spine-framework 0.4.3 → 0.5.0-rc.4

package/.framework/src/instructions/reference/functions/deliveries.md

@@ -0,0 +1,110 @@
+---
+title: deliveries
+audience: both
+kind: reference
+verified_against: f69c37d
+verified_on: 2026-10-30
+source_files:
+  - .framework/functions/deliveries.ts
+  - .framework/functions/integration-outbound.ts
+  - .framework/functions/_shared/middleware.ts
+  - .framework/migrations/004_core_rpc_and_broker.sql
+  - .framework/public/_redirects
+---
+
+# deliveries
+
+## What it does
+
+Read access to `public.webhook_deliveries` — the delivery log for
+inbound and outbound webhook traffic — plus `action=retry`, which
+re-executes a failed **outbound** delivery through
+`integration-outbound` (see
+[Integrations and webhooks](../../concepts/integrations-and-webhooks.md)).
+
+The table is `webhook_deliveries` — there is no `deliveries` table.
+
+## Route and methods
+
+`/.netlify/functions/deliveries`, reachable as `/api/deliveries`
+(`/api/*` → `/.netlify/functions/:splat`,
+`.framework/public/_redirects`). Dispatch is by `?action=` then method:
+
+| Call | Handler |
+|---|---|
+| `GET /api/deliveries` | `list` |
+| `GET /api/deliveries?id=<uuid>` | `get` |
+| `POST /api/deliveries?action=retry` | `retry` |
+
+Anything else (including `PATCH`/`DELETE`) → `Invalid action or method`.
+
+## Authentication
+
+Standard `createHandler` auth: `x-api-key` → machine principal,
+`x-internal-secret` → internal principal, `Bearer <jwt>` → human;
+anonymous → 401.
+
+- `list` requires `ctx.accountId` and always filters
+  `.eq('account_id', ctx.accountId)`.
+- `get`/`retry` have no explicit account check — for human callers RLS
+  on `ctx.db` bounds what the row lookup can see; machine principals run
+  `ctx.db` as the service role and can fetch any delivery by id.
+- No `sanitizeRecordData` is applied — rows are returned raw.
+
+## Request contract
+
+**list** (query): `direction`, `status`, `integration_id`,
+`api_key_id`, `limit` (default 100), `offset` (default 0). Ordered
+`created_at` desc; always account-scoped.
+
+**get** (query): `id`.
+
+**retry** (body): `id` — the delivery must exist and have
+`direction === 'outbound'` (inbound retries throw).
+
+## Response contract
+
+Standard `{ data, error, meta }` envelope.
+
+- `list`/`get` → raw `webhook_deliveries` row(s).
+- `retry` → the return value of `integration-outbound`'s `execute` for
+  `{ delivery_id }` — the delivery attempt result, not a new delivery
+  row.
+
+## Side effects
+
+- `retry` re-executes the delivery by calling
+  `.framework/functions/integration-outbound.ts` `execute` in-process
+  with `principal: SYSTEM_PRINCIPAL` and `accountId` taken from **the
+  delivery's** `account_id` (not the caller's). Outbound execution
+  writes its own attempt/status bookkeeping onto `webhook_deliveries`.
+- `emitLog` writes a `delivery.manual_retry` audit row to `public.logs`
+  with the delivery's `direction` and `integration_id`.
+- `list`/`get` are pure reads.
+
+## Errors
+
+| Condition | HTTP | Message |
+|---|---|---|
+| No account context (list) | 500 | `Account context required` |
+| Missing `id` | 500 | `Delivery ID is required` |
+| Missing row | 500 | `Delivery not found` / PostgREST message |
+| Retrying an inbound delivery | 500 | `Manual retry is only supported for outbound deliveries in this release` |
+| Bad action/method | 500 | `Invalid action or method` |
+
+Validation failures surface as HTTP 500 — thrown `Error`s without a
+`statusCode` map to 500 in the shared error path.
+
+## Agent notes
+
+- Retry runs the delivery under `SYSTEM_PRINCIPAL` — the caller's
+  identity is not propagated into the outbound execution, and the
+  delivery's own `account_id` decides the execution account. A caller
+  who can `get` a delivery id can retry it.
+- There is no redelivery of inbound payloads — inbound rows are
+  observability only; outbound retries go through the integration
+  adapter again and may produce real external HTTP calls.
