mergejudge-runner 0.2.0 → 0.2.1

package/README.md

@@ -10,7 +10,7 @@ Status: MVP. Design: [docs/RUNNERS_DESIGN.md](../../docs/RUNNERS_DESIGN.md).
 npx mergejudge-runner setup <code>   # code from the panel: Runners > Add runner
 ```
 
-Registers the runner, detects which of `claude` and `codex` are installed and signed in (`claude auth status`, `codex login status`; Codex counts only when signed in **with ChatGPT**, an API-key login is warned about and codex is not offered), asks yes/no about a GLM (z.ai) API key, and offers to start it at login (launchd / systemd user unit). Flags: `--yes`, `--no-autostart`, `--print-unit`, `--engine <claude-code|codex|glm>`, `--model [engine=]<model>` (repeatable), `--glm-key-env NAME`, `--no-glm`; `MERGEJUDGE_URL` overrides the server. See the [guide](https://mergejudge.si/guide/run-on-your-machine).
+Registers the runner, detects which of `claude` and `codex` are installed and signed in (`claude auth status`, `codex login status`; Codex counts only when signed in **with ChatGPT**, an API-key login is warned about and codex is not offered), asks yes/no about a GLM (z.ai) API key, and offers to start it at login (launchd / systemd user unit). Flags: `--yes`, `--no-autostart`, `--print-unit`, `--engine <claude-code|codex|glm>`, `--model [engine=]<model>` (repeatable), `--codex-effort <minimal|low|medium|high|xhigh>` (Codex reasoning effort, default `low`; stored as `engines.codex.effort` in the config), `--glm-key-env NAME`, `--no-glm`; `MERGEJUDGE_URL` overrides the server. See the [guide](https://mergejudge.si/guide/run-on-your-machine).
 
 ## Install (from source)
 
@@ -40,7 +40,7 @@ This stores the runner id and token in `$XDG_CONFIG_HOME/mergejudge-runner/confi
 | Engine | What runs | Billing |
 |---|---|---|
 | `claude-code` | `claude -p --tools ""` on your subscription | subscription |
-| `codex` | `codex exec --sandbox read-only --skip-git-repo-check --ephemeral --color never --ignore-user-config [-m MODEL] --output-schema <file> -o <file> -` (prompt on stdin, ChatGPT login) | subscription |
+| `codex` | `codex exec --sandbox read-only --skip-git-repo-check --ephemeral --color never --ignore-user-config [-m MODEL] -c model_reasoning_effort=EFFORT --output-schema <file> -o <file> -` (prompt on stdin, ChatGPT login) | subscription |
 | `glm` | `claude -p` with `ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic` and `ANTHROPIC_AUTH_TOKEN=<your key>` set only in the child env | **API key (billed per use)** |
 
 - The runner declares its engines, billing and models when it registers and on the runner heartbeat at every start. The config keeps `engine` (default), `engines: { name: { model, models } }` and, for GLM, `glmApiKey` (config is `0600`; `ZAI_API_KEY` in the environment works too). `models: ["*"]` offers any model.

package/dist/cli.js

@@ -27291,37 +27291,6 @@ function buildEngineSettings(enabled, defaultEngine, modelFlags) {
   return out;
 }
 
-// src/redact.ts
-var PATTERNS = [
-  /mjr_[ea]_[A-Za-z0-9_-]{6,}/g,
-  // MergeJudge runner / enrollment tokens
-  /sk-ant-[A-Za-z0-9_-]{6,}/g,
-  // Anthropic keys and OAuth tokens
-  /Bearer\s+[A-Za-z0-9._~+/=-]{8,}/gi,
-  /\b(ANTHROPIC_[A-Z_]+|MERGEJUDGE_[A-Z_]*(?:TOKEN|KEY|SECRET)[A-Z_]*|CLAUDE_CODE_OAUTH_TOKEN)=\S+/g,
-  /\bgh[pousr]_[A-Za-z0-9]{20,}/g,
-  /\bsk-(?:proj-)?[A-Za-z0-9_-]{20,}/g,
-  // OpenAI-style keys
-  /\b[0-9a-f]{32}\.[A-Za-z0-9]{8,}\b/g,
-  // z.ai API keys (<32 hex>.<secret>)
-  /\b(ZAI_[A-Z_]*|OPENAI_[A-Z_]*|CODEX_[A-Z_]*(?:KEY|TOKEN)[A-Z_]*)=\S+/g
-];
-var REDACTED = "[redacted]";
-function redact(text, secrets = []) {
-  let out = text;
-  for (const s of secrets) if (s && s.length >= 6) out = out.split(s).join(REDACTED);
