		}
+		items = append(items, item)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("iterate learnings: %w", err)
+	}
+	return items, nil
+}
+
+type learningScanner interface {
+	Scan(dest ...any) error
+}
+
+func scanLearning(row learningScanner) (thoughttypes.Learning, error) {
+	var learning thoughttypes.Learning
+	var sourceType pgtype.Text
+	var sourceRef pgtype.Text
+	var reviewedBy pgtype.Text
+	var tags []string
+
+	err := row.Scan(
+		&learning.ID,
+		&learning.Summary,
+		&learning.Details,
+		&learning.Category,
+		&learning.Area,
+		&learning.Status,
+		&learning.Priority,
+		&learning.Confidence,
+		&learning.ActionRequired,
+		&sourceType,
+		&sourceRef,
+		&learning.ProjectID,
+		&learning.RelatedThoughtID,
+		&learning.RelatedSkillID,
+		&reviewedBy,
+		&learning.ReviewedAt,
+		&learning.DuplicateOfLearningID,
+		&learning.SupersedesLearningID,
+		&tags,
+		&learning.CreatedAt,
+		&learning.UpdatedAt,
+	)
+	if err != nil {
+		return thoughttypes.Learning{}, err
+	}
+
+	learning.SourceType = sourceType.String
+	learning.SourceRef = sourceRef.String
+	if reviewedBy.Valid {
+		value := reviewedBy.String
+		learning.ReviewedBy = &value
+	}
+	if tags == nil {
+		learning.Tags = []string{}
+	} else {
+		learning.Tags = tags
+	}
+	return learning, nil
+}
+
+func nullableText(value string) *string {
+	trimmed := strings.TrimSpace(value)
+	if trimmed == "" {
+		return nil
+	}
+	return &trimmed
+}
+
+func nullableTextPtr(value *string) *string {
+	if value == nil {
+		return nil
+	}
+	trimmed := strings.TrimSpace(*value)
+	if trimmed == "" {
+		return nil
+	}
+	return &trimmed
+}

internal/store/projects.go

@@ -26,21 +26,42 @@ func (db *DB) CreateProject(ctx context.Context, name, description string) (thou
 }
 
 func (db *DB) GetProject(ctx context.Context, nameOrID string) (thoughttypes.Project, error) {
-	var row pgx.Row
+	lookup := strings.TrimSpace(nameOrID)
-	if parsedID, err := uuid.Parse(strings.TrimSpace(nameOrID)); err == nil {
+
-		row = db.pool.QueryRow(ctx, `
+	// Prefer guid lookup when input parses as UUID, but fall back to name lookup
+	// so UUID-shaped project names can still be resolved by name.
+	if parsedID, err := uuid.Parse(lookup); err == nil {
+		project, queryErr := db.getProjectByGUID(ctx, parsedID)
+		if queryErr == nil {
+			return project, nil
+		}
+		if queryErr != pgx.ErrNoRows {
+			return thoughttypes.Project{}, queryErr
+		}
+	}
+
+	return db.getProjectByName(ctx, lookup)
+}
+
+func (db *DB) getProjectByGUID(ctx context.Context, id uuid.UUID) (thoughttypes.Project, error) {
+	row := db.pool.QueryRow(ctx, `
 		select guid, name, description, created_at, last_active_at
 		from projects
 		where guid = $1
-		`, parsedID)
+	`, id)
-	} else {
+	return scanProject(row)
-		row = db.pool.QueryRow(ctx, `
+}
+
+func (db *DB) getProjectByName(ctx context.Context, name string) (thoughttypes.Project, error) {
+	row := db.pool.QueryRow(ctx, `
 		select guid, name, description, created_at, last_active_at
 		from projects
 		where name = $1
-		`, strings.TrimSpace(nameOrID))
+	`, name)
-	}
+	return scanProject(row)
+}
 
+func scanProject(row pgx.Row) (thoughttypes.Project, error) {
 	var project thoughttypes.Project
 	if err := row.Scan(&project.ID, &project.Name, &project.Description, &project.CreatedAt, &project.LastActiveAt); err != nil {
 		if err == pgx.ErrNoRows {

internal/store/thoughts.go

@@ -58,6 +58,12 @@ func (db *DB) InsertThought(ctx context.Context, thought thoughttypes.Thought, e
 		return thoughttypes.Thought{}, fmt.Errorf("commit thought insert: %w", err)
 	}
 
+	if len(thought.Embedding) > 0 {
+		created.EmbeddingStatus = "done"
+	} else {
+		created.EmbeddingStatus = "pending"
+	}
+
 	return created, nil
 }
 
@@ -576,7 +582,7 @@ func (db *DB) SearchThoughtsText(ctx context.Context, query string, limit int, p
 	args := []any{query}
 	conditions := []string{
 		"t.archived_at is null",
-		"to_tsvector('simple', t.content) @@ websearch_to_tsquery('simple', $1)",
+		"(to_tsvector('simple', t.content) || to_tsvector('simple', coalesce(p.name, ''))) @@ websearch_to_tsquery('simple', $1)",
 	}
 	if projectID != nil {
 		args = append(args, *projectID)
@@ -590,9 +596,10 @@ func (db *DB) SearchThoughtsText(ctx context.Context, query string, limit int, p
 
 	q := `
 		select t.guid, t.content, t.metadata,
-		       ts_rank_cd(to_tsvector('simple', t.content), websearch_to_tsquery('simple', $1)) as similarity,
+		       ts_rank_cd(to_tsvector('simple', t.content) || to_tsvector('simple', coalesce(p.name, '')), websearch_to_tsquery('simple', $1)) as similarity,
 		       t.created_at
 		from thoughts t
+		left join projects p on t.project_id = p.guid
 		where ` + strings.Join(conditions, " and ") + `
 		order by similarity desc
 		limit $` + fmt.Sprintf("%d", len(args))

internal/store/tool_annotations.go

@@ -0,0 +1,38 @@
+package store
+
+import (
+	"context"
+	"fmt"
+)
+
+func (db *DB) UpsertToolAnnotation(ctx context.Context, toolName, notes string) error {
+	_, err := db.pool.Exec(ctx, `
+		insert into tool_annotations (tool_name, notes)
+		values ($1, $2)
+		on conflict (tool_name) do update
+		  set notes      = excluded.notes,
+		      updated_at = now()
+	`, toolName, notes)
+	if err != nil {
+		return fmt.Errorf("upsert tool annotation: %w", err)
+	}
+	return nil
+}
+
+func (db *DB) GetToolAnnotations(ctx context.Context) (map[string]string, error) {
+	rows, err := db.pool.Query(ctx, `select tool_name, notes from tool_annotations`)
+	if err != nil {
+		return nil, fmt.Errorf("get tool annotations: %w", err)
+	}
+	defer rows.Close()
+
+	annotations := make(map[string]string)
+	for rows.Next() {
+		var toolName, notes string
+		if err := rows.Scan(&toolName, &notes); err != nil {
+			return nil, fmt.Errorf("scan tool annotation: %w", err)
+		}
+		annotations[toolName] = notes
+	}
+	return annotations, rows.Err()
+}

internal/tools/backfill.go

@@ -19,7 +19,7 @@ const backfillConcurrency = 4
 
 type BackfillTool struct {
 	store      *store.DB
-	provider ai.Provider
+	embeddings *ai.EmbeddingRunner
 	sessions   *session.ActiveProjects
 	logger     *slog.Logger
 }
@@ -47,8 +47,51 @@ type BackfillOutput struct {
 	Failures []BackfillFailure `json:"failures,omitempty"`
 }
 
-func NewBackfillTool(db *store.DB, provider ai.Provider, sessions *session.ActiveProjects, logger *slog.Logger) *BackfillTool {
+func NewBackfillTool(db *store.DB, embeddings *ai.EmbeddingRunner, sessions *session.ActiveProjects, logger *slog.Logger) *BackfillTool {
-	return &BackfillTool{store: db, provider: provider, sessions: sessions, logger: logger}
+	return &BackfillTool{store: db, embeddings: embeddings, sessions: sessions, logger: logger}
+}
+
+// QueueThought queues a single thought for background embedding generation.
+// It is used by capture when the embedding provider is temporarily unavailable.
+func (t *BackfillTool) QueueThought(ctx context.Context, id uuid.UUID, content string) {
+	go func() {
+		started := time.Now()
+		t.logger.Info("background embedding started",
+			slog.String("thought_id", id.String()),
+			slog.String("provider", t.embeddings.PrimaryProvider()),
+			slog.String("model", t.embeddings.PrimaryModel()),
+		)
+
+		result, err := t.embeddings.Embed(ctx, content)
+		if err != nil {
+			t.logger.Warn("background embedding error",
+				slog.String("thought_id", id.String()),
+				slog.String("provider", t.embeddings.PrimaryProvider()),
+				slog.String("model", t.embeddings.PrimaryModel()),
+				slog.String("stage", "embed"),
+				slog.Duration("duration", time.Since(started)),
+				slog.String("error", err.Error()),
+			)
+			return
+		}
+		if err := t.store.UpsertEmbedding(ctx, id, result.Model, result.Vector); err != nil {
+			t.logger.Warn("background embedding error",
+				slog.String("thought_id", id.String()),
+				slog.String("provider", t.embeddings.PrimaryProvider()),
+				slog.String("model", result.Model),
+				slog.String("stage", "upsert"),
+				slog.Duration("duration", time.Since(started)),
+				slog.String("error", err.Error()),
+			)
+			return
+		}
+		t.logger.Info("background embedding complete",
+			slog.String("thought_id", id.String()),
+			slog.String("provider", t.embeddings.PrimaryProvider()),
+			slog.String("model", result.Model),
+			slog.Duration("duration", time.Since(started)),
+		)
+	}()
 }
 
 func (t *BackfillTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in BackfillInput) (*mcp.CallToolResult, BackfillOutput, error) {
@@ -67,15 +110,15 @@ func (t *BackfillTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in
 		projectID = &project.ID
 	}
 
-	model := t.provider.EmbeddingModel()
+	primaryModel := t.embeddings.PrimaryModel()
 
-	thoughts, err := t.store.ListThoughtsMissingEmbedding(ctx, model, limit, projectID, in.IncludeArchived, in.OlderThanDays)
+	thoughts, err := t.store.ListThoughtsMissingEmbedding(ctx, primaryModel, limit, projectID, in.IncludeArchived, in.OlderThanDays)
 	if err != nil {
 		return nil, BackfillOutput{}, err
 	}
 
 	out := BackfillOutput{
-		Model:   model,
+		Model:   primaryModel,
 		Scanned: len(thoughts),
 		DryRun:  in.DryRun,
 	}
@@ -101,7 +144,7 @@ func (t *BackfillTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in
 			defer wg.Done()
 			defer sem.Release(1)
 
-			vec, embedErr := t.provider.Embed(ctx, content)
+			result, embedErr := t.embeddings.Embed(ctx, content)
 			if embedErr != nil {
 				mu.Lock()
 				out.Failures = append(out.Failures, BackfillFailure{ID: id.String(), Error: embedErr.Error()})
@@ -110,7 +153,7 @@ func (t *BackfillTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in
 				return
 			}
 
-			if upsertErr := t.store.UpsertEmbedding(ctx, id, model, vec); upsertErr != nil {
+			if upsertErr := t.store.UpsertEmbedding(ctx, id, result.Model, result.Vector); upsertErr != nil {
 				mu.Lock()
 				out.Failures = append(out.Failures, BackfillFailure{ID: id.String(), Error: upsertErr.Error()})
 				mu.Unlock()
@@ -130,7 +173,7 @@ func (t *BackfillTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in
 	out.Skipped = out.Scanned - out.Embedded - out.Failed
 
 	t.logger.Info("backfill completed",
-		slog.String("model", model),
+		slog.String("model", primaryModel),
 		slog.Int("scanned", out.Scanned),
 		slog.Int("embedded", out.Embedded),
 		slog.Int("failed", out.Failed),

internal/tools/capture.go

@@ -2,12 +2,10 @@ package tools
 
 import (
 	"context"
-	"log/slog"
 	"strings"
-	"time"
 
+	"github.com/google/uuid"
 	"github.com/modelcontextprotocol/go-sdk/mcp"
-	"golang.org/x/sync/errgroup"
 
 	"git.warky.dev/wdevs/amcs/internal/ai"
 	"git.warky.dev/wdevs/amcs/internal/config"
@@ -17,14 +15,24 @@ import (
 	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
 )
 
+// EmbeddingQueuer queues a thought for background embedding generation.
+type EmbeddingQueuer interface {
+	QueueThought(ctx context.Context, id uuid.UUID, content string)
+}
+
+// MetadataQueuer queues a thought for background metadata retry. Both
+// MetadataRetryer and EnrichmentRetryer satisfy this.
+type MetadataQueuer interface {
+	QueueThought(id uuid.UUID)
+}
+
 type CaptureTool struct {
 	store        *store.DB
-	provider        ai.Provider
+	embeddings   *ai.EmbeddingRunner
 	capture      config.CaptureConfig
 	sessions     *session.ActiveProjects
-	metadataTimeout time.Duration
+	retryer      MetadataQueuer
-	retryer         *MetadataRetryer
+	embedRetryer EmbeddingQueuer
-	log             *slog.Logger
 }
 
 type CaptureInput struct {
@@ -36,8 +44,8 @@ type CaptureOutput struct {
 	Thought thoughttypes.Thought `json:"thought"`
 }
 
-func NewCaptureTool(db *store.DB, provider ai.Provider, capture config.CaptureConfig, metadataTimeout time.Duration, sessions *session.ActiveProjects, retryer *MetadataRetryer, log *slog.Logger) *CaptureTool {
+func NewCaptureTool(db *store.DB, embeddings *ai.EmbeddingRunner, capture config.CaptureConfig, sessions *session.ActiveProjects, retryer MetadataQueuer, embedRetryer EmbeddingQueuer) *CaptureTool {
-	return &CaptureTool{store: db, provider: provider, capture: capture, sessions: sessions, metadataTimeout: metadataTimeout, retryer: retryer, log: log}
+	return &CaptureTool{store: db, embeddings: embeddings, capture: capture, sessions: sessions, retryer: retryer, embedRetryer: embedRetryer}
 }
 
 func (t *CaptureTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in CaptureInput) (*mcp.CallToolResult, CaptureOutput, error) {
@@ -51,61 +59,30 @@ func (t *CaptureTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in C
 		return nil, CaptureOutput{}, err
 	}
 
-	var embedding []float32
 	rawMetadata := metadata.Fallback(t.capture)
-	metadataNeedsRetry := false
+	rawMetadata.MetadataStatus = metadata.MetadataStatusPending
-
-	group, groupCtx := errgroup.WithContext(ctx)
-	group.Go(func() error {
-		vector, err := t.provider.Embed(groupCtx, content)
-		if err != nil {
-			return err
-		}
-		embedding = vector
-		return nil
-	})
-	group.Go(func() error {
-		metaCtx := groupCtx
-		attemptedAt := time.Now().UTC()
-		if t.metadataTimeout > 0 {
-			var cancel context.CancelFunc
-			metaCtx, cancel = context.WithTimeout(groupCtx, t.metadataTimeout)
-			defer cancel()
-		}
-		extracted, err := t.provider.ExtractMetadata(metaCtx, content)
-		if err != nil {
-			t.log.Warn("metadata extraction failed, using fallback", slog.String("provider", t.provider.Name()), slog.String("error", err.Error()))
-			rawMetadata = metadata.MarkMetadataPending(rawMetadata, t.capture, attemptedAt, err)
-			metadataNeedsRetry = true
-			return nil
-		}
-		rawMetadata = metadata.MarkMetadataComplete(extracted, t.capture, attemptedAt)
-		return nil
-	})
-
-	if err := group.Wait(); err != nil {
-		return nil, CaptureOutput{}, err
-	}
-
 	thought := thoughttypes.Thought{
 		Content:  content,
-		Embedding: embedding,
+		Metadata: rawMetadata,
-		Metadata:  metadata.Normalize(metadata.SanitizeExtracted(rawMetadata), t.capture),
 	}
 	if project != nil {
 		thought.ProjectID = &project.ID
 	}
 
