Harden Claude startup failure handling for empty-log Avenger runs

actions/setup/js/claude_harness.cjs

@@ -78,6 +78,7 @@ const MAX_TURNS_EXIT_PATTERN = /"subtype"\s*:\s*"error_max_turns"/;
 // this path must not be retried via --continue (fall back to a fresh run if budget remains).
 const NO_DEFERRED_MARKER_PATTERN = /No deferred tool marker found/i;
 const SIGNAL_TERMINATION_EXIT_CODES = new Set([137, 143]);
+const MAX_STARTUP_RETRIES = 2;
 
 /**
  * Emit a timestamped diagnostic log line to stderr.
@@ -99,6 +100,35 @@ function resolveRetryConfig(env = process.env, logger = log) {
   return resolveSharedRetryConfig(env, logger);
 }
 
+/**
+ * Parse bounded startup retries for zero-output startup failures.
+ * Retries are always fresh runs (never --continue).
+ *
+ * @param {NodeJS.ProcessEnv} [env]
+ * @param {(message: string) => void} [logger]
+ * @returns {number}
+ */
+function resolveStartupRetryLimit(env = process.env, logger = log) {
+  const raw = env.GH_AW_CLAUDE_STARTUP_RETRIES;
+  if (raw == null || raw === "") {
+    return 1;
+  }
+  if (!/^[+-]?\d+$/.test(raw)) {
+    logger(`invalid GH_AW_CLAUDE_STARTUP_RETRIES='${raw}' (expected integer); using default startup retries=1`);
+    return 1;
+  }
+  const parsed = Number.parseInt(raw, 10);
+  if (!Number.isFinite(parsed)) {
+    logger(`invalid GH_AW_CLAUDE_STARTUP_RETRIES='${raw}' (not finite); using default startup retries=1`);
+    return 1;
+  }
+  const clamped = Math.min(Math.max(parsed, 0), MAX_STARTUP_RETRIES);
+  if (clamped !== parsed) {
+    logger(`GH_AW_CLAUDE_STARTUP_RETRIES=${parsed} out of range; clamped to ${clamped}`);
+  }
+  return clamped;
+}
+
 /**
  * Determines if the collected output contains an Anthropic overload error.
  * @param {string} output - Collected stdout+stderr from the process
@@ -297,6 +327,7 @@ async function main() {
   const [, , command, ...args] = process.argv;
   const retryConfig = resolveRetryConfig(process.env, log);
   const { maxRetries, initialDelayMs, backoffMultiplier, maxDelayMs } = retryConfig;
+  const startupRetryLimit = resolveStartupRetryLimit(process.env, log);
 
   if (!command) {
     process.stderr.write("claude-harness: Usage: node claude_harness.cjs <command> [args...]\n");
@@ -347,6 +378,7 @@ async function main() {
   let lastExitCode = 1;
   let useContinueOnRetry = false;
   let continueDisabledPermanently = false;
+  let startupRetriesUsed = 0;
   const driverStartTime = Date.now();
   // Soft-timeout guard: polled at the top of the retry loop and after each backoff sleep.
   // It does not preempt a running attempt — if a single invocation runs past the soft
@@ -519,7 +551,19 @@ async function main() {
     if (attempt >= maxRetries) {
       log(`all ${maxRetries} retries exhausted — giving up (exitCode=${lastExitCode})`);
     } else {
-      log(`attempt ${attempt + 1}: no output produced — not retrying` + ` (possible causes: binary not found, permission denied, auth failure, or silent startup crash)`);
+      if (startupRetriesUsed < startupRetryLimit) {
+        startupRetriesUsed++;
+        useContinueOnRetry = false;
+        delay = initialDelayMs;
+        // attempt is 0-based; logs are 1-based, so "next attempt" is +2.
+        const nextAttemptNumber = attempt + 2;
+        const totalAttempts = maxRetries + 1;
+        log(`attempt ${attempt + 1}: no output produced — retrying startup as fresh run ` + `(startup retry ${startupRetriesUsed}/${startupRetryLimit}, next attempt ${nextAttemptNumber} of ${totalAttempts} total attempts)`);
+        continue;
+      }
+      log(
+        `attempt ${attempt + 1}: no output produced — not retrying (startup retry budget exhausted: ${startupRetriesUsed}/${startupRetryLimit}; possible causes: binary not found, permission denied, auth failure, or silent startup crash)`
+      );
     }
 