+- `webhook_deliveries` is created by
+  `.framework/migrations/004_core_rpc_and_broker.sql`; automatic retry
+  sweeps happen in `system-cron` via
+  `.framework/functions/_shared/retry-deliveries.ts`, not here.

package/.framework/src/instructions/reference/functions/embeddings.md

@@ -0,0 +1,166 @@
+---
+title: embeddings
+audience: both
+kind: reference
+verified_against: f69c37d
+verified_on: 2026-10-30
+source_files:
+  - .framework/functions/embeddings.ts
+  - .framework/functions/_shared/middleware.ts
+  - .framework/functions/_shared/llm.ts
+  - .framework/migrations/001_core_schema.sql
+  - .framework/migrations/044_match_embeddings_filtered_knn.sql
+  - .framework/public/_redirects
+---
+
+# embeddings
+
+## What it does
+
+Lifecycle management for `public.embeddings` — per-document vector rows
+(`UNIQUE(model_id, document_id, chunk_index)`, pgvector
+`embedding vector(1536)`) used by knowledge retrieval. Beyond basic
+CRUD it can **generate** an item's vector set from its type schema,
+**search** by vector similarity or full-text, and gate visibility with
+`approved_at`/`revoked_at`.
+
+## Route and methods
+
+`/.netlify/functions/embeddings`, reachable as `/api/embeddings`
+(`/api/*` → `/.netlify/functions/:splat`,
+`.framework/public/_redirects`). Dispatch is by `?action=` + method:
+
+| Call | Handler |
+|---|---|
+| `GET /api/embeddings` | `list` |
+| `GET /api/embeddings?id=<uuid>` | `get` |
+| `GET /api/embeddings?action=stats` | `getStats` |
+| `POST /api/embeddings` | `create` |
+| `PATCH /api/embeddings` (`id` in body or query) | `update` |
+| `POST /api/embeddings?action=generate` | `generate` |
+| `POST /api/embeddings?action=batch-create` | `batchCreate` |
+| `POST /api/embeddings?action=delete-document` | `deleteByDocument` |
+| `POST /api/embeddings?action=cleanup` | `cleanup` |
+| `POST /api/embeddings?action=search-similar` | `searchSimilar` |
+| `POST /api/embeddings?action=search-semantic` | `searchSemantic` |
+| `POST /api/embeddings?action=revoke` | `revoke` |
+| `POST /api/embeddings?action=approve` | `approve` |
+
+There is no `DELETE` route — `delete-document` is a POST action.
+
+## Authentication
+
+Standard `createHandler` auth: `x-api-key` → machine principal,
+`x-internal-secret` → internal principal, `Bearer <jwt>` → human;
+anonymous → 401.
+
+- `list`/`create`/`batchCreate`/`getStats`/`searchSemantic` require
+  `ctx.accountId`.
+- **No role check anywhere.** `generate`, `revoke`, and `approve` read
+  and write through `adminDb` (service role) for any authenticated
+  caller — including cross-account, since `generate` accepts an
+  `account_id_override`. Treat these actions as trusted-surface
+  operations.
+
+## Request contract
+
+**list** (query): `model_id`, `document_id`, `limit` (default 100),
+`offset` (default 0). Ordered `created_at` desc.
+
+**get** (query): `id`. **getStats**: none → `{ total_embeddings }`.
+
+**create** (body): `model_id`, `document_id`, `content` required;
+`chunk_index` (default 0), `metadata` `{}`. `account_id` stamped from
+context. (Does **not** compute a vector — it stores a row.)
+
+**update** (`id` + body): `content`, `metadata` only.
+
+**generate** (body): `item_id` required; `force_regenerate` (deletes and
+rebuilds all vectors for the document), `account_id_override` (defaults
+to the item's account, then the `spine-system` platform account).
+
+**batchCreate** (body): `embeddings_data` array — rows inserted verbatim
+with `account_id` stamped.
+
+**deleteByDocument** (body): `document_id` — hard-deletes all its rows.
+
+**cleanup** (body): `days_to_keep` (default 365) — hard-deletes rows
+older than the cutoff.
+
+**searchSimilar** (body): `query_embedding` required; `limit` (default
+10), `threshold` (default 0.7), `filter_account_ids`,
+`filter_vector_type`, `filter_item_types`, `filter_document_ids` — calls
+the `match_embeddings` RPC
+(`.framework/migrations/044_match_embeddings_filtered_knn.sql`).
