vantage-peers-mcp 2.8.0 → 2.9.0

package/CHANGELOG.md

@@ -1,5 +1,37 @@
 # Changelog
 
+## [2.9.0] — 2026-06-14 — Day 102 CRUD baseline PR-B: episode entity 5-op surface (mission k575kc1ryps0n8br95jw3q7d0x88m2v9)
+
+Mission `mcp-crud-baseline-standard` PR-B under T2. Second of 4 sub-PRs aligning the MCP surface with the Day 101 doctrine `j57dhrmkzjerjtssnr0z9ba57n88n7q7` ("5 ops per entity"). PR-B adds the missing read/list/search facades for the `episode` entity, completing the 5-op surface (the write side `store_episode` already existed since the 8-Sins doctrine).
+
+Architectural note: episodes are NOT a separate Convex table — per hotfix `7f958d0`, episodes are stored as memories with `type='episode'` (context/goal/action/outcome/insight + severity). The 4 new tools are thin wrappers that force `type='episode'` on the existing `memories:*` / `search:*` actions, so callers get an ergonomic episode-scoped surface without introducing a new backend table or duplicating index logic.
+
+### Added (4 canonical episode tools)
+
+- **`get_episode`** — fetch a single episode by memory ID. Calls `memories:getMemory` then asserts the returned row has `type='episode'`; otherwise returns a non-leaky "Episode not found" so wrong-type IDs do not leak existence of non-episode memories. Scope-aware via `scopeFilterGet`. Annotations: `readOnlyHint=true`, `openWorldHint=false`, `destructiveHint=false`.
+- **`list_episodes`** — list episodes ordered newest first. Calls `memories:listMemories` with `type='episode'` forced. Accepts `namespace?`, `createdBy?`, `limit?`, `fields?`, `cursor?` — same paging semantics as `list_memories`. Scope-aware via `scopeFilterList`. Envelope-capped via `capListResponseBytes("list_episodes")`.
+- **`search_episodes_by_keyword`** — BM25 search restricted to episodes. Calls `search:textSearch` with `type='episode'` forced. Same `guardRead` namespace gate, same 20-default / 200-cap limits.
+- **`search_episodes_by_semantic`** — semantic vector cosine search restricted to episodes. Calls `search:recall` with `type='episode'` forced. Same gate, same limits.
+
+All four are pure MCP-layer additions: no Convex schema change, no new index, no new action. tsc clean.
+
+### Why
+
+Episode = the 8-Sins / orchestrator-introspection memory type (severity + context/goal/action/outcome/insight). Until 2.9.0, recalling past episodes required `recall query='...' type='episode'` — discoverable only to callers who already knew the memory-side filter trick. Surfacing dedicated wrappers makes the episode lifecycle (write via `store_episode`, read via `get_episode`, browse via `list_episodes`, recall via the two search ops) consistent with the doctrine and self-documenting in any MCP client's tool list.
+
+`hybrid_search` remains entity-agnostic and is NOT mirrored for episodes (cross-cutting RRF tool per audit § 4).
+
+### Test fixture catch-up
+
+- `READ_ONLY_TOOLS` set in `mcp-server/src/__tests__/chatgpt-tool-annotations.test.ts` extended with the 4 new tool names.
+
+### Refs
+
+- Mission `k575kc1ryps0n8br95jw3q7d0x88m2v9` (MCP CRUD Baseline Standard, pilot Sigma + agents Sigma + Eta).
+- Pi authorization msg `jn74v2pkfz08agex4nfm2yfvfd88nzdw` — "chain T2 PR-B episode entity en autonomie scope mission".
+- Audit T1 deliverable: `analysis/mcp-crud-baseline-vp-audit-2026-06-14.md` row 7 (episode entity recommendation: add façade wrappers).
+- Architecture: hotfix `7f958d0` — episodes are memories with metadata, not a separate table.
+
 ## [2.8.0] — 2026-06-14 — Day 101 CRUD baseline PR-A: search_memories_by_keyword + search_memories_by_semantic (mission k575kc1ryps0n8br95jw3q7d0x88m2v9, task k1735qk9kx6agjjyt3e38rdvvh88mk0p)
 
 Mission `mcp-crud-baseline-standard` PR-A under T2 `[CRUD-T2] Implémentation gaps VP MCP`. First of 4 sub-PRs aligning the MCP surface with the Day 101 doctrine `j57dhrmkzjerjtssnr0z9ba57n88n7q7` ("5 ops per entity: get / list / search_by_keyword / search_by_semantic / create-or-upsert"). PR-A handles the convention drift on the `memories` entity — the only entity that already had both keyword + semantic search wired, but under non-canonical names.

package/README.md

@@ -242,7 +242,7 @@ Example:
 ### Session (1)
 `set_summary`
 
-## Compact payloads and status aliases (v2.8.0 — feature since v2.3.0)
+## Compact payloads and status aliases (v2.9.0 — feature since v2.3.0)
 
