instar 1.3.763 → 1.3.765

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "instar",
-  "version": "1.3.763",
+  "version": "1.3.765",
   "description": "Coherence infrastructure for self-evolving AI agents — on the Claude Code or Codex subscription you already have.",
   "type": "module",
   "main": "dist/index.js",

package/scripts/model-registry-freshness.manifest.json

@@ -91,6 +91,14 @@
       "file": "src/core/ModelTierEscalation.ts",
       "regex": "'claude-code':\\s*\\{\\s*default:\\s*'([^']+)',\\s*escalated:\\s*'([^']+)'",
       "note": "DEFAULT_TIER_ESCALATION.frameworks['claude-code'] default + escalated ids."
+    },
+    {
+      "id": "claude-clean-door-reviewer-default",
+      "door": "claude-code",
+      "tier": "capable",
+      "file": "src/core/crossModelReviewer.ts",
+      "regex": "CLAUDE_REVIEWER_DEFAULT_MODEL\\s*=\\s*'((?:claude|gpt|gemini)-[^']+)'",
+      "note": "CLAUDE_REVIEWER_DEFAULT_MODEL — the Anthropic clean-door reviewer pin (REVIEWER-DOOR-REWIRING §1.3). Extraction matches a frontier CONSTANT (not a capable:'…' tier decl), so the non-vacuity test (tests/unit/model-registry-freshness-reviewer-pin.test.ts) rots this constant and asserts the strict lint fails — closing the vacuous-tooth risk."
     }
   ],
   "$flaggedStaleNote": "Both prior flagged pins RESOLVED (operator-confirmed 2026-07-03): gemini capable swapped gemini-2.5-pro -> gemini-3.1-pro-preview (pin + topModels updated in the same change); codex capable stays gpt-5.5 (GA flagship — gpt-5.6-sol is preview-only/gov-gated/unreachable, NOT pinned). Nothing pending confirmation.",

package/skills/spec-converge/SKILL.md

@@ -115,12 +115,14 @@ The external "cross-model" pass is a set of independent non-Claude reads that si
      --spec <spec-path> --family gemini-cli \
      --context docs/foo.md --context docs/bar.md
    ```
-   Each emits a JSON `ReviewerResult` on stdout: `{ status, framework?, model?, verdict?, findings?, reason?, flag }`. Fold each `findings` into the round alongside the internal reviewers'. `status:'degraded'` (framework present but the call failed — timeout / error / rate-limited) is a *partial* cross-model pass for the round: fold in whatever came back and record the `degraded` flag — it does **not** collapse to `unavailable`. (Omitting `--family` keeps the old single-pass first-match behavior for existing callers.)
+   Each emits a JSON `ReviewerResult` on stdout: `{ status, framework?, model?, verdict?, findings?, reason?, flag, crossFamily }`. Fold each `findings` into the round alongside the internal reviewers'. `status:'degraded'` (framework present but the call failed — timeout / error / rate-limited) is a *partial* cross-model pass for the round: fold in whatever came back and record the `degraded` flag — it does **not** collapse to `unavailable`. (Omitting `--family` keeps the old single-pass first-match behavior for existing callers.)
+
+**Anthropic clean-door reviewer (`--family claude-code`) — a clean-door SECOND READ, NOT a cross-model opinion (REVIEWER-DOOR-REWIRING).** A third family reads the spec through the clean `claude -p` door on the strongest model (`claude-fable-5`), off the measured-penalized `opus × coding-harness` pair. It is **config-gated** (`specConverge.reviewers.anthropic.enabled` — omitted ⇒ live on a development agent, DARK on the fleet) and its result carries `crossFamily: false`, so it books into its OWN disclosure field **`clean-door-anthropic-review: claude-code:<model>`** — NEVER the `cross-model-review:` flag. A claude-only round therefore aggregates to `degraded-all-rounds`/`unavailable` exactly as before; the clean-door read can never launder a cross-model pass. Run it (when enabled) exactly like the others: `--family claude-code`. **D7 per-round-model disclosure:** the six INTERNAL reviewers run on the authoring session's model (opus, or `claude-fable-5` under tier-escalation — a quota-refused escalation silently drops them to opus-via-harness); record the ACTUAL per-round session/subagent model in the iteration log so a silent drop to opus is VISIBLE, not assumed.
 
 Two structural guards on the pass:
 
 - **Fail-loud model canary.** Model selection is dynamic (the `'capable'` tier resolved per framework — never a pinned name), and a tier word that falls through resolution (`'capable'` is not a model) is caught by `isConcreteReviewerModel` BEFORE the provider is invoked: the round degrades loudly with reason `model-resolution-canary` instead of silently selecting a dead reviewer.
-- **Trusted-provider allowlist.** `--family` only accepts frameworks on `TRUSTED_REVIEWER_FRAMEWORKS` (`codex-cli`, `gemini-cli` — first-party OAuth CLI adapters). The full spec text is handed to the reviewer, so it must NEVER be sent to a custom/base-URL endpoint; **pi-cli is deliberately excluded** from cross-model review for this reason (its provider may be a custom endpoint). An untrusted family is refused with `reason: 'untrusted-framework'`.
+- **Trusted-provider allowlist.** `--family` only accepts frameworks on `TRUSTED_REVIEWER_FRAMEWORKS` (`codex-cli`, `gemini-cli`, `claude-code` — first-party OAuth CLI adapters). The full spec text is handed to the reviewer, so it must NEVER be sent to a custom/base-URL endpoint; **pi-cli is deliberately excluded** from cross-model review for this reason (its provider may be a custom endpoint). An untrusted family is refused with `reason: 'untrusted-framework'`; a trusted-but-config-disabled family (the `claude-code` clean-door family on the fleet) is refused with `reason: 'no-supported-framework'` (trusted-egress ≠ enabled). The `claude-code` clean-door call is additionally hardened to codex-door parity for the untrusted spec text (empty allowed-tools, `--strict-mcp-config`, neutral scratch cwd, stdin prompt, env allowlist) — a spec under review is untrusted inbound text.
 