-  for (const re of PATTERNS) out = out.replace(re, REDACTED);
-  return out;
-}
-var tokenPrefix = (t) => t ? `${t.slice(0, 10)}\u2026` : "(none)";
-function createLogger(write, secrets = () => [], now = () => /* @__PURE__ */ new Date()) {
-  const emit = (level, msg, f) => {
-    const kv = f ? Object.entries(f).filter(([, v]) => v !== void 0).map(([k, v]) => ` ${k}=${typeof v === "string" && /[\s"]/.test(v) ? JSON.stringify(v) : String(v)}`).join("") : "";
-    write(redact(`${now().toISOString()} ${level} ${msg}${kv}`, secrets()));
-  };
-  return { info: (m, f) => emit("INFO", m, f), warn: (m, f) => emit("WARN", m, f), error: (m, f) => emit("ERROR", m, f) };
-}
-
 // src/engine.ts
 import { spawn } from "node:child_process";
 import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
@@ -27407,6 +27376,13 @@ var safeModel = (model) => {
   if (!MODEL_RE.test(model)) throw new EngineError("spawn", "model name has unsupported characters");
   return model;
 };
+var CODEX_EFFORTS = ["minimal", "low", "medium", "high", "xhigh"];
+var DEFAULT_CODEX_EFFORT = "low";
+var isCodexEffort = (s) => CODEX_EFFORTS.includes(s);
+var safeEffort = (e) => {
+  if (!isCodexEffort(e)) throw new EngineError("spawn", `unsupported codex reasoning effort (use ${CODEX_EFFORTS.join(", ")})`);
+  return e;
+};
 function claudeArgs(model) {
   const m = safeModel(model);
   return [
@@ -27423,6 +27399,7 @@ function claudeArgs(model) {
 }
 function codexArgs(o) {
   const m = safeModel(o.model);
+  const e = o.effort ? safeEffort(o.effort) : null;
   return [
     "exec",
     "--sandbox",
@@ -27433,6 +27410,7 @@ function codexArgs(o) {
     "never",
     "--ignore-user-config",
     ...m ? ["-m", m] : [],
+    ...e ? ["-c", `model_reasoning_effort=${e}`] : [],
     "--output-schema",
     o.schemaFile,
     "-o",
@@ -27606,7 +27584,7 @@ async function runCodex(prompt, signal, o = {}) {
   try {
     const schemaFile = join2(cwd, "schema.json"), outFile = join2(cwd, "last-message.txt");
     await writeFile(schemaFile, JSON.stringify(CODEX_OUTPUT_SCHEMA), { mode: 384 });
-    const args = codexArgs({ model: o.model, schemaFile, outFile });
+    const args = codexArgs({ model: o.model, effort: o.effort ?? DEFAULT_CODEX_EFFORT, schemaFile, outFile });
     await spawnCli({ bin: o.bin ?? "codex", args, env: engineEnv(o.env, "codex"), cwd, stdin: prompt, signal, killGraceMs: o.killGraceMs });
     let last;
     try {
@@ -27654,6 +27632,37 @@ ${req.user}`;
   }
 };
 
+// src/redact.ts
+var PATTERNS = [
+  /mjr_[ea]_[A-Za-z0-9_-]{6,}/g,
+  // MergeJudge runner / enrollment tokens
+  /sk-ant-[A-Za-z0-9_-]{6,}/g,
+  // Anthropic keys and OAuth tokens
+  /Bearer\s+[A-Za-z0-9._~+/=-]{8,}/gi,
+  /\b(ANTHROPIC_[A-Z_]+|MERGEJUDGE_[A-Z_]*(?:TOKEN|KEY|SECRET)[A-Z_]*|CLAUDE_CODE_OAUTH_TOKEN)=\S+/g,
+  /\bgh[pousr]_[A-Za-z0-9]{20,}/g,
+  /\bsk-(?:proj-)?[A-Za-z0-9_-]{20,}/g,
+  // OpenAI-style keys
+  /\b[0-9a-f]{32}\.[A-Za-z0-9]{8,}\b/g,
+  // z.ai API keys (<32 hex>.<secret>)
+  /\b(ZAI_[A-Z_]*|OPENAI_[A-Z_]*|CODEX_[A-Z_]*(?:KEY|TOKEN)[A-Z_]*)=\S+/g
+];
+var REDACTED = "[redacted]";
+function redact(text, secrets = []) {
+  let out = text;
+  for (const s of secrets) if (s && s.length >= 6) out = out.split(s).join(REDACTED);
+  for (const re of PATTERNS) out = out.replace(re, REDACTED);
+  return out;
+}
+var tokenPrefix = (t) => t ? `${t.slice(0, 10)}\u2026` : "(none)";
+function createLogger(write, secrets = () => [], now = () => /* @__PURE__ */ new Date()) {
+  const emit = (level, msg, f) => {
+    const kv = f ? Object.entries(f).filter(([, v]) => v !== void 0).map(([k, v]) => ` ${k}=${typeof v === "string" && /[\s"]/.test(v) ? JSON.stringify(v) : String(v)}`).join("") : "";
