jeo-code 0.8.21 → 0.8.23

package/src/agent/prompt-router.ts

@@ -48,7 +48,7 @@
  * contract for users who never opted in.
  */
 import { callLlm } from "./loop";
-import { resolveRoleModel, modelServableWithConfig } from "../ai/model-manager";
+import { resolveRoleModel, modelServableWithConfig, resolveProvider, oauthServesModel } from "../ai/model-manager";
 import { catalogMetadata, MODEL_CATALOG, compareReleaseDate, liveProviderCatalogModels, isLiveProviderModel, type CatalogModel, type ThinkLevel } from "../ai/model-catalog";
 import { priceForModel } from "../ai/pricing";
 import { tryExtractJsonObject } from "./json";
@@ -222,23 +222,97 @@ function hasAntigravityOauth(config: RoutingConfig): boolean {
   return !!(config.oauth?.antigravity || config.oauth?.gemini);
 }
 
+/** A model's classified credential-sharing SCOPE — models on the SAME scope share
+ *  ONE account-wide rate-limit/billing window, so an account-level failure (429,
+ *  usage/quota limit, credential rejection, billing block) on one model means every
+ *  OTHER model in the same scope is doomed the same way THIS instant. `"oauth-subscription"`:
+ *  an OAuth SUBSCRIPTION (Claude Pro/Max, ChatGPT/Codex, Kimi Code, Antigravity Cloud
+ *  Code Assist) serves this model — `key` is `${provider}:oauth` (bare `"antigravity"`
+ *  for that provider: all three re-exported companies route through the SAME Cloud
+ *  Code Assist account/credential, confirmed no per-company quota separation exists).
+ *  `"no-credential"`: the provider has ZERO usable credential of any kind — every
+ *  model on it is equally dead. */
+export type CredentialScope = { kind: "no-credential" | "oauth-subscription"; key: string };
+
+/** Classify `model` into the credential it actually resolves through, mirroring
+ *  `resolveCall`/`effectiveCredentialForProvider`'s real precedence (OAuth wins over
+ *  a configured API key whenever it serves the model) WITHOUT touching the network —
+ *  config-only, so this stays safely callable from a synchronous fast-path predicate
+ *  (e.g. the engine's per-retry `rateLimitFallbackAvailable` check).
+ *
+ *  `null` = API-key-served (or keyless/local) — an API key is typically its OWN
+ *  independent per-key/per-model budget (Groq/OpenAI/etc rate-limit each model
+ *  independently, unlike an OAuth subscription's single shared window), so a `null`
+ *  candidate is NEVER excluded as a group — only its exact model id is excluded on
+ *  failure, preserving the intentionally-tested "same-provider API-key fallback"
+ *  behavior (see test/launch-fallback-live-discovery.test.ts). */
+export function credentialScopeFor(model: string, config: RoutingConfig): CredentialScope | null {
+  const provider = resolveProvider(model);
+  if (provider === "ollama" || provider === "lmstudio") return null; // keyless, no shared account limit
+  if (provider === "antigravity") {
+    const hasOauth = !!(config.oauth?.antigravity || config.oauth?.gemini);
+    return hasOauth ? { kind: "oauth-subscription", key: "antigravity" } : { kind: "no-credential", key: "antigravity" };
+  }
+  const auth = provider as AuthProvider;
+  const hasApiKey = !!config.providers?.[auth];
+  const hasOauth = !!config.oauth?.[auth];
+  if (!hasApiKey && !hasOauth) return { kind: "no-credential", key: provider };
+  // OAuth wins over a configured API key whenever it actually serves this model
+  // (same precedence resolveCall/effectiveCredentialForProvider apply live) — that
+  // is the credential this call will really use, and its subscription window is
+  // what a 429 here reports on. When OAuth is present but does NOT serve this
+  // model (e.g. a non-Codex id under OpenAI OAuth), the call falls through to the
+  // API key (if any) — API-key-served, so `null` (independent budget).
+  if (hasOauth && oauthServesModel(auth, model)) return { kind: "oauth-subscription", key: `${provider}:oauth` };
+  // Defense-in-depth (currently unreachable in practice — every OAuth provider
+  // whose token doesn't serve a given model AND has no verified end-to-end backend
+  // is already excluded upstream from every auto-select pool, per
+  // isCoreEligible/hasAntigravityOauth): OAuth is configured, has NO API key
+  // fallback, and does not actually serve this model (e.g. a bare `gemini`-provider
+  // model under gemini-OAuth-only, since gemini's OAuth backend is not verified
+  // end-to-end) — genuinely zero usable credential for this specific model, not an
+  // independent API-key budget.
+  if (hasOauth && !hasApiKey) return { kind: "no-credential", key: provider };
+  return null;
+}
+
+/** Builds the routing catalog once per call: static `MODEL_CATALOG` rows need the
+ *  full `isAutoSelectCandidate` gate (including the `isLiveProviderModel` check,
+ *  when `openaiBaseUrl` is set), but rows already returned by
+ *  `liveProviderCatalogModels()` are live BY CONSTRUCTION — re-verifying each one
+ *  with `isLiveProviderModel` is pure O(n) waste (it dominated this function's
+ *  cost: ~1.4ms/call at 2000 live models). Live rows still need the shared
+ *  provider/OAuth gates (cloud provider, credential, antigravity/gemini), just not
+ *  the live-membership check specific to static rows. */
 function routingCatalog(config: RoutingConfig): CatalogModel[] {
-  const rows = [...MODEL_CATALOG, ...liveProviderCatalogModels(config)];
-  const eligible = rows.filter(m => isAutoSelectCandidate(m, config));
+  const staticEligible = MODEL_CATALOG.filter(m => isAutoSelectCandidate(m, config));
+  const liveEligible = liveProviderCatalogModels(config).filter(m => isAutoSelectCandidateLive(m, config));
   const byId = new Map<string, CatalogModel>();
-  for (const row of eligible) if (!byId.has(row.canonical)) byId.set(row.canonical, row);
+  for (const row of staticEligible) if (!byId.has(row.canonical)) byId.set(row.canonical, row);
+  for (const row of liveEligible) if (!byId.has(row.canonical)) byId.set(row.canonical, row);
   return [...byId.values()];
 }
 