     break;
@@ -551,6 +595,7 @@ if (typeof module !== "undefined" && module.exports) {
     hasNoopInSafeOutputs,
     hasExpectedSafeOutputs,
     resolveRetryConfig,
+    resolveStartupRetryLimit,
   };
 }
 

actions/setup/js/claude_harness.test.cjs

@@ -21,6 +21,7 @@ const {
   extractDeniedCommands,
   buildMissingToolPermissionIssuePayload,
   resolveRetryConfig,
+  resolveStartupRetryLimit,
 } = require("./claude_harness.cjs");
 
 const agentTempDir = "/tmp/gh-aw/agent";
@@ -30,7 +31,7 @@ function makeHarnessTempDir(name) {
   return fs.mkdtempSync(path.join(agentTempDir, name));
 }
 
-function runHarnessWithStub({ stubScript, prompt = "fix the bug", extraArgs = [] }) {
+function runHarnessWithStub({ stubScript, prompt = "fix the bug", extraArgs = [], extraEnv = {} }) {
   const tempDir = makeHarnessTempDir("claude-harness-");
   const stubPath = path.join(tempDir, "stub.cjs");
   const promptPath = path.join(tempDir, "prompt.txt");
@@ -40,7 +41,7 @@ function runHarnessWithStub({ stubScript, prompt = "fix the bug", extraArgs = []
 
   const result = spawnSync(process.execPath, ["claude_harness.cjs", process.execPath, stubPath, "--print", ...extraArgs, "--prompt-file", promptPath], {
     cwd: path.dirname(require.resolve("./claude_harness.cjs")),
-    env: { ...process.env, CLAUDE_HARNESS_STUB_CALLS: callsPath },
+    env: { ...process.env, ...extraEnv, CLAUDE_HARNESS_STUB_CALLS: callsPath },
     encoding: "utf8",
     timeout: 45000,
   });
@@ -448,6 +449,41 @@ process.exit(0);
       expect(result.stderr).toContain("failure_reason=cancelled_or_timed_out");
     }, 30000);
 
+    it("retries one no-output startup failure as a fresh run by default", () => {
+      const stubScript = `
+const fs = require("fs");
+const callsPath = process.env.CLAUDE_HARNESS_STUB_CALLS;
+const args = process.argv.slice(2);
+const priorCalls = fs.existsSync(callsPath) ? fs.readFileSync(callsPath, "utf8").trim().split("\\n").filter(Boolean).length : 0;
+fs.appendFileSync(callsPath, JSON.stringify({ args }) + "\\n", "utf8");
+if (priorCalls === 0) process.exit(1);
+process.stdout.write("startup retry succeeded\\n");
+process.exit(0);
+`;
+      const { result, calls } = runHarnessWithStub({ stubScript });
+      expect(result.status, result.stderr).toBe(0);
+      expect(calls.length).toBe(2);
+      expect(calls[1].args.includes("--continue")).toBe(false);
+      expect(result.stderr).toContain("no output produced — retrying startup as fresh run");
+    });
+
+    it("does not retry no-output startup failure when GH_AW_CLAUDE_STARTUP_RETRIES=0", () => {
+      const stubScript = `
+const fs = require("fs");
+const callsPath = process.env.CLAUDE_HARNESS_STUB_CALLS;
+fs.appendFileSync(callsPath, JSON.stringify({ args: process.argv.slice(2) }) + "\\n", "utf8");
+process.exit(1);
+`;
+      const { result, calls } = runHarnessWithStub({
+        stubScript,
+        extraEnv: { GH_AW_CLAUDE_STARTUP_RETRIES: "0" },
+      });
+      expect(result.status).toBe(1);
+      expect(calls.length).toBe(1);
+      expect(calls[0].args).toContain("fix the bug");
+      expect(result.stderr).toContain("startup retry budget exhausted: 0/0");
+    });
+
     it("returns true for normal partial-execution retry", () => {
       const result = shouldRetryWithContinue({
         attempt: 0,
@@ -559,6 +595,14 @@ process.exit(0);
       const config2 = resolveRetryConfig({ GH_AW_HARNESS_INITIAL_DELAY_MS: "0x10" });
       expect(config2.initialDelayMs).toBe(5000);
     });
+
+    it("uses startup retry default and clamps overrides to [0..2]", () => {
+      expect(resolveStartupRetryLimit({})).toBe(1);
+      expect(resolveStartupRetryLimit({ GH_AW_CLAUDE_STARTUP_RETRIES: "2" })).toBe(2);
+      expect(resolveStartupRetryLimit({ GH_AW_CLAUDE_STARTUP_RETRIES: "-5" })).toBe(0);
+      expect(resolveStartupRetryLimit({ GH_AW_CLAUDE_STARTUP_RETRIES: "9" })).toBe(2);
+      expect(resolveStartupRetryLimit({ GH_AW_CLAUDE_STARTUP_RETRIES: "nope" })).toBe(1);
+    });
   });
 
