@ask-llm/plugin 0.15.0 → 0.16.3

package/README.md

@@ -61,6 +61,8 @@ pi list
 
 Pi discovers the portable skills as `/skill:<name>` commands and registers native `ask-codex`, `ask-gemini`, `ask-grok`, `ask-ollama`, `ask-antigravity`, model-neutral `ask-cursor-agent`, and bounded concurrent `ask-multi` tools. Pi intentionally has no built-in MCP client; do not configure these as MCP servers in Pi. `fable-review` and `grok-pair` are excluded from Pi discovery; Grok pairing currently has Claude/Cursor adapters, while Pi retains its dedicated Codex pairing lifecycle.
 
+The plugin's `ask-gemini-run`, Gemini reviewer agent/skill, and Pi `ask-gemini` tool all delegate to the canonical Gemini executor: `gemini-3.1-pro-preview` remains primary and quota errors fall back to `gemini-3.8-flash` unless `ASK_GEMINI_FALLBACK_MODEL` overrides it.
+
 For codex-pair, create `.codex-pair/context.md`, ensure Pi trusts the project, then run interactive `/codex-pair` to grant user-owned canonical-project consent. The marker alone never authorizes data transfer/cost. Revoke with `/codex-pair revoke`. Pi findings are non-blocking; blocking Stop-gate and one-shot print parity are not available.
 
 ```bash

package/agents/brainstorm-coordinator.md

@@ -231,7 +231,7 @@ node "${CLAUDE_PLUGIN_ROOT}/dist/grok-run.js" "$(cat "$workdir/prompt.md")" > "$
 pid_grok=$!
 
 # Only include this line if ollama was requested:
-ollama run "${ASK_OLLAMA_MODEL:-qwen3.6:27b}" < "$workdir/prompt.md" > "$workdir/ollama.out" 2> "$workdir/ollama.err" &
+ollama run "${ASK_OLLAMA_MODEL:-qwen3.8:27b}" < "$workdir/prompt.md" > "$workdir/ollama.out" 2> "$workdir/ollama.err" &
 pid_ollama=$!
 
 # Wait for each by PID so we capture per-provider exit codes independently.

package/agents/gemini-reviewer.md

@@ -14,7 +14,7 @@ tools:
 <!-- PORTABLE-CONTRACT:START -->
 ## Portable contract
 
-Review only the supplied changes and context. Ask Gemini for concrete correctness, security, and regression concerns; validate every candidate against source; require file/line evidence and reproduction for behavior claims; omit style-only or speculative findings; report provider failures explicitly.
+Review only the supplied changes and context. Ask Gemini for concrete correctness, security, and regression concerns; validate every candidate against source; require file/line evidence and reproduction for behavior claims; omit style-only or speculative findings; report provider failures explicitly. Preserve the canonical `gemini-3.1-pro-preview` → `gemini-3.8-flash` quota fallback and disclose when fallback occurs.
 <!-- PORTABLE-CONTRACT:END -->
 
 <!-- HOST-ADAPTER:CLAUDE-CODE:START -->

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ask-llm/plugin",
-  "version": "0.15.0",
+  "version": "0.16.3",
   "description": "Ask LLM review, comparison, brainstorming, image, verification, and pairing workflows for Claude Code, Cursor Agent, and Pi",
   "type": "module",
   "main": "dist/index.js",
@@ -83,12 +83,12 @@
     ]
   },
   "dependencies": {
-    "@ask-llm/antigravity-mcp": "^0.7.2",
-    "@ask-llm/codex-mcp": "^0.7.5",
-    "@ask-llm/gemini-mcp": "^1.7.2",
-    "@ask-llm/grok-mcp": "^0.1.2",
-    "@ask-llm/mcp": "^0.8.0",
-    "@ask-llm/ollama-mcp": "^0.5.7"
+    "@ask-llm/antigravity-mcp": "^0.7.5",
+    "@ask-llm/codex-mcp": "^0.7.8",
+    "@ask-llm/gemini-mcp": "^1.9.1",
+    "@ask-llm/grok-mcp": "^0.1.5",
+    "@ask-llm/mcp": "^0.9.3",
+    "@ask-llm/ollama-mcp": "^0.6.2"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": "*",
@@ -99,15 +99,15 @@
     "node": ">=20.0.0"
   },
   "devDependencies": {
-    "@biomejs/biome": "^2.4.4",
-    "@earendil-works/pi-ai": "^0.83.0",
-    "@earendil-works/pi-coding-agent": "^0.83.0",
-    "@types/node": "^22.19.13",
-    "typebox": "^1.3.7",
+    "@biomejs/biome": "^2.5.10",
+    "@earendil-works/pi-ai": "^0.84.2",
+    "@earendil-works/pi-coding-agent": "^0.84.2",
+    "@types/node": "^22.20.1",
+    "typebox": "1.3.7",
     "typescript": "^7.0.2",