-	created, err := t.store.InsertThought(ctx, thought, t.provider.EmbeddingModel())
+	created, err := t.store.InsertThought(ctx, thought, t.embeddings.PrimaryModel())
 	if err != nil {
 		return nil, CaptureOutput{}, err
 	}
 	if project != nil {
 		_ = t.store.TouchProject(ctx, project.ID)
 	}
-	if metadataNeedsRetry && t.retryer != nil {
+
+	if t.retryer != nil {
 		t.retryer.QueueThought(created.ID)
 	}
+	if t.embedRetryer != nil {
+		t.embedRetryer.QueueThought(ctx, created.ID, content)
+	}
 
 	return nil, CaptureOutput{Thought: created}, nil
 }

internal/tools/context.go

@@ -16,7 +16,7 @@ import (
 
 type ContextTool struct {
 	store      *store.DB
-	provider ai.Provider
+	embeddings *ai.EmbeddingRunner
 	search     config.SearchConfig
 	sessions   *session.ActiveProjects
 }
@@ -41,8 +41,8 @@ type ProjectContextOutput struct {
 	Items   []ContextItem        `json:"items"`
 }
 
-func NewContextTool(db *store.DB, provider ai.Provider, search config.SearchConfig, sessions *session.ActiveProjects) *ContextTool {
+func NewContextTool(db *store.DB, embeddings *ai.EmbeddingRunner, search config.SearchConfig, sessions *session.ActiveProjects) *ContextTool {
-	return &ContextTool{store: db, provider: provider, search: search, sessions: sessions}
+	return &ContextTool{store: db, embeddings: embeddings, search: search, sessions: sessions}
 }
 
 func (t *ContextTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in ProjectContextInput) (*mcp.CallToolResult, ProjectContextOutput, error) {
@@ -72,7 +72,7 @@ func (t *ContextTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in P
 
 	query := strings.TrimSpace(in.Query)
 	if query != "" {
-		semantic, err := semanticSearch(ctx, t.store, t.provider, t.search, query, limit, t.search.DefaultThreshold, &project.ID, nil)
+		semantic, err := semanticSearch(ctx, t.store, t.embeddings, t.search, query, limit, t.search.DefaultThreshold, &project.ID, nil)
 		if err != nil {
 			return nil, ProjectContextOutput{}, err
 		}

internal/tools/describe.go

@@ -0,0 +1,89 @@
+package tools
+
+import (
+	"context"
+	"strings"
+
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+
+	"git.warky.dev/wdevs/amcs/internal/store"
+)
+
+// ToolEntry describes a single registered MCP tool.
+type ToolEntry struct {
+	Name        string
+	Description string
+	Category    string
+}
+
+// DescribeTool implements the describe_tools and annotate_tool MCP tools.
+type DescribeTool struct {
+	store   *store.DB
+	catalog []ToolEntry
+}
+
+func NewDescribeTool(db *store.DB, catalog []ToolEntry) *DescribeTool {
+	return &DescribeTool{store: db, catalog: catalog}
+}
+
+// describe_tools
+
+type DescribeToolsInput struct {
+	Category string `json:"category,omitempty" jsonschema:"filter results to a single category (e.g. thoughts, projects, files, skills, chat, meta)"`
+}
+
+type AnnotatedToolEntry struct {
+	Name        string `json:"name"`
+	Description string `json:"description"`
+	Category    string `json:"category"`
+	Notes       string `json:"notes,omitempty"`
+}
+
+type DescribeToolsOutput struct {
+	Tools []AnnotatedToolEntry `json:"tools"`
+}
+
+func (t *DescribeTool) Describe(ctx context.Context, _ *mcp.CallToolRequest, in DescribeToolsInput) (*mcp.CallToolResult, DescribeToolsOutput, error) {
+	annotations, err := t.store.GetToolAnnotations(ctx)
+	if err != nil {
+		return nil, DescribeToolsOutput{}, err
+	}
+
+	cat := strings.TrimSpace(strings.ToLower(in.Category))
+
+	entries := make([]AnnotatedToolEntry, 0, len(t.catalog))
+	for _, e := range t.catalog {
+		if cat != "" && e.Category != cat {
+			continue
+		}
+		entries = append(entries, AnnotatedToolEntry{
+			Name:        e.Name,
+			Description: e.Description,
+			Category:    e.Category,
+			Notes:       annotations[e.Name],
+		})
+	}
+
+	return nil, DescribeToolsOutput{Tools: entries}, nil
+}
+
+// annotate_tool
+
+type AnnotateToolInput struct {
+	ToolName string `json:"tool_name" jsonschema:"the exact name of the tool to annotate"`
+	Notes    string `json:"notes" jsonschema:"your usage notes, reminders, or gotchas for this tool; pass empty string to clear"`
+}
+
+type AnnotateToolOutput struct {
+	ToolName string `json:"tool_name"`
+}
+
+func (t *DescribeTool) Annotate(ctx context.Context, _ *mcp.CallToolRequest, in AnnotateToolInput) (*mcp.CallToolResult, AnnotateToolOutput, error) {
+	if strings.TrimSpace(in.ToolName) == "" {
+		return nil, AnnotateToolOutput{}, errRequiredField("tool_name")
+	}
+	if err := t.store.UpsertToolAnnotation(ctx, in.ToolName, in.Notes); err != nil {
+		return nil, AnnotateToolOutput{}, err
+	}
+	return nil, AnnotateToolOutput{ToolName: in.ToolName}, nil
+}

internal/tools/enrichment_retry.go

@@ -0,0 +1,227 @@
+package tools
+
+import (
+	"context"
+	"log/slog"
+	"sync"
+	"time"
+
+	"github.com/google/uuid"
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+	"golang.org/x/sync/semaphore"
+
+	"git.warky.dev/wdevs/amcs/internal/ai"
+	"git.warky.dev/wdevs/amcs/internal/config"
+	"git.warky.dev/wdevs/amcs/internal/metadata"
+	"git.warky.dev/wdevs/amcs/internal/session"
+	"git.warky.dev/wdevs/amcs/internal/store"
+	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
+)
+
+const enrichmentRetryConcurrency = 4
+const enrichmentRetryMaxAttempts = 5
+
+var enrichmentRetryBackoff = []time.Duration{
+	30 * time.Second,
+	2 * time.Minute,
+	10 * time.Minute,
+	30 * time.Minute,
+	2 * time.Hour,
+}
+
+type EnrichmentRetryer struct {
+	backgroundCtx   context.Context
+	store           *store.DB
+	metadata        *ai.MetadataRunner
+	capture         config.CaptureConfig
+	sessions        *session.ActiveProjects
+	metadataTimeout time.Duration
+	logger          *slog.Logger
+}
+
+type RetryEnrichmentTool struct {
+	retryer *EnrichmentRetryer
+}
+
+type RetryEnrichmentInput struct {
+	Project         string `json:"project,omitempty" jsonschema:"optional project name or id to scope the retry"`
+	Limit           int    `json:"limit,omitempty" jsonschema:"maximum number of thoughts to process in one call; defaults to 100"`
+	IncludeArchived bool   `json:"include_archived,omitempty" jsonschema:"whether to include archived thoughts; defaults to false"`
+	OlderThanDays   int    `json:"older_than_days,omitempty" jsonschema:"only retry thoughts whose last metadata attempt was at least N days ago; 0 means no restriction"`
+	DryRun          bool   `json:"dry_run,omitempty" jsonschema:"report counts without retrying metadata extraction"`
+}
+
+type RetryEnrichmentFailure struct {
+	ID    string `json:"id"`
+	Error string `json:"error"`
+}
+
+type RetryEnrichmentOutput struct {
+	Scanned  int                      `json:"scanned"`
+	Retried  int                      `json:"retried"`
+	Updated  int                      `json:"updated"`
+	Skipped  int                      `json:"skipped"`
+	Failed   int                      `json:"failed"`
+	DryRun   bool                     `json:"dry_run"`
+	Failures []RetryEnrichmentFailure `json:"failures,omitempty"`
+}
+
+func NewEnrichmentRetryer(backgroundCtx context.Context, db *store.DB, metadataRunner *ai.MetadataRunner, capture config.CaptureConfig, metadataTimeout time.Duration, sessions *session.ActiveProjects, logger *slog.Logger) *EnrichmentRetryer {
+	if backgroundCtx == nil {
+		backgroundCtx = context.Background()
+	}
+	return &EnrichmentRetryer{
+		backgroundCtx:   backgroundCtx,
+		store:           db,
+		metadata:        metadataRunner,
+		capture:         capture,
+		sessions:        sessions,
+		metadataTimeout: metadataTimeout,
+		logger:          logger,
+	}
+}
+
+func NewRetryEnrichmentTool(retryer *EnrichmentRetryer) *RetryEnrichmentTool {
+	return &RetryEnrichmentTool{retryer: retryer}
+}
+
+func (t *RetryEnrichmentTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in RetryEnrichmentInput) (*mcp.CallToolResult, RetryEnrichmentOutput, error) {
+	return t.retryer.Handle(ctx, req, in)
+}
+
+func (r *EnrichmentRetryer) QueueThought(id uuid.UUID) {
+	go func() {
+		started := time.Now()
+		r.logger.Info("background metadata started",
+			slog.String("thought_id", id.String()),
+			slog.String("provider", r.metadata.PrimaryProvider()),
+			slog.String("model", r.metadata.PrimaryModel()),
+		)
+		updated, err := r.retryOne(r.backgroundCtx, id)
+		if err != nil {
+			r.logger.Warn("background metadata error",
+				slog.String("thought_id", id.String()),
+				slog.String("provider", r.metadata.PrimaryProvider()),
+				slog.String("model", r.metadata.PrimaryModel()),
+				slog.Duration("duration", time.Since(started)),
+				slog.String("error", err.Error()),
+			)
+			return
+		}
+		r.logger.Info("background metadata complete",
+			slog.String("thought_id", id.String()),
+			slog.String("provider", r.metadata.PrimaryProvider()),
+			slog.String("model", r.metadata.PrimaryModel()),
+			slog.Bool("updated", updated),
+			slog.Duration("duration", time.Since(started)),
+		)
+	}()
+}
+
+func (r *EnrichmentRetryer) Handle(ctx context.Context, req *mcp.CallToolRequest, in RetryEnrichmentInput) (*mcp.CallToolResult, RetryEnrichmentOutput, error) {
+	limit := in.Limit
+	if limit <= 0 {
+		limit = 100
+	}
+
+	project, err := resolveProject(ctx, r.store, r.sessions, req, in.Project, false)
+	if err != nil {
+		return nil, RetryEnrichmentOutput{}, err
+	}
+
+	var projectID *uuid.UUID
+	if project != nil {
+		projectID = &project.ID
+	}
+
+	thoughts, err := r.store.ListThoughtsPendingMetadataRetry(ctx, limit, projectID, in.IncludeArchived, in.OlderThanDays)
+	if err != nil {
+		return nil, RetryEnrichmentOutput{}, err
+	}
+
+	out := RetryEnrichmentOutput{Scanned: len(thoughts), DryRun: in.DryRun}
+	if in.DryRun || len(thoughts) == 0 {
+		return nil, out, nil
+	}
+
+	sem := semaphore.NewWeighted(enrichmentRetryConcurrency)
+	var mu sync.Mutex
+	var wg sync.WaitGroup
+
+	for _, thought := range thoughts {
+		if ctx.Err() != nil {
+			break
+		}
+		if err := sem.Acquire(ctx, 1); err != nil {
+			break
+		}
+
+		wg.Add(1)
+		go func(thought thoughttypes.Thought) {
+			defer wg.Done()
+			defer sem.Release(1)
+
+			mu.Lock()
+			out.Retried++
+			mu.Unlock()
+
+			updated, err := r.retryOne(ctx, thought.ID)
+			if err != nil {
+				mu.Lock()
+				out.Failures = append(out.Failures, RetryEnrichmentFailure{ID: thought.ID.String(), Error: err.Error()})
+				mu.Unlock()
+				return
+			}
+			if updated {
+				mu.Lock()
+				out.Updated++
+				mu.Unlock()
+				return
+			}
+
+			mu.Lock()
+			out.Skipped++
+			mu.Unlock()
+		}(thought)
+	}
+
+	wg.Wait()
+	out.Failed = len(out.Failures)
+
+	return nil, out, nil
+}
+
+func (r *EnrichmentRetryer) retryOne(ctx context.Context, id uuid.UUID) (bool, error) {
+	thought, err := r.store.GetThought(ctx, id)
+	if err != nil {
+		return false, err
+	}
+	if thought.Metadata.MetadataStatus == metadata.MetadataStatusComplete {
+		return false, nil
+	}
+
+	attemptCtx := ctx
+	if r.metadataTimeout > 0 {
+		var cancel context.CancelFunc
+		attemptCtx, cancel = context.WithTimeout(ctx, r.metadataTimeout)
+		defer cancel()
+	}
+
+	attemptedAt := time.Now().UTC()
+	extracted, extractErr := r.metadata.ExtractMetadata(attemptCtx, thought.Content)
+	if extractErr != nil {
+		failedMetadata := metadata.MarkMetadataFailed(thought.Metadata, r.capture, attemptedAt, extractErr)
+		if _, updateErr := r.store.UpdateThoughtMetadata(ctx, thought.ID, failedMetadata); updateErr != nil {
+			return false, updateErr
+		}
+		return false, extractErr
+	}
+
+	completedMetadata := metadata.MarkMetadataComplete(metadata.SanitizeExtracted(extracted), r.capture, attemptedAt)
+	completedMetadata.Attachments = thought.Metadata.Attachments
+	if _, updateErr := r.store.UpdateThoughtMetadata(ctx, thought.ID, completedMetadata); updateErr != nil {
+		return false, updateErr
+	}
+
+	return true, nil
+}

internal/tools/helpers.go

@@ -87,6 +87,7 @@ func resolveProject(ctx context.Context, db *store.DB, sessions *session.ActiveP
 					Type:    mcperrors.TypeProjectNotFound,
 					Field:   "project",
 					Project: projectRef,
+					Hint:    fmt.Sprintf("project %q does not exist yet; call create_project with name=%q first, then retry", projectRef, projectRef),
 				},
 			)
 		}

