fix: detect max consecutive cache misses exceeded and surface dedicated failure context

actions/setup/js/ai_credits_context.cjs

@@ -18,6 +18,10 @@ const BUDGET_EXCEEDED_EVENT = "budget_exceeded";
 // The literal error type emitted by the AWF API proxy (HTTP 400) when maxAiCredits is active
 // and the requested model is not in the built-in pricing table.
 const UNKNOWN_MODEL_AI_CREDITS_TYPE = "unknown_model_ai_credits";
+// The literal error type emitted by the AWF API proxy (HTTP 403) when the consecutive cache
+// miss counter reaches the apiProxy.maxCacheMisses limit. Engine-agnostic: all engines share
+// the same proxy guardrail.
+const MAX_CACHE_MISSES_EXCEEDED_EVENT_TYPE = "max_cache_misses_exceeded";
 const MAX_AI_CREDITS_EXCEEDED_STDIO_RE = /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
@@ -425,6 +429,30 @@ function parseUnknownModelAICreditsAndModelFromAuditLog(auditJsonlPathOverride)
   );
 }
 
+/**
+ * Detects a `max_cache_misses_exceeded` event from the AWF API proxy event logs.
+ * The proxy emits this HTTP 403 error when the consecutive cache miss counter reaches
+ * the configured `apiProxy.maxCacheMisses` limit. Detection is engine-agnostic:
+ * all agentic engines share the same AWF API proxy guardrail.
+ * Structured entries emitted by the AWF API proxy look like:
+ *   { "type": "max_cache_misses_exceeded", "consecutive_cache_misses": 6, "max_cache_misses": 5 }
+ *
+ * @param {string} [eventLogPathOverride]
+ * @returns {boolean}
+ */
+function parseMaxCacheMissesExceededFromEventLog(eventLogPathOverride) {
+  return iterateJSONLFiles(
+    resolveUnknownModelAICreditsLogPaths(eventLogPathOverride),
+    false,
+    content => content.includes(MAX_CACHE_MISSES_EXCEEDED_EVENT_TYPE),
+    (acc, entry) => {
+      if (acc) return true; // already detected, short-circuit
+      return traverseObjectTree(entry, (_key, value) => value === MAX_CACHE_MISSES_EXCEEDED_EVENT_TYPE) || undefined;
+    },
+    acc => acc
+  );
+}
+
 /**
  * Single-pass combined read of the audit log, returning all AI credits fields at once.
  * Used by resolveAICreditsFailureState to avoid reading the same file twice.
@@ -552,5 +580,7 @@ module.exports = {
   parseMaxAICreditsExceededFromAuditLog,
   parseUnknownModelAICreditsFromAuditLog,
   parseUnknownModelAICreditsAndModelFromAuditLog,
+  parseMaxCacheMissesExceededFromEventLog,
   resolveAICreditsFailureState,
+  MAX_CACHE_MISSES_EXCEEDED_EVENT_TYPE,
 };

actions/setup/js/ai_credits_context.test.cjs

@@ -364,6 +364,60 @@ describe("ai_credits_context parseUnknownModelAICreditsAndModelFromAuditLog", ()
   });
 });
 