-    "vitest": "^4.1.10"
+    "vitest": "^4.1.11"
   },
   "publishConfig": {
     "access": "public"
   }
-}
+}

package/pi/extensions/provider-tools.ts

@@ -203,7 +203,7 @@ export function registerProviderTools(pi: ExtensionAPI): void {
     name: "ask-gemini",
     label: "Ask Gemini",
     description:
-      "Consult Gemini through Ask LLM's canonical executor, including its quota fallback, validation, sessions, and structured response. Output is bounded to Pi's 50KB/2000-line limits.",
+      "Consult Gemini through Ask LLM's canonical executor (`gemini-3.1-pro-preview` → `gemini-3.8-flash` on quota), including validation, sessions, and structured response. Output is bounded to Pi's 50KB/2000-line limits.",
     parameters: geminiSchema,
     provider: "gemini",
   });

package/scripts/benchmark/README.md

@@ -0,0 +1,114 @@
+# codex-pair prompt A/B benchmark (ADR-100)
+
+Empirical harness for validating prompt-template changes. Built initially
+to test ADR-099 (Karpathy baseline principles), but reusable for any
+future prompt change that wants empirical justification before shipping.
+
+## What this measures
+
+For each fixture, the driver:
+
+1. Renders the review prompt twice — once against the **pre-baseline**
+   template (current production prompt without the Karpathy block) and
+   once against the **baseline** template (ADR-099's version with the
+   block).
+2. Invokes `codex exec --json` with each rendered prompt against the
+   same fixture file + context.
+3. Parses the verdict JSON and matches findings against the fixture's
+   `probes.json` ground truth.
+4. Aggregates **recall** (true positives caught / total ground-truth
+   probes) and **extra findings** (findings not matching any probe —
+   either real concerns the probes didn't anticipate, or noise).
+
+The headline output is a markdown report with a per-fixture caught /
+missed / extra breakdown and an aggregate recall delta.
+
+## Decision rule
+
+Ship the prompt change to `prompts/review.txt` permanently if:
+
+- **Recall delta ≥ +10 percentage points** (the baseline catches a
+  meaningfully larger fraction of ground-truth probes than the
+  pre-baseline prompt does)
+- **Extra-finding delta ≤ +1 per fixture on average** (the baseline
+  doesn't add overwhelming noise)
+
+If recall regresses or noise jumps, the ADR-099 rollback path applies
+(two file edits + one test-assertion update; documented in ADR-099's
+Consequences section).
+
+## Usage
+
+```bash
+# Full run (all fixtures, both arms)
+node packages/claude-plugin/scripts/benchmark/prompt-ab.mjs \
+  --model gpt-5.5 \
+  --out benchmark-report.md
+
+# Single-fixture quick check
+node packages/claude-plugin/scripts/benchmark/prompt-ab.mjs \
+  --fixtures 01-overcomplication \
+  --out single-report.md
+
+# Cheaper model for sanity
+node packages/claude-plugin/scripts/benchmark/prompt-ab.mjs \
+  --model gpt-5.4-mini \
+  --timeout-ms 60000
+```
+
+## Cost
+
+- Per fixture per arm: one `codex exec` review (~$0.04–0.07 at current
+  gpt-5.5 pricing)
+- Four fixtures × two arms = ~$0.40 per full run
+- Variance is high (codex sometimes returns longer responses); budget
+  $1 per careful run to absorb retries
+
+## Output interpretation
+
+The report has three layers:
+
+1. **Aggregate table** — recall percentages per arm, plus a one-line
+   recall-delta and extra-finding-delta verdict line that mechanically
+   checks against the decision rule.
+2. **Per-fixture sections** — for each fixture and each arm, the list
+   of probes caught (✅), missed (❌), and extra findings (⚠️ — these
+   require manual judgment whether they're real concerns or noise).
+3. **Token usage** — the per-fixture input/output token counts. Useful
+   for cost analysis if the prompt grows further.
+
+## Limitations
+
+- **Keyword-based probe matching is crude.** A real finding that uses
+  different vocabulary than the probe's `keywords` array will be
+  flagged as missed. Tune keywords as needed when you see this happen.