internal/tools/learnings.go

@@ -0,0 +1,174 @@
+package tools
+
+import (
+	"context"
+	"strings"
+
+	"github.com/google/uuid"
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+
+	"git.warky.dev/wdevs/amcs/internal/config"
+	"git.warky.dev/wdevs/amcs/internal/session"
+	"git.warky.dev/wdevs/amcs/internal/store"
+	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
+)
+
+type LearningsTool struct {
+	store    *store.DB
+	sessions *session.ActiveProjects
+	cfg      config.SearchConfig
+}
+
+type AddLearningInput struct {
+	Summary               string     `json:"summary" jsonschema:"short curated learning summary"`
+	Details               string     `json:"details,omitempty" jsonschema:"optional detailed learning body"`
+	Category              string     `json:"category,omitempty"`
+	Area                  string     `json:"area,omitempty"`
+	Status                string     `json:"status,omitempty"`
+	Priority              string     `json:"priority,omitempty"`
+	Confidence            string     `json:"confidence,omitempty"`
+	ActionRequired        *bool      `json:"action_required,omitempty"`
+	SourceType            string     `json:"source_type,omitempty"`
+	SourceRef             string     `json:"source_ref,omitempty"`
+	Project               string     `json:"project,omitempty" jsonschema:"project name or id; falls back to active session project"`
+	RelatedThoughtID      *uuid.UUID `json:"related_thought_id,omitempty"`
+	RelatedSkillID        *uuid.UUID `json:"related_skill_id,omitempty"`
+	ReviewedBy            *string    `json:"reviewed_by,omitempty"`
+	DuplicateOfLearningID *uuid.UUID `json:"duplicate_of_learning_id,omitempty"`
+	SupersedesLearningID  *uuid.UUID `json:"supersedes_learning_id,omitempty"`
+	Tags                  []string   `json:"tags,omitempty"`
+}
+
+type AddLearningOutput struct {
+	Learning thoughttypes.Learning `json:"learning"`
+}
+
+type GetLearningInput struct {
+	ID uuid.UUID `json:"id" jsonschema:"learning id"`
+}
+
+type GetLearningOutput struct {
+	Learning thoughttypes.Learning `json:"learning"`
+}
+
+type ListLearningsInput struct {
+	Limit    int    `json:"limit,omitempty"`
+	Project  string `json:"project,omitempty" jsonschema:"project name or id; falls back to active session project"`
+	Category string `json:"category,omitempty"`
+	Area     string `json:"area,omitempty"`
+	Status   string `json:"status,omitempty"`
+	Priority string `json:"priority,omitempty"`
+	Tag      string `json:"tag,omitempty"`
+	Query    string `json:"query,omitempty"`
+}
+
+type ListLearningsOutput struct {
+	Learnings []thoughttypes.Learning `json:"learnings"`
+}
+
+func NewLearningsTool(db *store.DB, sessions *session.ActiveProjects, cfg config.SearchConfig) *LearningsTool {
+	return &LearningsTool{store: db, sessions: sessions, cfg: cfg}
+}
+
+func (t *LearningsTool) Add(ctx context.Context, req *mcp.CallToolRequest, in AddLearningInput) (*mcp.CallToolResult, AddLearningOutput, error) {
+	summary := strings.TrimSpace(in.Summary)
+	if summary == "" {
+		return nil, AddLearningOutput{}, errRequiredField("summary")
+	}
+
+	project, err := resolveProject(ctx, t.store, t.sessions, req, in.Project, false)
+	if err != nil {
+		return nil, AddLearningOutput{}, err
+	}
+
+	learning := thoughttypes.Learning{
+		Summary:               summary,
+		Details:               strings.TrimSpace(in.Details),
+		Category:              defaultString(strings.TrimSpace(in.Category), "insight"),
+		Area:                  defaultString(strings.TrimSpace(in.Area), "other"),
+		Status:                thoughttypes.LearningStatus(defaultString(strings.TrimSpace(in.Status), string(thoughttypes.LearningStatusPending))),
+		Priority:              thoughttypes.LearningPriority(defaultString(strings.TrimSpace(in.Priority), string(thoughttypes.LearningPriorityMedium))),
+		Confidence:            thoughttypes.LearningEvidenceLevel(defaultString(strings.TrimSpace(in.Confidence), string(thoughttypes.LearningEvidenceHypothesis))),
+		SourceType:            strings.TrimSpace(in.SourceType),
+		SourceRef:             strings.TrimSpace(in.SourceRef),
+		RelatedThoughtID:      in.RelatedThoughtID,
+		RelatedSkillID:        in.RelatedSkillID,
+		ReviewedBy:            in.ReviewedBy,
+		DuplicateOfLearningID: in.DuplicateOfLearningID,
+		SupersedesLearningID:  in.SupersedesLearningID,
+		Tags:                  normalizeStringSlice(in.Tags),
+	}
+	if in.ActionRequired != nil {
+		learning.ActionRequired = *in.ActionRequired
+	}
+	if project != nil {
+		learning.ProjectID = &project.ID
+	}
+
+	created, err := t.store.CreateLearning(ctx, learning)
+	if err != nil {
+		return nil, AddLearningOutput{}, err
+	}
+	return nil, AddLearningOutput{Learning: created}, nil
+}
+
+func (t *LearningsTool) Get(ctx context.Context, _ *mcp.CallToolRequest, in GetLearningInput) (*mcp.CallToolResult, GetLearningOutput, error) {
+	learning, err := t.store.GetLearning(ctx, in.ID)
+	if err != nil {
+		return nil, GetLearningOutput{}, err
+	}
+	return nil, GetLearningOutput{Learning: learning}, nil
+}
+
+func (t *LearningsTool) List(ctx context.Context, req *mcp.CallToolRequest, in ListLearningsInput) (*mcp.CallToolResult, ListLearningsOutput, error) {
+	project, err := resolveProject(ctx, t.store, t.sessions, req, in.Project, false)
+	if err != nil {
+		return nil, ListLearningsOutput{}, err
+	}
+
+	filter := thoughttypes.LearningFilter{
+		Limit:    normalizeLimit(in.Limit, t.cfg),
+		Category: strings.TrimSpace(in.Category),
+		Area:     strings.TrimSpace(in.Area),
+		Status:   strings.TrimSpace(in.Status),
+		Priority: strings.TrimSpace(in.Priority),
+		Tag:      strings.TrimSpace(in.Tag),
+		Query:    strings.TrimSpace(in.Query),
+	}
+	if project != nil {
+		filter.ProjectID = &project.ID
+	}
+
+	items, err := t.store.ListLearnings(ctx, filter)
+	if err != nil {
+		return nil, ListLearningsOutput{}, err
+	}
+	return nil, ListLearningsOutput{Learnings: items}, nil
+}
+
+func defaultString(value string, fallback string) string {
+	if value == "" {
+		return fallback
+	}
+	return value
+}
+
+func normalizeStringSlice(values []string) []string {
+	if len(values) == 0 {
+		return []string{}
+	}
+	out := make([]string, 0, len(values))
+	seen := map[string]struct{}{}
+	for _, value := range values {
+		trimmed := strings.TrimSpace(value)
+		if trimmed == "" {
+			continue
+		}
+		if _, ok := seen[trimmed]; ok {
+			continue
+		}
+		seen[trimmed] = struct{}{}
+		out = append(out, trimmed)
+	}
+	return out
+}

internal/tools/links.go

@@ -14,7 +14,7 @@ import (
 
 type LinksTool struct {
 	store      *store.DB
-	provider ai.Provider
+	embeddings *ai.EmbeddingRunner
 	search     config.SearchConfig
 }
 
@@ -47,8 +47,8 @@ type RelatedOutput struct {
 	Related []RelatedThought `json:"related"`
 }
 
-func NewLinksTool(db *store.DB, provider ai.Provider, search config.SearchConfig) *LinksTool {
+func NewLinksTool(db *store.DB, embeddings *ai.EmbeddingRunner, search config.SearchConfig) *LinksTool {
-	return &LinksTool{store: db, provider: provider, search: search}
+	return &LinksTool{store: db, embeddings: embeddings, search: search}
 }
 
 func (t *LinksTool) Link(ctx context.Context, _ *mcp.CallToolRequest, in LinkInput) (*mcp.CallToolResult, LinkOutput, error) {
@@ -117,7 +117,7 @@ func (t *LinksTool) Related(ctx context.Context, _ *mcp.CallToolRequest, in Rela
 	}
 
 	if includeSemantic {
-		semantic, err := semanticSearch(ctx, t.store, t.provider, t.search, thought.Content, t.search.DefaultLimit, t.search.DefaultThreshold, thought.ProjectID, &thought.ID)
+		semantic, err := semanticSearch(ctx, t.store, t.embeddings, t.search, thought.Content, t.search.DefaultLimit, t.search.DefaultThreshold, thought.ProjectID, &thought.ID)
 		if err != nil {
 			return nil, RelatedOutput{}, err
 		}

internal/tools/metadata_retry.go

@@ -23,17 +23,47 @@ const metadataRetryConcurrency = 4
 type MetadataRetryer struct {
 	backgroundCtx   context.Context
 	store           *store.DB
-	provider        ai.Provider
+	metadata        *ai.MetadataRunner
 	capture         config.CaptureConfig
 	sessions        *session.ActiveProjects
 	metadataTimeout time.Duration
 	logger          *slog.Logger
+	lock            *RetryLocker
 }
 
 type RetryMetadataTool struct {
 	retryer *MetadataRetryer
 }
 
+type RetryLocker struct {
+	mu    sync.Mutex
+	locks map[uuid.UUID]time.Time
+}
+
+func NewRetryLocker() *RetryLocker {
+	return &RetryLocker{locks: map[uuid.UUID]time.Time{}}
+}
+
+func (l *RetryLocker) Acquire(id uuid.UUID, ttl time.Duration) bool {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	if l.locks == nil {
+		l.locks = map[uuid.UUID]time.Time{}
+	}
+	now := time.Now()
+	if exp, ok := l.locks[id]; ok && exp.After(now) {
+		return false
+	}
+	l.locks[id] = now.Add(ttl)
+	return true
+}
+
+func (l *RetryLocker) Release(id uuid.UUID) {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	delete(l.locks, id)
+}
+
 type RetryMetadataInput struct {
 	Project         string `json:"project,omitempty" jsonschema:"optional project name or id to scope the retry"`
 	Limit           int    `json:"limit,omitempty" jsonschema:"maximum number of thoughts to process in one call; defaults to 100"`
@@ -57,18 +87,19 @@ type RetryMetadataOutput struct {
 	Failures []RetryMetadataFailure `json:"failures,omitempty"`
 }
 
-func NewMetadataRetryer(backgroundCtx context.Context, db *store.DB, provider ai.Provider, capture config.CaptureConfig, metadataTimeout time.Duration, sessions *session.ActiveProjects, logger *slog.Logger) *MetadataRetryer {
+func NewMetadataRetryer(backgroundCtx context.Context, db *store.DB, metadataRunner *ai.MetadataRunner, capture config.CaptureConfig, metadataTimeout time.Duration, sessions *session.ActiveProjects, logger *slog.Logger) *MetadataRetryer {
 	if backgroundCtx == nil {
 		backgroundCtx = context.Background()
 	}
 	return &MetadataRetryer{
 		backgroundCtx:   backgroundCtx,
 		store:           db,
-		provider:        provider,
+		metadata:        metadataRunner,
 		capture:         capture,
 		sessions:        sessions,
 		metadataTimeout: metadataTimeout,
 		logger:          logger,
+		lock:            NewRetryLocker(),
 	}
 }
 
@@ -82,9 +113,35 @@ func (t *RetryMetadataTool) Handle(ctx context.Context, req *mcp.CallToolRequest
 
 func (r *MetadataRetryer) QueueThought(id uuid.UUID) {
 	go func() {
-		if _, err := r.retryOne(r.backgroundCtx, id); err != nil {
+		started := time.Now()
-			r.logger.Warn("background metadata retry failed", slog.String("thought_id", id.String()), slog.String("error", err.Error()))
+		if !r.lock.Acquire(id, 15*time.Minute) {
+			return
 		}
+		defer r.lock.Release(id)
+
+		r.logger.Info("background metadata started",
+			slog.String("thought_id", id.String()),
+			slog.String("provider", r.metadata.PrimaryProvider()),
+			slog.String("model", r.metadata.PrimaryModel()),
+		)
+		updated, err := r.retryOne(r.backgroundCtx, id)
+		if err != nil {
+			r.logger.Warn("background metadata error",
+				slog.String("thought_id", id.String()),
+				slog.String("provider", r.metadata.PrimaryProvider()),
+				slog.String("model", r.metadata.PrimaryModel()),
+				slog.Duration("duration", time.Since(started)),
+				slog.String("error", err.Error()),
+			)
+			return
+		}
+		r.logger.Info("background metadata complete",
+			slog.String("thought_id", id.String()),
+			slog.String("provider", r.metadata.PrimaryProvider()),
+			slog.String("model", r.metadata.PrimaryModel()),
+			slog.Bool("updated", updated),
+			slog.Duration("duration", time.Since(started)),
+		)
 	}()
 }
 
@@ -138,7 +195,14 @@ func (r *MetadataRetryer) Handle(ctx context.Context, req *mcp.CallToolRequest,
 			out.Retried++
 			mu.Unlock()
 
+			if !r.lock.Acquire(thought.ID, 15*time.Minute) {
+				mu.Lock()
+				out.Skipped++
+				mu.Unlock()
+				return
+			}
 			updated, err := r.retryOne(ctx, thought.ID)
+			r.lock.Release(thought.ID)
 			if err != nil {
 				mu.Lock()
 				out.Failures = append(out.Failures, RetryMetadataFailure{ID: thought.ID.String(), Error: err.Error()})
@@ -181,7 +245,7 @@ func (r *MetadataRetryer) retryOne(ctx context.Context, id uuid.UUID) (bool, err
 	}
 
 	attemptedAt := time.Now().UTC()
-	extracted, extractErr := r.provider.ExtractMetadata(attemptCtx, thought.Content)
+	extracted, extractErr := r.metadata.ExtractMetadata(attemptCtx, thought.Content)
 	if extractErr != nil {
 		failedMetadata := metadata.MarkMetadataFailed(thought.Metadata, r.capture, attemptedAt, extractErr)
 		if _, updateErr := r.store.UpdateThoughtMetadata(ctx, thought.ID, failedMetadata); updateErr != nil {

internal/tools/recall.go

@@ -16,7 +16,7 @@ import (
 
 type RecallTool struct {
 	store      *store.DB
-	provider ai.Provider
+	embeddings *ai.EmbeddingRunner
 	search     config.SearchConfig
 	sessions   *session.ActiveProjects
 }
@@ -32,8 +32,8 @@ type RecallOutput struct {
 	Items   []ContextItem `json:"items"`
 }
 
-func NewRecallTool(db *store.DB, provider ai.Provider, search config.SearchConfig, sessions *session.ActiveProjects) *RecallTool {
+func NewRecallTool(db *store.DB, embeddings *ai.EmbeddingRunner, search config.SearchConfig, sessions *session.ActiveProjects) *RecallTool {
-	return &RecallTool{store: db, provider: provider, search: search, sessions: sessions}
+	return &RecallTool{store: db, embeddings: embeddings, search: search, sessions: sessions}
 }
 
 func (t *RecallTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in RecallInput) (*mcp.CallToolResult, RecallOutput, error) {
@@ -54,7 +54,7 @@ func (t *RecallTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in Re
 		projectID = &project.ID
 	}
 
-	semantic, err := semanticSearch(ctx, t.store, t.provider, t.search, query, limit, t.search.DefaultThreshold, projectID, nil)
+	semantic, err := semanticSearch(ctx, t.store, t.embeddings, t.search, query, limit, t.search.DefaultThreshold, projectID, nil)
 	if err != nil {
 		return nil, RecallOutput{}, err
 	}

internal/tools/reparse_metadata.go

@@ -23,7 +23,7 @@ const metadataReparseConcurrency = 4
 
 type ReparseMetadataTool struct {
 	store    *store.DB
-	provider ai.Provider
+	metadata *ai.MetadataRunner
 	capture  config.CaptureConfig
 	sessions *session.ActiveProjects
 	logger   *slog.Logger
@@ -53,8 +53,8 @@ type ReparseMetadataOutput struct {
 	Failures   []ReparseMetadataFailure `json:"failures,omitempty"`
 }
 
-func NewReparseMetadataTool(db *store.DB, provider ai.Provider, capture config.CaptureConfig, sessions *session.ActiveProjects, logger *slog.Logger) *ReparseMetadataTool {
+func NewReparseMetadataTool(db *store.DB, metadataRunner *ai.MetadataRunner, capture config.CaptureConfig, sessions *session.ActiveProjects, logger *slog.Logger) *ReparseMetadataTool {
-	return &ReparseMetadataTool{store: db, provider: provider, capture: capture, sessions: sessions, logger: logger}
+	return &ReparseMetadataTool{store: db, metadata: metadataRunner, capture: capture, sessions: sessions, logger: logger}
 }
 
 func (t *ReparseMetadataTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in ReparseMetadataInput) (*mcp.CallToolResult, ReparseMetadataOutput, error) {
@@ -107,7 +107,7 @@ func (t *ReparseMetadataTool) Handle(ctx context.Context, req *mcp.CallToolReque
 			normalizedCurrent := metadata.Normalize(thought.Metadata, t.capture)
 
 			attemptedAt := time.Now().UTC()
-			extracted, extractErr := t.provider.ExtractMetadata(ctx, thought.Content)
+			extracted, extractErr := t.metadata.ExtractMetadata(ctx, thought.Content)
 			normalizedTarget := normalizedCurrent
 			if extractErr != nil {
 				normalizedTarget = metadata.MarkMetadataFailed(normalizedCurrent, t.capture, attemptedAt, extractErr)

internal/tools/retrieval.go

@@ -11,12 +11,14 @@ import (
 	thoughttypes "git.warky.dev/wdevs/amcs/internal/types"
 )
 
