chore: sync actions from gh-aw@v0.92.4

setup/js/agent_execution.cjs

@@ -1,7 +1,7 @@
 // @ts-check
 
 const { detectErrors, buildOutputLines, isCAPIServerError } = require("./agent_error_patterns.cjs");
-const { detectNonRetryableHarnessGuard, isAuthenticationFailedError } = require("./harness_error_patterns.cjs");
+const { API_PROXY_GUARD_REJECTION_PATTERNS, detectNonRetryableHarnessGuard, isAuthenticationFailedError, selectNativeErrorDiagnostic, shouldSuppressGuardCategory } = require("./harness_error_patterns.cjs");
 const { crashSignalNameForExitCode } = require("./harness_crash_signals.cjs");
 const { ERR_VALIDATION } = require("./error_codes.cjs");
 const { sessionContext } = require("./agent_session.cjs");
@@ -243,11 +243,15 @@ function agentErrorSummaryText(content, eventsOrSourceEngine = [], rawSourceEngi
  */
 function collectNativeErrorEvidence(errors) {
   const diagnostics = [];
+  const guardDiagnostics = [];
   const codes = new Set();
   const types = new Set();
   for (const error of errors) {
-    diagnostics.push(JSON.stringify(error));
-    const queue = [error];
+    const diagnostic = selectNativeErrorDiagnostic(error);
+    const diagnosticText = JSON.stringify(diagnostic) ?? "";
+    diagnostics.push(diagnosticText);
+    guardDiagnostics.push(diagnosticText);
+    const queue = [diagnostic];
     for (let visited = 0; visited < 16 && queue.length; visited++) {
       const value = queue.shift();
       if (typeof value === "string") {
@@ -274,10 +278,9 @@ function collectNativeErrorEvidence(errors) {
       for (const key of ["error", "message", "errorMessage"]) {
         if (value[key] !== undefined) queue.push(value[key]);
       }
-      if (value.metadata?.raw !== undefined) queue.push(value.metadata.raw);
     }
   }
-  return { diagnostics, codes, types };
+  return { diagnostics, guardDiagnostics, codes, types };
 }
 
 /**
@@ -290,7 +293,7 @@ function collectAgentExecution({ content = "", events = [], categories = [], exi
   const categorySet = new Set(categories);
   const { diagnostics, errors: rawErrors } = collectAgentErrorEvidence(content, events, rawSourceEngine);
   const errors = [...events.flatMap(recordErrors), ...rawErrors];
-  const { diagnostics: nativeDiagnostics, codes, types } = collectNativeErrorEvidence(errors);
+  const { diagnostics: nativeDiagnostics, guardDiagnostics, codes, types } = collectNativeErrorEvidence(errors);
   let observedExit = exitCode;
   for (const data of observations) {
     validateAgentExecution(data);
@@ -305,6 +308,7 @@ function collectAgentExecution({ content = "", events = [], categories = [], exi
     const terminal = line.match(/^\[(?:copilot|claude|codex)-harness\].*?\bdone: exitCode=(\d+)\b/);
     if (terminal && exitCode === undefined && !observations.some(data => data.exitCode !== undefined)) observedExit = Number(terminal[1]);
   }
+  guardDiagnostics.push(...diagnostics);
   diagnostics.push(...nativeDiagnostics);
   const text = diagnostics.join("\n");
   for (const diagnostic of diagnostics) {
@@ -322,6 +326,12 @@ function collectAgentExecution({ content = "", events = [], categories = [], exi
   if (guards.awfAPIProxyBlockingRequests) categorySet.add("awf_api_proxy_blocking_requests");
   if (guards.goalAlreadyActive) categorySet.add("goal_already_active");
   if (guards.apiProxyGuardRejection) categorySet.add(guards.apiProxyGuardRejection.guard);
+  // Observation accepts attributed diagnostic text; retry decisions require transport envelopes.
+  for (const diagnostic of guardDiagnostics) {
+    for (const { guard, pattern } of API_PROXY_GUARD_REJECTION_PATTERNS) {
+      if (pattern.test(diagnostic) && !shouldSuppressGuardCategory(guard, diagnostic)) categorySet.add(guard);
+    }
+  }
   const crashSignal = observedExit === undefined ? null : crashSignalNameForExitCode(observedExit);
   if (crashSignal) {
     categorySet.add("sandbox_runtime_crash");

setup/js/ai_credits_context.cjs

@@ -3,6 +3,7 @@ require("./shim.cjs");
 
 const fs = require("fs");
 const path = require("path");
+const { parseAICreditsExceededProxyRejection } = require("./harness_error_patterns.cjs");
 