+    write(redact(`${now().toISOString()} ${level} ${msg}${kv}`, secrets()));
+  };
+  return { info: (m, f) => emit("INFO", m, f), warn: (m, f) => emit("WARN", m, f), error: (m, f) => emit("ERROR", m, f) };
+}
+
 // ../../node_modules/balanced-match/dist/esm/index.js
 var balanced = (a, b, str2) => {
   const ma = a instanceof RegExp ? maybeMatch(a, str2) : a;
@@ -30212,6 +30221,31 @@ function chunkById(snap, id) {
 var embeddingText = (c) => `${c.path}
 ${c.text}`.slice(0, 1500);
 
+// ../../src/budget.ts
+var DEFAULT_OUTPUT_TOKENS = 2500;
+var REASONING_OUTPUT_TOKENS = 8e3;
+var REASONING_MODEL = /(?:^|[/:@])(?:glm-?(?:5|4\.[5-9])|gpt-5|o[1-9](?:-|$)|deepseek-(?:r1|reasoner)|qwq|[^/]*thinking)/i;
+var isReasoningModel = (model) => !!model && REASONING_MODEL.test(model);
+var MIN_OUTPUT = 256;
+var MIN_INPUT = 1e3;
+function resolveOutputLimit(cfg) {
+  const total = cfg.budget.max_tokens_per_pr;
+  const explicit = cfg.budget.max_output_tokens;
+  if (explicit !== void 0) return Math.max(MIN_OUTPUT, Math.min(explicit, total - MIN_INPUT));
+  const want = isReasoningModel(cfg.provider.model) ? REASONING_OUTPUT_TOKENS : DEFAULT_OUTPUT_TOKENS;
+  return Math.max(MIN_OUTPUT, Math.min(want, Math.floor(total / 2)));
+}
+function retryOutputLimit(limit, room) {
+  return Math.max(limit, Math.min(limit * 2, room));
+}
+var COMPACT_JSON_INSTRUCTION = "Respond with compact JSON only, no reasoning text, keep findings short.";
+
+// ../../src/llm/provider.ts
+async function completeWithMeta(p, req) {
+  return p.completeEx ? p.completeEx(req) : { text: await p.complete(req), truncated: false };
+}
+var canUseTools = (p) => !!p && p.supportsTools === true && typeof p.chat === "function";
+
 // ../../src/rubrics.ts
 var import_yaml = __toESM(require_dist());
 var criterionSchema = external_exports.object({
@@ -30384,7 +30418,7 @@ function buildUserPrompt(input2, cfg, system = SYSTEM_PROMPT) {
   const ctxText = renderContext(input2.context ?? []);
   const budgetChars = Math.max(
     0,
-    cfg.budget.max_tokens_per_pr - cfg.budget.max_output_tokens - estimateTokens(system) - 100 - estimateTokens(ctxText)
+    cfg.budget.max_tokens_per_pr - resolveOutputLimit(cfg) - estimateTokens(system) - 100 - estimateTokens(ctxText)
   ) * 4;
   const title = truncate(sanitizeUntrusted(input2.title), 300);
   const body = truncate(sanitizeUntrusted(input2.body ?? ""), Math.min(2e3, Math.floor(budgetChars * 0.15)));
@@ -30458,14 +30492,68 @@ function parseTriageOutput(raw) {
   const a = t.indexOf("{");
   const b = t.lastIndexOf("}");
   if (a < 0 || b <= a) throw new Error("no JSON object in LLM output");
-  let json2 = t.slice(a, b + 1);
+  const json2 = t.slice(a, b + 1);
   try {
     return triageOutputSchema.parse(JSON.parse(json2));
-  } catch (e) {
-    json2 = json2.replace(/,\s*([}\]])/g, "$1");
-    return triageOutputSchema.parse(JSON.parse(json2));
+  } catch (first) {
+    const repaired = repairJson(t.slice(a));
+    if (repaired !== void 0) {
+      try {
+        return triageOutputSchema.parse(JSON.parse(repaired));
+      } catch {
+      }
+    }
+    throw first;
   }
 }
+function repairJson(s) {
+  let out = "";
+  const stack = [];
+  let inStr = false;
+  for (let i = 0; i < s.length; i++) {
+    const c = s[i];
+    if (inStr) {
+      if (c === "\\") {
+        const n = s[i + 1];
+        if (n !== void 0 && /["\\/bfnrtu]/.test(n)) {
+          out += c + n;
+          i++;
+        } else out += "\\\\";
+      } else if (c === '"') {
+        const rest = s.slice(i + 1).match(/^\s*(.)?/)?.[1];
+        const closes = rest === void 0 || rest === "," || rest === "}" || rest === "]" || rest === ":";