-function isAutoSelectCandidate(m: CatalogModel, config: RoutingConfig): boolean {
-  // Shared model-LEVEL credential gate (the full auth matrix is documented on
-  // modelServableWithConfig; keep aligned with resolveCall).
+/** Shared provider/OAuth eligibility gates, independent of live-membership. */
+function isCoreEligible(m: CatalogModel, config: RoutingConfig): boolean {
   if (!isCloudProvider(m.provider) || m.limitedAvailability || !modelServableWithConfig(m.provider, m.canonical, config)) return false;
-  if (m.provider === "openai" && config.openaiBaseUrl && !isLiveProviderModel("openai", m.canonical, config)) return false;
   if (hasAntigravityOauth(config) && m.provider === "gemini") return false;
   return true;
 }
 
+function isAutoSelectCandidate(m: CatalogModel, config: RoutingConfig): boolean {
+  if (!isCoreEligible(m, config)) return false;
+  if (m.provider === "openai" && config.openaiBaseUrl && !isLiveProviderModel("openai", m.canonical, config)) return false;
+  return true;
+}
+
+/** Same eligibility as `isAutoSelectCandidate`, for a row ALREADY sourced from
+ *  `liveProviderCatalogModels()` — skips the live-membership re-check (see
+ *  `routingCatalog`'s doc comment). */
+function isAutoSelectCandidateLive(m: CatalogModel, config: RoutingConfig): boolean {
+  return isCoreEligible(m, config);
+}
+
 