-// semanticSearch runs vector similarity search if embeddings exist for the active model
+// semanticSearch runs vector similarity search if embeddings exist for the
-// in the given scope, otherwise falls back to Postgres full-text search.
+// primary embedding model in the given scope, otherwise falls back to Postgres
+// full-text search. Search always uses the primary model so query vectors
+// match rows stored under the primary model name.
 func semanticSearch(
 	ctx context.Context,
 	db *store.DB,
-	provider ai.Provider,
+	embeddings *ai.EmbeddingRunner,
 	search config.SearchConfig,
 	query string,
 	limit int,
@@ -24,17 +26,18 @@ func semanticSearch(
 	projectID *uuid.UUID,
 	excludeID *uuid.UUID,
 ) ([]thoughttypes.SearchResult, error) {
-	hasEmbeddings, err := db.HasEmbeddingsForModel(ctx, provider.EmbeddingModel(), projectID)
+	model := embeddings.PrimaryModel()
+	hasEmbeddings, err := db.HasEmbeddingsForModel(ctx, model, projectID)
 	if err != nil {
 		return nil, err
 	}
 
 	if hasEmbeddings {
-		embedding, err := provider.Embed(ctx, query)
+		embedding, err := embeddings.EmbedPrimary(ctx, query)
 		if err != nil {
 			return nil, err
 		}
-		return db.SearchSimilarThoughts(ctx, embedding, provider.EmbeddingModel(), threshold, limit, projectID, excludeID)
+		return db.SearchSimilarThoughts(ctx, embedding, model, threshold, limit, projectID, excludeID)
 	}
 
 	return db.SearchThoughtsText(ctx, query, limit, projectID, excludeID)

internal/tools/search.go

@@ -16,7 +16,7 @@ import (
 
 type SearchTool struct {
 	store      *store.DB
-	provider ai.Provider
+	embeddings *ai.EmbeddingRunner
 	search     config.SearchConfig
 	sessions   *session.ActiveProjects
 }
@@ -32,8 +32,8 @@ type SearchOutput struct {
 	Results []thoughttypes.SearchResult `json:"results"`
 }
 
-func NewSearchTool(db *store.DB, provider ai.Provider, search config.SearchConfig, sessions *session.ActiveProjects) *SearchTool {
+func NewSearchTool(db *store.DB, embeddings *ai.EmbeddingRunner, search config.SearchConfig, sessions *session.ActiveProjects) *SearchTool {
-	return &SearchTool{store: db, provider: provider, search: search, sessions: sessions}
+	return &SearchTool{store: db, embeddings: embeddings, search: search, sessions: sessions}
 }
 
 func (t *SearchTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in SearchInput) (*mcp.CallToolResult, SearchOutput, error) {
@@ -56,7 +56,7 @@ func (t *SearchTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in Se
 		_ = t.store.TouchProject(ctx, project.ID)
 	}
 
-	results, err := semanticSearch(ctx, t.store, t.provider, t.search, query, limit, threshold, projectID, nil)
+	results, err := semanticSearch(ctx, t.store, t.embeddings, t.search, query, limit, threshold, projectID, nil)
 	if err != nil {
 		return nil, SearchOutput{}, err
 	}

internal/tools/summarize.go

@@ -15,7 +15,8 @@ import (
 
 type SummarizeTool struct {
 	store      *store.DB
-	provider ai.Provider
+	embeddings *ai.EmbeddingRunner
+	metadata   *ai.MetadataRunner
 	search     config.SearchConfig
 	sessions   *session.ActiveProjects
 }
@@ -32,8 +33,8 @@ type SummarizeOutput struct {
 	Count   int    `json:"count"`
 }
 
-func NewSummarizeTool(db *store.DB, provider ai.Provider, search config.SearchConfig, sessions *session.ActiveProjects) *SummarizeTool {
+func NewSummarizeTool(db *store.DB, embeddings *ai.EmbeddingRunner, metadata *ai.MetadataRunner, search config.SearchConfig, sessions *session.ActiveProjects) *SummarizeTool {
-	return &SummarizeTool{store: db, provider: provider, search: search, sessions: sessions}
+	return &SummarizeTool{store: db, embeddings: embeddings, metadata: metadata, search: search, sessions: sessions}
 }
 
 func (t *SummarizeTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in SummarizeInput) (*mcp.CallToolResult, SummarizeOutput, error) {
@@ -52,7 +53,7 @@ func (t *SummarizeTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in
 		if project != nil {
 			projectID = &project.ID
 		}
-		results, err := semanticSearch(ctx, t.store, t.provider, t.search, query, limit, t.search.DefaultThreshold, projectID, nil)
+		results, err := semanticSearch(ctx, t.store, t.embeddings, t.search, query, limit, t.search.DefaultThreshold, projectID, nil)
 		if err != nil {
 			return nil, SummarizeOutput{}, err
 		}
@@ -77,7 +78,7 @@ func (t *SummarizeTool) Handle(ctx context.Context, req *mcp.CallToolRequest, in
 
 	userPrompt := formatContextBlock("Summarize the following thoughts into concise prose with themes, action items, and notable people.", lines)
 	systemPrompt := "You summarize note collections. Be concise, concrete, and structured in plain prose."
-	summary, err := t.provider.Summarize(ctx, systemPrompt, userPrompt)
+	summary, err := t.metadata.Summarize(ctx, systemPrompt, userPrompt)
 	if err != nil {
 		return nil, SummarizeOutput{}, err
 	}

internal/tools/update.go

@@ -17,7 +17,8 @@ import (
 
 type UpdateTool struct {
 	store      *store.DB
-	provider ai.Provider
+	embeddings *ai.EmbeddingRunner
+	metadata   *ai.MetadataRunner
 	capture    config.CaptureConfig
 	log        *slog.Logger
 }
@@ -33,8 +34,8 @@ type UpdateOutput struct {
 	Thought thoughttypes.Thought `json:"thought"`
 }
 
-func NewUpdateTool(db *store.DB, provider ai.Provider, capture config.CaptureConfig, log *slog.Logger) *UpdateTool {
+func NewUpdateTool(db *store.DB, embeddings *ai.EmbeddingRunner, metadata *ai.MetadataRunner, capture config.CaptureConfig, log *slog.Logger) *UpdateTool {
-	return &UpdateTool{store: db, provider: provider, capture: capture, log: log}
+	return &UpdateTool{store: db, embeddings: embeddings, metadata: metadata, capture: capture, log: log}
 }
 
 func (t *UpdateTool) Handle(ctx context.Context, _ *mcp.CallToolRequest, in UpdateInput) (*mcp.CallToolResult, UpdateOutput, error) {
@@ -50,6 +51,7 @@ func (t *UpdateTool) Handle(ctx context.Context, _ *mcp.CallToolRequest, in Upda
 
 	content := current.Content
 	var embedding []float32
+	embeddingModel := ""
 	mergedMetadata := current.Metadata
 	projectID := current.ProjectID
 
@@ -58,11 +60,13 @@ func (t *UpdateTool) Handle(ctx context.Context, _ *mcp.CallToolRequest, in Upda
 		if content == "" {
 			return nil, UpdateOutput{}, errInvalidInput("content must not be empty")
 		}
-		embedding, err = t.provider.Embed(ctx, content)
+		embedResult, err := t.embeddings.Embed(ctx, content)
 		if err != nil {
 			return nil, UpdateOutput{}, err
 		}
-		extracted, extractErr := t.provider.ExtractMetadata(ctx, content)
+		embedding = embedResult.Vector
+		embeddingModel = embedResult.Model
+		extracted, extractErr := t.metadata.ExtractMetadata(ctx, content)
 		if extractErr != nil {
 			t.log.Warn("metadata extraction failed during update, keeping current metadata", slog.String("error", extractErr.Error()))
 			mergedMetadata = metadata.MarkMetadataFailed(mergedMetadata, t.capture, time.Now().UTC(), extractErr)
@@ -82,7 +86,7 @@ func (t *UpdateTool) Handle(ctx context.Context, _ *mcp.CallToolRequest, in Upda
 		projectID = &project.ID
 	}
 
-	updated, err := t.store.UpdateThought(ctx, id, content, embedding, t.provider.EmbeddingModel(), mergedMetadata, projectID)
+	updated, err := t.store.UpdateThought(ctx, id, content, embedding, embeddingModel, mergedMetadata, projectID)
 	if err != nil {
 		return nil, UpdateOutput{}, err
 	}

internal/types/learning.go

@@ -0,0 +1,68 @@
+package types
+
+import (
+	"time"
+
+	"github.com/google/uuid"
+)
+
+type LearningEvidenceLevel string
+
+const (
+	LearningEvidenceHypothesis LearningEvidenceLevel = "hypothesis"
+	LearningEvidenceObserved   LearningEvidenceLevel = "observed"
+	LearningEvidenceVerified   LearningEvidenceLevel = "verified"
+)
+
+type LearningStatus string
+
+const (
+	LearningStatusPending    LearningStatus = "pending"
+	LearningStatusInProgress LearningStatus = "in_progress"
+	LearningStatusResolved   LearningStatus = "resolved"
+	LearningStatusWontFix    LearningStatus = "wont_fix"
+	LearningStatusPromoted   LearningStatus = "promoted"
+)
+
+type LearningPriority string
+
+const (
+	LearningPriorityLow    LearningPriority = "low"
+	LearningPriorityMedium LearningPriority = "medium"
+	LearningPriorityHigh   LearningPriority = "high"
+)
+
+type Learning struct {
+	ID                    uuid.UUID             `json:"id"`
+	Summary               string                `json:"summary"`
+	Details               string                `json:"details"`
+	Category              string                `json:"category"`
+	Area                  string                `json:"area"`
+	Status                LearningStatus        `json:"status"`
+	Priority              LearningPriority      `json:"priority"`
+	Confidence            LearningEvidenceLevel `json:"confidence"`
+	ActionRequired        bool                  `json:"action_required"`
+	SourceType            string                `json:"source_type,omitempty"`
+	SourceRef             string                `json:"source_ref,omitempty"`
+	ProjectID             *uuid.UUID            `json:"project_id,omitempty"`
+	RelatedThoughtID      *uuid.UUID            `json:"related_thought_id,omitempty"`
+	RelatedSkillID        *uuid.UUID            `json:"related_skill_id,omitempty"`
+	ReviewedBy            *string               `json:"reviewed_by,omitempty"`
+	ReviewedAt            *time.Time            `json:"reviewed_at,omitempty"`
+	DuplicateOfLearningID *uuid.UUID            `json:"duplicate_of_learning_id,omitempty"`
+	SupersedesLearningID  *uuid.UUID            `json:"supersedes_learning_id,omitempty"`
+	Tags                  []string              `json:"tags"`
+	CreatedAt             time.Time             `json:"created_at"`
+	UpdatedAt             time.Time             `json:"updated_at"`
+}
+
+type LearningFilter struct {
+	Limit     int
+	ProjectID *uuid.UUID
+	Category  string
+	Area      string
+	Status    string
+	Priority  string
+	Tag       string
+	Query     string
+}

internal/types/thought.go

@@ -55,6 +55,7 @@ type Thought struct {
 	ID              uuid.UUID       `json:"id"`
 	Content         string          `json:"content"`
 	Embedding       []float32       `json:"embedding,omitempty"`
+	EmbeddingStatus string          `json:"embedding_status,omitempty"`
 	Metadata        ThoughtMetadata `json:"metadata"`
 	ProjectID       *uuid.UUID      `json:"project_id,omitempty"`
 	ArchivedAt      *time.Time      `json:"archived_at,omitempty"`

llm/learnings_schema.md

@@ -0,0 +1,77 @@
+# Structured Learnings Schema (v1)
+
+## Data Model
+
+| Field | Type | Description |
+|-------|------|-------------|
+| **ID** | string | Stable learning identifier |
+| **Category** | enum | `correction`, `insight`, `knowledge_gap`, `best_practice` |
+| **Area** | enum | `frontend`, `backend`, `infra`, `tests`, `docs`, `config`, `other` |
+| **Status** | enum | `pending`, `in_progress`, `resolved`, `wont_f` |
+| **Priority** | string | e.g., `low`, `medium`, `high` |
+| **Summary** | string | Brief description |
+| **Details** | string | Full description / context |
+| **ProjectID** | string (optional) | Reference to a project |
+| **ThoughtID** | string (optional) | Reference to a thought |
+| **SkillID** | string (optional) | Reference to a skill |
+| **CreatedAt** | timestamp | Creation timestamp |
+| **UpdatedAt** | timestamp | Last update timestamp |
+
+## Suggested SQL Definition
+
+```sql
+CREATE TABLE learnings (
+    id UUID PRIMARY KEY,
+    category TEXT NOT NULL,
+    area TEXT NOT NULL,
+    status TEXT NOT NULL,
+    priority TEXT,
+    summary TEXT,
+    details TEXT,
+    project_id UUID,
+    thought_id UUID,
+    skill_id UUID,
+    created_at TIMESTAMP WITH TIME ZONE DEFAULT NOW(),
+    updated_at TIMESTAMP WITH TIME ZONE DEFAULT NOW()
+);
+```
+
+## Tool Surface (MCP)
+
+- `create_learning` – insert a new learning record  
+- `list_learnings` – query with optional filters (category, area, status, project, etc.)  
+- `get_learning` – retrieve a single learning by ID  
+- `update_learning` – modify fields (e.g., status, priority) and/or links  
+
+## Enums (Go)
+
+```go
+type LearningCategory string
+const (
+    LearningCategoryCorrection LearningCategory = "correction"
+    LearningCategoryInsight    LearningCategory = "insight"
+    LearningCategoryKnowledgeGap LearningCategory = "knowledge_gap"
+    LearningCategoryBestPractice LearningCategory = "best_practice"
+)
+
+type LearningArea string
+const (
+    LearningAreaFrontend      LearningArea = "frontend"
+    LearningAreaBackend       LearningArea = "backend"
+    LearningAreaInfra         LearningArea = "infra"
+    LearningAreaTests         LearningArea = "tests"
+    LearningAreaDocs          LearningArea = "docs"
+    LearningAreaConfig        LearningArea = "config"
+    LearningAreaOther         LearningArea = "other"
+)
+
+type LearningStatus string
+const (
+    LearningStatusPending      LearningStatus = "pending"
+    LearningStatusInProgress   LearningStatus = "in_progress"
+    LearningStatusResolved     LearningStatus = "resolved"
+    LearningStatusWontF        LearningStatus = "wont_f"
+)
+```
+
+Let me know if this alignment works or if you’d like any adjustments before I proceed with the implementation.

llm/memory.md

@@ -2,6 +2,12 @@
 
 AMCS (Avalon Memory Crystal Server) is an MCP server for capturing and retrieving thoughts, memory, and project context. It is backed by Postgres with pgvector for semantic search.
 
+`amcs-cli` is a pre-built CLI that connects to the AMCS MCP server so agents do not need to implement their own HTTP MCP client. Download it from https://git.warky.dev/wdevs/amcs/releases
+
+The key command is `amcs-cli stdio`, which bridges the remote HTTP MCP server to a local stdio MCP transport. Register it as a stdio MCP server in your agent config and all AMCS tools are available immediately without any custom client code.
+
+Configure with `~/.config/amcs/config.yaml` (`server`, `token`), env vars `AMCS_URL` / `AMCS_TOKEN`, or `--server` / `--token` flags.
+
 You have access to an MCP memory server named AMCS.
 
 Use AMCS as memory with two scopes:
@@ -18,15 +24,30 @@ Use AMCS as memory with two scopes:
 6. If no strong project match exists, you may use global notebook memory with no project.
 7. If multiple projects plausibly match, ask the user before reading or writing project memory.
 