+
+**searchSemantic** (body): `query` required; `model_id`,
+`document_ids`, `limit` (default 10) — Postgres `textSearch`
+(`websearch`, `english`) over `content`, scoped to `ctx.accountId`.
+
+**revoke** / **approve** (body): `document_id`.
+
+## Response contract
+
+Standard `{ data, error, meta }` envelope.
+
+- `generate` → `{ item_id, embeddings_created, embeddings_skipped,
+  errors }` (unchanged content is skipped by hash, not re-embedded).
+- `searchSimilar` → `{ results, count, model_id, threshold }`;
+  `searchSemantic` → `{ results, count, method: 'text_search', query }`.
+- `revoke` → `{ document_id, revoked_count }`; `approve` →
+  `{ document_id, approved_count }`; `deleteByDocument`/`cleanup` →
+  `{ deleted_count }`; `batchCreate` → inserted rows.
+- `list`/`get` sanitize via `sanitizeRecordData` — full row for system
+  admins, `{ id }` for non-admins (rows carry no `design_schema`).
+
+## Side effects
+
+- `generate` reads the item + its type's `design_schema` (`vector_types`
+  — a flat array or a per-`kb_type` map defaulting to
+  `['semantic','structure']` — and `embedding_fields.structure`), builds
+  one text per vector type via `CONTENT_BUILDERS`, calls the OpenAI
+  embeddings API (`text-embedding-3-small`), and upserts
+  `public.embeddings` rows keyed by fixed `chunk_index` slots
+  (`VECTOR_TYPE_SLOTS`: semantic 0, structure 1, context 2, …) so the
+  `(model_id, document_id, chunk_index)` unique constraint never
+  collides across vector types.
+- `generate` deletes stale rows left over from a previously larger
+  vector-type set, and skips re-embedding rows whose `content_hash`
+  matches (16-char SHA-256 prefix).
+- `generate` sets `approved_at` only when the item's `status` is
+  `'published'`; `metadata` carries `vector_type`, `item_type`,
+  `kb_type`, `security_level`, `source_item_id`.
+- A `'summary'` vector type is generated via `callOpenAIChat`
+  (`LLM_DEFAULT_MODEL` or `gpt-4o-mini`) with a sync title+description
+  fallback.
+- `revoke` sets `revoked_at` on all non-revoked rows for the document;
+  `approve` sets `approved_at` and clears `revoked_at` — both via
+  `adminDb` regardless of caller role.
+- `emitLog` writes `embedding.created`, `embedding.updated`,
+  `embeddings.generated`, `embeddings.batch_created`,
+  `embeddings.document_deleted`, `embeddings.cleaned`,
+  `embeddings.revoked`, `embeddings.approved` to `public.logs`.
+
+## Errors
+
+| Condition | HTTP | Message |
+|---|---|---|
+| Missing fields | 500 | `model_id, document_id, and content are required` / `item_id is required` / `document_id is required` / `embeddings_data array is required` |
+| Item missing (generate) | 500 | `Item <id> not found` |
+| Embedding API failure | 500 | `Embedding API call failed: <message>` |
+| Missing search inputs | 500 | `query_embedding is required for similarity search` / `query is required for semantic search` |
+| Bad action/method | 500 | `Invalid action or method` |
+
+## Agent notes
+
+- Vectors are stored as the pgvector string literal `[v1,v2,…]`
+  produced by `callEmbeddingApi` — the `embedding` column is
+  `vector(1536)` (`.framework/migrations/001_core_schema.sql`).
+- `metadata.vector_type` is the discriminator retrieval filters use;
+  `chunk_index` is a slot convention, not a sequence of text chunks.
+- `generate`'s `account_id_override` plus unconditional `adminDb` means
+  any authenticated caller can regenerate embeddings into another
+  account — and `revoke`/`approve`/`cleanup`/`deleteByDocument` are
+  equally ungated. Do not expose this endpoint to untrusted principals.
+- Retrieval-time search is `match_embeddings` (vector) or `textSearch`
+  (keyword) — semantic ranking lives in the RPC, not here.