 /** Size class jeo infers for a catalog model, used to group EQUIVALENT models
  *  across DIFFERENT providers for pool-based routing (see `tierModelPool`) — the
@@ -353,17 +427,19 @@ export function selectFromPool(pool: readonly string[], sessionId: string | unde
  *  model has a known price (caller falls back to `defaultModel`). Tiebreak on
  *  an exact cost tie: NEWER `releaseDate` wins (a same-priced newer model is
  *  likely a refined successor), then canonical id (deterministic fallback
- *  when recency also ties or is unconfirmed on both sides). */
-export function cheapestCredentialed(config: RoutingConfig): string | null {
+ *  when recency also ties or is unconfirmed on both sides). Optional `filter`
+ *  narrows the candidate pool BEFORE price comparison (e.g. vision-capable
+ *  only, matching `strongestCredentialed`'s filter param) — a cheap model that
+ *  fails the filter is never picked over a pricier one that passes it. */
+export function cheapestCredentialed(config: RoutingConfig, filter?: (m: CatalogModel) => boolean): string | null {
   let best: CatalogModel | null = null;
   let bestCost = Infinity;
   for (const m of routingCatalog(config)) {
-    // Candidate eligibility is centralized in `isAutoSelectCandidate`: it keeps
-    // OAuth-backed Antigravity routing on provider-qualified 3.1+ Gemini rows and
-    // prevents public `gemini` catalog rows from winning when Antigravity OAuth is
-    // the available Gemini lane. The real veto gate in launch.ts's `runTurn` still
-    // re-verifies full readiness (including live OAuth expiry) before a turn uses it.
-    if (!isAutoSelectCandidate(m, config)) continue;
+    if (filter && !filter(m)) continue;
+    // `routingCatalog(config)` already filtered to eligible rows (static rows via
+    // `isAutoSelectCandidate`, live rows via `isAutoSelectCandidateLive`) — no need
+    // to re-check here. See cheapestCredentialed's own doc comment for the
+    // Antigravity/Gemini OAuth rationale behind that eligibility gate.
     const price = priceForModel(m.canonical);
     if (!price) continue;
     const cost = price.inPerM + price.outPerM;
@@ -388,6 +464,27 @@ export function cheapestCredentialed(config: RoutingConfig): string | null {
   return best?.canonical ?? null;
 }
 
+/** Cheap-tier model for a GRADER/VERIFIER call (goal-verifier, vision-verify,
+ *  `critic` role) — the blog-post "Haiku-for-graders" principle: a verdict call
+ *  reads an artifact + a rubric, it does not need frontier reasoning strength,
+ *  and defaulting graders to the SAME expensive model driving the turn they are
+ *  grading both wastes budget and (for self-graded turns) reintroduces the very
+ *  self-preference bias an independent grader exists to avoid. Resolution order:
+ *  `roles.smol` (explicit user pin, but only when it also satisfies `requireImages`)
+ *  → live-cheapest credentialed (filtered the same way) → `defaultModel` (never
+ *  null — every caller needs a concrete model id to pass to `callLlm`).
+ *  `opts.requireImages` (the browser vision-verify caller) excludes catalog rows
+ *  with `images: false` — a text-only cheap model silently ignores an attached
+ *  screenshot rather than erroring, which would surface as a hallucinated verdict
+ *  instead of a loud failure, so the filter runs BEFORE cost/pin selection, not after. */
+export function resolveVerifierModel(config: RoutingConfig, opts: { requireImages?: boolean } = {}): string {
+  const pinned = config.roles?.smol;
+  const pinnedOk = pinned && (!opts.requireImages || catalogMetadata(pinned)?.images !== false);
+  if (pinnedOk) return pinned!;
+  const filter = opts.requireImages ? (m: CatalogModel) => m.images !== false : undefined;
+  return cheapestCredentialed(config, filter) || config.defaultModel;
+}
+
 /** Most capable cloud model jeo has a credential for, computed LIVE off
  *  `MODEL_CATALOG` — ranks by full (xhigh) thinking support first, then max
  *  output tokens, then context window, then NEWER `releaseDate` (a same-specs
@@ -406,9 +503,8 @@ export function strongestCredentialed(
 ): string | null {
   let best: CatalogModel | null = null;
   for (const m of routingCatalog(config)) {
-    // Candidate eligibility is centralized in `isAutoSelectCandidate` (see
+    // `routingCatalog(config)` already filtered to eligible rows (see
     // cheapestCredentialed above for the Antigravity/Gemini OAuth rationale).
-    if (!isAutoSelectCandidate(m, config)) continue;
     if (filter && !filter(m)) continue;
     if (!best) { best = m; continue; }
     const xhighM = m.thinking.includes("xhigh") ? 1 : 0;
@@ -433,7 +529,7 @@ export function strongestCredentialed(
  *  a genuinely mid-class model like `claude-sonnet-4-6`. `null` when no
  *  mid-class-suffixed model is credentialed (caller falls through to
  *  `roles.medium`/`defaultModel` — never to `complex`'s flagship pick). */
-function strongestMidTierCredentialed(config: RoutingConfig): string | null {
+export function strongestMidTierCredentialed(config: RoutingConfig): string | null {
   const midClass = routingCatalog(config).filter(
     m => sizeClassFor(m) === "mid",
   );