 ### `fields=lite` — reduced token payloads
 

package/dist/server.js

@@ -102,7 +102,7 @@ const convexUrl = loadConvexUrl();
 const convex = new ConvexHttpClient(convexUrl);
 const server = new McpServer({
     name: "vantage-peers",
-    version: "2.8.0",
+    version: "2.9.0",
 });
 // ─────────────────────────────────────────────────────────────────────────────
 // Helper: structured error response for MCP tool handlers

package/dist/src/tools.js

@@ -1240,6 +1240,225 @@ export function registerTools(server, convex, oauthCtx) {
             return mcpConvexError(error);
         }
     });
+    // ── get_episode ────────────────────────────────────────────────────────────
+    // Day 102 v2.9.0 — episode entity 5-op surface (PR-B).
+    // Thin wrapper: episodes are stored as memories with type='episode'
+    // (no separate table — see hotfix 7f958d0). This calls memories:getMemory
+    // and asserts type='episode' so callers get a non-leaky 404 on wrong-type IDs.
+    server.tool("get_episode", "Fetch a single episode by its memory document ID. Episodes are memories with type='episode' carrying context/goal/action/outcome/insight + severity. " +
+        "WHEN: use when you have an episodeId from store_episode or a prior search and need the full record. " +
+        "EXAMPLE: get_episode episodeId='j57dy3049btafda9m2f5d2ggk987ph3f'.", {
+        episodeId: z.string().describe("Episode (memory) document ID"),
+    }, {
+        readOnlyHint: true,
+        openWorldHint: false,
+        destructiveHint: false,
+        title: "Get episode",
+    }, async ({ episodeId }) => {
+        try {
+            const memory = await convex.query("memories:getMemory", {
+                memoryId: episodeId,
+            });
+            const filtered = scopeFilterGet(oauthCtx, memory);
+            if (filtered === null) {
+                return mcpError(`Episode not found: ${episodeId}`);
+            }
+            if (filtered?.type !== "episode") {
+                return mcpError(`Episode not found: ${episodeId}`);
+            }
+            return {
+                content: [{ type: "text", text: JSON.stringify(filtered, null, 2) }],
+            };
+        }
+        catch (error) {
+            return mcpConvexError(error);
+        }
+    });
+    // ── list_episodes ──────────────────────────────────────────────────────────
+    // Day 102 v2.9.0 — episode entity 5-op surface (PR-B).
+    // Thin wrapper on memories:listMemories with type='episode' forced.
+    server.tool("list_episodes", "List episodes (memories with type='episode') ordered newest first. " +
+        "WHEN: use to enumerate episodes by namespace or creator before recall/audit. " +
+        "EXAMPLE: list_episodes namespace='orchestrator/sigma' limit=20.", {
+        namespace: z
+            .string()
+            .optional()
+            .describe("Filter to a specific namespace — omit to list across all"),
+        createdBy: z
+            .string()
+            .optional()
+            .describe("Filter by creator role (e.g. 'sigma', 'pi')"),
+        limit: z
+            .number()
+            .int()
+            .min(1)
+            .max(200)
+            .optional()
+            .describe("Max items to return. Default 20 (envelope-safe). Cap 200."),
+        fields: z
+            .enum(["lite", "full"])
+            .optional()
+            .describe("'lite' returns compact payload (less tokens), 'full' is default."),
+        cursor: z
+            .string()
+            .optional()
+            .describe("S3.3 B8 — opaque pagination cursor from a prior call's `nextCursor`."),
+    }, {
+        readOnlyHint: true,
+        openWorldHint: false,
+        destructiveHint: false,
+        title: "List episodes",
+    }, async ({ namespace, createdBy, limit, fields, cursor }) => {
+        try {
+            const nsDenied = guardRead(namespace);
+            if (nsDenied)
+                return nsDenied;
+            let backendCursor;
+            if (cursor !== undefined && cursor !== "") {
+                try {
+                    const decoded = decodeCursor(cursor);
+                    if (decoded && "backendCursor" in decoded) {
+                        backendCursor = decoded.backendCursor;
+                    }
+                }
+                catch (err) {
+                    return mcpError(err?.message ?? "invalid cursor");
+                }
+            }
+            const effectiveLimit = limit === undefined ? undefined : clampLimit(limit);
+            const queryArgs = {
+                namespace,
+                type: "episode",
+                createdBy,
+                limit: effectiveLimit ?? 20,
+                fields: fields ?? "lite",
+            };
+            if (backendCursor !== undefined) {
+                queryArgs.paginationOpts = {
+                    numItems: effectiveLimit ?? 50,
+                    cursor: backendCursor,
+                };
+            }
+            const memories = await convex.query("memories:listMemories", queryArgs);
+            const rawList = Array.isArray(memories)
+                ? memories
+                : Array.isArray(memories?.page)
+                    ? memories.page
+                    : [];
+            const filteredList = scopeFilterList(oauthCtx, rawList);