+        if (closes) {
+          out += c;
+          inStr = false;
+        } else out += '\\"';
+      } else if (c === "\n") out += "\\n";
+      else if (c === "\r") out += "\\r";
+      else if (c === "	") out += "\\t";
+      else if (c < " ") out += " ";
+      else out += c;
+      continue;
+    }
+    if (c === '"') {
+      inStr = true;
+      out += c;
+    } else if (c === "{" || c === "[") {
+      stack.push(c === "{" ? "}" : "]");
+      out += c;
+    } else if (c === "}" || c === "]") {
+      out = out.replace(/,\s*$/, "");
+      if (stack[stack.length - 1] === c) {
+        stack.pop();
+        out += c;
+        if (!stack.length) return out;
+      }
+    } else out += c;
+  }
+  if (!stack.length) return out;
+  if (inStr) out += '"';
+  out = out.replace(/,\s*$/, "").replace(/:\s*$/, ": null");
+  out = out.replace(/,\s*"(?:[^"\\]|\\.)*"\s*$/, "");
+  return out + stack.reverse().join("");
+}
 var MAX_FINDINGS = 20;
 function validateAgainstFiles(out, files, criteria = []) {
   const known = new Set(files.map((f) => f.path));
@@ -30515,8 +30603,15 @@ function tokenBreakdown(opts) {
     skipped_diff_files: opts.skipped
   };
 }
+function outputLimitWarning(o) {
+  return `LLM output limit: ${o.ref ?? "?"} model=${o.model ?? "?"} limit=${o.limit} output_tokens=${o.outputTokens} truncated=${o.truncated} retry=${o.retry}${o.retryLimit ? ` (limit ${o.retryLimit}, compact JSON)` : ""}`;
+}
+function retryInstruction(truncated, why) {
+  const reason = truncated ? "Your previous reply was cut off at the output limit" : `Your previous reply was not valid JSON (${why})`;
+  return `REPAIR: ${reason}. ${COMPACT_JSON_INSTRUCTION} Reply with ONLY a single JSON object matching the schema. No markdown fences, no prose.`;
+}
 async function runTriage(provider, input2, cfg, ctx = {}, log = () => {
-}) {
+}, warn = log) {
   const system = buildSystemPrompt(ctx) + (input2.context?.length ? `
 
 ${CONTEXT_RULES}` : "");
@@ -30526,6 +30621,9 @@ ${CONTEXT_RULES}` : "");
   let raw = "";
   let retried = false;
   let lastErr = "unknown parse error";
+  const limit = resolveOutputLimit(cfg);
+  let retryLimit = limit;
+  let firstTruncated = false, firstOut = 0, lastTruncated = false;
   const mkBreakdown = (outTok) => ({
     ...tokenBreakdown({ systemBase: system, policyText: ctx.policyText, criteria: ctx.criteria, diffTokens: diffOnly, outputTokens: outTok, skipped }),
     context: contextTokens,
@@ -30534,20 +30632,38 @@ ${CONTEXT_RULES}` : "");
   try {
     for (let attempt = 0; attempt < 2; attempt++) {
       const repair = attempt === 1;
-      if (repair) retried = true;
+      if (repair) {
+        retried = true;
+        retryLimit = retryOutputLimit(limit, cfg.budget.max_tokens_per_pr - estimateTokens(system) - estimateTokens(prompt) - 100);
+      }
       try {
-        raw = await provider.complete({
+        const r = await completeWithMeta(provider, {
           system,
           user: repair ? `${prompt}
 
-REPAIR: Your previous reply was not valid JSON (${lastErr}). Reply with ONLY a single JSON object matching the schema. No markdown fences, no prose.` : prompt,
-          maxOutputTokens: cfg.budget.max_output_tokens,
+${retryInstruction(firstTruncated, lastErr)}` : prompt,
+          maxOutputTokens: repair ? retryLimit : limit,
           jsonMode: true
         });
-        const llm = validateAgainstFiles(parseTriageOutput(raw), input2.files, ctx.criteria ?? []);
+        raw = r.text;
+        lastTruncated = r.truncated;
+        if (!repair) {
+          firstTruncated = r.truncated;
+          firstOut = r.outputTokens ?? estimateTokens(raw);
+        }