package/src/agent/skill-lessons.ts

@@ -0,0 +1,315 @@
+/**
+ * Skill self-compounding: the durable link the bundled/user skill docs were
+ * missing. `recordFailedAttempt` (memory.ts) already writes stalled turns into
+ * OKF concept memory, but nothing ever fed that signal BACK into the skill
+ * files themselves — a skill that repeatedly leads an agent into the same dead
+ * end never learns to warn about it. This module closes that loop:
+ *
+ *   - `matchLessonToSkill` deterministically (no LLM) maps free text to the
+ *     bundled skill it most plausibly concerns, conservatively preferring
+ *     `undefined` over a wrong guess.
+ *   - `appendSkillLesson`/`recordSkillLesson` append a lesson bullet into a
+ *     `## Known Failure Modes` or `## Anti-Patterns (do NOT do)` section of the
+ *     project-level `.jeo/skills/<name>.md` file (seeded from the bundled
+ *     skill on first write), idempotent on exact-title re-append.
+ *   - `evalSkillLessons` is the self-eval half: ONE batched LLM call judges
+ *     whether each recorded lesson is still "covered" by the skill's current
+ *     guidance or has gone "stale" (guidance drifted, entry looks obsolete),
+ *     so a human/agent can prune or fold lessons back into the skill body.
+ */
+import * as fs from "node:fs/promises";
+import * as path from "node:path";
+import { SKILLS, getSkill, bundledSkillFileContent } from "../skills/catalog";
+import { callLlm } from "./loop";
+import { resolveVerifierModel } from "./prompt-router";
+import { readGlobalConfig } from "./state";
+import { tryExtractJsonObject } from "./json";
+
+export interface SkillLesson {
+  skill: string;
+  kind: "failure-mode" | "anti-pattern";
+  title: string;
+  detail: string;
+}
+
+const SECTION_HEADER: Record<SkillLesson["kind"], string> = {
+  "failure-mode": "## Known Failure Modes",
+  "anti-pattern": "## Anti-Patterns (do NOT do)",
+};
+
+/** Hand-picked keyword sets per bundled skill, used ONLY for the deterministic
+ *  free-text -> skill matcher below. Kept as one small record (not inline in
+ *  the function) so the mapping is easy to audit/extend as skills are added. */
+const SKILL_KEYWORDS: Record<string, string[]> = {
+  "deep-interview": ["ambiguity", "interview", "clarify", "clarification", "requirements", "socratic", "vague"],
+  "deep-dive": ["root cause", "root-cause", "diagnose", "diagnosis", "trace", "hypothesis", "causal"],
+  "ralplan": ["critic", "consensus", "[okay]", "[iterate]", "plan draft", "plan critique", "iterate verdict"],
+  "team": ["parallel_group", "worktree", "team executor", "plan step", "coordinated agents", "shared task list"],
+  "ultragoal": ["verification", "acceptance criteria", "goal verifier", "goal state", "verdict: met", "not_met"],
+};
+
+/**
+ * Deterministic (no LLM) free-text -> bundled-skill match. Scores each SKILLS
+ * entry by counting keyword/substring hits (the skill's own `name` plus its
+ * hand-picked related terms from {@link SKILL_KEYWORDS}) against the lowercased
+ * input text, and returns the highest-scoring skill name — or `undefined` when
+ * nothing scores at all, since a wrong guess is worse than no guess here.
+ */
+export function matchLessonToSkill(text: string): string | undefined {
+  const haystack = text.toLowerCase();
+  if (!haystack.trim()) return undefined;
+
+  let bestName: string | undefined;
+  let bestScore = 0;
+  for (const skill of SKILLS) {
+    const terms = [skill.name.toLowerCase(), ...(SKILL_KEYWORDS[skill.name] ?? [])];
+    let score = 0;
+    for (const term of terms) {
+      if (term && haystack.includes(term)) score++;
+    }
+    if (score > bestScore) {
+      bestScore = score;
+      bestName = skill.name;
+    }
+  }
+  return bestScore > 0 ? bestName : undefined;
+}
+
+/** Project-level skill file path — `<cwd>/.jeo/skills/<name>.md`, the same
+ *  highest-precedence directory `skillDirs()` (catalog.ts) resolves last, so a
+ *  lesson written here transparently overrides/extends the bundled version by
+ *  name the next time skills are loaded. */
+export function skillFilePath(cwd: string, skillName: string): string {
+  return path.join(cwd, ".jeo", "skills", `${skillName}.md`);
+}
+
+/** Locate the target H2 section within `content`'s lines. Returns the line
+ *  index range `[headerLine, bulletsEnd)` where `bulletsEnd` is the index to
+ *  insert a new bullet at (immediately before the next `## ` header, or
+ *  end-of-file, trimming any trailing blank lines inside the section so the
+ *  new bullet lands directly after the last existing one). `null` when the
+ *  section header isn't present yet. */
+function findSection(lines: string[], header: string): { headerLine: number; insertAt: number } | null {
+  const headerLine = lines.findIndex(l => l.trim() === header);