+            const filteredEnvelope = Array.isArray(memories)
+                ? filteredList
+                : { ...memories, page: filteredList };
+            const text = capListResponseBytes(filteredEnvelope, JSON.stringify(filteredEnvelope, null, 2), "list_episodes");
+            return {
+                content: [{ type: "text", text }],
+            };
+        }
+        catch (error) {
+            return mcpConvexError(error);
+        }
+    });
+    // ── search_episodes_by_keyword ─────────────────────────────────────────────
+    // Day 102 v2.9.0 — episode entity 5-op surface (PR-B).
+    // Thin wrapper on search:textSearch with type='episode' forced.
+    server.tool("search_episodes_by_keyword", "BM25 full-text keyword search restricted to episodes (memories with type='episode'). " +
+        "WHEN: use when search_episodes_by_semantic returns too-broad results and you need an exact phrase or ID inside an episode field. " +
+        "EXAMPLE: search_episodes_by_keyword query='convex deploy schema' namespace='orchestrator/sigma' limit=10.", {
+        query: z.string().describe("Search query text"),
+        namespace: z
+            .string()
+            .optional()
+            .describe("Namespace filter (e.g. 'orchestrator/sigma')"),
+        limit: z
+            .number()
+            .int()
+            .min(1)
+            .max(200)
+            .optional()
+            .describe("Max items to return. Default 20 (envelope-safe). Cap 200."),
+        fields: z
+            .enum(["lite", "full"])
+            .optional()
+            .describe("'lite' returns compact payload (less tokens), 'full' is default."),
+    }, {
+        readOnlyHint: true,
+        openWorldHint: false,
+        destructiveHint: false,
+        title: "Search episodes by keyword (BM25)",
+    }, async ({ query, namespace, limit, fields }) => {
+        try {
+            const nsDenied = guardRead(namespace);
+            if (nsDenied)
+                return nsDenied;
+            const results = await convex.action("search:textSearch", {
+                query,
+                namespace,
+                type: "episode",
+                limit: limit ?? 20,
+                fields: fields ?? "lite",
+            });
+            return {
+                content: [{ type: "text", text: JSON.stringify(results, null, 2) }],
+            };
+        }
+        catch (error) {
+            return mcpConvexError(error);
+        }
+    });
+    // ── search_episodes_by_semantic ────────────────────────────────────────────
+    // Day 102 v2.9.0 — episode entity 5-op surface (PR-B).
+    // Thin wrapper on search:recall with type='episode' forced.
+    server.tool("search_episodes_by_semantic", "Semantic vector search restricted to episodes (memories with type='episode'), ranked by cosine similarity. " +
+        "WHEN: use to recall structured past events by intent — failure modes, lessons, similar contexts. " +
+        "EXAMPLE: search_episodes_by_semantic query='hook false positive blocked publish' namespace='orchestrator/sigma' limit=20.", {
+        query: z
+            .string()
+            .describe("Natural language query to search for relevant episodes"),
+        namespace: z
+            .string()
+            .optional()
+            .describe("Filter to a specific namespace — omit to search all"),
+        limit: z
+            .number()
+            .int()
+            .min(1)
+            .max(200)
+            .optional()
+            .describe("Max items to return. Default 20 (envelope-safe). Cap 200."),
+        fields: z
+            .enum(["lite", "full"])
+            .optional()
+            .describe("'lite' returns compact payload (less tokens), 'full' is default."),
+    }, {
+        readOnlyHint: true,
+        openWorldHint: false,
+        destructiveHint: false,
+        title: "Search episodes by semantic (vector cosine)",
+    }, async ({ query, namespace, limit, fields }) => {
+        try {
+            const nsDenied = guardRead(namespace);
+            if (nsDenied)
+                return nsDenied;
+            const results = await convex.action("search:recall", {
+                query,
+                namespace,
+                type: "episode",
+                limit: limit ?? 20,
+                fields: fields ?? "lite",
+            });
+            return {
+                content: [
+                    {
+                        type: "text",
+                        text: JSON.stringify(results, null, 2),
+                    },
+                ],
+            };
+        }
+        catch (error) {
+            return mcpConvexError(error);
+        }
+    });
     // ── get_profile ─────────────────────────────────────────────────────────────
     server.tool("get_profile", "Fetch an orchestrator profile with static identity and dynamic session state fields. " +
         "WHEN: use to check peer status, capabilities, or current task before assigning work. " +

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "vantage-peers-mcp",
-  "version": "2.8.0",
+  "version": "2.9.0",
   "description": "MCP server for VantagePeers — shared memory, messaging, and task coordination for AI agent teams",
   "type": "module",
   "main": "./dist/server.js",