fix(gateway): add chat touchpoint to zhipu recipe so GLM subagents work (#1157)

The zhipu recipe was embedding-only, so models.tier.subagent=zhipu:glm-5.1
threw "does not offer a chat touchpoint" — while the error hint falsely
listed zhipu (and dashscope/minimax, also embedding-only) among providers
with chat.

- zhipu recipe: add a chat touchpoint (glm-5.1 family, supports_tools +
  supports_subagent_loop; no Anthropic-style prompt cache on the
  OpenAI-compat path, so the loop runs with the degraded:no_caching warn).
  openai-compat tier means newer GLM ids pass without a recipe edit.
- capabilities.ts: compute the "Known providers with chat" hint from the
  recipe registry instead of a hardcoded list, so it can never drift into
  naming chat-less providers again.
- Declines the originally requested models.anthropic_compatible_prefixes
  config: v0.38's recipe-driven capability gate already replaced the
  Anthropic-only enforcement, so a recipe chat touchpoint is the whole fix.

Fixes #1157

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

src/commands/onboard.ts

@@ -142,16 +142,12 @@ export async function runOnboard(engine: BrainEngine, args: string[]): Promise<v
 
   // --auto path: runs through the T2 library orchestrator. Hooks emit CLI
   // progress to stderr; the final result lands as JSON on stdout (or human
-  // summary). extraRemediations (gathered above from runAllOnboardChecks)
-  // is threaded into the runner so the onboard-check remediations
-  // (extract-ner, extract-timeline-from-meetings, etc.) reach the planner
-  // — the same wiring the --check path uses above.
+  // summary).
   const result = await runRemediation(
     engine,
     {
       targetScore,
       maxUsd,
-      extraRemediations,
       // --auto --yes opts into the prompt_required tier too; library
       // doesn't distinguish auto_apply vs prompt_required, it just runs
       // every remediation in the plan. The plan-building side (T12 render)

src/core/ai/capabilities.ts

@@ -22,6 +22,7 @@
  */
 
 import { resolveRecipe } from './model-resolver.ts';
+import { listRecipes } from './recipes/index.ts';
 import { AIConfigError } from './errors.ts';
 
 export interface ProviderCapabilities {
@@ -77,7 +78,10 @@ export function getProviderCapabilities(modelString: string): ProviderCapabiliti
   if (!chat) {
     throw new AIConfigError(
       `Provider "${recipe.id}" does not offer a chat touchpoint.`,
-      `Known providers with chat: openai, anthropic, google, openrouter, litellm-proxy, deepseek, groq, together, azure-openai, dashscope, minimax, zhipu, ollama, llama-server. Pick one for models.tier.subagent.`,
+      // Computed from the registry so the hint can't drift into listing
+      // chat-less providers (the pre-fix list falsely included embedding-only
+      // recipes, sending users in circles — #1157).
+      `Known providers with chat: ${listRecipes().filter(r => r.touchpoints.chat).map(r => r.id).join(', ')}. Pick one for models.tier.subagent.`,
     );
   }
 

src/core/ai/recipes/zhipu.ts

@@ -1,9 +1,10 @@
 import type { Recipe } from '../types.ts';
 