+  Future enhancement: LLM-judge mode that asks gemini "did this finding
+  match this probe?" — costs another ~$0.01 per evaluation but better
+  signal.
+- **Four fixtures is small.** Sample size is sufficient to see large
+  effects but won't reveal subtle improvements. Add more fixtures as
+  you encounter representative bug patterns in the wild — see
+  `fixtures/README.md` for the contribution shape.
+- **Codex is non-deterministic.** Run the benchmark twice and you'll
+  see modest variance per fixture. Wide swings (>30% recall delta on a
+  single fixture) suggest the fixture is borderline and should be
+  refined rather than accepted as signal.
+- **The fixtures are not the ADR-077 four-task benchmark.** ADR-077's
+  fixtures live on `experiment/codex-pair-poc` and exercise *bug
+  detection* (float-money precision etc.). These ADR-100 fixtures
+  exercise *prompt-rule detection* (does the Karpathy baseline change
+  what codex flags?). Different questions; both empirical.
+
+## Adding fixtures
+
+See `fixtures/README.md`. Three files per fixture: `code.ts`,
+`context.md`, `probes.json`. The driver auto-discovers any directory
+under `fixtures/` containing all three.
+
+## Related ADRs
+
+- **ADR-099** — codex-pair Karpathy baseline principles in review prompt
+- **ADR-077** — original codex-pair POC + four-task benchmark
+- **ADR-089** — externalized prompt template + golden fixture invariant
+- **ADR-100** — this benchmark harness (the document you should read for
+  methodology rationale)

package/scripts/benchmark/fixtures/README.md

@@ -0,0 +1,29 @@
+# Benchmark fixtures (ADR-100)
+
+Each fixture is a self-contained scenario crafted to exercise one of
+the Karpathy-baseline rules from ADR-099. Each directory contains:
+
+- **`code.ts`** — the file content sent to codex for review. Starts
+  with a comment block restating the task that was supposedly given to
+  Claude, so codex has the same context a real edit would have.
+- **`context.md`** — marker-file content sent as `projectContext` in
+  the rendered prompt.
+- **`probes.json`** — the ground-truth expectations: `should_flag`
+  entries are findings codex SHOULD raise (true positives); `should_not_flag`
+  entries are things codex should NOT raise (negative probes for noise).
+
+## Fixture index
+
+| Fixture | Tests | Why this exercises the rule |
+|---|---|---|
+| `01-overcomplication` | Simplicity | Implements `getFullName` via a single-use interface, factory class, runtime null checks against TypeScript types, and unused configurability params — every kind of "code beyond what was asked" |
+| `02-drive-by-refactor` | Surgical scope | Bug fix to `parsePagination` correctly applied, but the diff also includes an unrelated function rename, a reformatted constants block, and a history comment — all changes that should be in a separate PR |
+| `03-orphan-imports` | Surgical scope | Removes `formatV1` correctly, but leaves `decodeBase64` and `logger` imports orphaned (no longer used by anything in the file) |
+| `04-hidden-assumption` | Hidden assumptions | Implements `firstAvailableSlot` correctly for an already-sorted input, but the task asked for the "earliest" slot — depends on an unstated sort-order invariant. Should either sort internally or document the requirement. |
+
+## Adding a new fixture
+
+Create a new directory under `fixtures/` with the three files. The
+driver auto-discovers any directory containing all three. Naming
+convention: `NN-short-slug/` where `NN` is a two-digit sequence and
+`slug` describes the rule being exercised.