+describe("ai_credits_context parseMaxCacheMissesExceededFromEventLog", () => {
+  let tmpDir;
+  let parseMaxCacheMissesExceededFromEventLog;
+
+  beforeEach(async () => {
+    vi.resetModules();
+    tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), "aic-cache-misses-test-"));
+    delete process.env.GH_AW_AGENT_OUTPUT;
+    const mod = await import("./ai_credits_context.cjs");
+    const exports = mod.default || mod;
+    parseMaxCacheMissesExceededFromEventLog = exports.parseMaxCacheMissesExceededFromEventLog;
+  });
+
+  afterEach(() => {
+    vi.restoreAllMocks();
+    fs.rmSync(tmpDir, { recursive: true, force: true });
+    delete process.env.GH_AW_AGENT_OUTPUT;
+  });
+
+  function writeEventLog(lines, filename = "event-logs.jsonl") {
+    const logDir = path.join(tmpDir, "sandbox", "firewall", "logs", "api-proxy-logs");
+    fs.mkdirSync(logDir, { recursive: true });
+    const logPath = path.join(logDir, filename);
+    fs.writeFileSync(logPath, lines.map(l => JSON.stringify(l)).join("\n") + "\n", "utf8");
+    process.env.GH_AW_AGENT_OUTPUT = path.join(tmpDir, "output.json");
+    return logPath;
+  }
+
+  it("detects max_cache_misses_exceeded in event-logs.jsonl", () => {
+    writeEventLog([{ type: "max_cache_misses_exceeded", consecutive_cache_misses: 6, max_cache_misses: 5 }]);
+    expect(parseMaxCacheMissesExceededFromEventLog()).toBe(true);
+  });
+
+  it("detects max_cache_misses_exceeded in events.jsonl fallback", () => {
+    writeEventLog([{ type: "max_cache_misses_exceeded", consecutive_cache_misses: 7, max_cache_misses: 5 }], "events.jsonl");
+    expect(parseMaxCacheMissesExceededFromEventLog()).toBe(true);
+  });
+
+  it("returns false when no matching event is present", () => {
+    writeEventLog([{ type: "response", status: 200 }]);
+    expect(parseMaxCacheMissesExceededFromEventLog()).toBe(false);
+  });
+
+  it("returns false for missing event log", () => {
+    process.env.GH_AW_AGENT_OUTPUT = path.join(tmpDir, "output.json");
+    expect(parseMaxCacheMissesExceededFromEventLog("/nonexistent/path/event-logs.jsonl")).toBe(false);
+  });
+
+  it("does not detect other error types", () => {
+    writeEventLog([{ type: "unknown_model_ai_credits" }]);
+    expect(parseMaxCacheMissesExceededFromEventLog()).toBe(false);
+  });
+});
+
 describe("ai_credits_context parseMaxAICreditsFromAuditLog", () => {
   let tmpDir;
   /** @type {(path?: string) => string} */

actions/setup/js/detect_agent_errors.cjs

@@ -45,7 +45,7 @@
 
 const fs = require("fs");
 const { MAX_RUNS_EXCEEDED_PATTERNS, isMaxRunsExceededError } = require("./harness_retry_guard.cjs");
-const { parseUnknownModelAICreditsAndModelFromAuditLog } = require("./ai_credits_context.cjs");
+const { parseUnknownModelAICreditsAndModelFromAuditLog, parseMaxCacheMissesExceededFromEventLog } = require("./ai_credits_context.cjs");
 
 const LOG_FILE = "/tmp/gh-aw/agent-stdio.log";
 
@@ -121,6 +121,17 @@ function buildCombinedPattern(patterns) {
 // The pooled per-run invocation budget is saturated — retries cannot make progress.
 const INVOCATION_CAP_EXCEEDED_PATTERN = buildCombinedPattern(MAX_RUNS_EXCEEDED_PATTERNS);
 
+// Pattern: AWF API proxy consecutive cache miss limit exceeded.
+// The AWF API proxy (engine-agnostic) enforces a configurable limit on back-to-back
+// requests that miss the prompt cache (apiProxy.maxCacheMisses, default 5). When the
+// limit is reached it rejects further requests with HTTP 403 and error type
+// "max_cache_misses_exceeded". Two observable forms:
+//   1) JSON error type in provider API response:  "max_cache_misses_exceeded"
+//   2) Human-readable message from SDK wrapper:   "Maximum consecutive cache misses exceeded"
+// Structured events from the AWF API proxy event log are checked separately via
+// parseMaxCacheMissesExceededFromEventLog().
+const MAX_CACHE_MISSES_EXCEEDED_PATTERN = /(?:\bmax_cache_misses_exceeded\b|\bmaximum\s+consecutive\s+cache\s+misses\s+exceeded\b)/i;
+
 /**
  * Determines if the collected output contains the observed Copilot/CAPI quota exhaustion error.
  * @param {string} output - Collected stdout+stderr from the process
@@ -142,6 +153,17 @@ function isInvocationCapExceededError(output) {
   return isMaxRunsExceededError(output);
 }
 