 const MAX_AI_CREDITS_FIELDS = new Set(["max_ai_credits", "maxAiCredits"]);
 const AI_CREDITS_FIELDS = new Set(["ai_credits", "aiCredits"]);
@@ -32,10 +33,88 @@ const API_PROXY_GUARD_REJECTION_EVENT_TYPES = [MAX_CACHE_MISSES_EXCEEDED_EVENT_T
 // Counter fields carried by proxy guard rejection events, reported alongside the guard name
 // so the step log states what limit was hit and at which value.
 const API_PROXY_GUARD_COUNTER_FIELDS = ["consecutive_cache_misses", "max_cache_misses", "effective_tokens", "max_effective_tokens", "permission_denied_count", "max_permission_denied", "model"];
-const MAX_AI_CREDITS_EXCEEDED_STDIO_RE = /maximum ai credits exceeded(?:\s*\((\d+(?:\.\d+)?)\s*\/\s*(\d+(?:\.\d+)?)\))?/i;
+// Raw stderr fallback assumes unprefixed subprocess diagnostics; only explicitly allowlisted wrappers are unwrapped below.
+const MAX_AI_CREDITS_EXCEEDED_STDIO_RE =
+  /^(?:(?:Execution failed:\s*)?CAPIError:\s*429\s+|Failed to get response from the AI model; retried \d+ times(?: \(total retry wait time:[^)]+\))?(?: \(Request-ID [^)]+\))?\.? Last error: CAPIError:\s*429\s+|Fatal:\s*)Maximum AI credits exceeded(?:\s*\((\d+(?:\.\d+)?)\s*\/\s*(\d+(?:\.\d+)?)\))?\.?$/i;
+const MAX_AI_CREDITS_EXCEEDED_NATIVE_ERROR_RE = /^(?:(?:API Error:\s*)?403(?::\s*|\s+)(?:Forbidden[.:]?\s*)?)?Maximum AI credits exceeded(?:\s*\((\d+(?:\.\d+)?)\s*\/\s*(\d+(?:\.\d+)?)\))?\.?$/i;
 const DEFAULT_AGENT_STDIO_LOG = "/tmp/gh-aw/agent-stdio.log";
 const AGENT_STDIO_LOG_MAX_TAIL = 64 * 1024; // 64 KB — sufficient for any realistic error block
 
