loopctl-mcp-server 2.86.0 → 2.87.0

package/README.md

@@ -342,7 +342,8 @@ it is enforced server-side and a no-op for a non-superadmin key — see below.)
 | `memory_list` | List your own long-term memories, newest first, paginated with `meta.total_count/limit/offset` (the true scoped count, never silently capped by `limit`). Optional: `limit`, `offset`, `include_superseded`, `all_subjects` (superadmin only; ignored for non-superadmin keys). |
 | `memory_forget` | Delete one of your own long-term memories by id. A foreign-subject, foreign-tenant, or unknown id returns 404 (no existence leak). Required: `id`. |
 | `memory_promote` | Call at session end to compile this session's short-term (`session`-tier) memory into durable `long_term` memory — unlike `memory_remember` (a single explicit write), this compiles the whole session in one shot; fire it once at session end, not per turn. Returns 202 with `{session_id, status: "enqueued"}` — promotion runs asynchronously, so the resulting memory is recallable via `memory_recall` only after the worker drains. You can only promote your own sessions (scope resolved server-side from your key). Required: `session_id`. |
-| `recall_context` | ONE round-trip returning the re-ranked `global ∪ active-project` union of long-term MEMORY **and** KNOWLEDGE for `query` — what you previously assembled by calling `memory_recall` and `knowledge_search` separately. Pass `project_id` (from `resolve_project`) to merge global with that project on both sides; absent → global-only. The knowledge half is combined-search *summaries* (not full bodies — use `knowledge_context` for those). Response carries merged `results` (each tagged `source: memory\|knowledge`) plus the untouched per-source `memory`/`knowledge` envelopes; `meta.degraded?` flags a one-sided degrade (the other side is still returned — never a 500). Each per-source envelope's `meta.ann_iterative_scan` describes only THAT half's vector read, and the two are resolved independently, so they may differ. A blank query, or one over 500 chars, is a `422` up front. Required: `query`. Optional: `project_id`, `limit`. The top-level `meta.outcome` classifies the whole endpoint and the `memory` envelope carries its own; `meta.degraded_reason` names the strongest-remedy half when both degrade, and `meta.search_mode` names the lane that half actually served (`keyword_only`) or is `null` when it served nothing. |
+| `recall_context` | ONE round-trip returning the re-ranked `global ∪ active-project` union of long-term MEMORY **and** KNOWLEDGE for `query` — what you previously assembled by calling `memory_recall` and `knowledge_search` separately. Pass `project_id` (from `resolve_project`) to merge global with that project on both sides; absent → global-only. The knowledge half is combined-search *summaries* (not full bodies — use `knowledge_context` for those). Response carries merged `results` (each tagged `source: memory\|knowledge`) plus the untouched per-source `memory`/`knowledge` envelopes; `meta.degraded?