package/.framework/src/instructions/reference/functions/eval-harness.md

@@ -0,0 +1,149 @@
+---
+title: eval-harness
+audience: both
+kind: reference
+verified_against: f69c37d
+verified_on: 2026-10-30
+source_files:
+  - .framework/functions/eval-harness.ts
+  - .framework/functions/_shared/llm.ts
+  - .framework/functions/_shared/middleware.ts
+  - .framework/functions/_shared/resolve-ids.ts
+  - .framework/functions/admin-data.ts
+  - .framework/migrations/034_match_embeddings_knn.sql
+  - public/_redirects
+---
+
+# eval-harness
+
+## What it does
+
+`eval-harness` (`.framework/functions/eval-harness.ts`) is an internal
+retrieval-quality harness. It stores test cases as `eval_test_case` `items`
+rows, replays each question against the `embeddings` corpus via the
+`match_embeddings` RPC, scores the retrieval, and records the run as a
+`pipeline_executions` row so results show up in the executions UI.
+
+This is a dev/QA tool, not a production path — nothing else calls it.
+
+Actions (dispatched on `?action=`):
+
+| Action | Method | Purpose |
+|---|---|---|
+| `create_test_case` | POST | Insert an `eval_test_case` item via admin-data. |
+| `list_test_cases` | GET | List active cases (default action for any GET). |
+| `run_eval` | POST | Score all (or one `test_type` of) active cases. |
+| `get_results` | GET | Recent eval runs (`pipeline_executions` rows). |
+| `get_coverage` | GET | Case count per `test_type`, with gaps listed. |
+
+`test_type` values: `easy`, `medium`, `hard`, `security`, `hallucination`,
+`stale`, `account_gating`.
+
+## Route and methods
+
+- `/api/eval-harness?action=<action>` — `public/_redirects` maps `/api/*` to
+  `/.netlify/functions/:splat`. The method guard is `ctx.query.method`
+  (populated from the HTTP method): creates/runs are POST, reads are GET.
+- GET with no `action` (or an unrecognized action on GET) falls through to
+  `list_test_cases`; anything else throws `Invalid action or method`.
+
+## Authentication
+
+`createHandler` resolves a principal and rejects anonymous requests with 401.
+`run_eval` and `get_results` additionally require `ctx.accountId`
+(`Account context required`). There is **no system-admin gate** — any
+authenticated principal can create cases and run evals, which is part of why
+this page treats the surface as internal-facing.
+
+## Request contract
+
+`create_test_case` (POST):
+
+| Field | Default | Notes |
+|---|---|---|
+| `question` | — | Required; also becomes the item `title` (120 chars). |
+| `expected_tier` | `'medium'` | `high` / `medium` / `low`. |
+| `expected_escalate` | `false` | Stored, not currently scored. |
+| `expected_source_ids` | `[]` | Document ids retrieval should hit. |
+| `excluded_source_ids` | `[]` | Document ids retrieval must not hit (boundary). |
+| `test_type` | `'medium'` | One of the categories above. |
+
+`list_test_cases` (GET): query params `test_type`, `limit` (200), `offset`.
+
+`run_eval` (POST): `{ test_type?, similarity_threshold = 0.15,
+match_count = 8 }`.
+
+`get_results` (GET): query param `limit` (5).
+
+`get_coverage` (GET): no inputs.
+
+## Response contract
+
+- `create_test_case` → the created `items` record (admin-data `create`
+  result).
+- `run_eval` →
+  `{ run_id, summary, results }` where `summary` is `{ total, passed,
+  failed, skipped, pass_rate, run_at, account_id }` and each result is
+  `{ id, question, test_type, status, retrieval_recall, boundary_safe,
+  boundary_violations, tier_correct, inferred_tier, expected_tier,
+  mean_similarity, retrieved_count }`. With no cases it returns
+  `{ run_id: null, message: 'No test cases found', total: 0 }`.
+- `get_results` → recent `pipeline_executions` rows (`id`, `status`,
+  `result`, `started_at`, `completed_at`, `trigger_data`) filtered to
+  `trigger_data->>eval_run = 'true'`.
+- `get_coverage` → `{ total_cases, coverage: [{ test_type, case_count,
+  has_cases }], recommendation: [<test_types with no cases>] }`.
+
+Scoring per case: **boundary safe** means no `excluded_source_ids` appear in
+the retrieved set; **retrieval recall** is the fraction of
+`expected_source_ids` retrieved (`null` when none declared); **inferred
+tier** maps mean match similarity to `high` ≥ 0.85, `medium` ≥ 0.60, else
+`low`. A case passes when boundary-safe AND (recall is null OR ≥ 0.5) AND
+inferred tier equals expected tier. Cases with `data.needs_source_review`
+are skipped — they're diagnostic captures awaiting human promotion to
+golden cases. Embed-API errors skip the case (`skipped`, reason
+`embed_api_error`), as does any per-case exception.
