@bli-cockpit/mcp 0.1.22 → 0.1.24

package/README.md

@@ -34,7 +34,7 @@ about what an agent can reach.
 
 <!-- BEGIN GENERATED verb census — `npm run mcp:readme` -->
 
-**76 of 79 Tower verbs have an MCP twin.**
+**77 of 80 Tower verbs have an MCP twin.**
 Each tool goes through the SAME door its CLI verb calls, with the same
 collector device token — never a second route and never a service-role
 reader. `src/verb-census.test.ts` fails when a verb is in none of the
@@ -81,6 +81,7 @@ three tables below.
 | `cockpit memory log` | `memory_experience` | `POST /api/memory/experience` |
 | `cockpit model set` | `model_set` | `POST /api/settings/jarvis-model` |
 | `cockpit model show` | `model_show` | `GET /api/settings/jarvis-model` |
+| `cockpit models compare` | `models_compare` | `GET /api/models/cards?compare=` |
 | `cockpit models list` | `models_list` | `GET /api/models/cards` |
 | `cockpit models show` | `models_card` | `GET /api/models/cards?id=` |
 | `cockpit msg channels` | `msg_channels` | `GET /api/msg/channels` |

package/dist/models-tools.js

@@ -46,9 +46,13 @@ export function registerModelsTools(server, deps) {
     }));
     register("models_card", {
         title: "Read one model card",
-        description: "One model in full: every documented value with the provider page it was read from, the quirks quoted "
-            + "verbatim, our measured latency and tokens/s, each probe verdict, and any stale mark. A field the "
-            + "provider does not state reads unknown with the reason — never a guess.",
+        description: "One model in full, as six questions: what's going on (window, price, today's spend and turns), what's "
+            + "wrong (stale fields, refused probes, today's failures), what's right (probe verdicts, bench record), "
+            + "what should we do / double down / forget about it — each a verbatim quotation from a committed bench "
+            + "report or the words `no bench verdict — not recommended either way`, never advice this tool composed. "
+            + "The provider's claim and our measurement stay in separate columns, and a legend names the window and "
+            + "row count behind every measured figure. A field the provider does not state reads unknown with the "
+            + "reason — never a guess.",
         inputSchema: {
             id: z
                 .string()
@@ -59,6 +63,9 @@ export function registerModelsTools(server, deps) {
     }, async (args) => withSession(deps, async (session) => {
         const params = new URLSearchParams();
         params.set("id", String(args.id));
+        // The six sections (BLI-3918) — the same answer the model page renders
+        // and `cockpit models show` prints.
+        params.set("sections", "1");
         const response = await callAgentDoor(session, deps.fetchImpl, "GET", `/api/models/cards${queryString(params)}`);
         // A 404 here is the door's own sentence naming the id and the command
         // that would add it, which is more useful than anything this file
@@ -67,7 +74,63 @@ export function registerModelsTools(server, deps) {
             return errorResult(doorFailureText("models_card", response));
         const body = response.body;
         const card = (body.card ?? {});
-        const lines = (Array.isArray(body.lines) ? body.lines : []);
-        return textResult(lines.join("\n") || `Model card ${card.id ?? args.id}.`, { card });
+        const sections = (body.sections ?? null);
+        const sectionLines = (Array.isArray(body.section_lines) ? body.section_lines : []);
+        // A server that predates the sections still answers the whole card, and
+        // an agent gets that rather than nothing.
+        const lines = sectionLines.length > 0
+            ? sectionLines
+            : (Array.isArray(body.lines) ? body.lines : []);
+        return textResult(lines.join("\n") || `Model card ${card.id ?? args.id}.`, {
+            card,
+            ...(sections ? { sections } : {}),
+        });
+    }));
+    register("models_compare", {
+        title: "Compare two to four model cards",
+        description: "Two to four models side by side — models are columns, facts are rows — over the same door and the same "
+            + "builder the page at /settings/models/compare renders and `cockpit models compare` prints. Groups: "
+            + "overview, price (including `$ per 100 JARVIS turns`, our own measured tokens a turn times the "
+            + "provider's price, labelled with the window it was measured over — `not measured` where no profile "
+            + "exists, never a per-million price relabelled), our measurements, the bench and the quirks. The bench "
+            + "rows appear only when every model was judged in the SAME run against the SAME baseline; otherwise the "
+            + "group says `benched separately — not a head-to-head` rather than subtracting two win rates. A group "
+            + "with nothing in it for one column says so once, in that column; a single missing cell is `—`, and a "
+            + "measured one with no rows behind it is `not measured` — the two are never blurred. Read-only.",
+        inputSchema: {
+            ids: z
+                .array(z.string().min(1).max(200))
+                .min(2)
+                .max(4)
+                .describe("Two to four card ids, `provider:model` (e.g. openai:gpt-5.6-luna), in column order."),
+            highlight: z
+                .boolean()
+                .optional()
+                .describe("Mark the best cell in each scored row (lowest wins on price and latency, highest on tokens/s, "
+                + "most wins on the bench). Off by default; a row where any column is missing or unmeasured is "
+                + "never scored, and says why."),
+        },
+    }, async (args) => withSession(deps, async (session) => {
+        const ids = args.ids.map((id) => String(id));
+        const params = new URLSearchParams();
+        params.set("compare", ids.join(","));
+        if (args.highlight === true)
+            params.set("highlight", "1");
+        const response = await callAgentDoor(session, deps.fetchImpl, "GET", `/api/models/cards${queryString(params)}`);
+        // The door's own refusal names the id nothing answers to, or the column
+        // count it would not draw — both more useful than anything composed here.
+        if (!response.ok)
+            return errorResult(doorFailureText("models_compare", response));
+        const body = response.body;
+        const compare = (body.compare ?? {});
+        const lines = (Array.isArray(body.compare_lines) ? body.compare_lines : []);
+        // Every column names its PROVIDER as well as its model, because the
+        // same weights are reachable through two doors and two columns headed
+        // alike would be indistinguishable.
+        const headers = (compare.columns ?? [])
+            .map((column) => column.header ?? column.id ?? "")
+            .filter((header) => header.length > 0);
+        return textResult(lines.join("\n")
+            || `Compared ${ids.length} model cards: ${(headers.length > 0 ? headers : ids).join(", ")}.`, { compare });
     }));
 }

