Fix credit-ceiling handling: exit 0 when AI credits budget is enforced

.github/workflows/daily-max-ai-credits-test.md

@@ -1,9 +1,8 @@
 ---
 private: true
 emoji: "🧪"
-description: "⚠️ INTENTIONALLY FAILS — Tests that max-ai-credits: 1 is enforced by the AWF firewall and that the per-run budget guardrail cuts off the agent."
+description: "Tests that max-ai-credits: 1 is enforced by the AWF firewall and that the per-run budget guardrail cuts off the agent. Concludes success when the credit limit is reached."
 features:
-  intentional-failure: true
   gh-aw-detection: true
 on:
   schedule: daily around 10:30
@@ -30,32 +29,35 @@ safe-outputs:
     close-older-key: "daily-max-ai-credits-test"
     labels: [automation, testing]
     max: 1
+  report-failure-as-issue:
+    - "!ai_credits_rate_limit_error"
+    - "!max_ai_credits_exceeded"
   messages:
-    run-started: "🧪 [{workflow_name}]({run_url}) — per-run AI credit limit test running (intentionally fails, limit: 1 AI credit/run)."
-    run-success: "⚠️ [{workflow_name}]({run_url}) completed without hitting the per-run limit of 1 AI credit — verify that max-ai-credits enforcement is working."
-    run-failure: "🚫 [{workflow_name}]({run_url}) {status} — expected: the per-run AI credit limit of 1 was reached and the AWF firewall cut off the agent."
+    run-started: "🧪 [{workflow_name}]({run_url}) — per-run AI credit limit test running (limit: 1 AI credit/run)."
+    run-success: "🧪 [{workflow_name}]({run_url}) — expected: the per-run AI credit limit of 1 was reached and the AWF firewall cut off the agent."
+    run-failure: "⚠️ [{workflow_name}]({run_url}) {status} — completed without hitting the per-run limit of 1 AI credit — verify that max-ai-credits enforcement is working."
 ---
 
-### Daily Max AI Credits Test (Intentionally Fails)
+### Daily Max AI Credits Test
 
 **Report Formatting**: Use h3 (###) or lower for all headers in your report
 to maintain proper document hierarchy. Wrap long sections in
 `<details><summary>View Full Details</summary>` tags to improve readability.
 
 
-> ⚠️ **This workflow is intentionally broken.** It exists solely to verify that
+> 🧪 **This workflow tests the per-run AI credits guardrail.** It verifies that
 > `max-ai-credits: 1` is enforced by the AWF firewall and that the agent is
-> cut off when the per-run budget is exhausted.
-> **Do not fix this workflow.**
+> cut off when the per-run budget is exhausted. The run **concludes success**
+> when the credit limit is reached as expected.
 
 #### What This Tests
 
 1. The AWF firewall enforces the `max-ai-credits` per-run budget.
 2. Once the agent consumes more than 1 AI credit in a single run, the firewall cuts off the LLM API.
 3. The prompt forces multiple turns and multiple large-file reads so the run reliably burns credits.
-4. The run is expected to be cut off by the per-run budget before all jobs can complete.
+4. The run is expected to be cut off by the per-run budget before all turns complete.
 
-#### Task (broken by design)
+#### Task
 
 Use **at least four separate assistant turns**. Do not combine all work into one response.
 
@@ -70,7 +72,7 @@ Turn 4: Call `noop` with the message: "Completed max-ai-credits multi-turn guard
 After each job, briefly summarize what was read, then continue to the next turn.
 
 Since the per-run budget is `max-ai-credits: 1`, the AWF firewall should cut off the agent
-before all turns complete. That failure is the expected and correct outcome.
+before all turns complete. That is the expected and correct outcome — the run will conclude success.
 
 If the workflow somehow completes without hitting the per-run limit, call `noop` with the message:
 "Per-run credit limit not exceeded — verify that max-ai-credits: 1 is enforced by the AWF firewall."

actions/setup/js/check_daily_aic_workflow_guardrail.cjs

@@ -703,12 +703,16 @@ async function main() {
     } catch (summaryError) {
       core.warning(`Failed to write daily AIC summary: ${getErrorMessage(summaryError)}`);
     }
-    core.warning(`Daily workflow AIC guardrail exceeded for ${workflowName}: ${totalAIC}/${threshold}.`);
-    core.setFailed(`Daily workflow AIC guardrail exceeded for ${workflowName}: ${totalAIC}/${threshold}.`);
+    // Log as info so the activation job succeeds. The daily_ai_credits_exceeded output
+    // is already set to "true"; the agent job's condition (daily_ai_credits_exceeded != 'true')
+    // will skip the agent, and the conclusion job will handle reporting via the
+    // daily_ai_credits_exceeded flag. Failing the activation job here causes the overall
+    // workflow to fail even though hitting the daily limit is an expected, graceful outcome.
+    core.info(`Daily workflow AIC guardrail exceeded for ${workflowName}: ${totalAIC}/${threshold}.`);
   } catch (error) {
     // Treat unexpected guardrail execution errors as non-blocking skips so transient
     // API/runtime issues do not fail activation. The output stays at the default "false",
-    // allowing the agent to run. Legitimate threshold exceedance still fails via setFailed.
+    // allowing the agent to run.
     core.warning(`Daily workflow AI Credits guardrail encountered an unexpected error and will be skipped: ${getErrorMessage(error)}`);
   }
 }