` flags a one-sided degrade (the other side is still returned — never a 500). Each per-source envelope's `meta.ann_iterative_scan` describes only THAT half's vector read, and the two are resolved independently, so they may differ. A blank query, or one over 500 chars, is a `422` up front. Required: `query`. Optional: `project_id`, `limit`. The top-level `meta.outcome` classifies the whole endpoint and the `memory` envelope carries its own; `meta.degraded_reason` names the strongest-remedy half when both degrade, and `meta.search_mode` names the lane that half actually served (`keyword_only`) or is `null` when it served nothing. Response also carries a SELECTION LEDGER: per item `rank`, `selection_reason` (`keyword`\|`semantic`\|`keyword+semantic`\|`keyword_fallback` for knowledge, `semantic`\|`ilike_fallback` for memory) and `tokens_estimate` (bytes/4, an estimate); in `meta`, `recall_id`, `candidates_considered`, `selected_count`, `tokens_selected`, `tokens_candidates`, `tokens_saved_vs_candidates`. The merged order is deterministic (score DESC, then source, then id), so an unchanged corpus renders a byte-identical `data` array — cache that, not the whole response, since `meta.recall_id` is new on every call. Keep `meta.recall_id` for `recall_referenced`. |
+| `recall_referenced` | Record which of the articles a recall SURFACED you actually USED — the third funnel stage (surfaced -> opened -> referenced), and the only one nothing else records. Pass the `meta.recall_id` from a `recall_context` response plus the ids you referenced -- the `article.id` of the `source: knowledge` items, since a `memory` item's id is not an article and is not referenceable; call it once, after writing your answer. Only ids THAT recall surfaced, in your own tenant, are accepted — anything else is `422` `not_surfaced` and NOTHING is written. Your key is stamped server-side, at most 50 ids per call, and repeats cannot inflate the metric (it counts distinct `(recall, article)` pairs). These rows are deliberately not reads: they never feed the heat index or any ranking. Required: `recall_id`, `article_ids`. Optional: `project_id`. |
 | `memory_graduate` | Graduate ONE of your long-term memories into a durable Knowledge Wiki article — the explicit, on-demand version of the hourly graduation sweep. Use when a private memory has proven valuable enough to become durable knowledge. **Visibility**: the graduated article stays **owner-visible** (`metadata.visibility: "owner"`, keyed to your subject) — discoverable by YOU, NOT peer-readable (graduation does not share a memory to teammates; `re_scope: "global"` widens only the project scope, not visibility). Scope is key-derived (you can only graduate your OWN memory; a foreign/unknown `memory_id` → 404). DEDUPED by the novelty gate: `data.verdict` is `created` (novel → published) or `gated_to_draft` (near-dup → review draft) with a new article (**201**), or `duplicate`/`deduplicated` (already represented → canonical article, nothing created) (**200**). By default the article inherits the memory's project scope; pass `re_scope: "global"` to promote a PROJECT memory to a tenant-wide article — only valid on its FIRST graduation, and only if the hourly sweep hasn't graduated it project-scoped first (`409` `already_graduated` otherwise). An already-graduated global memory re-graduates idempotently (**200**). `503` `gate_unavailable` if the embedding backend is down — retry later. Required: `memory_id`. Optional: `re_scope` (`inherit`\|`global`). |
 