package/dist/verb-census.js

@@ -241,6 +241,8 @@ export const MCP_TWINS = {
     // the org setting that picks from it.
     "models list": { tool: "models_list", door: "GET /api/models/cards" },
     "models show": { tool: "models_card", door: "GET /api/models/cards?id=" },
+    // BLI-3919: two to four cards side by side, on the same door.
+    "models compare": { tool: "models_compare", door: "GET /api/models/cards?compare=" },
     "scout board": { tool: "scout_board", door: "GET /api/cockpit/scout" },
     "scout start": { tool: "scout_start", door: "POST /api/cockpit/scout (start)" },
     "scout dismiss": { tool: "scout_dismiss", door: "POST /api/cockpit/scout (dismiss)" },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bli-cockpit/mcp",
-  "version": "0.1.22",
+  "version": "0.1.24",
   "private": false,
   "description": "bli-tower \u2014 an MCP server over BLI Cockpit's agent doors: JARVIS (jarvis_*), documents (docs_*), channels (msg_*), issues (work_*), the daily page (brief_*), meeting notes (notes_*), the ops board (ops_status/slack_*), settings/team/model, Scout and the workbook, plus the legacy event-stream tools (emit_event, get_ticket_timeline, get_active_tickets).",
   "type": "module",
@@ -30,7 +30,7 @@
     "readme": "npm run build && node scripts/write-readme-census.mjs"
   },
   "dependencies": {
-    "@bli-cockpit/telemetry-core": "0.1.37",
+    "@bli-cockpit/telemetry-core": "0.1.38",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "zod": "^4.3.6"
   },