package/scripts/codex-pair-log.mjs

@@ -19,7 +19,7 @@ import { readFileSync } from "node:fs";
 import { access } from "node:fs/promises";
 import { homedir } from "node:os";
 import { dirname, join, resolve } from "node:path";
-import { CONTEXT_FILENAME, logPath as resolveLogPath, PAIR_ROOT_DIR } from "./lib/state.mjs";
+import { CONTEXT_FILENAME, PAIR_ROOT_DIR, logPath as resolveLogPath } from "./lib/state.mjs";
 
 const MARKER_FILE = join(PAIR_ROOT_DIR, CONTEXT_FILENAME);
 
@@ -171,17 +171,10 @@ function showSummary(entries) {
     }
     if (e.fellBack) fallbackCount++;
     if (e.verdict === "cached") cachedCount++;
-    if (
-      e.verdict === "none" ||
-      e.verdict === "concerns" ||
-      e.verdict === "cached"
-    ) {
+    if (e.verdict === "none" || e.verdict === "concerns" || e.verdict === "cached") {
       runCount++;
     }
-    if (
-      e.durationMs != null &&
-      (e.verdict === "none" || e.verdict === "concerns")
-    ) {
+    if (e.durationMs != null && (e.verdict === "none" || e.verdict === "concerns")) {
       durationSum += e.durationMs;
       durationCount++;
     }
@@ -240,9 +233,7 @@ async function main() {
   }
   const markerDir = await findMarkerUp(process.cwd());
   if (!markerDir) {
-    process.stderr.write(
-      `codex-pair-log: no .codex-pair/context.md marker found in cwd or parents\n`,
-    );
+    process.stderr.write(`codex-pair-log: no .codex-pair/context.md marker found in cwd or parents\n`);
     process.exit(1);
   }
   const logPath = resolveLogPath(markerDir);

package/scripts/codex-pair-prompt-drain.mjs

@@ -13,9 +13,9 @@
 import { access } from "node:fs/promises";
 import { homedir } from "node:os";
 import { dirname, join, resolve } from "node:path";
-import { CONTEXT_FILENAME, PAIR_ROOT_DIR } from "./lib/state.mjs";
 import { drainPending, joinPendingForSurface } from "./lib/debounce-state.mjs";
 import { collectSessionMarkers } from "./lib/session-registry.mjs";
+import { CONTEXT_FILENAME, PAIR_ROOT_DIR } from "./lib/state.mjs";
 
 const MARKER_FILE = join(PAIR_ROOT_DIR, CONTEXT_FILENAME);
 

package/scripts/codex-pair-session.mjs

@@ -13,16 +13,16 @@ import { homedir } from "node:os";
 import { dirname, join, resolve } from "node:path";
 import { bootstrapBroker, clearStaleBrokerState, teardownBroker } from "./lib/broker-lifecycle.mjs";
 import { clearAllDebounceState } from "./lib/debounce-state.mjs";
+import { clearSession } from "./lib/session-registry.mjs";
 import {
   appendLog,
-  clearAutoPause,
   CONTEXT_FILENAME,
+  clearAutoPause,
   PAIR_ROOT_DIR,
   readPauseInfo,
   readPluginVersion,
   resolveAutoResume,
 } from "./lib/state.mjs";
-import { clearSession } from "./lib/session-registry.mjs";
 
 const MARKER_FILE = join(PAIR_ROOT_DIR, CONTEXT_FILENAME);
 

package/scripts/codex-pair-stop-gate.mjs

@@ -7,17 +7,18 @@
 // findMarkerUp is duplicated by design (zero-workspace-imports; see prompt-drain).
 
 import { execFileSync } from "node:child_process";
-import { existsSync, readFileSync, readdirSync, realpathSync, statSync } from "node:fs";
+import { existsSync, readdirSync, readFileSync, realpathSync, statSync } from "node:fs";
 import { homedir } from "node:os";
 import { dirname, join, resolve } from "node:path";
 import { debounceRoot, drainPending, joinPendingForSurface, reviewingRoot } from "./lib/debounce-state.mjs";
+import { collectSessionMarkers } from "./lib/session-registry.mjs";
 import {
   CONTEXT_FILENAME,
-  INFLIGHT_TTL_MIN_MS,
-  PAIR_ROOT_DIR,
   contextPath,
+  INFLIGHT_TTL_MIN_MS,
   inflightRoot,
   logPath,
+  PAIR_ROOT_DIR,
   readAcks,
 } from "./lib/state.mjs";
 import {
@@ -28,7 +29,6 @@ import {
   parseGitPorcelain,
   selectLatestEntries,
 } from "./lib/stop-gate.mjs";
-import { collectSessionMarkers } from "./lib/session-registry.mjs";
 
 const MARKER_FILE = join(PAIR_ROOT_DIR, CONTEXT_FILENAME);
 
@@ -131,9 +131,7 @@ function readInFlightInputs(markerDir) {
 function inflightFreshMs(markerDir) {
   const fmTimeout = Number(readMarkerScalar(markerDir, "timeoutMs"));
   const timeout =
-    Number.isFinite(fmTimeout) && fmTimeout > 0
-      ? fmTimeout
-      : Number(process.env.ASK_CODEX_TIMEOUT_MS ?? 800_000);
+    Number.isFinite(fmTimeout) && fmTimeout > 0 ? fmTimeout : Number(process.env.ASK_CODEX_TIMEOUT_MS ?? 800_000);
   return Math.max(timeout, INFLIGHT_TTL_MIN_MS) + 60_000;
 }
 
@@ -266,6 +264,8 @@ async function main() {
 }
 
 main().catch((err) => {
-  process.stderr.write(`[codex-pair] WARNING: stop-gate failed (${err?.message ?? err}). Allowing turn end — HIGH findings may remain.\n`);
+  process.stderr.write(
+    `[codex-pair] WARNING: stop-gate failed (${err?.message ?? err}). Allowing turn end — HIGH findings may remain.\n`,
+  );
   process.exit(0);
 });

package/scripts/codex-pair-watch.mjs

@@ -17,12 +17,11 @@
 // invocation is inlined; semantics mirror `codexExecutor.ts` deliberately.