+        let llm;
+        try {
+          llm = validateAgainstFiles(parseTriageOutput(raw), input2.files, ctx.criteria ?? []);
+        } catch (pe) {
+          if (!r.truncated) throw pe;
+          throw new Error(`output truncated at ${repair ? retryLimit : limit} tokens (${pe.message})`);
+        }
         const outTok = estimateTokens(raw);
-        if (retried) log("LLM JSON parse succeeded on repair attempt.");
-        return { llm, tokensEstimate: estimateTokens(system) + diffTokens + outTok, breakdown: mkBreakdown(outTok), retried };
+        if (retried) {
+          log("LLM JSON parse succeeded on repair attempt.");
+          warn(outputLimitWarning({ ref: ctx.ref, model: cfg.provider.model, limit, retryLimit, outputTokens: firstOut, truncated: firstTruncated, retry: "succeeded" }));
+        }
+        return { llm, tokensEstimate: estimateTokens(system) + diffTokens + outTok, breakdown: mkBreakdown(outTok), retried, ...firstTruncated ? { truncated: true } : {}, outputLimit: limit };
       } catch (e) {
         lastErr = e.message.slice(0, 160);
         log(`LLM JSON attempt ${attempt + 1} failed: ${lastErr}`);
@@ -30558,12 +30674,15 @@ REPAIR: Your previous reply was not valid JSON (${lastErr}). Reply with ONLY a s
   } catch (e) {
     const outTok = estimateTokens(raw);
     log(`LLM triage falling back to rules-only after JSON failure: ${e.message.slice(0, 160)}`);
+    if (retried) warn(outputLimitWarning({ ref: ctx.ref, model: cfg.provider.model, limit, retryLimit, outputTokens: firstOut, truncated: firstTruncated || lastTruncated, retry: "failed" }));
     return {
       llm: null,
       error: e.message.slice(0, 200),
       tokensEstimate: estimateTokens(system) + diffTokens + outTok,
       breakdown: mkBreakdown(outTok),
-      retried
+      retried,
+      ...firstTruncated || lastTruncated ? { truncated: true } : {},
+      outputLimit: limit
     };
   }
 }
@@ -30652,7 +30771,7 @@ function deepRules(cfg, tools) {
 }
 var turnTokens = (m) => estimateTokens(m.content + (m.role === "assistant" && m.toolCalls ? JSON.stringify(m.toolCalls) : ""));
 async function runDeepTriage(provider, input2, cfg, ctx = {}, opts, log = () => {
-}) {
+}, warn = log) {
   const now = opts.now ?? Date.now;
   const started = now();
   const system = buildSystemPrompt(ctx) + `
@@ -30668,19 +30787,26 @@ ${deepRules(cfg, opts.tools)}`;
   let stopped;
   let retried = false;
   let raw = "";
-  const turn = async (tools, jsonMode) => {
+  const limit = resolveOutputLimit(cfg);
+  let retryLimit = limit;
+  let lastTruncated = false, lastOut = 0, firstTruncated = false, firstOut = 0;
+  const turn = async (tools, jsonMode, maxOutputTokens = limit) => {
     const inEst = sysTok + messages.reduce((n, m) => n + turnTokens(m), 0);
-    const t = await provider.chat({ system, messages, tools, maxOutputTokens: cfg.budget.max_output_tokens, jsonMode });
+    const t = await provider.chat({ system, messages, tools, maxOutputTokens, jsonMode });
     modelCalls++;
     const o = t.usage?.output ?? estimateTokens(t.text + JSON.stringify(t.toolCalls));
     used += (t.usage?.input ?? inEst) + o;
     outTok += o;
+    lastTruncated = !!t.truncated;
+    lastOut = o;
     return t;
   };
   const result = (llm, error62) => ({
     llm,
     ...error62 ? { error: error62 } : {},
     retried,
+    outputLimit: limit,
+    ...firstTruncated || lastTruncated ? { truncated: true } : {},
     tokensEstimate: used,
     breakdown: {
       ...tokenBreakdown({ systemBase: system, policyText: ctx.policyText, criteria: ctx.criteria, diffTokens, outputTokens: outTok, skipped }),
@@ -30697,7 +30823,7 @@ ${deepRules(cfg, opts.tools)}`;