+
+## Side effects
+
+- `create_test_case` inserts an `items` row through admin-data — type
+  `eval_test_case`, `status: 'active'` — so validation, audit, and triggers
+  run.
+- `run_eval` calls `callOpenAIEmbeddings` (`text-embedding-3-small`) once per
+  case — that routes through `integration-outbound` with the `openai`
+  integration — then `match_embeddings` with `filter_vector_type:
+  'semantic'` over the caller's account plus the `spine-system` platform
+  account.
+- `run_eval` inserts a `pipeline_executions` row with the sentinel
+  `pipeline_id` `00000000-0000-0000-0000-000000000000`,
+  `trigger_data.eval_run: true`, `status: 'completed'` when zero failures
+  else `'failed'`, and the full `{ summary, results }` in `result` — then
+  emits an `eval.run_completed` audit event.
+
+## Errors
+
+| Condition | Error |
+|---|---|
+| Missing question | `question is required` |
+| Type not installed | `eval_test_case type not found — run app install first` |
+| No account context | `Account context required` |
+| Bad dispatch | `Invalid action or method` |
+
+## Agent notes
+
+- **Internal-facing.** This is a QA harness shipped with the framework, not
+  a product surface: no admin gate, results land in `pipeline_executions`
+  rather than a dedicated table, and `run_eval` synchronously embeds up to
+  500 questions in one request — slow and costly on large suites. Run it in
+  dev/staging or behind your own gate.
+- `expected_escalate` is stored but never read by scoring — only
+  recall/boundary/tier count.
+- Recall uses `document_id` equality against `expected_source_ids` — fill
+  those with `embeddings.document_id` values (items UUIDs for knowledge
+  records), not item titles.
+- Eval runs share the `spine-system` platform corpus with search — account
+  isolation is tested by `account_gating` cases whose
+  `excluded_source_ids` cover other tenants' documents.
+- The `eval_test_case` item type must be installed for creation to work;
+  `resolveTypeId` failure throws the "run app install first" error rather
+  than skipping.

package/.framework/src/instructions/reference/functions/index.md

@@ -0,0 +1,103 @@
+---
+title: Functions index
+audience: both
+kind: index
+verified_against: f69c37d
+verified_on: 2026-10-30
+source_files:
+  - .framework/functions/
+  - .framework/public/_redirects
+---
+
+# Functions index
+
+All 47 Netlify functions in `.framework/functions/`, grouped by domain.
+Every function is reachable publicly as `/api/<name>` (`/api/*` →
+`/.netlify/functions/:splat`, `.framework/public/_redirects`), and all
+JSON responses use the `{ data, error, meta }` envelope from
+`.framework/functions/_shared/middleware.ts`.
+
+Pages linked below are full references; everything else is a one-line
+summary.
+
+## Data & access
+
+| Function | What it does |
+|---|---|
+| [admin-data](./admin-data.md) | Generic CRUD + stats over the entity allowlist, with record permissions, schema stamping, and trigger firing. |
+| [account-nodes](./account-nodes.md) | One account plus its ancestors, children, and descendants for hierarchy views. |
+| [api-keys](./api-keys.md) | Mint, validate, revoke, and list `sk_…` machine API keys plus usage logs. |
+| [auth](./auth.md) | The caller's user context — person, role, permissions, accessible accounts — plus a health probe. |
+| [invites](./invites.md) | Registration placement, invite lifecycle, join-request approval, and member management. |
+| [item-progress](./item-progress.md) | Per-person pipeline progress on items with enforced status transitions. |
+
+## Config CRUD
+
+| Function | What it does |
+|---|---|
+| [apps](./apps.md) | App registry CRUD plus schema, availability, and version actions. |
+| [prompt-configs](./prompt-configs.md) | Prompt configuration CRUD — model, context template, escalation, output fields. |
+| [roles](./roles.md) | Role CRUD (`system_admin`-gated). |
+| [types](./types.md) | Type CRUD plus `getBySlug`/`getSchema`; derives `validation_schema` from `design_schema`. |
+
+## Automation
+
+| Function | What it does |
+|---|---|
+| [pipelines](./pipelines.md) | Pipeline CRUD plus `toggle`, `by-trigger` lookup, and execution listing. |