   describe("noop pre-flight and retry guard", () => {

actions/setup/js/log_parser_bootstrap.cjs

@@ -4,6 +4,90 @@
 const { generatePlainTextSummary, generateCopilotCliStyleSummary, wrapAgentLogInSection, formatSafeOutputsPreview } = require("./log_parser_shared.cjs");
 const { getErrorMessage } = require("./error_helpers.cjs");
 const { ERR_API, ERR_CONFIG, ERR_VALIDATION } = require("./error_codes.cjs");
+const INFERENCE_ACCESS_ERROR_PATTERN = /Access denied by policy settings|invalid access to inference/i;
+const CLAUDE_RATE_LIMIT_PATTERN = /rate_limit_error|429 Too Many Requests|"api_error_status"\s*:\s*429|request rejected \(429\)|rate limit/i;
+const CLAUDE_OVERLOAD_PATTERN = /overloaded_error|"overloaded"/i;
+/** Matches HTTP protocol/status lines such as "HTTP/1.1 503". */
+const CLAUDE_HTTP_5XX_PROTOCOL_PATTERN = /HTTP(?:\/\d\.\d)?\s+5\d{2}\b/;
+/** Matches structured status fields such as "status: 502" or "status code=500". */
+const CLAUDE_HTTP_5XX_STATUS_FIELD_PATTERN = /status(?:\s+code)?\s*[:=]\s*5\d{2}\b/;
+/** Matches request/fetch/http error prose that carries an explicit 5xx code. */
+const CLAUDE_HTTP_5XX_REQUEST_ERROR_PATTERN = /(?:http|fetch|request)\s+(?:failed|error)[^\n]*?\b5\d{2}\b/;
+const CLAUDE_HTTP_5XX_STATUS_PATTERN = new RegExp([CLAUDE_HTTP_5XX_PROTOCOL_PATTERN.source, CLAUDE_HTTP_5XX_STATUS_FIELD_PATTERN.source, CLAUDE_HTTP_5XX_REQUEST_ERROR_PATTERN.source].join("|"), "i");
+const STARTUP_DIAGNOSTIC_LINE_PATTERN = /(?:ERR_|Error:|CAPIError|Authentication failed|rate[_ -]?limit|429|\b5\d{2}\b|overloaded|inference)/i;
+const MAX_DIAGNOSTIC_TAIL_LINES = 8;
+
+/**
+ * Build startup diagnostics for Claude failures with no structured entries.
+ * @param {string} rawContent
+ * @returns {{
+ *   exitCode: string,
+ *   inferenceAccessError: boolean,
+ *   aiCreditsRateLimitError: boolean,
+ *   transientInferenceAvailabilityError: boolean,
+ *   summaryLine: string,
+ *   summaryMarkdown: string
+ * }}
+ */
+function buildClaudeStartupDiagnostics(rawContent) {
+  const content = typeof rawContent === "string" ? rawContent : "";
+  const lines = content
+    .split(/\r?\n/)
+    .map(line => line.trim())
+    .filter(Boolean);
+  const startupLines = lines.filter(line => line.includes("[claude-harness]") || STARTUP_DIAGNOSTIC_LINE_PATTERN.test(line));
+  // Intentionally avoid falling back to arbitrary raw log lines here. For empty-structured-log
+  // failures, the unmatched tail can contain unrelated stdout/stderr, prompts, or secrets; when
+  // no startup/diagnostic lines are recognized, we emit no tail section rather than risk leaking it.
+  const tailLines = startupLines.slice(-MAX_DIAGNOSTIC_TAIL_LINES);
+  const tailText = tailLines.join("\n");
+  const safeTailText = escapeHtml(tailText);
+
+  let exitCode = "unknown";
+  const donePrefix = "done: exitCode=";
+  const doneIdx = content.lastIndexOf(donePrefix);
+  if (doneIdx !== -1) {
+    const doneTail = content.slice(doneIdx + donePrefix.length);
+    const doneCode = doneTail.match(/^(\d+)/);
+    if (doneCode) {
+      exitCode = doneCode[1];
+    }
+  } else {
+    const failedPrefix = "failed: exitCode=";
+    const failedIdx = content.lastIndexOf(failedPrefix);
+    if (failedIdx !== -1) {
+      const failedTail = content.slice(failedIdx + failedPrefix.length);
+      const failedCode = failedTail.match(/^(\d+)/);
+      if (failedCode) {
+        exitCode = failedCode[1];
+      }
+    }
+  }
+
+  const inferenceAccessError = INFERENCE_ACCESS_ERROR_PATTERN.test(content);
+  const aiCreditsRateLimitError = CLAUDE_RATE_LIMIT_PATTERN.test(content) || CLAUDE_OVERLOAD_PATTERN.test(content);
+  const transientInferenceAvailabilityError = aiCreditsRateLimitError || CLAUDE_HTTP_5XX_STATUS_PATTERN.test(content);
+  const summaryLine = `Claude startup failed before structured logging (exitCode=${exitCode}).`;
+  const summaryMarkdown = safeTailText ? `<details><summary>Claude startup diagnostics</summary>\n\n<pre><code>${safeTailText}</code></pre>\n</details>` : "";
+
+  return {
+    exitCode,
+    inferenceAccessError,
+    aiCreditsRateLimitError,
+    transientInferenceAvailabilityError,
+    summaryLine,
+    summaryMarkdown,
+  };
+}
+
+/**
+ * @param {string} text
+ * @returns {string}
+ */
+function escapeHtml(text) {
+  const htmlEntities = { '"': "&quot;", "'": "&#39;", "&": "&amp;", "<": "&lt;", ">": "&gt;" };
+  return text.replace(/["'&<>]/g, char => htmlEntities[char]);
+}
 