+/**
+ * Determines if the collected output indicates the AWF API proxy cache miss limit is exceeded.
+ * Checks the agent stdio log for the text-form signal. The structured AWF API proxy event log
+ * is checked separately in detectErrors() via parseMaxCacheMissesExceededFromEventLog().
+ * @param {string} output - Collected stdout+stderr from the process
+ * @returns {boolean}
+ */
+function isMaxCacheMissesExceededError(output) {
+  return MAX_CACHE_MISSES_EXCEEDED_PATTERN.test(output);
+}
+
 /**
  * Normalize model names to a single safe line for GitHub Actions outputs and issue titles.
  * @param {string} value
@@ -164,7 +186,7 @@ function extractMissingModelPricingModelName(logContent) {
 /**
  * Detect known error patterns in a log string and return detection results.
  * @param {string} logContent - Contents of the agent stdio log
- * @returns {{ inferenceAccessError: boolean, mcpPolicyError: boolean, agenticEngineTimeout: boolean, modelNotSupportedError: boolean, http400ResponseError: boolean, capiQuotaExceededError: boolean, invocationCapExceeded: boolean, missingModelPricingError: boolean, missingModelPricingModelName: string }}
+ * @returns {{ inferenceAccessError: boolean, mcpPolicyError: boolean, agenticEngineTimeout: boolean, modelNotSupportedError: boolean, http400ResponseError: boolean, capiQuotaExceededError: boolean, invocationCapExceeded: boolean, maxCacheMissesExceeded: boolean, missingModelPricingError: boolean, missingModelPricingModelName: string }}
  */
 function detectErrors(logContent) {
   const missingModelPricingModelName = extractMissingModelPricingModelName(logContent);
@@ -176,14 +198,15 @@ function detectErrors(logContent) {
     http400ResponseError: HTTP_400_RESPONSE_ERROR_PATTERN.test(logContent),
     capiQuotaExceededError: isCAPIQuotaExceededError(logContent),
     invocationCapExceeded: isInvocationCapExceededError(logContent),
+    maxCacheMissesExceeded: isMaxCacheMissesExceededError(logContent),
     missingModelPricingError: missingModelPricingModelName !== "",
     missingModelPricingModelName,
   };
 }
 
 /**
  * Build GitHub Actions output lines from detection results.
- * @param {{ inferenceAccessError: boolean, mcpPolicyError: boolean, agenticEngineTimeout: boolean, modelNotSupportedError: boolean, http400ResponseError: boolean, capiQuotaExceededError: boolean, invocationCapExceeded: boolean, missingModelPricingError: boolean, missingModelPricingModelName: string }} results
+ * @param {{ inferenceAccessError: boolean, mcpPolicyError: boolean, agenticEngineTimeout: boolean, modelNotSupportedError: boolean, http400ResponseError: boolean, capiQuotaExceededError: boolean, invocationCapExceeded: boolean, maxCacheMissesExceeded: boolean, missingModelPricingError: boolean, missingModelPricingModelName: string }} results
  * @returns {string[]}
  */
 function buildOutputLines(results) {
@@ -196,14 +219,15 @@ function buildOutputLines(results) {
     `http_400_response_error=${results.http400ResponseError}`,
     `capi_quota_exceeded_error=${effectiveCAPIQuotaExceeded}`,
     `invocation_cap_exceeded=${results.invocationCapExceeded}`,
+    `max_cache_misses_exceeded=${results.maxCacheMissesExceeded}`,
     `missing_model_pricing_error=${results.missingModelPricingError}`,
     `missing_model_pricing_model_name=${results.missingModelPricingModelName}`,
   ];
 }
 
 /**
  * Write GitHub Actions outputs to $GITHUB_OUTPUT.
- * @param {{ inferenceAccessError: boolean, mcpPolicyError: boolean, agenticEngineTimeout: boolean, modelNotSupportedError: boolean, http400ResponseError: boolean, capiQuotaExceededError: boolean, invocationCapExceeded: boolean, missingModelPricingError: boolean, missingModelPricingModelName: string }} results
+ * @param {{ inferenceAccessError: boolean, mcpPolicyError: boolean, agenticEngineTimeout: boolean, modelNotSupportedError: boolean, http400ResponseError: boolean, capiQuotaExceededError: boolean, invocationCapExceeded: boolean, maxCacheMissesExceeded: boolean, missingModelPricingError: boolean, missingModelPricingModelName: string }} results
  */
 function writeOutputs(results) {
   const outputFile = process.env.GITHUB_OUTPUT;
@@ -244,8 +268,17 @@ function main() {
     process.stderr.write(`[detect-agent-errors] Detected missing model pricing from firewall structured log: model "${auditModelName}" has no AI credits pricing configured\n`);
   }
 