       const convTok = sysTok + messages.reduce((n, m) => n + turnTokens(m), 0);
       if (steps >= cfg.deep.max_steps) stopped = "steps";
       else if (now() - started >= cfg.deep.max_seconds * 1e3) stopped = "time";
-      else if (used + convTok + cfg.budget.max_output_tokens + headroom > cfg.deep.max_tokens) stopped = "tokens";
+      else if (used + convTok + limit + headroom > cfg.deep.max_tokens) stopped = "tokens";
       if (stopped) break;
       const t = await turn(opts.tools.defs, false);
       if (!t.toolCalls.length) {
@@ -30732,13 +30858,20 @@ ${deepRules(cfg, opts.tools)}`;
       const why = e.message.slice(0, 160);
       log(`LLM JSON attempt 1 failed: ${why}`);
       retried = true;
+      firstTruncated = lastTruncated;
+      firstOut = lastOut;
+      const convNow = sysTok + messages.reduce((n, m) => n + turnTokens(m), 0);
+      retryLimit = retryOutputLimit(limit, cfg.deep.max_tokens - used - convNow);
       messages.push({ role: "assistant", content: raw });
-      messages.push({ role: "user", content: `REPAIR: Your previous reply was not valid JSON (${why}). Reply with ONLY a single JSON object matching the schema. No markdown fences, no prose.` });
-      raw = (await turn(void 0, true)).text;
-      return result(validateAgainstFiles(parseTriageOutput(raw), input2.files, ctx.criteria ?? []));
+      messages.push({ role: "user", content: retryInstruction(firstTruncated, why) });
+      raw = (await turn(void 0, true, retryLimit)).text;
+      const llm = validateAgainstFiles(parseTriageOutput(raw), input2.files, ctx.criteria ?? []);
+      warn(outputLimitWarning({ ref: ctx.ref, model: cfg.provider.model, limit, retryLimit, outputTokens: firstOut, truncated: firstTruncated, retry: "succeeded" }));
+      return result(llm);
     }
   } catch (e) {
     log(`LLM triage falling back to rules-only after deep-mode failure: ${e.message.slice(0, 160)}`);
+    if (retried) warn(outputLimitWarning({ ref: ctx.ref, model: cfg.provider.model, limit, retryLimit, outputTokens: firstOut, truncated: firstTruncated || lastTruncated, retry: "failed" }));
     return result(null, e.message.slice(0, 200));
   }
 }
@@ -30847,9 +30980,6 @@ function semanticSearch(e, base, snap, changedPaths) {
   };
 }
 
-// ../../src/llm/provider.ts
-var canUseTools = (p) => !!p && p.supportsTools === true && typeof p.chat === "function";
-
 // ../../src/policy.ts
 var POLICY_MAX_EACH = 8 * 1024;
 var POLICY_MAX_TOTAL = 8 * 1024;
@@ -30968,8 +31098,8 @@ async function produceReviewResult(o) {
       context = g.pieces;
     }
     const prompt = { title: pr.title ?? "", body: pr.body ?? "", files: kept, context };
-    const tctx = { policyText: policies.text, criteria, skippedDiffFiles: skipped };
-    const res = mode === "deep" && tools && canUseTools(provider) ? await runDeepTriage(provider, prompt, cfg, tctx, { tools }, log.info) : await runTriage(provider, prompt, cfg, tctx, log.info);
+    const tctx = { policyText: policies.text, criteria, skippedDiffFiles: skipped, ref: o.ref ?? `#${pr.number}` };
+    const res = mode === "deep" && tools && canUseTools(provider) ? await runDeepTriage(provider, prompt, cfg, tctx, { tools }, log.info, log.warning) : await runTriage(provider, prompt, cfg, tctx, log.info, log.warning);
     usedMode = mode;
     tokens = res.tokensEstimate;
     llm = res.llm;
@@ -30977,7 +31107,7 @@ async function produceReviewResult(o) {
     log.info(
       `Tokens ~${res.breakdown.total} (system ${res.breakdown.system} \xB7 policy ${res.breakdown.policy} \xB7 rubric ${res.breakdown.rubric} \xB7 diff ${res.breakdown.diff}${res.breakdown.context ? ` \xB7 context ${res.breakdown.context}` : ""}${res.breakdown.tools ? ` \xB7 tools ${res.breakdown.tools} (${res.breakdown.tool_calls} calls)` : ""} \xB7 out ${res.breakdown.output}) \xB7 mode ${mode}` + (res.retried ? " \xB7 JSON repair retry used" : "")