 
 import { spawn } from "node:child_process";
-import { access, readFile } from "node:fs/promises";
 import { readFileSync } from "node:fs";
-import { dirname, join, resolve } from "node:path";
+import { access, readFile } from "node:fs/promises";
 import { homedir } from "node:os";
+import { dirname, join, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
-import { IS_WINDOWS, terminateProcessTree } from "./lib/process.mjs";
 // M4: broker integration. Importing initializeBroker + isBrokerEnabled +
 // submitReview from broker.mjs transitively pulls in broker-transport,
 // broker-rpc, broker-lifecycle. ESM-static cost is paid on every hook
@@ -30,15 +29,14 @@ import { IS_WINDOWS, terminateProcessTree } from "./lib/process.mjs";
 // isn't set, so the per-edit fast path is unaffected.
 import { initializeBroker, isBrokerEnabled, readBrokerState, submitReview } from "./lib/broker.mjs";
 import {
-  DEFAULT_DEBOUNCE_MS,
-  DEFAULT_DEBOUNCE_MAX_MS,
   bumpEditRecord,
+  DEFAULT_DEBOUNCE_MAX_MS,
+  DEFAULT_DEBOUNCE_MS,
   drainPending,
   joinPendingForSurface,
   markReviewed,
   sweepStaleDebounce,
 } from "./lib/debounce-state.mjs";
-import { buildReviewPrompt } from "./lib/prompt.mjs";
 import {
   buildVerdictMessage,
   DEFAULT_SURFACE_THRESHOLD,
@@ -48,20 +46,23 @@ import {
   VALID_THRESHOLDS,
   VERDICT_PREFIXES,
 } from "./lib/parser.mjs";
+import { IS_WINDOWS, terminateProcessTree } from "./lib/process.mjs";
+import { buildReviewPrompt } from "./lib/prompt.mjs";
+import { registerMarker } from "./lib/session-registry.mjs";
 import {
-  appendLog,
   AUTOPAUSE_FAILURE_THRESHOLD,
+  appendLog,
+  CONTEXT_FILENAME,
   clearAutoPause,
   clearReviewFailures,
   computeCacheKey,
-  CONTEXT_FILENAME,
   contextPath,
   getBlockingFromShard,
   getCachedConcerns,
   hashConcernBody,
+  INFLIGHT_TTL_MIN_MS,
   ignorePath,
   includePath,
-  INFLIGHT_TTL_MIN_MS,
   logPath,
   PAIR_ROOT_DIR,
   readPauseInfo,
@@ -74,7 +75,6 @@ import {
   updateRepetitions,
   writeAutoPause,
 } from "./lib/state.mjs";
-import { registerMarker } from "./lib/session-registry.mjs";
 
 const SCRIPT_DIR = dirname(fileURLToPath(import.meta.url));
 const DEFAULTS_PATH = join(SCRIPT_DIR, "..", "codex-pair-defaults.json");
@@ -293,10 +293,7 @@ function parseFrontmatter(content) {
       valueRaw = valueRaw.slice(0, inlineComment.index);
     }
     let value = valueRaw.trim();
-    if (
-      (value.startsWith('"') && value.endsWith('"')) ||
-      (value.startsWith("'") && value.endsWith("'"))
-    ) {
+    if ((value.startsWith('"') && value.endsWith('"')) || (value.startsWith("'") && value.endsWith("'"))) {
       value = value.slice(1, -1);
     }
     if (value === "true") frontmatter[key] = true;
@@ -505,23 +502,13 @@ function resolveConfig(frontmatter) {
   return {
     model: typeof fm.model === "string" && fm.model.length > 0 ? fm.model : DEFAULT_MODEL,
     fallbackModel:
-      typeof fm.fallbackModel === "string" && fm.fallbackModel.length > 0
-        ? fm.fallbackModel
-        : FALLBACK_MODEL,
-    timeoutMs:
-      typeof fm.timeoutMs === "number" && fm.timeoutMs > 0 ? fm.timeoutMs : DEFAULT_TIMEOUT_MS,
-    maxFileBytes:
-      typeof fm.maxFileBytes === "number" && fm.maxFileBytes > 0
-        ? fm.maxFileBytes
-        : MAX_FILE_BYTES,
