MCP introspection & observability

Modules:

apps/mcp/src/resources.ts — read-only resource templates.
apps/mcp/src/prompts.ts — best-practice prompt catalogue.
apps/mcp/src/metrics.ts — in-memory Prometheus registry.
apps/mcp/src/routes/introspection.ts — sidecar HTTP routes.

Resources

Four URI templates served via the MCP resources/list + resources/read standard endpoints:

URI	Returns
`drobek://plans/{id}`	plan row + tasks + edges
`drobek://tasks/{id}`	full task entity
`drobek://knowledge/{id}`	full knowledge body (vs the search snippet)
`drobek://orgs/{id}/stats`	entitlements + current-period usage

Cross-org reads (token from org A → resource of org B) → 403 auth.cross_org. Auth flows through an AsyncLocalStorage populated by the MCP request middleware in server.ts — resources cannot bypass it.

Prompts

Three canned templates served via prompts/list + prompts/get:

start-work-session — the canonical agent loop.
summarize-knowledge — guidance for learnings in complete_task.
review-needs-review — workflow for clearing needs_review tasks.

Placeholders {{agent_name}}, {{org_slug}}, {{task_id}} are substituted server-side from the request's auth context before the body is shipped to the client.

Sidecar routes

Route	Auth	Returns
`GET /metrics`	scope `audit:read`	Prometheus text
`GET /audit?limit=&cursor=`	scope `audit:read`	`{ items, nextCursor }`
`GET /mcp/usage`	any valid token	`{ rateLimit, entitlements, usage }`

Metric names:

drobek_mcp_tool_calls_total{tool, status} — counter.
drobek_mcp_tool_duration_ms_bucket{tool, le} + _sum + _count —

histogram, buckets [5, 10, 25, 50, 100, 250, 500, 1000, 2500].

Audit cursor format: base64url-encoded {at: ISO, id: uuid} — descending order by (createdAt, id). The +1 overflow row is used to determine nextCursor without a second round-trip.