 The detection + invocation + parsing live in the unit-tested `src/core/crossModelReviewer.ts` module (built to `dist/core/crossModelReviewer.js`); the script is a thin file-I/O wrapper. codex and gemini are the supported frameworks in an extensible registry (`SUPPORTED_REVIEWER_FRAMEWORKS`) — further first-party CLIs plug in there with **no skill change**.
 

package/skills/spec-converge/scripts/cross-model-review.mjs

@@ -125,9 +125,33 @@ function readRepoFile(rel) {
   return fs.readFileSync(abs, 'utf-8');
 }
 
+/**
+ * Read the minimal reviewer config the module needs (REVIEWER-DOOR-REWIRING
+ * §1.5): the developmentAgent gate flag + the `specConverge.reviewers` block, so
+ * the Anthropic clean-door family resolves live-on-dev / dark-fleet. Best-effort:
+ * a missing/unparseable `.instar/config.json` ⇒ `{}` ⇒ fleet-dark (byte-identical
+ * `[codex, gemini]`). `.instar/config.json` is machine-local (no config
+ * replication in instar) — enabling the family is a per-machine edit.
+ */
+function loadReviewerConfig() {
+  try {
+    const cfgPath = path.join(ROOT, '.instar', 'config.json');
+    if (!fs.existsSync(cfgPath)) return {};
+    const cfg = JSON.parse(fs.readFileSync(cfgPath, 'utf-8'));
+    const out = {};
+    if (typeof cfg.developmentAgent === 'boolean') out.developmentAgent = cfg.developmentAgent;
+    if (cfg.specConverge && typeof cfg.specConverge === 'object') out.specConverge = cfg.specConverge;
+    return out;
+  } catch {
+    return {};
+  }
+}
+
 async function main() {
   const { spec, context, detectOnly, hashOnly, family, stateDir, timeoutMs } = parseArgs();
   const mod = await loadModule();
+  // Config-gate the Anthropic clean-door family (§1.5). Absent config ⇒ fleet-dark.
+  const reviewerConfig = loadReviewerConfig();
 
   // ── --hash-only: the delta-gating hash of the spec's reviewable body ──
   if (hashOnly) {
@@ -139,9 +163,9 @@ async function main() {
 
   // ── --detect-only: report ALL available families (Piece 3) ──
   if (detectOnly) {
-    const all = mod.detectAllCrossModelReviewers();
+    const all = mod.detectAllCrossModelReviewers({}, reviewerConfig);
     // Back-compat: keep the old single-framework fields (first-match shape).
-    const first = mod.detectCrossModelReviewer();
+    const first = mod.detectCrossModelReviewer({}, reviewerConfig);
     const report = {
       available: all.length > 0,
       frameworks: all,
@@ -152,10 +176,11 @@ async function main() {
     // Record the activation observation into the durable standing-framework
     // baseline when a state dir was provided. A record failure is surfaced in
     // the JSON (fail-loud), never silently swallowed — a missing baseline
-    // would quietly weaken the externals-mandatory check.
+    // would quietly weaken the externals-mandatory check. Iterate the ACTIVE
+    // set so a config-disabled claude family is not recorded as present.
     if (stateDir) {
       const frameworks = {};
-      for (const entry of mod.SUPPORTED_REVIEWER_FRAMEWORKS) {
+      for (const entry of mod.resolveActiveReviewerFrameworks(reviewerConfig)) {
         frameworks[entry.id] = all.some((d) => d.framework === entry.id);
       }
       try {
@@ -186,7 +211,10 @@ async function main() {
       );
       process.exit(0);
     }
-    familyEntry = mod.SUPPORTED_REVIEWER_FRAMEWORKS.find((f) => f.id === family) ?? null;
+    // Config gate (§1.5): a trusted-but-disabled family (the claude clean-door
+    // family on the fleet) is refused here — trusted-egress ≠ enabled.
+    const active = mod.resolveActiveReviewerFrameworks(reviewerConfig);
+    familyEntry = active.find((f) => f.id === family) ?? null;
     if (!familyEntry) {
       process.stdout.write(
         JSON.stringify({
@@ -199,7 +227,7 @@ async function main() {
     }
   }
 
-  const detection = familyEntry ? familyEntry.detect() : mod.detectCrossModelReviewer();
+  const detection = familyEntry ? familyEntry.detect() : mod.detectCrossModelReviewer({}, reviewerConfig);
 
   // Unavailable → print the unavailable flag, exit 0. Never block.
   if (!detection.available) {
@@ -227,9 +255,11 @@ async function main() {
         promptText: assembled.promptText,
         timeoutMs: Number.isFinite(timeoutMs) ? timeoutMs : mod.REVIEW_TIMEOUT_MS,
         detectionOverride: detection,
+        reviewerConfig,
       })
     : await mod.runCrossModelReview({
         assembled,
+        config: reviewerConfig,
         ...(Number.isFinite(timeoutMs) ? { timeoutMs } : {}),
       });