+## Session Startup
+
+At the very start of any session with AMCS:
+
+1. Call `describe_tools` to get the full list of available tools with their categories and any notes you have previously annotated. Read the notes before using a tool — they contain accumulated gotchas, workflow patterns, and field-ordering requirements you have recorded from prior sessions.
+
 ## Project Session Startup
 
-At the start of every project session, after setting the active project:
+After setting the active project:
 
 1. Call `list_project_skills` to load any saved agent behaviour instructions for the project.
 2. Call `list_project_guardrails` to load any saved agent constraints for the project.
 3. Apply all returned skills and guardrails immediately and for the duration of the session.
 4. Only generate or define new skills and guardrails if none are returned. If you do create new ones, store them with `add_skill` or `add_guardrail` and link them to the project with `add_project_skill` or `add_project_guardrail` so they persist for future sessions.
 
+## Project Does Not Exist Yet
+
+If any tool returns a `project_not_found` error:
+
+1. Call `create_project` with the same name you tried to use.
+2. Immediately retry the original tool call with that project.
+
+Do not abandon the project scope or retry without a project. The project simply needs to be created first.
+
 ## Project Memory Rules
 
 - Use project memory for code decisions, architecture, TODOs, debugging findings, and context specific to the current repo or workstream.
@@ -59,6 +80,16 @@ At the start of every project session, after setting the active project:
 - Do not base64-encode a file to pass it to `save_file` if an `amcs://files/{id}` URI is already available from a prior `upload_file` or HTTP upload.
 - When saving, choose the narrowest correct scope: project if project-specific, global if not.
 
+## Tool Annotations
+
+As you learn non-obvious behaviours, gotchas, or workflow patterns for individual tools, persist them with `annotate_tool`:
+
+```json
+{ "tool_name": "capture_thought", "notes": "Always pass project explicitly — session state is unreliable in this client." }
+```
+
+Notes are returned by `describe_tools` in future sessions. Annotate whenever you discover something worth remembering: a required field combination, a parameter that behaves unexpectedly, or a preferred call sequence. Pass an empty string to clear a note.
+
 ## Skills and Guardrails
 
 - **Skills** are reusable agent behaviour instructions (e.g. output formatting rules, reasoning strategies, workflow conventions).
@@ -69,4 +100,4 @@ At the start of every project session, after setting the active project:
 
 ## Short Operational Form
 
-Use AMCS memory in project scope when the current work matches a known project. If no clear project matches, global notebook memory is allowed for non-project-specific information. At the start of every project session call `list_project_skills` and `list_project_guardrails` and apply what is returned; only create new skills or guardrails if none exist. If your MCP client does not preserve sessions across calls, pass `project` explicitly instead of relying on `set_active_project`. Store durable notes with `capture_thought`. For binary files or files larger than 10 MB, call `upload_file` with `content_path` to stage the file and get an `amcs://files/{id}` URI, then pass that URI to `save_file` as `content_uri` to link it to a thought. For small files, use `save_file` or `upload_file` with `content_base64` directly. Browse stored files with `list_files`, and load them with `load_file` only when their contents are needed. Stored files can also be read as raw binary via MCP resources at `amcs://files/{id}`. Never store project-specific memory globally when a matching project exists, and never store memory in the wrong project. If project matching is ambiguous, ask the user.
+At the start of every session, call `describe_tools` to read the full tool list and any accumulated usage notes. Use AMCS memory in project scope when the current work matches a known project; if no clear project matches, global notebook memory is allowed for non-project-specific information. At the start of every project session call `list_project_skills` and `list_project_guardrails` and apply what is returned; only create new skills or guardrails if none exist. If your MCP client does not preserve sessions across calls, pass `project` explicitly instead of relying on `set_active_project`. Store durable notes with `capture_thought`. For binary files or files larger than 10 MB, call `upload_file` with `content_path` to stage the file and get an `amcs://files/{id}` URI, then pass that URI to `save_file` as `content_uri` to link it to a thought. For small files, use `save_file` or `upload_file` with `content_base64` directly. Browse stored files with `list_files`, and load them with `load_file` only when their contents are needed. Stored files can also be read as raw binary via MCP resources at `amcs://files/{id}`. Never store project-specific memory globally when a matching project exists, and never store memory in the wrong project. If project matching is ambiguous, ask the user. If a tool returns `project_not_found`, call `create_project` with that name and retry — never drop the project scope. Whenever you discover a non-obvious tool behaviour, gotcha, or workflow pattern, record it with `annotate_tool` so future sessions benefit.