     );
-    if (!llm) notes.push(`LLM triage failed (${res.error}); rules-only.`);
+    if (!llm) notes.push(`LLM triage failed (${res.error}${res.truncated ? `; output cut off at the ${res.outputLimit}-token limit, retry failed` : ""}); rules-only.`);
     else if (o.llmCache) {
       await o.llmCache.put({ llm, tokens, breakdown: res.breakdown, policiesUsed: policies.used }).catch((e) => log.warning(`Could not cache LLM result: ${e.message}`));
     }
@@ -31232,7 +31362,7 @@ function installAutostart(d, bundle) {
 }
 
 // src/version.ts
-var VERSION = "0.2.0";
+var VERSION = "0.2.1";
 
 // src/run.ts
 var MAX_RESULT_BYTES = 512 * 1024;
@@ -31321,7 +31451,7 @@ async function runLoop(o) {
   const caps = runnerCaps(cfg, engEnv, new Set(Object.keys(blocked)));
   const engineOptions = (engine) => {
     const base = { env: engEnv, killGraceMs: o.claude?.killGraceMs };
-    if (engine === "codex") return { ...base, bin: o.codex?.bin ?? cfg.codexBin ?? void 0, ...o.codex?.killGraceMs !== void 0 ? { killGraceMs: o.codex.killGraceMs } : {} };
+    if (engine === "codex") return { ...base, effort: cfg.engines?.codex?.effort ?? void 0, bin: o.codex?.bin ?? cfg.codexBin ?? void 0, ...o.codex?.killGraceMs !== void 0 ? { killGraceMs: o.codex.killGraceMs } : {} };
     return { ...base, bin: o.claude?.bin ?? cfg.claudeBin ?? void 0, ...engine === "glm" ? { apiKey: glmKeyOf(cfg, engEnv) } : {} };
   };
   let nextAnnounce = 0;
@@ -31549,8 +31679,8 @@ async function runLoop(o) {
 var USAGE = `mergejudge-runner ${VERSION}
 
 Usage:
-  mergejudge-runner register --code <enrollment-code> --url <https://app.mergejudge.si> [--name <name>] [--engine claude-code|codex|glm] [--model [engine=]<model>]... [--glm-key-env NAME] [--claude-bin <path>] [--codex-bin <path>] [--concurrency 1] [--force]
-  mergejudge-runner setup [code] [--yes] [--no-autostart] [--print-unit] [--name <name>] [--url <url>] [--engine <name>] [--model [engine=]<model>]... [--glm-key-env NAME | --no-glm] [--force]
+  mergejudge-runner register --code <enrollment-code> --url <https://app.mergejudge.si> [--name <name>] [--engine claude-code|codex|glm] [--model [engine=]<model>]... [--codex-effort minimal|low|medium|high|xhigh] [--glm-key-env NAME] [--claude-bin <path>] [--codex-bin <path>] [--concurrency 1] [--force]
+  mergejudge-runner setup [code] [--yes] [--no-autostart] [--print-unit] [--name <name>] [--url <url>] [--engine <name>] [--model [engine=]<model>]... [--codex-effort minimal|low|medium|high|xhigh] [--glm-key-env NAME | --no-glm] [--force]
   mergejudge-runner run [--concurrency 1] [--once] [--claude-bin <path>] [--codex-bin <path>] [--job-timeout-min 10]
   mergejudge-runner status
 
@@ -31559,8 +31689,8 @@ setup registers this machine (code from the panel, or prompted), detects which o
 Common: --config <path> (default: $XDG_CONFIG_HOME or ~/.config /mergejudge-runner/config.json)
 `;
 var COMMAND_USAGE = {
-  register: "mergejudge-runner register --code <enrollment-code> --url <https://app.mergejudge.si> [--name <name>] [--engine claude-code|codex|glm] [--model [engine=]<model>]... [--glm-key-env NAME] [--claude-bin <path>] [--codex-bin <path>] [--concurrency 1] [--force]",
-  setup: "mergejudge-runner setup [code] [--yes] [--no-autostart] [--print-unit] [--name <name>] [--url <url>] [--engine <name>] [--model [engine=]<model>]... [--glm-key-env NAME | --no-glm] [--force]\n\nRegisters this machine (code from the panel, or prompted), detects which of claude and codex are installed and signed in, asks whether to add a GLM (z.ai, billed per use) API key (--glm-key-env NAME reads it from that environment variable; --no-glm skips the question), and offers to start the runner at login. MERGEJUDGE_URL overrides the default server.",
