fix(autofix): count only Gate failures with failed jobs

.github/sync-manifest.yml

@@ -48,7 +48,7 @@ workflows:
       - stranske/Template
 
   - source: .github/workflows/autofix.yml
-    description: "Autofix workflow - automatically fixes lint/format issues"
+    description: "Autofix workflow - repairs failed/timed-out checks; cancellations do not trigger repairs"
 
   - source: .github/workflows/pr-46-dependency-repair-contract.yml
     description: "Dependency repair contract - keeps bot PRs bot-owned and validates agent repair promotion provenance"
@@ -60,7 +60,7 @@ workflows:
     description: "PR event hub - consolidates PR meta, bot comments, and verify-to-issue handlers"
 
   - source: .github/workflows/agents-81-gate-followups.yml
-    description: "Gate followups hub - consolidates keepalive and autofix followups"
+    description: "Gate followups hub - keepalive and autofix; budgets count Gate failures with failed jobs"
 
   - source: .github/workflows/agents-keepalive-sweep.yml
     description: "Keepalive sweep - periodic level-based resync; dispatches the loop for open agent PRs so silent zero-commit stalls resurface"

.github/workflows/agents-autofix-loop.yml

@@ -354,8 +354,10 @@ jobs:
             outputs.gate_conclusion = String(run?.conclusion || run?.status || '');
             outputs.gate_run_id = String(run?.id || manualInputs.runId || '');
 
-            if ((run.conclusion || '').toLowerCase() === 'success') {
-              return stop('upstream Gate succeeded');
+            const isFailure = (value) =>
+              ['failure', 'timed_out'].includes(String(value || '').toLowerCase());
+            if (!isFailure(run.conclusion)) {
+              return stop('upstream Gate has no counted failure', 'gate_not_failed');
             }
 
             if ((run.event || '').toLowerCase() !== 'pull_request') {
@@ -494,54 +496,6 @@ jobs:
               ? configMatch[1].toLowerCase() === 'true'
               : hasExplicitAgentLabel;
 
-            // Auto-escalation: Escalate to Codex CLI when Gate fails
-            // Triggers if: (1) basic autofix ran but insufficient, OR (2) no basic autofix applied
-            // Note: We do NOT add agent:codex label here because that triggers external Codex UI
-            // which would conflict with our internal Codex CLI run. Only add autofix:escalated.
-            if (!autofixEnabled && !configMatch) {
-              const hasAutofixLabel =
-                labels.includes('autofix:applied') || labels.includes('autofix');
-              const hasEscalatedLabel = labels.includes('autofix:escalated');
-              const gateConclusion = (run.conclusion || '').toLowerCase();
-              const gateFailed = gateConclusion === 'failure';
-
-              // Escalate if Gate failed and we haven't already escalated
-              if (gateFailed && !hasEscalatedLabel) {
-                const reason = hasAutofixLabel
-                  ? 'Basic autofix ran but Gate still failing'
-                  : 'No basic autofix (non-Python PR?) and Gate failing';
-                core.info(`🔄 Auto-escalation: ${reason}. Escalating to Codex CLI...`);
-                try {
-                  await withRetry(() =>
-                    github.rest.issues.addLabels({
-                      owner,
-                      repo,
-                      issue_number: prNumber,
-                      labels: ['autofix:escalated'],
-                    })
-                  );
-                  core.info(
-                    '✅ Added autofix:escalated label - Codex CLI will run in this workflow'
-                  );
-                  autofixEnabled = true;
-                } catch (error) {
-                  core.warning(`Failed to add escalation labels: ${error.message}`);
-                }
-              }
-            }
-            if (!autofixEnabled) {
-              return stop('autofix disabled for this pull request');
-            }
-
-            // Phase 2: Support both Codex and Claude autofix
-            const supportedAgents = ['codex', 'claude'];
-            if ((outputs.agent_type || '') && !supportedAgents.includes(outputs.agent_type)) {
-              return stop(
-                `unsupported agent type for autofix loop: ${outputs.agent_type}`,
-                'unsupported_agent'
-              );
-            }
-
             const jobs = await paginateWithRetry(
               github,
               github.rest.actions.listJobsForWorkflowRun,
@@ -553,7 +507,8 @@ jobs:
               }
             );
 
-            const workflowFile = 'agents-autofix-loop.yml';
+            // Count failed Gate executions, not unrelated follow-up workflow runs.
+            const workflowFile = run.workflow_id;
             // Reduce attempts for auto-escalated PRs (they weren't agent-initiated)
             const isEscalated = labels.includes('autofix:escalated');
             const maxAttempts = isEscalated
@@ -572,7 +527,82 @@ jobs:
               }
             );
 