+  // Also check the AWF API proxy event logs for the `max_cache_misses_exceeded` structured
+  // event. This covers all engines since the proxy guardrail fires independently of the
+  // underlying AI engine.
+  const eventLogCacheMissesExceeded = parseMaxCacheMissesExceededFromEventLog();
+  if (eventLogCacheMissesExceeded && !stdioResults.maxCacheMissesExceeded) {
+    process.stderr.write("[detect-agent-errors] Detected max cache misses exceeded from AWF API proxy event log\n");
+  }
+
   const results = {
     ...stdioResults,
+    maxCacheMissesExceeded: stdioResults.maxCacheMissesExceeded || eventLogCacheMissesExceeded,
     missingModelPricingError: stdioResults.missingModelPricingError || auditMissingPricing,
     missingModelPricingModelName: stdioResults.missingModelPricingModelName || sanitizeModelName(auditModelName),
   };
@@ -271,6 +304,9 @@ function main() {
   if (results.invocationCapExceeded) {
     process.stderr.write("[detect-agent-errors] Detected invocation cap exhaustion: the pooled per-run LLM invocation budget is fully saturated\n");
   }
+  if (results.maxCacheMissesExceeded) {
+    process.stderr.write("[detect-agent-errors] Detected max cache misses exceeded: the AWF API proxy consecutive cache miss limit was reached\n");
+  }
   if (results.missingModelPricingError && !auditMissingPricing) {
     process.stderr.write(`[detect-agent-errors] Detected missing model pricing: model "${results.missingModelPricingModelName}" has no AI credits pricing configured\n`);
   }
@@ -287,13 +323,15 @@ module.exports = {
   extractMissingModelPricingModelName,
   isCAPIQuotaExceededError,
   isInvocationCapExceededError,
+  isMaxCacheMissesExceededError,
   INFERENCE_ACCESS_ERROR_PATTERN,
   MCP_POLICY_BLOCKED_PATTERN,
   AGENTIC_ENGINE_TIMEOUT_PATTERN,
   MODEL_NOT_SUPPORTED_PATTERN,
   HTTP_400_RESPONSE_ERROR_PATTERN,
   CAPI_QUOTA_EXCEEDED_PATTERN,
   INVOCATION_CAP_EXCEEDED_PATTERN,
+  MAX_CACHE_MISSES_EXCEEDED_PATTERN,
   MISSING_MODEL_PRICING_PATTERN,
   buildOutputLines,
 };

actions/setup/js/detect_agent_errors.test.cjs

@@ -4,13 +4,15 @@ const {
   detectErrors,
   isCAPIQuotaExceededError,
   isInvocationCapExceededError,
+  isMaxCacheMissesExceededError,
   INFERENCE_ACCESS_ERROR_PATTERN,
   MCP_POLICY_BLOCKED_PATTERN,
   AGENTIC_ENGINE_TIMEOUT_PATTERN,
   MODEL_NOT_SUPPORTED_PATTERN,
   HTTP_400_RESPONSE_ERROR_PATTERN,
   CAPI_QUOTA_EXCEEDED_PATTERN,
   INVOCATION_CAP_EXCEEDED_PATTERN,
+  MAX_CACHE_MISSES_EXCEEDED_PATTERN,
   MISSING_MODEL_PRICING_PATTERN,
   extractMissingModelPricingModelName,
   buildOutputLines,
@@ -318,6 +320,7 @@ describe("detect_agent_errors.cjs", () => {
       expect(result.http400ResponseError).toBe(false);
       expect(result.capiQuotaExceededError).toBe(false);
       expect(result.invocationCapExceeded).toBe(false);
+      expect(result.maxCacheMissesExceeded).toBe(false);
       expect(result.missingModelPricingError).toBe(false);
       expect(result.missingModelPricingModelName).toBe("");
     });
@@ -484,6 +487,7 @@ describe("detect_agent_errors.cjs", () => {
       expect(result.http400ResponseError).toBe(false);
       expect(result.capiQuotaExceededError).toBe(false);
       expect(result.invocationCapExceeded).toBe(false);
+      expect(result.maxCacheMissesExceeded).toBe(false);
       expect(result.missingModelPricingError).toBe(false);
     });
 