+  if (headerLine < 0) return null;
+  let end = lines.length;
+  for (let i = headerLine + 1; i < lines.length; i++) {
+    if (lines[i]!.startsWith("## ")) { end = i; break; }
+  }
+  let insertAt = end;
+  while (insertAt > headerLine + 1 && lines[insertAt - 1]!.trim() === "") insertAt--;
+  return { headerLine, insertAt };
+}
+
+/** Does a `- **title**: ...` (or bare `- title...`) bullet with this exact
+ *  title (case-insensitive) already exist between `[start, end)`? */
+function sectionHasTitle(lines: string[], start: number, end: number, title: string): boolean {
+  const needle = title.trim().toLowerCase();
+  for (let i = start; i < end; i++) {
+    const m = lines[i]!.match(/^-\s+\*\*(.+?)\*\*\s*:/);
+    if (m && m[1]!.trim().toLowerCase() === needle) return true;
+  }
+  return false;
+}
+
+/**
+ * Core compounding write: append `lesson` as a markdown bullet into the
+ * project-level skill file's failure-mode/anti-pattern section, seeding the
+ * file from the bundled skill's own on-disk text on first write so existing
+ * guidance is never lost. Idempotent on exact-title re-append (case-
+ * insensitive) — returns `{ appended: false, reason: "duplicate" }` rather
+ * than writing a second copy. Best-effort: any I/O error is caught and
+ * reported, never thrown (mirrors `recordFailedAttempt`'s fail-open contract).
+ */
+export async function appendSkillLesson(cwd: string, lesson: SkillLesson): Promise<{ appended: boolean; reason?: string }> {
+  const title = lesson.title.trim();
+  if (!title) return { appended: false, reason: "empty title" };
+  const detail = lesson.detail.trim();
+  const header = SECTION_HEADER[lesson.kind];
+
+  try {
+    const filePath = skillFilePath(cwd, lesson.skill);
+    let content: string;
+    try {
+      content = await fs.readFile(filePath, "utf-8");
+    } catch {
+      const bundled = getSkill(lesson.skill);
+      content = bundled ? bundledSkillFileContent(bundled) : `# ${lesson.skill}\n\n`;
+    }
+
+    let lines = content.replace(/\r\n/g, "\n").split("\n");
+    let section = findSection(lines, header);
+    if (!section) {
+      // Append the section header at the end of the file (blank-line separated).
+      if (lines.length > 0 && lines[lines.length - 1]!.trim() !== "") lines.push("");
+      if (!(lines.length > 0 && lines[lines.length - 1] === "")) lines.push("");
+      lines.push(header, "");
+      section = { headerLine: lines.length - 2, insertAt: lines.length };
+    } else if (sectionHasTitle(lines, section.headerLine + 1, section.insertAt, title)) {
+      return { appended: false, reason: "duplicate" };
+    }
+
+    const bullet = `- **${title}**: ${detail}`;
+    lines.splice(section.insertAt, 0, bullet);
+
+    const finalContent = lines.join("\n").replace(/\n{3,}$/, "\n\n");
+    await fs.mkdir(path.dirname(filePath), { recursive: true });
+    const tmpPath = `${filePath}.tmp-${process.pid}`;
+    await fs.writeFile(tmpPath, finalContent, "utf-8");
+    await fs.rename(tmpPath, filePath);
+    return { appended: true };
+  } catch (err) {
+    return { appended: false, reason: err instanceof Error ? err.message : String(err) };
+  }
+}
+
+/**
+ * Convenience wrapper for the automatic (stall-guard) path: matches `freeText`
+ * to a bundled skill via {@link matchLessonToSkill} and, only on a match,
+ * writes the lesson via {@link appendSkillLesson}. Unrelated text (no skill
+ * match) correctly writes nothing.
+ */
+export async function recordSkillLesson(
+  cwd: string,
+  freeText: string,
+  lesson: Omit<SkillLesson, "skill">,
+): Promise<{ appended: boolean; skill?: string; reason?: string }> {
+  const skill = matchLessonToSkill(freeText);
+  if (!skill) return { appended: false, reason: "no matching skill" };
+  const result = await appendSkillLesson(cwd, { ...lesson, skill });
+  return { ...result, skill };
+}
+
+export interface LessonEvalResult {
+  skill: string;
+  total: number;
+  covered: number;
+  stale: Array<{ title: string; reason: string }>;
+}
+
+interface ParsedBullet {
+  title: string;
+  detail: string;
+}
+
+/** Extract every bullet from `[start, end)`, splitting the `- **title**:
+ *  detail` format {@link appendSkillLesson} produces — a plain `- detail`
+ *  bullet (no bold title) tolerantly uses the whole bullet text as both. */
+function extractBullets(lines: string[], start: number, end: number): ParsedBullet[] {
+  const out: ParsedBullet[] = [];
+  for (let i = start; i < end; i++) {
+    const line = lines[i]!;
+    if (!line.startsWith("- ")) continue;
+    const m = line.match(/^-\s+\*\*(.+?)\*\*\s*:\s*(.*)$/);
+    if (m) {
+      out.push({ title: m[1]!.trim(), detail: m[2]!.trim() });
+    } else {
+      const text = line.slice(2).trim();
+      if (text) out.push({ title: text, detail: text });