-            const attemptCount = previousRuns.length + 1;
+            const listJobsForGateAttempt = async (runId, attemptNumber, latestJobs = null) => {
+              const currentAttempt = Math.max(1, Number(run.run_attempt) || 1);
+              if (runId === run.id && attemptNumber === currentAttempt && latestJobs) {
+                return latestJobs;
+              }
+              const listAttemptJobs = github.rest.actions.listJobsForWorkflowRunAttempt;
+              if (typeof listAttemptJobs === 'function') {
+                return await paginateWithRetry(github, listAttemptJobs, {
+                  owner,
+                  repo,
+                  run_id: runId,
+                  attempt_number: attemptNumber,
+                  per_page: 100,
+                });
+              }
+              if (attemptNumber === 1) {
+                return await paginateWithRetry(
+                  github,
+                  github.rest.actions.listJobsForWorkflowRun,
+                  {
+                    owner,
+                    repo,
+                    run_id: runId,
+                    per_page: 100,
+                  }
+                );
+              }
+              return [];
+            };
+
+            const countFailingAttemptsForRun = async (gateRun, latestJobs = null) => {
+              const maxAttempt = Math.max(1, Number(gateRun.run_attempt) || 1);
+              let count = 0;
+              for (let attempt = 1; attempt <= maxAttempt; attempt += 1) {
+                const attemptJobs = await listJobsForGateAttempt(
+                  gateRun.id,
+                  attempt,
+                  gateRun.id === run.id ? latestJobs : null
+                );
+                if (attemptJobs.some((job) => isFailure(job.conclusion))) {
+                  count += 1;
+                }
+              }
+              return count;
+            };
+
+            const gateRunsById = new Map();
+            for (const candidate of previousRuns) {
+              if (candidate.head_sha !== run.head_sha || !isFailure(candidate.conclusion)) {
+                continue;
+              }
+              const existing = gateRunsById.get(candidate.id);
+              const candidateAttempt = Number(candidate.run_attempt) || 1;
+              const existingAttempt = existing ? Number(existing.run_attempt) || 1 : 0;
+              if (!existing || candidateAttempt >= existingAttempt) {
+                gateRunsById.set(candidate.id, candidate);
+              }
+            }
+            gateRunsById.set(run.id, {
+              ...run,
+              run_attempt: Math.max(
+                Number(run.run_attempt) || 1,
+                Number(gateRunsById.get(run.id)?.run_attempt) || 0
+              ),
+            });
+
+            let attemptsWithFailingJobs = 0;
+            let gateAttemptsExamined = 0;
+            for (const gateRun of gateRunsById.values()) {
+              gateAttemptsExamined += Math.max(1, Number(gateRun.run_attempt) || 1);
+              attemptsWithFailingJobs += await countFailingAttemptsForRun(
+                gateRun,
+                gateRun.id === run.id ? jobs : null
+              );
+            }
+            const attemptCount = attemptsWithFailingJobs;
             outputs.attempts = String(attemptCount);
             outputs.max_attempts = String(maxAttempts);
 
@@ -581,7 +611,7 @@ jobs:
             let triggerStep = null;
             for (const job of jobs) {
               const conclusion = (job.conclusion || job.status || '').toLowerCase();
-              if (!conclusion || ['success', 'skipped'].includes(conclusion)) {
+              if (!isFailure(conclusion)) {
                 continue;
               }
 
@@ -590,7 +620,7 @@ jobs:
                   ? job.steps
                       .filter((step) => {
                         const stepConclusion = (step.conclusion || step.status || '').toLowerCase();
-                        return stepConclusion && !['success', 'skipped'].includes(stepConclusion);
+                        return isFailure(stepConclusion);
                       })
                       .map(
                         (step) =>
@@ -609,7 +639,7 @@ jobs:
                 const failingStep = Array.isArray(job.steps)
                   ? job.steps.find((step) => {
                       const stepConclusion = (step.conclusion || step.status || '').toLowerCase();
-                      return stepConclusion && !['success', 'skipped'].includes(stepConclusion);
+                      return isFailure(stepConclusion);
                     })
                   : null;
                 triggerStep = failingStep || null;
@@ -645,6 +675,7 @@ jobs:
               `PR: #${prNumber}`,
               `Head SHA: ${headSha}`,
               `Autofix attempts for this head: ${attemptCount} / ${maxAttempts}`,
+              `Gate runs examined: ${gateRunsById.size}; gate attempts examined: ${gateAttemptsExamined}; attempts with failing jobs: ${attemptsWithFailingJobs}`,
               'Fix scope: src/, tests/, tools/, scripts/, agents/, templates/, .github/',
             ];
 