+  register: "mergejudge-runner register --code <enrollment-code> --url <https://app.mergejudge.si> [--name <name>] [--engine claude-code|codex|glm] [--model [engine=]<model>]... [--codex-effort minimal|low|medium|high|xhigh] [--glm-key-env NAME] [--claude-bin <path>] [--codex-bin <path>] [--concurrency 1] [--force]",
+  setup: "mergejudge-runner setup [code] [--yes] [--no-autostart] [--print-unit] [--name <name>] [--url <url>] [--engine <name>] [--model [engine=]<model>]... [--codex-effort minimal|low|medium|high|xhigh] [--glm-key-env NAME | --no-glm] [--force]\n\nRegisters this machine (code from the panel, or prompted), detects which of claude and codex are installed and signed in, asks whether to add a GLM (z.ai, billed per use) API key (--glm-key-env NAME reads it from that environment variable; --no-glm skips the question), and offers to start the runner at login. MERGEJUDGE_URL overrides the default server.",
   run: "mergejudge-runner run [--concurrency 1] [--once] [--claude-bin <path>] [--codex-bin <path>] [--job-timeout-min 10]",
   status: "mergejudge-runner status"
 };
@@ -31595,6 +31725,7 @@ async function main(argv, deps = {}) {
         engine: { type: "string" },
         "claude-bin": { type: "string" },
         "codex-bin": { type: "string" },
+        "codex-effort": { type: "string" },
         model: { type: "string", multiple: true },
         "glm-key-env": { type: "string" },
         "no-glm": { type: "boolean" },
@@ -31631,6 +31762,12 @@ ${USAGE}`);
     return e instanceof ConfigError || e.code?.startsWith("ERR_PARSE_ARGS") ? 2 : 1;
   }
 }
+function applyCodexEffort(settings, effort) {
+  if (effort === void 0) return;
+  if (!isCodexEffort(effort)) throw new ConfigError(`--codex-effort must be one of ${CODEX_EFFORTS.join(", ")}`);
+  if (!settings.codex) throw new ConfigError("--codex-effort needs the codex engine (codex installed and signed in, or --engine codex)");
+  settings.codex.effort = effort;
+}
 async function register(v, configPath, deps, out) {
   const code = v.code, urlRaw = v.url;
   if (!code || !urlRaw) throw new ConfigError("register needs --code and --url");
@@ -31638,6 +31775,7 @@ async function register(v, configPath, deps, out) {
   const engine = v.engine ?? "claude-code";
   const glmApiKey = readGlmKeyEnv(v["glm-key-env"], env);
   const settings = buildEngineSettings([engine], pickDefaultEngine(engine, engine === "glm" && !glmApiKey && !env[GLM_KEY_ENV] ? [] : [engine]), v.model ?? []);
+  applyCodexEffort(settings, v["codex-effort"]);
   const cfg = await registerRunner({
     code,
     url: urlRaw,
@@ -31784,6 +31922,7 @@ async function setup(v, positionals, configPath, deps, out, err) {
     const requested = v.engine;
     const engine = requested && !enabled.includes(requested) && enabled.length === 0 && requested !== "glm" ? requested : pickDefaultEngine(requested, enabled);
     const settings = buildEngineSettings(enabled.includes(engine) ? enabled : [...enabled, engine], engine, v.model ?? []);
+    applyCodexEffort(settings, v["codex-effort"]);
     cfg = await registerRunner({
       code: code2,
       url: url2,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mergejudge-runner",
-  "version": "0.2.0",
+  "version": "0.2.1",
   "license": "MIT",
   "type": "module",
   "description": "Run MergeJudge reviews on your own machine with your own Claude Code login (runner protocol /runner/v1)",