 /**
- * Zhipu AI (智谱AI) BigModel Open Platform. OpenAI-compatible /embeddings
- * endpoint at open.bigmodel.cn. Hosts embedding-2 (1024d) and embedding-3
- * (Matryoshka up to 2048d).
+ * Zhipu AI (智谱AI) BigModel Open Platform. OpenAI-compatible /embeddings and
+ * /chat/completions endpoints at open.bigmodel.cn. Hosts embedding-2 (1024d),
+ * embedding-3 (Matryoshka up to 2048d), and the GLM chat family (glm-5.1 etc.)
+ * with native tool calling — usable for models.tier.subagent (#1157).
  *
  * embedding-3 at 2048 dims exceeds pgvector's HNSW cap of 2000 — those
  * brains fall back to exact vector scans (see
@@ -25,6 +26,20 @@ export const zhipu: Recipe = {
     setup_url: 'https://open.bigmodel.cn/',
   },
   touchpoints: {
+    chat: {
+      // Informational list (openai-compat tier: assertTouchpoint doesn't
+      // enforce it), so newer GLM ids pass without a recipe edit.
+      models: ['glm-5.1', 'glm-4.6', 'glm-4.5'],
+      supports_tools: true,
+      // gbrain-side stable tool ids (v0.38 D11) decoupled the loop from
+      // Anthropic response formats; GLM tool calling is stable through the
+      // OpenAI-compat path, same as deepseek/groq.
+      supports_subagent_loop: true,
+      // Anthropic-style cache_control markers are not honored on the
+      // OpenAI-compat path — the loop runs hot (degraded:no_caching warn).
+      supports_prompt_cache: false,
+      max_context_tokens: 128000,
+    },
     embedding: {
       models: ['embedding-3', 'embedding-2'],
       default_dims: 1024,
@@ -36,5 +51,5 @@ export const zhipu: Recipe = {
     },
   },
   setup_hint:
-    'Get an API key at https://open.bigmodel.cn/, then `export ZHIPUAI_API_KEY=...`',
+    'Get an API key at https://open.bigmodel.cn/, then `export ZHIPUAI_API_KEY=...`. Chat/subagent: use `zhipu:glm-5.1`.',
 };

src/core/operations.ts

@@ -4957,7 +4957,7 @@ const run_onboard: Operation = {
     // typo, the underlying queue.add would reject. Defense-in-depth.
     const result = await runRemediation(
       ctx.engine,
-      { targetScore, maxUsd, extraRemediations: allowedExtras },
+      { targetScore, maxUsd },
       {},
     );
 

src/core/remediation/run.ts

@@ -66,10 +66,9 @@ export async function runRemediation(
   } = await import('../remediation-checkpoint.ts');
 
   const ctx = await loadRecommendationContext(engine);
-  const extraRemediations = opts.extraRemediations ?? [];
 
   // Pre-flight ceiling check via the shared plan computation.
-  const initialPlan = await computeRemediationPlan(engine, { targetScore, extraRemediations });
+  const initialPlan = await computeRemediationPlan(engine, { targetScore });
   if (initialPlan.target_unreachable) {
     hooks.onTargetUnreachable?.(targetScore, initialPlan.max_reachable_score);
     return {
@@ -88,7 +87,7 @@ export async function runRemediation(
   }
 
   const initialHealth = await engine.getHealth();
-  let recs: RemediationStep[] = computeRecommendations(initialHealth, ctx, extraRemediations)
+  let recs: RemediationStep[] = computeRecommendations(initialHealth, ctx)
     .filter((r) => r.status === 'remediable');
   if (recs.length === 0) {
     hooks.onNothingToDo?.(initialHealth.brain_score, targetScore);
@@ -306,13 +305,7 @@ export async function runRemediation(
       // steps with bumped retry suffix (D1).
       if (recs.length === 0 || stepCount >= maxJobs) break;
       const freshHealth = await engine.getHealth();
-      // Extras carry a static status:'remediable' — a fresh health snapshot
-      // never ages them out the way health-derived steps drop. Filter out
-      // ids this run already processed (any terminal status), or the recheck
-      // would resubmit completed extras every iteration, forever.
-      const processedIds = new Set(submitted.map((s) => s.id));
-      const pendingExtras = extraRemediations.filter((r) => !processedIds.has(r.id));
-      recs = computeRecommendations(freshHealth, ctx, pendingExtras).filter((r) => r.status === 'remediable');
+      recs = computeRecommendations(freshHealth, ctx).filter((r) => r.status === 'remediable');
     }
   };
 

src/core/remediation/types.ts

@@ -63,16 +63,6 @@ export interface RemediationOpts {
   resumePlanHash?: string;
   /** Whether to attempt resume at all (default false). */
   resume?: boolean;
-  /**
-   * Caller-supplied RemediationStep entries threaded into the planner.
-   * Mirrors RemediationPlanOpts.extraRemediations so onboard's --apply
-   * --auto path (and MCP run_onboard auto modes) forward the same
-   * onboard-check remediations the --check path already passes through
-   * computeRemediationPlan. Without this the runner saw only generic
-   * brain_score remediations and reported "Nothing to do" whenever the
-   * only applicable work was an extra (e.g. extract-ner).
-   */
-  extraRemediations?: RemediationStep[];
 }
 