+    }
+  }
+  return out;
+}
+
+/** The skill body with BOTH lesson sections (header + their bullets) removed —
+ *  i.e. just the skill's core guidance, which is what the eval judge should
+ *  compare each lesson against. */
+function coreBodyWithoutLessons(lines: string[]): string {
+  const headers = new Set(Object.values(SECTION_HEADER));
+  const kept: string[] = [];
+  let skipping = false;
+  for (const line of lines) {
+    if (headers.has(line.trim())) { skipping = true; continue; }
+    if (skipping && line.startsWith("## ")) skipping = false;
+    if (!skipping) kept.push(line);
+  }
+  return kept.join("\n").trim();
+}
+
+/**
+ * Self-eval half of the compounding loop: reads the project-level skill file
+ * and, for every recorded failure-mode/anti-pattern bullet, asks an
+ * independent LLM call whether the skill's CURRENT core guidance still
+ * addresses it ("covered") or has drifted ("stale"). Zero bullets short-
+ * circuits to a zero-cost result (no LLM call). `null` when the project-level
+ * file doesn't exist at all (no lessons ever recorded for this skill).
+ * Parse failure fails conservatively: every entry is treated as "covered"
+ * rather than risk falsely flagging a lesson as obsolete.
+ */
+export async function evalSkillLessons(
+  cwd: string,
+  skillName: string,
+  opts: { model?: string } = {},
+): Promise<LessonEvalResult | null> {
+  const filePath = skillFilePath(cwd, skillName);
+  let content: string;
+  try {
+    content = await fs.readFile(filePath, "utf-8");
+  } catch {
+    return null;
+  }
+
+  const lines = content.replace(/\r\n/g, "\n").split("\n");
+  const failureSection = findSection(lines, SECTION_HEADER["failure-mode"]);
+  const antiSection = findSection(lines, SECTION_HEADER["anti-pattern"]);
+  const bullets: ParsedBullet[] = [
+    ...(failureSection ? extractBullets(lines, failureSection.headerLine + 1, failureSection.insertAt) : []),
+    ...(antiSection ? extractBullets(lines, antiSection.headerLine + 1, antiSection.insertAt) : []),
+  ];
+
+  if (bullets.length === 0) {
+    return { skill: skillName, total: 0, covered: 0, stale: [] };
+  }
+
+  const model = opts.model ?? resolveVerifierModel(await readGlobalConfig());
+  const coreBody = coreBodyWithoutLessons(lines);
+  const lessonList = bullets.map((b, i) => `${i + 1}. "${b.title}": ${b.detail}`).join("\n");
+
+  const systemPrompt = `You are an independent Skill Auditor. You are shown a skill's CURRENT core guidance text and a list of lessons (failure modes / anti-patterns) previously recorded against earlier versions of this skill.
+
+For EACH lesson, judge:
+- "covered": the skill's current guidance still clearly addresses this lesson (a reader following the current guidance would avoid/handle it).
+- "stale": the guidance has drifted and no longer clearly addresses this lesson, OR the lesson itself looks obsolete/no-longer-applicable.
+
+Respond with ONLY a JSON object: {"results": [{"title": string, "verdict": "covered" | "stale", "reason": string}]}. Include EVERY lesson by its exact title. No markdown, no code fences, no other text.`;
+
+  const userMessage = `Skill's current core guidance:\n\n${coreBody}\n\nRecorded lessons:\n\n${lessonList}\n\nJudge each lesson.`;
+
+  const verdictByTitle = new Map<string, { verdict: "covered" | "stale"; reason: string }>();
+  try {
+    const response = await callLlm(
+      [
+        { role: "system", content: systemPrompt },
+        { role: "user", content: userMessage },
+      ],
+      { model, jsonMode: true, maxTokens: 1500, reasoningEffort: "none" },
+    );
+    const parsed = tryExtractJsonObject<{ results?: unknown }>(response);
+    if (parsed && Array.isArray(parsed.results)) {
+      for (const entry of parsed.results) {
+        if (
+          entry && typeof entry === "object" &&
+          "title" in entry && typeof entry.title === "string" &&
+          "verdict" in entry && (entry.verdict === "covered" || entry.verdict === "stale")
+        ) {
+          const reason = "reason" in entry && typeof entry.reason === "string" ? entry.reason : "";
+          verdictByTitle.set(entry.title.trim().toLowerCase(), { verdict: entry.verdict, reason });
+        }
+      }
+    }
+  } catch {
+    // fail conservatively below: an empty verdictByTitle treats every bullet as "covered"
+  }
+
+  let covered = 0;
+  const stale: Array<{ title: string; reason: string }> = [];
+  for (const b of bullets) {
+    const v = verdictByTitle.get(b.title.trim().toLowerCase());
+    if (v && v.verdict === "stale") {
+      stale.push({ title: b.title, reason: v.reason || "flagged stale by skill auditor" });
+    } else {
+      covered++;
+    }
+  }
+  return { skill: skillName, total: bullets.length, covered, stale };
+}