+| [pipeline-executions](./pipeline-executions.md) | Run bookkeeping — list/get, start/complete/cancel, `running`, `stats`, `cleanup`. |
+| [pipeline-runner-background](./pipeline-runner-background.md) | Netlify background function that executes one dispatched pipeline run. |
+| [system-cron](./system-cron.md) | Cron tick: sweeps stale executions, fires due schedules and timers, evaluates threshold alerts, retries deliveries, cleans logs. |
+| [timers](./timers.md) | Timer CRUD + `toggle` for scheduled pipeline firing. |
+| [trigger-scheduler](./trigger-scheduler.md) | Netlify scheduled function (`* * * * *`) that bridges Clockwork invocations to `system-cron`. |
+| [triggers](./triggers.md) | Trigger CRUD plus `toggle`, `listByEvent`, and execution listing. |
+
+## Integrations & AI
+
+| Function | What it does |
+|---|---|
+| [ai-agents](./ai-agents.md) | Agent CRUD plus `action=run` multi-turn inference. |
+| [ai-interactions](./ai-interactions.md) | Feedback and interaction capture (`submit_feedback`, `create_interaction`, CIL learning proposals). |
+| [deliveries](./deliveries.md) | `webhook_deliveries` list/get plus outbound `action=retry`. |
+| [embeddings](./embeddings.md) | Embedding lifecycle — `generate`, `search-similar`/`search-semantic`, `revoke`, `approve`, `cleanup`. |
+| [eval-harness](./eval-harness.md) | AI eval cases, runs, results, and coverage reporting. |
+| [integration-outbound](./integration-outbound.md) | Executes outbound deliveries through integration adapters with retry bookkeeping. |
+| [integration-routes](./integration-routes.md) | Public inbound webhook receiver — machine-principal auth, handler registry, record creation. |
+| [integrations](./integrations.md) | Integration registry CRUD (providers, credentials, config). |
+| [kb](./kb.md) | Knowledge-base `ingest`, `search`, `publish`, `reindex`, `revoke`. |
+
+## Observability & system
+
+| Function | What it does |
+|---|---|
+| [admin-funnel](./admin-funnel.md) | Visitor→user→install funnel metrics. Dormant: no `handler` export — nothing routes to it. |
+| [admin-install-funnel](./admin-install-funnel.md) | Install-funnel metrics plus `exportData`. Dormant: no `handler` export. |
+| [admin-signals](./admin-signals.md) | Product-signal metrics and audit listing. Dormant: no `handler` export. |
+| [debug-auth](./debug-auth.md) | Auth diagnostics — resolved principal type and env flags. Debug endpoint; not gated beyond authentication. |
+| [logs](./logs.md) | Log query API — `account`/`target`/`person`/`stats`/`search`/`cleanup` — plus log writes. |
+| [observability](./observability.md) | Metrics API — event volume, error rate, latency percentiles, pipeline stats, top actors, AI support metrics, cleanup. |
+| [security](./security.md) | Flagged accounts, ban/unban across blocklist + Supabase + edge firewall, blocklist management. |
+| [system](./system.md) | Deployment `manifest`, `health`, and `openapi` descriptors. |
+| [tests](./tests.md) | `test_runs`/`test_results` reporting — list, get, 30-day per-suite stats. |
+
+## Internal — `cil-*` tooling
+
+The Code Intelligence Layer functions. These are internal tooling for
+Spine's code-understanding index (chunks, docs, links, embeddings stored
+as runtime records); they are not part of the app-facing API and have no
+reference pages.
+
+| Function | What it does |
+|---|---|
+| cil-agent | LLM tool-use agent over the CIL index (chunk lookup, contracts, blast radius, hash validation). |
+| cil-agent-ingest | Ingests agent session/doc content into CIL chunk fields. |
+| cil-callmap | Derives file-level call/import links into the CIL graph. |
+| cil-config | CIL project config extraction and status (`run`/`extract`/`status`). |
+| cil-context | Context-brief assembly for a chunk (contract, callers, tests, dependencies). |
+| cil-doc | LLM-backed doc generation/sync for chunks. |
+| cil-embeddings | Embedding generation for CIL chunks. |
+| cil-graph | Graph queries over CIL `links` (callers/callees, blast radius). |
+| cil-ingest | AST chunking and marker sync for source files. |
+| cil-pipeline | CIL pipeline status and per-step file readiness. |
+| cil-project | CIL project CRUD (method-dispatched). |
+| cil-validate | Doc/freshness validation and hash updates for CIL chunks. |