actions/setup/js/check_daily_aic_workflow_guardrail.test.cjs

@@ -478,7 +478,7 @@ describe("check_daily_aic_workflow_guardrail", () => {
     }
   });
 
-  it("main() marks the step failed when the daily AI Credits guardrail is exceeded", async () => {
+  it("main() does not mark the step failed when the daily AI Credits guardrail is exceeded", async () => {
     const getRunAICSpy = vi.spyOn(exports, "getRunAIC").mockResolvedValue(200);
 
     const coreOutputs = {};
@@ -558,8 +558,7 @@ describe("check_daily_aic_workflow_guardrail", () => {
       expect(coreOutputs["daily_ai_credits_exceeded"]).toBe("true");
       expect(coreOutputs["daily_ai_credits_total_effective_tokens"]).toBe("200");
       expect(coreOutputs["daily_ai_credits_threshold"]).toBe("100");
-      expect(setFailed).toHaveBeenCalledTimes(1);
-      expect(setFailed.mock.calls[0][0]).toMatch(/guardrail exceeded/i);
+      expect(setFailed).not.toHaveBeenCalled();
     } finally {
       delete global.core;
       delete global.github;

actions/setup/js/claude_harness.cjs

@@ -54,6 +54,7 @@ const { countPermissionDeniedIssues, hasNumerousPermissionDeniedIssues, extractD
 const { detectNonRetryableHarnessGuard, buildSoftTimeoutGuard, emitSoftTimeoutSignal, isAuthenticationFailedError } = require("./harness_retry_guard.cjs");
 const { MODEL_NOT_SUPPORTED_PATTERN: INVALID_MODEL_ERROR_PATTERN } = require("./detect_agent_errors.cjs");
 const { applyModelFallback } = require("./model_fallback.cjs");
+const { parseMaxAICreditsExceededFromAuditLog } = require("./ai_credits_context.cjs");
 
 // Pattern to detect Anthropic API overload errors (HTTP 529).
 // Matches "overloaded_error" from the Anthropic error type field, and the
@@ -479,12 +480,26 @@ async function main() {
     }
 
     const nonRetryableGuard = detectNonRetryableHarnessGuard(result.output);
-    if (nonRetryableGuard.aiCreditsExceeded || nonRetryableGuard.awfAPIProxyBlockingRequests || nonRetryableGuard.maxRunsExceeded) {
+    const trustedAICreditsExceeded = nonRetryableGuard.aiCreditsExceeded && parseMaxAICreditsExceededFromAuditLog();
+    if (nonRetryableGuard.aiCreditsExceeded && !trustedAICreditsExceeded) {
+      log(`attempt ${attempt + 1}: AI credits marker found in CLI output without trusted firewall audit confirmation — preserving normal failure handling`);
+    }
+    const shouldTreatAICreditsExceededAsSuccess = trustedAICreditsExceeded && !isAuthenticationFailed;
+    if (shouldTreatAICreditsExceededAsSuccess || nonRetryableGuard.awfAPIProxyBlockingRequests || nonRetryableGuard.maxRunsExceeded) {
       const reasons = [];
-      if (nonRetryableGuard.aiCreditsExceeded) reasons.push("AI credits budget exceeded");
+      if (shouldTreatAICreditsExceededAsSuccess) reasons.push("AI credits budget exceeded");
       if (nonRetryableGuard.awfAPIProxyBlockingRequests) reasons.push("AWF API proxy is blocking requests");
       if (nonRetryableGuard.maxRunsExceeded) reasons.push("maximum LLM invocations exceeded");
       log(`attempt ${attempt + 1}: ${reasons.join(" and ")} — not retrying (non-retryable guard condition)`);
+      // When the per-run AI credits budget is exceeded the AWF firewall intentionally
+      // stopped the agent — this is controlled budget enforcement, not an unexpected
+      // error.  Exit 0 so the agent step and job succeed; the ai_credits_rate_limit_error
+      // output surfaced by parse-mcp-gateway will inform downstream handlers (e.g.
+      // handle_agent_failure) of the budget exceedance.
+      if (shouldTreatAICreditsExceededAsSuccess) {
+        log(`attempt ${attempt + 1}: AI credits budget enforced — exiting 0 (budget control, not an error)`);
+        lastExitCode = 0;
+      }
       break;
     }