package/src/agent/subagents.ts

@@ -16,7 +16,7 @@ import architectPrompt from "../prompts/agents/architect.md" with { type: "text"
 import criticPrompt from "../prompts/agents/critic.md" with { type: "text" };
 import executorPrompt from "../prompts/agents/executor.md" with { type: "text" };
 import plannerPrompt from "../prompts/agents/planner.md" with { type: "text" };
-import type { PromptTier } from "./prompt-router";
+import { strongestCredentialed, cheapestCredentialed, type PromptTier } from "./prompt-router";
 
 const PROMPT_ROUTE_TIERS = ["trivial", "standard", "high", "complex"] as const satisfies readonly PromptTier[];
 
@@ -176,9 +176,6 @@ export function resolveSubagentModel(
   const entry = config.subagents?.[normalized];
   if (entry?.model) return entry.model;
 
-  // Exception: load strongestCredentialed dynamically using require to prevent eager loading and circular dependency issues that break test mocking.
-  // eslint-disable-next-line @typescript-eslint/no-require-imports
-  const { strongestCredentialed } = require("./prompt-router");
   if (normalized === "executor") {
     return config.roles?.xhigh || config.roles?.slow || strongestCredentialed(config) || config.defaultModel;
   }
@@ -189,11 +186,15 @@ export function resolveSubagentModel(
     return config.roles?.high || strongestCredentialed(config) || config.defaultModel;
   }
   if (normalized === "critic") {
+    // Grader role (plan reviewer): prefer an explicit mid/cheap pin, else the
+    // live-cheapest credentialed model — a GENERAL search (not 2 hardcoded ids),
+    // so critic tracks catalog drift the same way every other role's fallback
+    // does instead of silently collapsing to defaultModel when neither o3-mini
+    // nor gemini-2.5-flash happens to be credentialed.
     return (
       config.roles?.medium ||
       config.roles?.smol ||
-      strongestCredentialed(config, (m: CatalogModel) => m.canonical === "o3-mini") ||
-      strongestCredentialed(config, (m: CatalogModel) => m.canonical === "gemini-2.5-flash") ||
+      cheapestCredentialed(config) ||
       config.defaultModel
     );
   }