+/**
+ * Compatibility for native diagnostics alongside the pair-only harness parser.
+ * Never walk arbitrary JSON fields: only engine-owned error slots are eligible.
+ * @param {string} line
+ * @returns {{ aiCredits: string, maxAICredits: string } | null}
+ */
+function parseNativeAICreditsRejection(line) {
+  let diagnostic = line;
+  let nativeEnvelope = false;
+  if (line.trimStart().startsWith("{")) {
+    let record;
+    try {
+      record = JSON.parse(line);
+    } catch {
+      return null;
+    }
+    if (!record || typeof record !== "object" || Array.isArray(record)) return null;
+    if (record.role || record.data?.role || record.error?.role || record.parentToolUseId || record.parent_tool_use_id || record.parentToolCallId || record.agentId || record.data?.parentToolCallId || record.data?.agentId) return null;
+    if (["warning", "info"].includes(record.data?.severity || record.severity)) return null;
+    if (record.message?.role && !["message_end", "turn_end", "assistant"].includes(record.type)) return null;
+    if (record.message?.role && record.message.role !== "assistant") return null;
+    const rejection = parseAICreditsExceededProxyRejection(line);
+    if (rejection) return { aiCredits: String(rejection.aiCredits), maxAICredits: String(rejection.maxAICredits) };
+    if (["message_end", "turn_end"].includes(record.type) && record.message?.role === "assistant" && record.message.stopReason === "error") {
+      diagnostic = record.message.errorMessage;
+    } else if (record.type === "error" && typeof record.error === "string" && record.message === undefined) {
+      diagnostic = record.error;
+    } else if (record.type === "session.error" && record.data) {
+      diagnostic = record.data.message;
+    } else {
+      return null;
+    }
+    nativeEnvelope = true;
+  } else {
+    const sdkError = /^\[(?:copilot-sdk-driver|sdk-driver)\]\s+(?:\[sdk-driver\]\s+)?(?:error|unhandled error):\s*(?:Error:\s*)?([\s\S]*)$/i.exec(line);
+    if (sdkError) {
+      diagnostic = sdkError[1];
+      nativeEnvelope = true;
+    }
+    // Pi strips the HTTP status from its JSON-quoted error field.
+    const providerError = /^\[gh-aw\/pi-provider\]\s+\S+\s+provider_error provider=.+? model=.+? api=\S+ status=403 method=\S+ url=.+? response_headers=\S+ error=("(?:[^"\\]|\\.)*")$/.exec(line);
+    if (providerError) {
+      try {
+        diagnostic = JSON.parse(providerError[1]);
+      } catch {
+        return null;
+      }
+      nativeEnvelope = true;
+    }
+  }
+  if (typeof diagnostic !== "string") return null;
+  let match = MAX_AI_CREDITS_EXCEEDED_STDIO_RE.exec(diagnostic);
+  if (!match && nativeEnvelope) {
+    // Goose's terminal error wraps the response and appends fixed retry advice.
+    diagnostic = diagnostic.replace(/^Ran into this error: Authentication error: Status: 403 Forbidden\. Response: /, "").replace(/\n\nPlease retry if you think this is a transient or recoverable error\.$/, "");
+    // Pi may retain the provider's JSON error body after the HTTP status.
+    const jsonBody = /^(?:403(?::\s*|\s+))?(\{.*\})$/.exec(diagnostic);
+    if (jsonBody) {
+      try {
+        const body = JSON.parse(jsonBody[1]);
+        if (body.role || body.error?.role || ["assistant", "user", "tool", "tool_result", "toolResponse", "message"].includes(body.type)) return null;
+        diagnostic = typeof body.error?.message === "string" ? body.error.message : body.message;
+      } catch {
+        return null;
+      }
+    }
+    match = typeof diagnostic === "string" ? MAX_AI_CREDITS_EXCEEDED_NATIVE_ERROR_RE.exec(diagnostic) : null;
+  }
+  if (!match) return null;
+  const aiCredits = match[1] || "";
+  const maxAICredits = match[2] || "";
+  if (aiCredits && (!Number.isFinite(Number(aiCredits)) || !Number.isFinite(Number(maxAICredits)) || Number(maxAICredits) <= 0 || Number(aiCredits) < Number(maxAICredits))) return null;
+  return { aiCredits, maxAICredits };
+}
+
 /**
  * @param {unknown} value
  * @returns {string}
@@ -716,25 +795,21 @@ function parseAICreditsExceededFromAgentStdio() {
       fs.closeSync(fd);
     }
     const content = buf.toString("utf8");
-    const proxyHTTP403 = content.split(/\r?\n/).some(line => {
-      const proxyFailure = line.match(/Authentication failed with provider at (\S+) \(HTTP 403\)\.?/i);
+    const lines = content.split(/\r?\n/);
+    const proxyHTTP403 = lines.some(line => {
+      const proxyFailure = line.match(/^Authentication failed with provider at (\S+) \(HTTP 403\)\.?$/i);
       return proxyFailure !== null && isLikelyAWFAPIProxyURL(proxyFailure[1]);
     });
-    // Use matchAll and take the last occurrence — in retried runs the final
-    // entry carries the authoritative (highest) credit values.
-    const RE_G = new RegExp(MAX_AI_CREDITS_EXCEEDED_STDIO_RE.source, "gi");
-    const allMatches = [...content.matchAll(RE_G)];
-    const match = allMatches.at(-1);
-    if (!match) return { ...EMPTY_AGENT_STDIO_STATE, proxyHTTP403 };
-    const aiCredits = parsePositiveNumberString(match[1] || "");
-    const maxAICredits = parsePositiveNumberString(match[2] || "");
-    return {
-      aiCredits,
-      maxAICredits,
-      rateLimitError: true,
-      maxAICreditsExceeded: true,
-      proxyHTTP403,
-    };
+    // JSON tool/assistant records can quote rejection examples. Only inspect
+    // authenticated transport errors or a complete native stderr diagnostic.
+    let result = { ...EMPTY_AGENT_STDIO_STATE, proxyHTTP403 };
+    for (const line of lines) {
+      const rejection = parseNativeAICreditsRejection(line);
+      if (!rejection) continue;
+      const { aiCredits, maxAICredits } = rejection;
+      result = { aiCredits, maxAICredits, rateLimitError: true, maxAICreditsExceeded: true, proxyHTTP403 };
+    }
+    return result;
   } catch {
     return EMPTY_AGENT_STDIO_STATE;
   }

setup/js/checkout_pr_branch.cjs

@@ -36,6 +36,7 @@ const { detectForkPR } = require("./pr_helpers.cjs");
 const { parseAllowedBots, isAllowedBot, canonicalizeBotIdentifier } = require("./check_permissions_utils.cjs");
 const { ERR_API, ERR_PERMISSION } = require("./error_codes.cjs");
 const TRUSTED_CHECKOUT_PERMISSIONS = ["write", "maintain", "admin"];
+const TRUSTED_BOT_PR_ACTIONS = ["opened", "synchronize", "ready_for_review", "reopened"];
 // Centralized command/label routing uses the repository GITHUB_TOKEN, whose
 // GitHub-authenticated actor is fixed to this platform identity.
 const CENTRALIZED_ROUTER_ACTOR = "github-actions[bot]";
@@ -295,7 +296,8 @@ async function assertTrustedCheckoutRuntime(awContext, pullRequest) {
   // events require the collaborator check unless explicitly allow-listed below.
   if (
     (context.eventName === "pull_request" || context.eventName === "pull_request_target") &&
-    (context.payload.action === "opened" || context.payload.action === "synchronize") &&
+    typeof context.payload.action === "string" &&
+    TRUSTED_BOT_PR_ACTIONS.includes(context.payload.action) &&
     senderType === "Bot" &&
     context.payload.sender?.login === actor &&
     context.payload.repository !== undefined &&