llm/options_to_openclaw.md

@@ -0,0 +1,826 @@
+# AMCS → OpenClaw Alternative: Gap Analysis & Roadmap
+
+## Context
+
+AMCS is a **passive** MCP memory server. OpenClaw's key differentiator is that it's an **always-on autonomous agent** — it proactively acts, monitors, and learns without human prompting. AMCS has the data model and search foundation; it's missing the execution engine and channel integrations that make OpenClaw compelling.
+
+OpenClaw's 3 pillars AMCS lacks:
+1. **Autonomous heartbeat** — scheduled jobs that run without user prompts
+2. **Channel integrations** — 25+ messaging platforms (Telegram, Slack, Discord, email, etc.)
+3. **Self-improving memory** — knowledge graph distillation, daily notes, living summary (MEMORY.md)
+
+---
+
+## Phase 1: Autonomous Heartbeat Engine (Critical — unlocks everything else)
+
+### 1a. Add `Complete()` to AI Provider
+
+The current `Provider` interface in `internal/ai/provider.go` only supports `Summarize(ctx, systemPrompt, userPrompt)`. An autonomous agent needs a stateful multi-turn call with tool awareness.
+
+**Extend the interface:**
+
+```go
+// internal/ai/provider.go
+
+type CompletionRole string
+
+const (
+    RoleSystem    CompletionRole = "system"
+    RoleUser      CompletionRole = "user"
+    RoleAssistant CompletionRole = "assistant"
+)
+
+type CompletionMessage struct {
+    Role    CompletionRole `json:"role"`
+    Content string         `json:"content"`
+}
+
+type CompletionResult struct {
+    Content    string `json:"content"`
+    StopReason string `json:"stop_reason"` // "stop" | "length" | "error"
+    Model      string `json:"model"`
+}
+
+type Provider interface {
+    Embed(ctx context.Context, input string) ([]float32, error)
+    ExtractMetadata(ctx context.Context, input string) (thoughttypes.ThoughtMetadata, error)
+    Summarize(ctx context.Context, systemPrompt, userPrompt string) (string, error)
+    Complete(ctx context.Context, messages []CompletionMessage) (CompletionResult, error)
+    Name() string
+    EmbeddingModel() string
+}
+```
+
+**Implement in `internal/ai/compat/client.go`:**
+
+`Complete` is a simplification of the existing `extractMetadataWithModel` path — same OpenAI-compatible `/chat/completions` endpoint, same auth headers, no JSON schema coercion. Add a `chatCompletionsRequest` type (reuse or extend the existing unexported struct) and a `Complete` method on `*Client` that:
+1. Builds the request body from `[]CompletionMessage`
+2. POSTs to `c.baseURL + "/chat/completions"` with `c.metadataModel`
+3. Reads the first choice's `message.content`
+4. Returns `CompletionResult{Content, StopReason, Model}`
+
+Error handling mirrors the metadata path: on HTTP 429/503 mark the model unhealthy (`c.modelHealth`), return a wrapped error. No fallback model chain needed for agent calls — callers should retry on next heartbeat tick.
+
+---
+
+### 1b. Heartbeat Engine Package
+
+**New package: `internal/agent/`**
+
+#### `internal/agent/job.go`
+
+```go
+package agent
+
+import (
+    "context"
+    "time"
+)
+
+// Job is a single scheduled unit of autonomous work.
+type Job interface {
+    Name() string
+    Interval() time.Duration
+    Run(ctx context.Context) error
+}
+```
+
+#### `internal/agent/engine.go`
+
+The engine manages a set of jobs and fires each on its own ticker. It mirrors the pattern already used for `runBackfillPass` and `runMetadataRetryPass` in `internal/app/app.go`, but generalises it.
+
+```go
+package agent
+
+import (
+    "context"
+    "log/slog"
+    "sync"
+    "time"
+)
+
+type Engine struct {
+    jobs   []Job
+    store  JobStore  // persists agent_job_runs rows
+    logger *slog.Logger
+}
+
+func NewEngine(store JobStore, logger *slog.Logger, jobs ...Job) *Engine {
+    return &Engine{jobs: jobs, store: store, logger: logger}
+}
+
+// Run starts all job tickers and blocks until ctx is cancelled.
+func (e *Engine) Run(ctx context.Context) {
+    var wg sync.WaitGroup
+    for _, job := range e.jobs {
+        wg.Add(1)
+        go func(j Job) {
+            defer wg.Done()
+            e.runLoop(ctx, j)
+        }(job)
+    }
+    wg.Wait()
+}
+
+func (e *Engine) runLoop(ctx context.Context, j Job) {
+    ticker := time.NewTicker(j.Interval())
+    defer ticker.Stop()
+    for {
+        select {
+        case <-ctx.Done():
+            return
+        case <-ticker.C:
+            e.runOnce(ctx, j)
+        }
+    }
+}
+
+func (e *Engine) runOnce(ctx context.Context, j Job) {
+    runID, err := e.store.StartRun(ctx, j.Name())
+    if err != nil {
+        e.logger.Error("agent: failed to start job run record",
+            slog.String("job", j.Name()), slog.String("error", err.Error()))
+        return
+    }
+    if err := j.Run(ctx); err != nil {
+        e.logger.Error("agent: job failed",
+            slog.String("job", j.Name()), slog.String("error", err.Error()))
+        _ = e.store.FinishRun(ctx, runID, "failed", "", err.Error())
+        return
+    }
+    _ = e.store.FinishRun(ctx, runID, "ok", "", "")
+    e.logger.Info("agent: job complete", slog.String("job", j.Name()))
+}
+```
+
+**Deduplication / double-run prevention:** `StartRun` should check for an existing `running` row younger than `2 * j.Interval()` and return `ErrAlreadyRunning` — the caller skips that tick.
+
+#### `internal/agent/distill.go`
+
+```go
+// DistillJob clusters semantically related thoughts and promotes
+// durable insights into knowledge nodes.
+type DistillJob struct {
+    store     store.ThoughtQuerier
+    provider  ai.Provider
+    cfg       AgentDistillConfig
+    projectID *uuid.UUID // nil = all projects
+}
+
+func (j *DistillJob) Name() string             { return "distill" }
+func (j *DistillJob) Interval() time.Duration  { return j.cfg.Interval }
+
+func (j *DistillJob) Run(ctx context.Context) error {
+    // 1. Fetch recent thoughts not yet distilled (metadata.distilled != true)
+    //    using store.ListThoughts with filter Days = cfg.MinAgeHours/24
+    // 2. Group into semantic clusters via SearchSimilarThoughts
+    // 3. For each cluster > MinClusterSize:
+    //    a. Call provider.Summarize with insight extraction prompt
+    //    b. InsertThought with type="insight", metadata.knowledge_node=true
+    //    c. InsertLink from each cluster member to the insight, relation="distilled_from"
+    //    d. UpdateThought on each source to set metadata.distilled=true
+    // 4. Return nil; partial failures are logged but do not abort the run
+}
+```
+
+Prompt used in step 3a:
+```
+System: You extract durable knowledge from a cluster of related notes.
+        Return a single paragraph (2-5 sentences) capturing the core insight.
+        Do not reference the notes themselves. Write in third person.
+User: [concatenated thought content, newest first, max 4000 tokens]
+```
+
+#### `internal/agent/daily_notes.go`
+
+Runs at a configured hour each day (checked by comparing `time.Now().Hour()` against `cfg.Hour` inside the loop — skip if already ran today by querying `agent_job_runs` for a successful `daily_notes` run with `started_at >= today midnight`).
+
+Collects:
+- Thoughts created today (`store.ListThoughts` with `Days=1`)
+- CRM interactions logged today
+- Calendar activities for today
+- Maintenance logs from today
+
+Formats into a structured markdown string and calls `store.InsertThought` with `type=daily_note`.
+
+#### `internal/agent/living_summary.go`
+
+Regenerates `MEMORY.md` from the last N daily notes + all knowledge nodes. Calls `provider.Summarize` and upserts the result via `store.UpsertFile` using a fixed name `MEMORY.md` scoped to the project (or global if no project).
+
+---
+
+### 1c. Config Structs
+
+Add to `internal/config/config.go`:
+
+```go
+type Config struct {
+    // ... existing fields ...
+    Agent    AgentConfig    `yaml:"agent"`
+    Channels ChannelsConfig `yaml:"channels"`
+    Shell    ShellConfig    `yaml:"shell"`
+}
+
+type AgentConfig struct {
+    Enabled       bool                  `yaml:"enabled"`
+    Distill       AgentDistillConfig    `yaml:"distill"`
+    DailyNotes    AgentDailyNotesConfig `yaml:"daily_notes"`
+    LivingSummary AgentLivingSummary    `yaml:"living_summary"`
+    Archival      AgentArchivalConfig   `yaml:"archival"`
+    Model         string                `yaml:"model"` // override for agent calls; falls back to AI.Metadata.Model
+}
+
+type AgentDistillConfig struct {
+    Enabled        bool          `yaml:"enabled"`
+    Interval       time.Duration `yaml:"interval"`        // default: 24h
+    BatchSize      int           `yaml:"batch_size"`      // thoughts per run; default: 50
+    MinClusterSize int           `yaml:"min_cluster_size"` // default: 3
+    MinAgeHours    int           `yaml:"min_age_hours"`   // ignore thoughts younger than this; default: 6
+}
+
+type AgentDailyNotesConfig struct {
+    Enabled bool `yaml:"enabled"`
+    Hour    int  `yaml:"hour"` // 0-23 UTC; default: 23
+}
+
+type AgentLivingSummary struct {
+    Enabled  bool          `yaml:"enabled"`
+    Interval time.Duration `yaml:"interval"`   // default: 24h
+    MaxDays  int           `yaml:"max_days"`   // daily notes lookback; default: 30
+}
+
+type AgentArchivalConfig struct {
+    Enabled           bool          `yaml:"enabled"`
+    Interval          time.Duration `yaml:"interval"`             // default: 168h (weekly)
+    ArchiveOlderThan  int           `yaml:"archive_older_than_days"` // default: 90
+}
+```
+
+**Full YAML reference (`configs/dev.yaml` additions):**
+
+```yaml
+agent:
+  enabled: false
+  model: ""                        # leave blank to reuse ai.metadata.model
+  distill:
+    enabled: false
+    interval: 24h
+    batch_size: 50
+    min_cluster_size: 3
+    min_age_hours: 6
+  daily_notes:
+    enabled: false
+    hour: 23                       # UTC hour to generate (0–23)
+  living_summary:
+    enabled: false
+    interval: 24h
+    max_days: 30
+  archival:
+    enabled: false
+    interval: 168h
+    archive_older_than_days: 90
+```
+
+---
+
+### 1d. Wire into `internal/app/app.go`
+
+After the existing `MetadataRetry` goroutine block:
+
+```go
+if cfg.Agent.Enabled {
+    jobStore := store.NewJobStore(db)
+    var jobs []agent.Job
+    if cfg.Agent.Distill.Enabled {
+        jobs = append(jobs, agent.NewDistillJob(db, provider, cfg.Agent.Distill, nil))
+    }
+    if cfg.Agent.DailyNotes.Enabled {
+        jobs = append(jobs, agent.NewDailyNotesJob(db, provider, cfg.Agent.DailyNotes))
+    }
+    if cfg.Agent.LivingSummary.Enabled {
+        jobs = append(jobs, agent.NewLivingSummaryJob(db, provider, cfg.Agent.LivingSummary))
+    }
+    if cfg.Agent.Archival.Enabled {
+        jobs = append(jobs, agent.NewArchivalJob(db, cfg.Agent.Archival))
+    }
+    engine := agent.NewEngine(jobStore, logger, jobs...)
+    go engine.Run(ctx)
+}
+```
+
+---
+
+### 1e. New MCP Tools — `internal/tools/agent.go`
+
+```go
+// list_agent_jobs
+// Returns all registered jobs with: name, interval, last_run (status, started_at, finished_at), next_run estimate.
+
+// trigger_agent_job
+// Input: { "job": "distill" }
+// Fires the job immediately in a goroutine; returns a run_id for polling.
+
+// get_agent_job_history
+// Input: { "job": "distill", "limit": 20 }
+// Returns rows from agent_job_runs ordered by started_at DESC.
+```
+
+Register in `internal/app/app.go` routes by adding `Agent tools.AgentTool` to `mcpserver.ToolSet` and wiring `tools.NewAgentTool(engine)`.
+
+---
+
+### 1f. Migration — `migrations/021_agent_jobs.sql`
+
+```sql
+CREATE TABLE agent_job_runs (
+  id          uuid        PRIMARY KEY DEFAULT gen_random_uuid(),
+  job_name    text        NOT NULL,
+  started_at  timestamptz NOT NULL DEFAULT now(),
+  finished_at timestamptz,
+  status      text        NOT NULL DEFAULT 'running', -- running | ok | failed | skipped
+  output      text,
+  error       text,
+  metadata    jsonb       NOT NULL DEFAULT '{}'
+);
+
+CREATE INDEX idx_agent_job_runs_lookup
+  ON agent_job_runs (job_name, started_at DESC);
+```
+
+**`JobStore` interface (`internal/store/agent.go`):**
+
+```go
+type JobStore interface {
+    StartRun(ctx context.Context, jobName string) (uuid.UUID, error)
+    FinishRun(ctx context.Context, id uuid.UUID, status, output, errMsg string) error
+    LastRun(ctx context.Context, jobName string) (*AgentJobRun, error)
+    ListRuns(ctx context.Context, jobName string, limit int) ([]AgentJobRun, error)
+}
+```
+
+---
+
+## Phase 2: Knowledge Graph Distillation
+
+Builds on Phase 1's distillation job. `thought_links` already exists with typed `relation` — the missing piece is a way to mark and query promoted knowledge nodes.
+
+### 2a. Extend `ThoughtMetadata`
+
+In `internal/types/thought.go`, add two fields to `ThoughtMetadata`:
+
+```go
+type ThoughtMetadata struct {
+    // ... existing fields ...
+    KnowledgeNode   bool `json:"knowledge_node,omitempty"`   // true = promoted insight
+    KnowledgeWeight int  `json:"knowledge_weight,omitempty"` // number of source thoughts that fed this node
+    Distilled       bool `json:"distilled,omitempty"`        // true = this thought has been processed by distill job
+}
+```
+
+These are stored in the existing `metadata jsonb` column — no schema migration needed.
+
+### 2b. Store Addition
+
+In `internal/store/thoughts.go` add:
+
+```go
+// ListKnowledgeNodes returns thoughts where metadata->>'knowledge_node' = 'true',
+// ordered by knowledge_weight DESC, then created_at DESC.
+func (db *DB) ListKnowledgeNodes(ctx context.Context, projectID *uuid.UUID, limit int) ([]types.Thought, error)
+```
+
+SQL:
+```sql
+SELECT id, content, metadata, project_id, archived_at, created_at, updated_at
+FROM thoughts
+WHERE (metadata->>'knowledge_node')::boolean = true
+  AND ($1::uuid IS NULL OR project_id = $1)
+  AND archived_at IS NULL
+ORDER BY (metadata->>'knowledge_weight')::int DESC, created_at DESC
+LIMIT $2
+```
+
+### 2c. New MCP Tools — `internal/tools/knowledge.go`
+
+```go
+// get_knowledge_graph
+// Input: { "project_id": "uuid|null", "limit": 50 }
+// Returns: { nodes: [Thought], edges: [ThoughtLink] }
+// Fetches ListKnowledgeNodes + their outgoing/incoming links via store.GetThoughtLinks.
+
+// distill_now
+// Input: { "project_id": "uuid|null", "batch_size": 20 }
+// Triggers the distillation job synchronously (for on-demand use); returns { insights_created: N }
+```
+
+---
+
+## Phase 3: Channel Integrations — Telegram First
+
+### 3a. Channel Adapter Interface — `internal/channels/channel.go`
+
+```go
+package channels
+
+import (
+    "context"
+    "time"
+)
+
+type Attachment struct {
+    Name      string
+    MediaType string
+    Data      []byte
+}
+
+type InboundMessage struct {
+    ChannelID   string      // e.g. telegram chat ID as string
+    SenderID    string      // e.g. telegram user ID as string
+    SenderName  string      // display name
+    Text        string
+    Attachments []Attachment
+    Timestamp   time.Time
+    Raw         any         // original platform message for debug/logging
+}
+
+type Channel interface {
+    Name() string
+    Start(ctx context.Context, handler func(InboundMessage)) error
+    Send(ctx context.Context, channelID string, text string) error
+}
+```
+
+### 3b. Telegram Implementation — `internal/channels/telegram/bot.go`
+
+Uses `net/http` only (no external Telegram SDK). Long-polling loop:
+
+```go
+type Bot struct {
+    token      string
+    allowedIDs map[int64]struct{} // empty = all allowed
+    baseURL    string             // https://api.telegram.org/bot{token}
+    client     *http.Client
+    offset     int64
+    logger     *slog.Logger
+}
+
+func (b *Bot) Name() string { return "telegram" }
+
+func (b *Bot) Start(ctx context.Context, handler func(channels.InboundMessage)) error {
+    for {
+        updates, err := b.getUpdates(ctx, b.offset, 30 /*timeout seconds*/)
+        if err != nil {
+            if ctx.Err() != nil { return nil }
+            // transient error: log and back off 5s
+            time.Sleep(5 * time.Second)
+            continue
+        }
+        for _, u := range updates {
+            b.offset = u.UpdateID + 1
+            if u.Message == nil { continue }
+            if !b.isAllowed(u.Message.Chat.ID) { continue }
+            handler(b.toInbound(u.Message))
+        }
+    }
+}
+
+func (b *Bot) Send(ctx context.Context, channelID string, text string) error {
+    // POST /sendMessage with chat_id and text
+    // Splits messages > 4096 chars automatically
+}
+```
+
+**Error handling:**
+- HTTP 401 (bad token): return fatal error, engine stops channel
+- HTTP 429 (rate limit): respect `retry_after` from response body, sleep, retry
+- HTTP 5xx: exponential backoff (5s → 10s → 30s → 60s), max 3 retries then sleep 5 min
+
+### 3c. Channel Router — `internal/channels/router.go`
+
+```go
+type Router struct {
+    store    store.ContactQuerier
+    thoughts store.ThoughtInserter
+    provider ai.Provider
+    channels map[string]channels.Channel
+    cfg      config.ChannelsConfig
+    logger   *slog.Logger
+}
+
+func (r *Router) Handle(msg channels.InboundMessage) {
+    // 1. Resolve sender → CRM contact (by channel_identifiers->>'telegram' = senderID)
+    //    If not found: create a new professional_contact with the sender name + channel identifier
+    // 2. Capture message as thought:
+    //    content = msg.Text
+    //    metadata.source = "telegram"
+    //    metadata.type = "observation"
+    //    metadata.people = [senderName]
+    //    metadata (extra, stored in JSONB): channel="telegram", channel_id=msg.ChannelID, sender_id=msg.SenderID
+    // 3. If cfg.Telegram.Respond:
+    //    a. Load recent context via store.SearchSimilarThoughts(msg.Text, limit=10)
+    //    b. Build []CompletionMessage with system context + recent thoughts + user message
+    //    c. Call provider.Complete(ctx, messages)
+    //    d. Capture response as thought (type="assistant_response", source="telegram")
+    //    e. Send reply via channel.Send(ctx, msg.ChannelID, result.Content)
+    //    f. Save chat history via store.InsertChatHistory
+}
+```
+
+**Agent response system prompt (step 3b):**
+```
+You are a personal assistant with access to the user's memory.
+Relevant context from memory:
+{joined recent thought content}
+
+Respond concisely. If you cannot answer from memory, say so.
+```
+
+### 3d. Config — full YAML reference
+
+```yaml
+channels:
+  telegram:
+    enabled: false
+    bot_token: ""
+    allowed_chat_ids: []           # empty = all chats allowed
+    capture_all: true              # save every inbound message as a thought
+    respond: true                  # send LLM reply back to sender
+    response_model: ""             # blank = uses agent.model or ai.metadata.model
+    poll_timeout_seconds: 30       # Telegram long-poll timeout (max 60)
+    max_message_length: 4096       # split replies longer than this
+  discord:
+    enabled: false
+    bot_token: ""
+    guild_ids: []                  # empty = all guilds
+    capture_all: true
+    respond: true
+  slack:
+    enabled: false
+    bot_token: ""
+    app_token: ""                  # for socket mode
+    capture_all: true
+    respond: true
+  email:
+    enabled: false
+    imap_host: ""
+    imap_port: 993
+    smtp_host: ""
+    smtp_port: 587
+    username: ""
+    password: ""
+    poll_interval: 5m
+    capture_all: true
+    folders: ["INBOX"]
+```
+
+### 3e. Schema Migration — `migrations/022_channel_contacts.sql`
+
+```sql
+-- Store per-channel identity handles on CRM contacts
+ALTER TABLE professional_contacts
+  ADD COLUMN IF NOT EXISTS channel_identifiers jsonb NOT NULL DEFAULT '{}';
+
+-- e.g. {"telegram": "123456789", "discord": "user#1234", "slack": "U01234567"}
+CREATE INDEX idx_contacts_telegram_id
+  ON professional_contacts ((channel_identifiers->>'telegram'))
+  WHERE channel_identifiers->>'telegram' IS NOT NULL;
+```
+
+### 3f. New MCP Tools — `internal/tools/channels.go`
+
+```go
+// send_channel_message
+// Input: { "channel": "telegram", "channel_id": "123456789", "text": "Hello" }
+// Sends a message on the named channel. Returns { sent: true, channel: "telegram" }
+
+// list_channel_conversations
+// Input: { "channel": "telegram", "limit": 20, "days": 7 }
+// Lists chat histories filtered by channel metadata. Wraps store.ListChatHistories.
+
+// get_channel_status
+// Returns: [{ channel: "telegram", connected: true, uptime_seconds: 3600 }, ...]
+```
+
+### 3g. Future Channel Adapters
+
+Each is a new subdirectory implementing `channels.Channel`. No router or MCP tool changes needed.
+
+| Channel | Package | Approach |
+|---------|---------|----------|
+| Discord | `internal/channels/discord/` | Gateway WebSocket (discord.com/api/gateway); or use `discordgo` lib |
+| Slack | `internal/channels/slack/` | Socket Mode WebSocket (no public URL needed) |
+| Email (IMAP) | `internal/channels/email/` | IMAP IDLE or poll; SMTP for send |
+| Signal | `internal/channels/signal/` | Wrap `signal-cli` JSON-RPC subprocess |
+| WhatsApp | `internal/channels/whatsapp/` | Meta Cloud API webhook (requires public URL) |
+
+---
+
+## Phase 4: Shell / Computer Access
+
+### 4a. Shell Tool — `internal/tools/shell.go`
+
+```go
+type ShellInput struct {
+    Command    string `json:"command"`
+    WorkingDir string `json:"working_dir,omitempty"` // override default; must be within allowed prefix
+    Timeout    string `json:"timeout,omitempty"`     // e.g. "30s"; overrides config default
+    CaptureAs  string `json:"capture_as,omitempty"`  // thought type for stored output; default "shell_output"
+    SaveOutput bool   `json:"save_output"`           // store stdout/stderr as a thought
+}
+
+type ShellOutput struct {
+    Stdout   string `json:"stdout"`
+    Stderr   string `json:"stderr"`
+    ExitCode int    `json:"exit_code"`
+    ThoughtID *uuid.UUID `json:"thought_id,omitempty"` // set if save_output=true
+}
+```
+
+**Execution model:**
+1. Validate `command` against `cfg.Shell.AllowedCommands` (if non-empty) and `cfg.Shell.BlockedCommands`
+2. `exec.CommandContext(ctx, "sh", "-c", command)` with `Dir` set to working dir
+3. Capture stdout + stderr into `bytes.Buffer`
+4. On timeout: kill process group (`syscall.Kill(-cmd.Process.Pid, syscall.SIGKILL)`), return exit code -1
+5. If `SaveOutput`: call `store.InsertThought` with content = truncated stdout (max 8KB) + stderr summary
+
+**Security controls:**
+
+```yaml
+shell:
+  enabled: false
+  working_dir: "/tmp/amcs-agent"  # all commands run here unless overridden
+  allowed_working_dirs:           # if set, working_dir overrides must be within one of these
+    - "/tmp/amcs-agent"
+    - "/home/user/projects"
+  timeout: 30s
+  max_output_bytes: 65536         # truncate captured output beyond this
+  allowed_commands: []            # empty = all; non-empty = exact binary name allowlist
+  blocked_commands:               # checked before allowed_commands
+    - "rm"
+    - "sudo"
+    - "su"
+    - "curl"
+    - "wget"
+  save_output_by_default: false
+```
+
+The tool is registered with `mcp.Tool.Annotations` `Destructive: true` so MCP clients prompt for confirmation.
+
+### 4b. File Bridge Tools
+
+Also in `internal/tools/shell.go`:
+
+```go
+// read_file_from_path
+// Input: { "path": "/abs/path/file.txt", "link_to_thought": "uuid|null" }
+// Reads file from server filesystem → stores as AMCS file via store.InsertFile
+// Returns: { file_id: "uuid", size_bytes: N, media_type: "text/plain" }
+
+// write_file_to_path
+// Input: { "file_id": "uuid", "path": "/abs/path/output.txt" }
+// Loads AMCS file → writes to filesystem path
+// Path must be within cfg.Shell.AllowedWorkingDirs if set
+```
+
+---
+
+## Phase 5: Self-Improving Memory
+
+### 5a. Skill Discovery Job — `internal/agent/skill_discovery.go`
+
+Runs weekly. Algorithm:
+
+1. Load last 30 days of `chat_histories` via `store.ListChatHistories(days=30)`
+2. Extract assistant message patterns with `provider.Complete`:
+   ```
+   System: Identify reusable behavioural patterns or preferences visible in these conversations.
+           Return a JSON array of { "name": "...", "description": "...", "tags": [...] }.
+           Only include patterns that would be useful across future sessions.
+   User: [last N assistant + user messages, newest first]
+   ```
+3. For each discovered pattern, call `store.InsertSkill` with tag `auto-discovered` and the current date
+4. Link to all projects via `store.LinkSkillToProject`
+
+Deduplication: before inserting, call `store.SearchSkills(pattern.name)` — if similarity > 0.9, skip.
+
+### 5b. Thought Archival Job — `internal/agent/archival.go`
+
+```go
+func (j *ArchivalJob) Run(ctx context.Context) error {
+    // 1. ListThoughts older than cfg.ArchiveOlderThanDays with no knowledge_node link
+    //    SQL: thoughts where created_at < now() - interval '$N days'
+    //         AND metadata->>'knowledge_node' IS DISTINCT FROM 'true'
+    //         AND archived_at IS NULL
+    //         AND id NOT IN (SELECT thought_id FROM thought_links WHERE relation = 'distilled_from')
+    // 2. For each batch: store.ArchiveThought(ctx, id)
+    // 3. Log count
+}
+```
+
+Uses the existing `ArchiveThought` store method — no new SQL needed.
+
+---
+
+## End-to-End Agent Loop Flow
+
+```
+Telegram message arrives
+        │
+        ▼
+channels/telegram/bot.go (long-poll goroutine)
+        │  InboundMessage{}
+        ▼
+channels/router.go Handle()
+        ├── Resolve sender → CRM contact (store.SearchContacts by channel_identifiers)
+        ├── store.InsertThought (source="telegram", type="observation")
+        ├── store.SearchSimilarThoughts (semantic context retrieval)
+        ├── ai.Provider.Complete (build messages → LLM call)
+        ├── store.InsertThought (source="telegram", type="assistant_response")
+        ├── store.InsertChatHistory (full turn saved)
+        └── channels.Channel.Send (reply dispatched to Telegram)
+
+Meanwhile, every 24h:
+agent/engine.go ticker fires DistillJob
+        ├── store.ListThoughts (recent, not yet distilled)
+        ├── store.SearchSimilarThoughts (cluster by semantic similarity)
+        ├── ai.Provider.Summarize (insight extraction prompt)
+        ├── store.InsertThought (type="insight", knowledge_node=true)
+        └── store.InsertLink (relation="distilled_from" for each source)
+
+After distill:
+agent/living_summary.go
+        ├── store.ListKnowledgeNodes
+        ├── store.ListThoughts (type="daily_note", last 30 days)
+        ├── ai.Provider.Summarize (MEMORY.md regeneration)
+        └── store.UpsertFile (name="MEMORY.md", linked to project)
+```
+
+---
+
+## Error Handling & Retry Strategy
+
+| Scenario | Handling |
+|----------|----------|
+| LLM returns 429 | Mark model unhealthy in `modelHealth` map (existing pattern), return error, engine logs and skips tick |
+| LLM returns 5xx | Same as 429 |
+| Telegram 429 | Read `retry_after` from response, sleep exact duration, retry immediately |
+| Telegram 5xx | Exponential backoff: 5s → 10s → 30s → 60s, reset after success |
+| Telegram disconnects | Long-poll timeout naturally retries; context cancel exits cleanly |
+| Agent job panics | `engine.runOnce` wraps in `recover()`, logs stack trace, marks run `failed` |
+| Agent double-run | `store.StartRun` checks for `running` row < `2 * interval` old → returns `ErrAlreadyRunning`, tick skipped silently |
+| Shell command timeout | `exec.CommandContext` kills process group via SIGKILL, returns exit_code=-1 and partial output |
+| Distillation partial failure | Each cluster processed independently; failure of one cluster logged and skipped, others continue |
+
+---
+
+## Critical Files
+
+| File | Change |
+|------|--------|
+| `internal/ai/provider.go` | Add `Complete()`, `CompletionMessage`, `CompletionResult` |
+| `internal/ai/compat/client.go` | Implement `Complete()` on `*Client` |
+| `internal/config/config.go` | Add `AgentConfig`, `ChannelsConfig`, `ShellConfig` |
+| `internal/types/thought.go` | Add `KnowledgeNode`, `KnowledgeWeight`, `Distilled` to `ThoughtMetadata` |
+| `internal/store/thoughts.go` | Add `ListKnowledgeNodes()` |
+| `internal/store/agent.go` | New: `JobStore` interface + implementation |
+| `internal/app/app.go` | Wire agent engine + channel router goroutines |
+| `internal/mcpserver/server.go` | Add `Agent`, `Knowledge`, `Channels`, `Shell` to `ToolSet` |
+| `internal/agent/` | New package: engine, job, distill, daily_notes, living_summary, archival, skill_discovery |
+| `internal/channels/` | New package: channel interface, router, telegram/ |
+| `internal/tools/agent.go` | New: list_agent_jobs, trigger_agent_job, get_agent_job_history |
+| `internal/tools/knowledge.go` | New: get_knowledge_graph, distill_now |
+| `internal/tools/channels.go` | New: send_channel_message, list_channel_conversations, get_channel_status |
+| `internal/tools/shell.go` | New: run_shell_command, read_file_from_path, write_file_to_path |
+| `migrations/021_agent_jobs.sql` | New table: agent_job_runs |
+| `migrations/022_channel_contacts.sql` | ALTER professional_contacts: add channel_identifiers jsonb |
+
+---
+
+## Sequence / Parallelism
+
+```
+Phase 1 (Heartbeat Engine) ──► Phase 2 (Knowledge Graph)
+                           └──► Phase 5 (Self-Improving)
+
+Phase 3 (Telegram) ──► Phase 3g (Discord / Slack / Email)
+
+Phase 4 (Shell) [fully independent — no dependencies on other phases]
+```
+
+**Minimum viable OpenClaw competitor = Phase 1 + Phase 3** (autonomous scheduling + Telegram channel).
+
+---
+
+## Verification
+
+| Phase | Test |
+|-------|------|
+| 1 — Heartbeat | Set `distill.interval: 1m` in dev config. Capture 5+ related thoughts. Wait 1 min. Query `thought_links` for `relation=distilled_from` rows. Check `agent_job_runs` has a `status=ok` row. |
+| 1 — Daily notes | Set `daily_notes.hour` to current UTC hour. Restart server. Within 1 min, `list_thoughts` should return a `type=daily_note` entry for today. |
+| 2 — Knowledge graph | Call `get_knowledge_graph` MCP tool. Verify `nodes` array contains `type=insight` thoughts with `knowledge_node=true`. Verify edges list `distilled_from` links. |
+| 3 — Telegram inbound | Send a message to the configured bot. Call `search_thoughts` with the message text — should appear with `source=telegram`. |
+| 3 — Telegram response | Send a question to the bot. Verify a reply arrives in Telegram. Call `list_chat_histories` — should contain the turn. |
+| 4 — Shell | Call `run_shell_command` with `{"command": "echo hello", "save_output": true}`. Verify `stdout=hello\n`, `exit_code=0`, and a new thought with `type=shell_output`. |
+| 4 — Blocked command | Call `run_shell_command` with `{"command": "sudo whoami"}`. Verify error returned without execution. |
+| 5 — Skill discovery | Run `trigger_agent_job` with `{"job": "skill_discovery"}`. Verify new rows in `agent_skills` with tag `auto-discovered`. |
+| Full loop | Send Telegram message → agent responds → distill job runs → knowledge node created from conversation → MEMORY.md regenerated with new insight. |