 /**
  * Bootstrap helper for log parser entry points.
@@ -344,7 +428,26 @@ async function runLogParser(options) {
           `Claude produced no structured log entries, but agent completed with ${safeOutputEntriesCount} safe output ${safeOutputEntriesCount === 1 ? "entry" : "entries"} — treating as non-fatal post-completion infrastructure failure`
         );
       } else {
-        core.setFailed(`${ERR_CONFIG}: Claude execution failed: no structured log entries were produced. This usually indicates a startup or configuration error before tool execution.`);
+        const diagnostics = buildClaudeStartupDiagnostics(content);
+        if (diagnostics.summaryMarkdown) {
+          await core.summary.addRaw(diagnostics.summaryMarkdown).write();
+        }
+
+        if (diagnostics.inferenceAccessError) {
+          core.setOutput("inference_access_error", "true");
+        }
+        if (diagnostics.aiCreditsRateLimitError) {
+          core.setOutput("ai_credits_rate_limit_error", "true");
+        }
+
+        const errorCode = diagnostics.inferenceAccessError || diagnostics.transientInferenceAvailabilityError ? ERR_API : ERR_CONFIG;
+        const failureKind = diagnostics.inferenceAccessError
+          ? "inference access denied by policy"
+          : diagnostics.transientInferenceAvailabilityError
+            ? "transient inference availability signal detected"
+            : "startup/configuration failure detected";
+
+        core.setFailed(`${errorCode}: Claude execution failed: no structured log entries were produced. ${diagnostics.summaryLine} ${failureKind}.`);
       }
     }