+      typeof fm.fallbackModel === "string" && fm.fallbackModel.length > 0 ? fm.fallbackModel : FALLBACK_MODEL,
+    timeoutMs: typeof fm.timeoutMs === "number" && fm.timeoutMs > 0 ? fm.timeoutMs : DEFAULT_TIMEOUT_MS,
+    maxFileBytes: typeof fm.maxFileBytes === "number" && fm.maxFileBytes > 0 ? fm.maxFileBytes : MAX_FILE_BYTES,
     surfaceThreshold:
-      surfaceCandidate && VALID_THRESHOLDS.has(surfaceCandidate)
-        ? surfaceCandidate
-        : DEFAULT_SURFACE_THRESHOLD,
-    debounceMs:
-      typeof fm.debounceMs === "number" && fm.debounceMs >= 0 ? fm.debounceMs : DEBOUNCE_MS,
-    debounceMaxMs:
-      typeof fm.debounceMaxMs === "number" && fm.debounceMaxMs > 0 ? fm.debounceMaxMs : DEBOUNCE_MAX_MS,
+      surfaceCandidate && VALID_THRESHOLDS.has(surfaceCandidate) ? surfaceCandidate : DEFAULT_SURFACE_THRESHOLD,
+    debounceMs: typeof fm.debounceMs === "number" && fm.debounceMs >= 0 ? fm.debounceMs : DEBOUNCE_MS,
+    debounceMaxMs: typeof fm.debounceMaxMs === "number" && fm.debounceMaxMs > 0 ? fm.debounceMaxMs : DEBOUNCE_MAX_MS,
   };
 }
 
@@ -729,9 +716,7 @@ function spawnCodex({ prompt, model, timeoutMs }) {
       setTimeout(() => {
         terminateProcessTree(child, "SIGKILL");
       }, 5000);
-      rejectCall(
-        taggedError(`codex exec timed out after ${Math.round(timeoutMs / 1000)}s`, "timeout"),
-      );
+      rejectCall(taggedError(`codex exec timed out after ${Math.round(timeoutMs / 1000)}s`, "timeout"));
     }, timeoutMs);
 
     child.on("error", (err) => {
@@ -757,9 +742,7 @@ function spawnCodex({ prompt, model, timeoutMs }) {
         // tail (often just the stdin banner) — #176.
         const errorEvents = extractJsonlErrorEvents(stdout);
         const reason =
-          errorEvents.length > 0
-            ? errorEvents[errorEvents.length - 1]
-            : stderrTail(stderr) || `codex exit ${code}`;
+          errorEvents.length > 0 ? errorEvents[errorEvents.length - 1] : stderrTail(stderr) || `codex exit ${code}`;
         rejectCall(taggedError(reason, "error"));
       }
     });
@@ -1209,9 +1192,7 @@ async function main() {
       verdict: "skipped",
       reason: `unreadable: ${err.message}`,
     });
-    await emitSystemMessage(
-      `codex-pair ${VERDICT_PREFIXES.skipped}: ${filePath} — unreadable (${err.message})`,
-    );
+    await emitSystemMessage(`codex-pair ${VERDICT_PREFIXES.skipped}: ${filePath} — unreadable (${err.message})`);
     process.exit(0);
   }
 

package/skills/gemini-review/SKILL.md

@@ -6,7 +6,7 @@ description: Get a second opinion from Gemini on your current code changes. Anal
 <!-- PORTABLE-CONTRACT:START -->
 ## Portable contract
 
-Gather the relevant staged, unstaged, and untracked code changes; build a bounded context brief; request a Gemini review; verify each reported finding against source; and return only prioritized, source-supported findings. Preserve provider fallback, timeout, and explicit failure disclosure.
+Gather the relevant staged, unstaged, and untracked code changes; build a bounded context brief; request a Gemini review; verify each reported finding against source; and return only prioritized, source-supported findings. Preserve the canonical `gemini-3.1-pro-preview` → `gemini-3.8-flash` quota fallback, timeout, and explicit failure disclosure.
 <!-- PORTABLE-CONTRACT:END -->
 
 ## Host adapters