 ### Knowledge Management Tools (orchestrator key)

package/index.js

@@ -1881,6 +1881,35 @@ async function recallContext({ query, project_id, limit }) {
   return withRemediationNotice(result);
 }
 
+async function recallReferenced({ recall_id, article_ids, project_id }) {
+  // The third funnel stage. recall_id is a PATH segment: the server verifies that every
+  // article id was surfaced by THAT recall before recording anything, so a wrong id fails
+  // the whole call rather than recording a half-truth. The recording key is derived
+  // server-side from this key — nothing about the identity is sent.
+  //
+  // Path-injection guard, the same `UUID_RE` check knowledgeAgentUsage runs. `format:
+  // "uuid"` in the input schema is advisory — MCP does not enforce it — so a model that
+  // hallucinates or mis-copies an id containing `/` or `..` would otherwise have it
+  // spliced raw into the path, where URL normalisation sends the POST somewhere other
+  // than the endpoint this tool describes.
+  if (typeof recall_id !== "string" || !UUID_RE.test(recall_id.trim())) {
+    throw new Error(
+      "recall_id must be the meta.recall_id UUID from a recall_context response.",
+    );
+  }
+
+  const payload = { article_ids };
+  if (project_id) payload.project_id = project_id;
+
+  const result = await apiCall(
+    "POST",
+    `/api/v1/recall/${encodeURIComponent(recall_id.trim())}/referenced`,
+    payload,
+    process.env.LOOPCTL_AGENT_KEY,
+  );
+  return toContent(result);
+}
+
 async function memoryList({ limit, offset, include_superseded, all_subjects }) {
   // all_subjects is superadmin-only server-side; a non-superadmin key sending
   // this is ignored (falls back to its own subject) rather than erroring — the
@@ -5493,7 +5522,19 @@ const TOOLS = [
       "meta.degraded is true — never a hard failure. Each envelope's " +
       "`meta.ann_iterative_scan` discloses whether THAT half's vector read ran with " +
       "pgvector's iterative scan (`unavailable` ⇒ possibly incomplete); the two halves " +
-      "are resolved independently and may differ.",
+      "are resolved independently and may differ. SELECTION LEDGER: every merged item " +
+      "also carries `rank` (its position in THIS list), `selection_reason` (which lane " +
+      "put it there — keyword|semantic|keyword+semantic|keyword_fallback for knowledge, " +
+      "semantic|ilike_fallback for memory) and `tokens_estimate` (bytes/4 — an estimate, " +
+      "not a tokenizer count), and `meta` carries `recall_id`, `candidates_considered`, " +
+      "`selected_count`, `tokens_selected`, `tokens_candidates` and " +
+      "`tokens_saved_vs_candidates`, so you can explain your own context assembly. The " +
+      "merged order is deterministic (score DESC, then source, then id), so an unchanged " +
+      "corpus renders a byte-identical `data` array between turns — cache that array, not " +
+      "the whole response, since `meta.recall_id` is new on every call. KEEP " +
+      "`meta.recall_id`: after you " +
+      "answer, pass it to recall_referenced with the ids you actually used — that is the " +
+      "third funnel stage and nothing else records it.",
     inputSchema: {
       type: "object",
       properties: {
@@ -5517,6 +5558,47 @@ const TOOLS = [
       required: ["query"],
     },
   },
+  {
+    name: "recall_referenced",
+    description:
+      "Record which of the articles a recall SURFACED you actually USED in your answer — " +
+      "the third funnel stage (surfaced → opened → referenced). Pass the `meta.recall_id` " +
+      "from a recall_context response plus the article ids you referenced. This is the " +
+      "only signal that distinguishes 'the KB answered the question' from 'the KB was " +
+      "searched'; surfaced-to-opened follow-through is measured at 1.67% and what " +
+      "happened after an open was never recorded at all. Cheap and safe to call: only " +
+      "ids that THIS recall surfaced, in your own tenant, are accepted (anything else is " +
+      "a 422 not_surfaced and NOTHING is written), your key is stamped server-side, and " +
+      "repeating a call cannot inflate the metric — it counts distinct (recall, article) " +
+      "pairs. These rows are deliberately NOT reads: they never feed the heat index or " +
+      "any ranking, because a ranking that consumed a self-report could be gamed by one. " +
+      "Call it once, after you have written your answer.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        recall_id: {
+          type: "string",
+          format: "uuid",
+          description: "The `meta.recall_id` of the recall_context call that surfaced these articles.",
+        },
+        article_ids: {
+          type: "array",
+          items: { type: "string", format: "uuid" },
+          description:
+            "The ids you actually used, from that recall's `data`: the `article.id` of " +
+            "each item whose `source` is `knowledge`. A `memory` item's id is NOT an " +
+            "article and is not referenceable — including one fails the whole call. " +
+            "Non-empty, at most one recall page's worth.",
+        },
+        project_id: {
+          type: "string",
+          format: "uuid",
+          description: "Optional: attribution only, the same project scope you recalled under.",
+        },
+      },
+      required: ["recall_id", "article_ids"],
+    },
+  },
   {
     name: "memory_list",
     description:
@@ -8020,6 +8102,8 @@ server.setRequestHandler(CallToolRequestSchema, async (request) => {
       return await memoryRecall(args);
     case "recall_context":
       return await recallContext(args);
+    case "recall_referenced":
+      return await recallReferenced(args);
 
     case "memory_list":
       return await memoryList(args);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "loopctl-mcp-server",
-  "version": "2.86.0",
+  "version": "2.87.0",
   "description": "MCP server for loopctl — structural trust for AI development loops",
   "type": "module",
   "main": "index.js",