setup/js/collect_ndjson_output.cjs

@@ -9,7 +9,7 @@ const { AGENT_OUTPUT_FILENAME, TMP_GH_AW_PATH } = require("./constants.cjs");
 const { ERR_API, ERR_PARSE } = require("./error_codes.cjs");
 const { parseIntTemplatable } = require("./templatable.cjs");
 const { isProbingNoopMessage } = require("./intent_probe.cjs");
-const { buildEmptyOutputOutcome } = require("./empty_output_outcome.cjs");
+const { buildEmptyOutputOutcome, collectOutcomeSession, hasFailedFinalResult } = require("./empty_output_outcome.cjs");
 
 const MENTION_AWARE_OUTPUT_TYPES = new Set(["add_comment", "close_discussion", "create_discussion", "create_issue", "create_pull_request", "create_pull_request_review_comment", "reply_to_pull_request_review_comment"]);
 
@@ -374,8 +374,19 @@ async function main() {
     let collectorDriverExitCode;
     let collectorRetryCount;
     let collectorEngineErrorType;
-    if (!queueScope && !parsedItems.some(item => !["missing_tool", "missing_data"].includes(item.type))) {
-      const incompleteOutcome = buildEmptyOutputOutcome(errors);
+    const hasTerminalOutput = parsedItems.some(item => !["missing_tool", "missing_data"].includes(item.type));
+    let failedFinalResult = false;
+    let session;
+    if (!queueScope && hasTerminalOutput) {
+      try {
+        session = collectOutcomeSession(TMP_GH_AW_PATH);
+        failedFinalResult = hasFailedFinalResult(session.events);
+      } catch {
+        core.warning("Runtime diagnostics could not be collected.");
+      }
+    }
+    if (!queueScope && (failedFinalResult || !hasTerminalOutput)) {
+      const incompleteOutcome = buildEmptyOutputOutcome(errors, undefined, session);
       parsedItems.push(incompleteOutcome);
       collectorEmptyOutputCause = incompleteOutcome.reason;
       collectorFailureCause = incompleteOutcome.failureCause;

setup/js/create_prompt.cjs

@@ -162,14 +162,15 @@ function renderWorkClaim(payload) {
     return "";
   }
 
-  return assignment.claims
+  const assignmentJSON = JSON.stringify({ work_queue_assignment: assignment }).replace(/</g, "\\u003c");
+  const claims = assignment.claims
     .map(member => {
-      const work = JSON.stringify({ id: member.work_id, payload: member.work, result_refs: member.result_refs }).replace(/</g, "\\u003c");
       const claim = JSON.stringify({ dispatch_id: assignment.dispatch_id, id: member.claim_id, work_id: member.work_id, claim_handle: member.handle }).replace(/</g, "\\u003c");
       const handle = JSON.stringify(member.handle).replace(/</g, "\\u003c");
-      return `<work-claim>\n${work}\n${claim}\nWhen the work is complete, call work_queue_claim_finish with outcome "completed" and claim_handle ${handle} to record the finish intent. Set keep_open: true only with completed to finish this Claim while retaining the Work in the queue and leaving its backing Issue open; the Work can be claimed again only after delivery is verified. keep_open defaults to false. If you cannot complete the work, call it with outcome "cancelled" and the same claim_handle. Always include this claim_handle on safe outputs and queue-control calls. The claim is only complete after trusted reconciliation verifies it; recording intent alone does not authorize safe outputs.\n</work-claim>\n`;
+      return `<work-claim>\n${claim}\nWhen the work is complete, call work_queue_claim_finish with outcome "completed" and claim_handle ${handle} to record the finish intent. Set keep_open: true only with completed to finish this Claim while retaining the Work in the queue and leaving its backing Issue open; the Work can be claimed again only after delivery is verified. keep_open defaults to false. If you cannot complete the work, call it with outcome "cancelled" and the same claim_handle. Always include this claim_handle on safe outputs and queue-control calls. The claim is only complete after trusted reconciliation verifies it; recording intent alone does not authorize safe outputs.\n</work-claim>\n`;
     })
     .join("");
+  return `<work-queue-assignment>\n${assignmentJSON}\n</work-queue-assignment>\n${claims}`;
 }
 
 /**