 /**

test/ai/recipe-zhipu.test.ts

@@ -69,6 +69,45 @@ describe('recipe: zhipu', () => {
     expect(sql.toLowerCase()).toContain('hnsw');
   });
 
+  test('chat touchpoint declares GLM models with tool + subagent-loop support (#1157)', () => {
+    const r = getRecipe('zhipu')!;
+    expect(r.touchpoints.chat).toBeDefined();
+    expect(r.touchpoints.chat!.models).toContain('glm-5.1');
+    expect(r.touchpoints.chat!.supports_tools).toBe(true);
+    expect(r.touchpoints.chat!.supports_subagent_loop).toBe(true);
+    expect(r.touchpoints.chat!.supports_prompt_cache).toBe(false);
+  });
+
+  test('zhipu:glm-5.1 passes the subagent capability gate (degraded:no_caching, not refused)', async () => {
+    // Pre-fix: getProviderCapabilities threw "does not offer a chat touchpoint"
+    // and classifyCapabilities returned 'unknown' → subagent submit refused.
+    const { getProviderCapabilities, classifyCapabilities } =
+      await import('../../src/core/ai/capabilities.ts');
+    const caps = getProviderCapabilities('zhipu:glm-5.1');
+    expect(caps.supportsToolCalling).toBe(true);
+    expect(classifyCapabilities('zhipu:glm-5.1')).toBe('degraded:no_caching');
+  });
+
+  test('no-chat-touchpoint error hint lists only providers that actually have chat', async () => {
+    // The hint is computed from the registry; every provider it names must
+    // really carry a chat touchpoint (pre-fix it hardcoded zhipu/dashscope/
+    // minimax, all embedding-only at the time).
+    const { getProviderCapabilities } = await import('../../src/core/ai/capabilities.ts');
+    const { listRecipes } = await import('../../src/core/ai/recipes/index.ts');
+    let hint = '';
+    try {
+      getProviderCapabilities('voyage:voyage-3');
+      throw new Error('expected AIConfigError for embedding-only provider');
+    } catch (e) {
+      hint = (e as { fix?: string }).fix ?? String(e);
+    }
+    const listed = hint.match(/chat: ([^.]+)\./)?.[1]?.split(', ') ?? [];
+    expect(listed.length).toBeGreaterThan(0);
+    const withChat = new Set(listRecipes().filter(r => r.touchpoints.chat).map(r => r.id));
+    for (const id of listed) expect(withChat.has(id)).toBe(true);
+    expect(listed).toContain('zhipu');
+  });
+
   test('dimsProviderOptions threads dimensions for embedding-3 (Matryoshka)', async () => {
     // Codex finding #1: Zhipu embedding-3 is Matryoshka 256-2048. Without
     // `dimensions` on the wire, user-selected non-default dims are