llm/sample_learning.json

@@ -0,0 +1,14 @@
+{
+  "id": "123e4567-e89b-12d3-a456-426614174000",
+  "category": "insight",
+  "area": "frontend",
+  "status": "pending",
+  "priority": "high",
+  "summary": "Understanding React hooks lifecycle",
+  "details": "React hooks provide a way to use state and other React features without writing a class. This learning note captures key insights about hooks lifecycle and common pitfalls.",
+  "project_id": "proj-001",
+  "thought_id": "th-001",
+  "skill_id": "skill-001",
+  "created_at": "2026-04-05T19:30:00Z",
+  "updated_at": "2026-04-05T19:30:00Z"
+}

llm/structured-learnings/README.md

@@ -0,0 +1,7 @@
+# Structured Learnings
+
+This directory is intended to hold structured learning modules and resources.
+
+---
+
+*Add your learning materials here.*

llm/todo.md

@@ -1,450 +1,172 @@
 # AMCS TODO
-## Auto Embedding Backfill Tool
 
-## Objective
+## Future Plugin: Lifestyle Tools (calendar, meals, household, CRM)
 
-Add an MCP tool that automatically backfills missing embeddings for existing thoughts so semantic search keeps working after:
+The following tool groups have been removed from the core server and are candidates for a separate optional plugin or extension server. The store/tool implementations remain in the codebase but are no longer registered.
 
-* embedding model changes
+### calendar
-* earlier capture or update failures
+- `add_family_member` — Add a family member to the household.
-* import or migration of raw thoughts without vectors
+- `list_family_members` — List all family members.
+- `add_activity` — Schedule a one-time or recurring family activity.
+- `get_week_schedule` — Get all activities scheduled for a given week.
+- `search_activities` — Search activities by title, type, or family member.
+- `add_important_date` — Track a birthday, anniversary, deadline, or other important date.
+- `get_upcoming_dates` — Get important dates coming up in the next N days.
 
-The tool should be safe to run repeatedly, should not duplicate work, and should make it easy to restore semantic coverage without rewriting existing thoughts.
+### meals
+- `add_recipe` — Save a recipe with ingredients and instructions.
+- `search_recipes` — Search recipes by name, cuisine, tags, or ingredient.
+- `update_recipe` — Update an existing recipe.
+- `create_meal_plan` — Set the weekly meal plan; replaces existing.
+- `get_meal_plan` — Get the meal plan for a given week.
+- `generate_shopping_list` — Generate shopping list from the weekly meal plan.
+
+### household
+- `add_household_item` — Store a household fact (paint, appliance, measurement, etc.).
+- `search_household_items` — Search household items by name, category, or location.
+- `get_household_item` — Retrieve a household item by id.
+- `add_vendor` — Add a service provider (plumber, electrician, landscaper, etc.).
+- `list_vendors` — List household service vendors, optionally filtered by service type.
+
+### crm
+- `add_professional_contact` — Add a professional contact to the CRM.
+- `search_contacts` — Search professional contacts by name, company, title, notes, or tags.
+- `log_interaction` — Log an interaction with a professional contact.
+- `get_contact_history` — Get full history (interactions and opportunities) for a contact.
+- `create_opportunity` — Create a deal, project, or opportunity linked to a contact.
+- `get_follow_ups_due` — List contacts with a follow-up date due within the next N days.
+- `link_thought_to_contact` — Append a stored thought to a contact's notes.
+
+**Implementation notes:**
+- Store implementations: `internal/tools/calendar.go`, `internal/tools/meals.go`, `internal/tools/household.go`, `internal/tools/crm.go`
+- DB store layers: `internal/store/calendar.go`, `internal/store/meals.go`, `internal/store/household.go`, `internal/store/crm.go`
+- Re-register via `mcpserver.ToolSet` fields: `Household`, `Calendar`, `Meals`, `CRM`
+- Re-add `registerHouseholdTools`, `registerCalendarTools`, `registerMealTools`, `registerCRMTools` to the register slice in `NewHandlers`
+- Add catalog entries back in `BuildToolCatalog`
+
+---
+## Embedding Backfill and Text-Search Fallback Audit
+
+This file originally described the planned `backfill_embeddings` work and semantic-to-text fallback behavior. Most of that work is now implemented. This document now tracks what landed, what still needs verification, and what follow-up work remains.
+
+For current operator-facing behavior, prefer `README.md`.
 
 ---
 
-## Desired outcome
+## Status summary
 
-After this work:
+### Implemented
 
-* raw thought text remains the source of truth
+The main work described in this file is already present in the repo:
-* embeddings are treated as derived data per model
+
-* search continues to query only embeddings from the active embedding model
+- `backfill_embeddings` MCP tool exists
-* when no embeddings exist for the active model and scope, search falls back to Postgres text search
+- missing-embedding selection helpers exist in the store layer
-* operators or MCP clients can trigger a backfill for the current model
+- embedding upsert helpers exist in the store layer
-* AMCS can optionally auto-run a limited backfill pass on startup or on a schedule later
+- semantic retrieval falls back to Postgres full-text search when the active model has no embeddings in scope
+- fallback behavior is wired into the main query-driven tools
+- a full-text index migration exists
+- optional automatic backfill runner exists in config/startup flow
+- retry and reparse maintenance tooling also exists around metadata quality
+
+### Still worth checking or improving
+
+The broad feature is done, but some implementation-depth items are still worth tracking:
+
+- test coverage around fallback/backfill behavior
+- whether configured backfill batching is used consistently end-to-end
+- observability depth beyond logs
+- response visibility into which retrieval mode was used
 
 ---
 
-## Why this is needed
+## What is already implemented
 
-Current search behavior is model-specific:
+### Backfill tool
 
-* query text is embedded with the configured provider model
+Implemented:
-* results are filtered by `embeddings.model`
-* thoughts with no embedding for that model are invisible to semantic search
 
-This means a model switch leaves old thoughts searchable only by listing and metadata filters until new embeddings are generated.
+- `backfill_embeddings`
+- project scoping
+- archived-thought filtering
+- age filtering
+- dry-run mode
+- bounded concurrency
+- best-effort per-item failure handling
+- idempotent embedding upsert behavior
 
-To avoid that dead zone, AMCS should also support a lexical fallback path backed by native Postgres text-search indexing.
+### Search fallback
+
+Implemented:
+
+- full-text fallback when no embeddings exist for the active model in scope
+- fallback helper shared by query-based tools
+- full-text index migration on thought content
+
+### Tools using fallback
+
+Implemented fallback coverage for:
+
+- `search_thoughts`
+- `recall_context`
+- `get_project_context` when a query is provided
+- `summarize_thoughts` when a query is provided
+- semantic neighbors in `related_thoughts`
+
+### Optional automatic behavior
+
+Implemented:
+
+- config-gated startup backfill pass
+- config-gated periodic backfill loop
 
 ---
 
-## Tool proposal
+## Remaining follow-ups
 
-### New MCP tool
+### 1. Expose retrieval mode in responses
 
-`backfill_embeddings`
+Still outstanding.
 
-Purpose:
+Why it matters:
+- callers currently benefit from fallback automatically
+- but debugging is easier if responses explicitly say whether retrieval was `semantic` or `text`
 
-* find thoughts missing an embedding for the active model
+Suggested shape:
-* generate embeddings in batches
+- add a machine-readable field such as `retrieval_mode: semantic|text`
-* write embeddings with upsert semantics
+- keep it consistent across all query-based tools that use shared retrieval logic
-* report counts for scanned, embedded, skipped, and failed thoughts
 
-### Input
+### 2. Verify and improve tests
 
-```json
+Still worth auditing.
-{
-  "project": "optional project name or id",
-  "limit": 100,
-  "batch_size": 20,
-  "include_archived": false,
-  "older_than_days": 0,
-  "dry_run": false
-}
-```
 
-Notes:
+Recommended checks:
+- no-embedding scope falls back to text search
+- project-scoped fallback only searches within project scope
+- archived thoughts remain excluded by default
+- `related_thoughts` falls back correctly when semantic vectors are unavailable
+- backfill creates embeddings that later restore semantic search
 
-* `project` scopes the backfill to a project when desired
+### 3. Re-embedding / migration ergonomics
-* `limit` caps total thoughts processed in one tool call
-* `batch_size` controls provider load
-* `include_archived` defaults to `false`
-* `older_than_days` is optional and mainly useful to avoid racing with fresh writes
-* `dry_run` returns counts and sample IDs without calling the embedding provider
 
-### Output
+Still optional future work.
 
-```json
+Potential additions:
-{
+- count missing embeddings by project
-  "model": "openai/text-embedding-3-small",
+- add `missing_embeddings` stats to `thought_stats`
-  "scanned": 100,
+- add a controlled re-embed or reindex flow for model migrations
-  "embedded": 87,
-  "skipped": 13,
-  "failed": 0,
-  "dry_run": false,
-  "failures": []
-}
-```
-
-Optional:
-
-* include a short `next_cursor` later if we add cursor-based paging
 
 ---
 
-## Backfill behavior
+## Notes for maintainers
 
-### Core rules
+Do not read this file as an untouched future roadmap item anymore. The repo has already implemented the core work described here.
 
-* Backfill only when a thought is missing an embedding row for the active model.
+If more backfill/fallback work is planned, append it as concrete follow-ups against the current codebase rather than preserving the old speculative rollout order.
-* Do not recompute embeddings that already exist for that model unless an explicit future `force` flag is added.
-* Keep embeddings per model side by side in the existing `embeddings` table.
-* Use `insert ... on conflict (thought_id, model) do update` so retries stay idempotent.
-
-### Selection query
-
-Add a store query that returns thoughts where no embedding exists for the requested model.
-
-Shape:
-
-* from `thoughts t`
-* left join `embeddings e on e.thought_id = t.guid and e.model = $model`
-* filter `e.id is null`
-* optional filters for project, archived state, age
-* order by `t.created_at asc`
-* limit by requested batch
-
-Ordering oldest first is useful because it steadily restores long-tail recall instead of repeatedly revisiting recent writes.
-
-### Processing loop
-
-For each selected thought:
-
-1. read `content`
-2. call `provider.Embed(content)`
-3. upsert embedding row for `thought_id + model`
-4. continue on per-item failure and collect errors
-
-Use bounded concurrency instead of fully serial processing so large backfills complete in reasonable time without overwhelming the provider.
-
-Recommended first pass:
-
-* one tool invocation handles batches internally
-* concurrency defaults to a small fixed number like `4`
-* `batch_size` and concurrency are kept server-side defaults at first, even if only `limit` is exposed in MCP input
 
 ---
 
-## Search fallback behavior
+## Historical note
 
-### Goal
+The original long-form proposal was replaced during the repo audit because it described work that is now largely complete and was causing issue/document drift.
 