@@ -515,6 +519,67 @@ commentary" has no AI credits pricing`;
       expect(result.missingModelPricingError).toBe(false);
       expect(result.missingModelPricingModelName).toBe("");
     });
+
+    it("detects max cache misses exceeded (JSON error type form)", () => {
+      const result = detectErrors('{"error":{"type":"max_cache_misses_exceeded","message":"Maximum consecutive cache misses exceeded (6 / 5).","consecutive_cache_misses":6,"max_cache_misses":5}}');
+      expect(result.maxCacheMissesExceeded).toBe(true);
+      expect(result.inferenceAccessError).toBe(false);
+      expect(result.invocationCapExceeded).toBe(false);
+    });
+
+    it("detects max cache misses exceeded (human-readable message form)", () => {
+      const result = detectErrors("Maximum consecutive cache misses exceeded");
+      expect(result.maxCacheMissesExceeded).toBe(true);
+      expect(result.inferenceAccessError).toBe(false);
+      expect(result.invocationCapExceeded).toBe(false);
+    });
+
+    it("detects max cache misses exceeded in a production log line", () => {
+      const log = '2026-07-30T06:14:50.000Z [ERROR] Error in API request: 403 {"error":{"type":"max_cache_misses_exceeded","message":"Maximum consecutive cache misses exceeded (6 / 5).","consecutive_cache_misses":6,"max_cache_misses":5}}';
+      const result = detectErrors(log);
+      expect(result.maxCacheMissesExceeded).toBe(true);
+    });
+
+    it("does not false-positive on unrelated cache miss content", () => {
+      const result = detectErrors("Cache miss for key: model-output-xyz");
+      expect(result.maxCacheMissesExceeded).toBe(false);
+    });
+  });
+
+  describe("MAX_CACHE_MISSES_EXCEEDED_PATTERN", () => {
+    it("matches max_cache_misses_exceeded error type", () => {
+      expect(MAX_CACHE_MISSES_EXCEEDED_PATTERN.test('{"type":"max_cache_misses_exceeded"}')).toBe(true);
+    });
+
+    it("matches Maximum consecutive cache misses exceeded message", () => {
+      expect(MAX_CACHE_MISSES_EXCEEDED_PATTERN.test("Maximum consecutive cache misses exceeded")).toBe(true);
+    });
+
+    it("is case-insensitive", () => {
+      expect(MAX_CACHE_MISSES_EXCEEDED_PATTERN.test("MAXIMUM CONSECUTIVE CACHE MISSES EXCEEDED")).toBe(true);
+    });
+
+    it("does not match unrelated cache miss content", () => {
+      expect(MAX_CACHE_MISSES_EXCEEDED_PATTERN.test("Cache miss for key: output")).toBe(false);
+    });
+  });
+
+  describe("isMaxCacheMissesExceededError", () => {
+    it("returns false for empty input", () => {
+      expect(isMaxCacheMissesExceededError("")).toBe(false);
+    });
+
+    it("detects max_cache_misses_exceeded JSON error type", () => {
+      expect(isMaxCacheMissesExceededError('{"error":{"type":"max_cache_misses_exceeded"}}')).toBe(true);
+    });
+
+    it("detects human-readable message form", () => {
+      expect(isMaxCacheMissesExceededError("Maximum consecutive cache misses exceeded")).toBe(true);
+    });
+
+    it("returns false for unrelated content", () => {
+      expect(isMaxCacheMissesExceededError("Some unrelated error message")).toBe(false);
+    });
   });
 
   describe("buildOutputLines", () => {
@@ -542,6 +607,7 @@ commentary" has no AI credits pricing`;
         http400ResponseError: false,
         capiQuotaExceededError: false,
         invocationCapExceeded: false,
+        maxCacheMissesExceeded: false,
         missingModelPricingError: true,
         missingModelPricingModelName: "claude-opus-5",
       });
@@ -559,12 +625,47 @@ commentary" has no AI credits pricing`;
         http400ResponseError: false,
         capiQuotaExceededError: false,
         invocationCapExceeded: false,
+        maxCacheMissesExceeded: false,
         missingModelPricingError: false,
         missingModelPricingModelName: "",
       });
 
       expect(lines).toContain("missing_model_pricing_error=false");
       expect(lines).toContain("missing_model_pricing_model_name=");
     });
+
+    it("emits max_cache_misses_exceeded=true when detected", () => {
+      const lines = buildOutputLines({
+        inferenceAccessError: false,
+        mcpPolicyError: false,
+        agenticEngineTimeout: false,
+        modelNotSupportedError: false,
+        http400ResponseError: false,
+        capiQuotaExceededError: false,
+        invocationCapExceeded: false,
+        maxCacheMissesExceeded: true,
+        missingModelPricingError: false,
+        missingModelPricingModelName: "",
+      });
+
+      expect(lines).toContain("max_cache_misses_exceeded=true");
+    });
+
+    it("emits max_cache_misses_exceeded=false when not detected", () => {
+      const lines = buildOutputLines({
+        inferenceAccessError: false,
+        mcpPolicyError: false,
+        agenticEngineTimeout: false,
+        modelNotSupportedError: false,
+        http400ResponseError: false,
+        capiQuotaExceededError: false,
+        invocationCapExceeded: false,
+        maxCacheMissesExceeded: false,
+        missingModelPricingError: false,
+        missingModelPricingModelName: "",
+      });
+
+      expect(lines).toContain("max_cache_misses_exceeded=false");
+    });
   });
 });