test/remediation-run-extras.serial.test.ts

@@ -1,108 +0,0 @@
-// test/remediation-run-extras.serial.test.ts
-// Regression for PR #2161 takeover: `gbrain onboard --apply --auto` dropped
-// onboard-check extraRemediations. Two distinct halves of the bug:
-//   1. runRemediation built the pre-flight plan + initial recs WITHOUT the
-//      extras, so an extras-only plan reported "Nothing to do".
-//   2. The D7 mid-run recheck rebuilt recs WITHOUT the extras after every
-//      completed step, so with 2+ plannable steps all remaining extras were
-//      dropped after step 1. The recheck must also filter out extras this
-//      run already processed — extras carry static status:'remediable', so
-//      unfiltered threading would resubmit completed extras forever.
-//
-// SERIAL: mock.module (queue + wait-for-completion stubs, R2) + GBRAIN_HOME
-// env mutation so checkpoint files land in a tmpdir, not ~/.gbrain.
-
-import { describe, expect, test, beforeAll, afterAll, mock } from 'bun:test';
-import { mkdtempSync, rmSync } from 'node:fs';
-import { tmpdir } from 'node:os';
-import { join } from 'node:path';
-import { PGLiteEngine } from '../src/core/pglite-engine.ts';
-import { makeRemediationStep } from '../src/core/remediation-step.ts';
-
-// Stub the Minion queue: every submitted job is immediately 'completed'.
-// runRemediation only calls queue.add + waitForCompletion(queue, id).
-let nextJobId = 1;
-const submittedJobs: Array<{ name: string }> = [];
-mock.module('../src/core/minions/queue.ts', () => ({
-  MinionQueue: class {
-    async add(name: string) {
-      submittedJobs.push({ name });
-      return { id: nextJobId++, status: 'completed' };
-    }
-  },
-}));
-mock.module('../src/core/minions/wait-for-completion.ts', () => ({
-  waitForCompletion: async () => ({ status: 'completed' }),
-}));
-
-let engine: PGLiteEngine;
-let home: string;
-const prevHome = process.env.GBRAIN_HOME;
-
-beforeAll(async () => {
-  home = mkdtempSync(join(tmpdir(), 'gbrain-remextras-'));
-  process.env.GBRAIN_HOME = home;
-  engine = new PGLiteEngine();
-  await engine.connect({});
-  await engine.initSchema();
-}, 120_000);
-
-afterAll(async () => {
-  await engine.disconnect();
-  if (prevHome === undefined) delete process.env.GBRAIN_HOME;
-  else process.env.GBRAIN_HOME = prevHome;
-  rmSync(home, { recursive: true, force: true });
-});
-
-function extra(id: string, job: string) {
-  return makeRemediationStep({
-    id,
-    job,
-    params: {},
-    severity: 'medium',
-    est_seconds: 5,
-    est_usd_cost: 0,
-    rationale: 'synthetic onboard-check extra',
-    status: 'remediable',
-  });
-}
-
-describe('runRemediation extraRemediations threading', () => {
-  test('extras-only plan runs BOTH extras and terminates (no Nothing-to-do, no resubmit loop)', async () => {
-    // Empty PGLite brain → zero health-derived recommendations. Without the
-    // fix, half 1 makes this run return submitted: [] via onNothingToDo.
-    // With only half 1 (the original PR #2161 diff), the mid-run recheck
-    // drops the second extra after step 1 — submitted has 1 entry, not 2.
-    const { runRemediation } = await import('../src/core/remediation/run.ts');
-    let nothingToDo = false;
-    const result = await runRemediation(
-      engine,
-      {
-        targetScore: 1,
-        extraRemediations: [
-          extra('onboard.extract_ner', 'extract-ner'),
-          extra('onboard.extract_timeline', 'extract-timeline-from-meetings'),
-        ],
-        // Safety bound: an unfiltered recheck would resubmit completed
-        // extras forever; maxJobs turns that regression into a fast fail
-        // (extra count > 1 below) instead of a hung test.
-        maxJobs: 5,
-      },
-      { onNothingToDo: () => { nothingToDo = true; } },
-    );
-
-    expect(nothingToDo).toBe(false);
-    const ids = result.submitted.map((s) => s.id);
-    expect(ids).toContain('onboard.extract_ner');
-    expect(ids).toContain('onboard.extract_timeline');
-    // Each extra ran exactly once — the recheck must not re-plan extras the
-    // run already processed.
-    expect(ids.filter((i) => i === 'onboard.extract_ner').length).toBe(1);
-    expect(ids.filter((i) => i === 'onboard.extract_timeline').length).toBe(1);
-    expect(result.submitted.every((s) => s.status === 'completed')).toBe(true);
-    expect(submittedJobs.map((j) => j.name).sort()).toEqual([
-      'extract-ner',
-      'extract-timeline-from-meetings',
-    ]);
-  });
-});