-If semantic retrieval cannot run because no embeddings exist for the active model in the selected scope, AMCS should fall back to Postgres text search instead of returning empty semantic results by default.
+If needed, recover the older version from git history.
-
-### Fallback rules
-
-* If embeddings exist for the active model, keep using vector search as the primary path.
-* If no embeddings exist for the active model in scope, run Postgres text search against raw thought content.
-* Fallback should apply to:
-
-  * `search_thoughts`
-  * `recall_context`
-  * `get_project_context` when `query` is provided
-  * `summarize_thoughts` when `query` is provided
-  * semantic neighbors in `related_thoughts`
-
-* Fallback should not mutate data. It is retrieval-only.
-* Backfill remains the long-term fix; text search is the immediate safety net.
-
-### Postgres search approach
-
-Add a native full-text index on thought content and query it with a matching text-search configuration.
-
-Recommended first pass:
-
-* add a migration creating a GIN index on `to_tsvector('simple', content)`
-* use `websearch_to_tsquery('simple', $query)` for user-entered text
-* rank results with `ts_rank_cd(...)`
-* continue excluding archived thoughts by default
-* continue honoring project scope
-
-Using the `simple` configuration is a safer default for mixed prose, identifiers, and code-ish text than a language-specific stemmer.
-
-### Store additions for fallback
-
-Add store methods such as:
-
-* `HasEmbeddingsForModel(ctx, model string, projectID *uuid.UUID) (bool, error)`
-* `SearchThoughtsText(ctx, query string, limit int, projectID *uuid.UUID, excludeID *uuid.UUID) ([]SearchResult, error)`
-
-These should be used by a shared retrieval helper in `internal/tools` so semantic callers degrade consistently.
-
-### Notes on ranking
-
-Text-search scores will not be directly comparable to vector similarity scores.
-
-That is acceptable in v1 because:
-
-* each request will use one retrieval mode at a time
-* fallback is only used when semantic search is unavailable
-* response payloads can continue to return `similarity` as a generic relevance score
-
----
-
-## Auto behavior
-
-The user asked for an auto backfill tool, so define two layers:
-
-### Layer 1: explicit MCP tool
-
-Ship `backfill_embeddings` first.
-
-This is the lowest-risk path because:
-
-* it is observable
-* it is rate-limited by the caller
-* it avoids surprise provider cost on startup
-
-### Layer 2: optional automatic runner
-
-Add a config-gated background runner after the tool exists and is proven stable.
-
-Config sketch:
-
-```yaml
-backfill:
-  enabled: false
-  run_on_startup: false
-  interval: "15m"
-  batch_size: 20
-  max_per_run: 100
-  include_archived: false
-```
-
-Behavior:
-
-* on startup, if enabled and `run_on_startup=true`, run a small bounded backfill pass
-* if `interval` is set, periodically backfill missing embeddings for the active configured model
-* log counts and failures, but never block server startup on backfill failure
-
-This keeps the first implementation simple while still giving us a clean path to true automation.
-
----
-
-## Store changes
-
-Add store methods focused on missing-model coverage.
-
-### New methods
-
-* `ListThoughtsMissingEmbedding(ctx, model string, limit int, projectID *uuid.UUID, includeArchived bool, olderThanDays int) ([]Thought, error)`
-* `UpsertEmbedding(ctx, thoughtID uuid.UUID, model string, embedding []float32) error`
-
-### Optional later methods
-
-* `CountThoughtsMissingEmbedding(ctx, model string, projectID *uuid.UUID, includeArchived bool) (int, error)`
-* `ListThoughtIDsMissingEmbeddingPage(...)` for cursor-based paging on large datasets
-
-### Why separate `UpsertEmbedding`
-
-`InsertThought` and `UpdateThought` already contain embedding upsert logic, but a dedicated helper will:
-
-* reduce duplication
-* let backfill avoid full thought updates
-* make future re-embedding jobs cleaner
-
----
-
-## Tooling changes
-
-### New file
-
-`internal/tools/backfill.go`
-
-Responsibilities:
-
-* parse input
-* resolve project if provided
-* select missing thoughts
-* run bounded embedding generation
-* record per-item failures without aborting the whole batch
-* return summary counts
-
-### MCP registration
-
-Add the tool to:
-
-* `internal/mcpserver/server.go`
-* `internal/mcpserver/schema.go` and tests if needed
-* `internal/app/app.go` wiring
-
-Suggested tool description:
-
-* `Generate missing embeddings for stored thoughts using the active embedding model.`
-
----
-
-## Config changes
-
-No config is required for the first manual tool beyond the existing embedding provider settings.
-
-For the later automatic runner, add:
-
-* `backfill.enabled`
-* `backfill.run_on_startup`
-* `backfill.interval`
-* `backfill.batch_size`
-* `backfill.max_per_run`
-* `backfill.include_archived`
-
-Validation rules:
-
-* `batch_size > 0`
-* `max_per_run >= batch_size`
-* `interval` must parse when provided
-
----
-
-## Failure handling
-
-The backfill tool should be best-effort, not all-or-nothing.
-
-Rules:
-
-* one thought failure does not abort the full run
-* provider errors are captured and counted
-* database upsert failures are captured and counted
-* final tool response includes truncated failure details
-* full details go to logs
-
-Failure payloads should avoid returning raw thought content to the caller if that would create noisy or sensitive responses. Prefer thought IDs plus short error strings.
-
----
-
-## Observability
-
-Add structured logs for:
-
-* selected model
-* project scope
-* scan count
-* success count
-* failure count
-* duration
-
-Later, metrics can include:
-
-* `amcs_backfill_runs_total`
-* `amcs_backfill_embeddings_total`
-* `amcs_backfill_failures_total`
-* `amcs_thoughts_missing_embeddings`
-
----
-
-## Concurrency and rate limiting
-
-Keep the first version conservative.
-
-Plan:
-
-* use a worker pool with a small fixed concurrency
-* keep batch sizes small by default
-* stop fetching new work once `limit` is reached
-* respect `ctx` cancellation so long backfills can be interrupted cleanly
-
-Do not add provider-specific rate-limit logic in v1 unless real failures show it is needed.
-
----
-
-## Security and safety
-
-* Reuse existing MCP auth.
-* Do not expose a broad `force=true` option in v1.
-* Default to non-archived thoughts only.
-* Do not mutate raw thought text or metadata during backfill.
-* Treat embeddings as derived data that may be regenerated safely.
-
----
-
-## Testing plan
-
-### Store tests
-
-Add tests for:
-
-* listing thoughts missing embeddings for a model
-* project-scoped missing-embedding queries
-* archived thought filtering
-* idempotent upsert behavior
-
-### Tool tests
-
-Add tests for:
-
-* dry-run mode
-* successful batch embedding
-* partial provider failures
-* empty result set
-* project resolution
-* context cancellation
-
-### Integration tests
-
-Add a flow covering:
-
-1. create thoughts without embeddings for a target model
-2. run `backfill_embeddings`
-3. confirm rows exist in `embeddings`
-4. confirm `search_thoughts` can now retrieve them when using that model
-
-### Fallback search tests
-
-Add coverage for:
-
-* no embeddings for model -> `search_thoughts` uses Postgres text search
-* project-scoped queries only search matching project thoughts
-* archived thoughts stay excluded by default
-* `related_thoughts` falls back to text search neighbors when semantic vectors are unavailable
-* once embeddings exist, semantic search remains the primary path
-
----
-
-## Rollout order
-
-1. Add store helpers for missing-embedding selection and embedding upsert.
-2. Add Postgres full-text index migration and text-search store helpers.
-3. Add shared semantic-or-text fallback retrieval logic for query-based tools.
-4. Add `backfill_embeddings` MCP tool and wire it into the server.
-5. Add unit and integration tests.
-6. Document usage in `README.md`.
-7. Add optional background auto-runner behind config.
-8. Consider a future `force` or `reindex_model` path only after v1 is stable.
-
----
-
-## Open questions
-
-* Should the tool expose `batch_size` to clients, or should batching stay internal?
-* Should the first version support only the active model, or allow a `model` override for admins?
-* Should archived thoughts be backfilled by default during startup jobs but not MCP calls?
-* Do we want a separate CLI/admin command for large one-time reindex jobs outside MCP?
-
-Recommended answers for v1:
-
-* keep batching mostly internal
-* use only the active configured model
-* exclude archived thoughts by default everywhere
-* postpone a dedicated CLI until volume justifies it
-
----
-
-## Nice follow-ups
-
-* add a `missing_embeddings` stat to `thought_stats`
-* expose a read-only tool for counting missing embeddings by project
-* add a re-embed path for migrating from one model to another in controlled waves
-* add metadata extraction backfill as a separate job if imported content often lacks metadata
-* expose the retrieval mode in responses for easier debugging of semantic vs text fallback

migrations/019_tool_annotations.sql

@@ -0,0 +1,14 @@
+-- Migration: 019_tool_annotations
+-- Adds a table for model-authored usage notes per tool.
+
+create table if not exists tool_annotations (
+  id         bigserial   primary key,
+  tool_name  text        not null,
+  notes      text        not null default '',
+  created_at timestamptz not null default now(),
+  updated_at timestamptz not null default now(),
+  constraint tool_annotations_tool_name_unique unique (tool_name)
+);
+
+grant all on table public.tool_annotations to amcs;
+grant usage, select on sequence tool_annotations_id_seq to amcs;

schema/README.md

@@ -0,0 +1,35 @@
+# Schema workflow
+
+The `schema/*.dbml` files are the database schema source of truth.
+
+## Generate SQL migrations
+
+Run:
+
+```bash
+make generate-migrations
+```
+
+This uses `relspec` to convert the DBML files into PostgreSQL SQL and writes the generated schema migration to:
+
+- `migrations/020_generated_schema.sql`
+
+## Check schema drift
+
+Run:
+
+```bash
+make check-schema-drift
+```
+
+This regenerates the SQL from `schema/*.dbml` and compares it with `migrations/020_generated_schema.sql`.
+If the generated output differs, the command fails so CI can catch schema drift.
+
+## Workflow
+
+1. Update the DBML files in `schema/`
+2. Run `make generate-migrations`
+3. Review the generated SQL
+4. Commit both the DBML changes and the generated migration
+
+Existing handwritten migrations stay in place. Going forward, update the DBML first and regenerate the SQL from there.

schema/calendar.dbml

@@ -0,0 +1,44 @@
+Table family_members {
+  id uuid [pk, default: `gen_random_uuid()`]
+  name text [not null]
+  relationship text
+  birth_date date
+  notes text
+  created_at timestamptz [not null, default: `now()`]
+}
+
+Table activities {
+  id uuid [pk, default: `gen_random_uuid()`]
+  family_member_id uuid [ref: > family_members.id]
+  title text [not null]
+  activity_type text
+  day_of_week text
+  start_time time
+  end_time time
+  start_date date
+  end_date date
+  location text
+  notes text
+  created_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    day_of_week
+    family_member_id
+    (start_date, end_date)
+  }
+}
+
+Table important_dates {
+  id uuid [pk, default: `gen_random_uuid()`]
+  family_member_id uuid [ref: > family_members.id]
+  title text [not null]
+  date_value date [not null]
+  recurring_yearly boolean [not null, default: false]
+  reminder_days_before int [not null, default: 7]
+  notes text
+  created_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    date_value
+  }
+}

schema/core.dbml

@@ -0,0 +1,48 @@
+Table thoughts {
+  id bigserial [pk]
+  guid uuid [unique, not null, default: `gen_random_uuid()`]
+  content text [not null]
+  metadata jsonb [default: `'{}'::jsonb`]
+  created_at timestamptz [default: `now()`]
+  updated_at timestamptz [default: `now()`]
+  project_id uuid [ref: > projects.guid]
+  archived_at timestamptz
+}
+
+Table projects {
+  id bigserial [pk]
+  guid uuid [unique, not null, default: `gen_random_uuid()`]
+  name text [unique, not null]
+  description text
+  created_at timestamptz [default: `now()`]
+  last_active_at timestamptz [default: `now()`]
+}
+
+Table thought_links {
+  from_id bigint [not null, ref: > thoughts.id]
+  to_id bigint [not null, ref: > thoughts.id]
+  relation text [not null]
+  created_at timestamptz [default: `now()`]
+
+  indexes {
+    (from_id, to_id, relation) [pk]
+    from_id
+    to_id
+  }
+}
+
+Table embeddings {
+  id bigserial [pk]
+  guid uuid [unique, not null, default: `gen_random_uuid()`]
+  thought_id uuid [not null, ref: > thoughts.guid]
+  model text [not null]
+  dim int [not null]
+  embedding vector [not null]
+  created_at timestamptz [default: `now()`]
+  updated_at timestamptz [default: `now()`]
+
+  indexes {
+    (thought_id, model) [unique]
+    thought_id
+  }
+}

schema/crm.dbml

@@ -0,0 +1,53 @@
+Table professional_contacts {
+  id uuid [pk, default: `gen_random_uuid()`]
+  name text [not null]
+  company text
+  title text
+  email text
+  phone text
+  linkedin_url text
+  how_we_met text
+  tags "text[]" [not null, default: `'{}'`]
+  notes text
+  last_contacted timestamptz
+  follow_up_date date
+  created_at timestamptz [not null, default: `now()`]
+  updated_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    last_contacted
+    follow_up_date
+  }
+}
+
+Table contact_interactions {
+  id uuid [pk, default: `gen_random_uuid()`]
+  contact_id uuid [not null, ref: > professional_contacts.id]
+  interaction_type text [not null]
+  occurred_at timestamptz [not null, default: `now()`]
+  summary text [not null]
+  follow_up_needed boolean [not null, default: false]
+  follow_up_notes text
+  created_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    (contact_id, occurred_at)
+  }
+}
+
+Table opportunities {
+  id uuid [pk, default: `gen_random_uuid()`]
+  contact_id uuid [ref: > professional_contacts.id]
+  title text [not null]
+  description text
+  stage text [not null, default: 'identified']
+  value "decimal(12,2)"
+  expected_close_date date
+  notes text
+  created_at timestamptz [not null, default: `now()`]
+  updated_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    stage
+  }
+}

schema/files.dbml

@@ -0,0 +1,25 @@
+Table stored_files {
+  id bigserial [pk]
+  guid uuid [unique, not null, default: `gen_random_uuid()`]
+  thought_id uuid [ref: > thoughts.guid]
+  project_id uuid [ref: > projects.guid]
+  name text [not null]
+  media_type text [not null]
+  kind text [not null, default: 'file']
+  encoding text [not null, default: 'base64']
+  size_bytes bigint [not null]
+  sha256 text [not null]
+  content bytea [not null]
+  created_at timestamptz [not null, default: `now()`]
+  updated_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    thought_id
+    project_id
+    sha256
+  }
+}
+
+// Cross-file refs (for relspecgo merge)
+Ref: stored_files.thought_id > thoughts.guid [delete: set null]
+Ref: stored_files.project_id > projects.guid [delete: set null]

schema/household.dbml

@@ -0,0 +1,31 @@
+Table household_items {
+  id uuid [pk, default: `gen_random_uuid()`]
+  name text [not null]
+  category text
+  location text
+  details jsonb [not null, default: `'{}'`]
+  notes text
+  created_at timestamptz [not null, default: `now()`]
+  updated_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    category
+  }
+}
+
+Table household_vendors {
+  id uuid [pk, default: `gen_random_uuid()`]
+  name text [not null]
+  service_type text
+  phone text
+  email text
+  website text
+  notes text
+  rating int
+  last_used date
+  created_at timestamptz [not null, default: `now()`]
+
+  indexes {
+    service_type
+  }
+}