@@ -656,6 +687,55 @@ jobs:
 
             outputs.appendix = appendixLines.join('\n');
 
+            if (failingJobs.length === 0) {
+              return stop('Gate has no failing jobs', 'no_failing_jobs');
+            }
+
+            // Auto-escalation: only after confirming real failed/timed-out Gate jobs.
+            // Note: We do NOT add agent:codex label here because that triggers external Codex UI
+            // which would conflict with our internal Codex CLI run. Only add autofix:escalated.
+            if (!autofixEnabled && !configMatch) {
+              const hasAutofixLabel =
+                labels.includes('autofix:applied') || labels.includes('autofix');
+              const hasEscalatedLabel = labels.includes('autofix:escalated');
+              const gateFailed = isFailure((run.conclusion || '').toLowerCase());
+
+              if (gateFailed && !hasEscalatedLabel) {
+                const reason = hasAutofixLabel
+                  ? 'Basic autofix ran but Gate still failing'
+                  : 'No basic autofix (non-Python PR?) and Gate failing';
+                core.info(`🔄 Auto-escalation: ${reason}. Escalating to Codex CLI...`);
+                try {
+                  await withRetry(() =>
+                    github.rest.issues.addLabels({
+                      owner,
+                      repo,
+                      issue_number: prNumber,
+                      labels: ['autofix:escalated'],
+                    })
+                  );
+                  core.info(
+                    '✅ Added autofix:escalated label - Codex CLI will run in this workflow'
+                  );
+                  autofixEnabled = true;
+                } catch (error) {
+                  core.warning(`Failed to add escalation labels: ${error.message}`);
+                }
+              }
+            }
+            if (!autofixEnabled) {
+              return stop('autofix disabled for this pull request');
+            }
+
+            // Phase 2: Support both Codex and Claude autofix
+            const supportedAgents = ['codex', 'claude'];
+            if ((outputs.agent_type || '') && !supportedAgents.includes(outputs.agent_type)) {
+              return stop(
+                `unsupported agent type for autofix loop: ${outputs.agent_type}`,
+                'unsupported_agent'
+              );
+            }
+
             if (attemptCount > maxAttempts) {
               return stop(
                 `autofix attempt limit reached (${attemptCount} > ${maxAttempts})`,

.github/workflows/autofix.yml

@@ -56,7 +56,7 @@ jobs:
           custom-predicate: >-
             (event_name == 'pull_request_target' && action == 'labeled' &&
              (label.name == 'autofix' || label.name == 'autofix:clean')) ||
-            (event_name == 'workflow_run' && workflow_run.conclusion == 'failure' &&
+            (event_name == 'workflow_run' && (workflow_run.conclusion == 'failure' || workflow_run.conclusion == 'timed_out') &&
              length(workflow_run.pull_requests) > 0)
 
       - name: Checkout for API helpers
@@ -124,7 +124,7 @@ jobs:
               );
 
               const failedChecks = (data.check_runs || []).filter(
-                (cr) => (cr.conclusion || '').toLowerCase() === 'failure'
+                (cr) => ['failure', 'timed_out'].includes((cr.conclusion || '').toLowerCase())
               );
 
               const relevantFailures = failedChecks.filter((cr) => {
@@ -157,7 +157,7 @@ jobs:
                 { maxRetries: 3 }
               );
               const failedJobs = jobs.filter(
-                (job) => (job.conclusion || '').toLowerCase() === 'failure'
+                (job) => ['failure', 'timed_out'].includes((job.conclusion || '').toLowerCase())
               );
               const relevantFailures = failedJobs.filter((job) => {
                 const name = String(job.name || '').toLowerCase();
@@ -222,7 +222,7 @@ jobs:
               const triggerHeadSha = String(run?.head_sha || run?.head_commit?.id || '');
               const runId = Number(run?.id || 0);
 
-              if (run.conclusion !== 'failure') {
+              if (!['failure', 'timed_out'].includes(run.conclusion)) {
                 core.info(
                   `${workflowName} concluded '${run.conclusion}' — no autofix needed`
                 );

config/template-drift-allowlist.txt

@@ -189,11 +189,11 @@ fingerprint_refreshed = 2026-08-24
 [pair.17]
 main = .github/workflows/autofix.yml
 template = templates/consumer-repo/.github/workflows/autofix.yml
-main_sha256 = 81f2268c29cc8b5da388a44eab46a8613df153d9e4e83449ab3a5070c03558a7
-template_sha256 = 88cf48708d972137ef8e9fb3efc2323ee93c8cab16eaba0fc22c8c11ea6695c4
-divergence = Intentional divergence reviewed 2026-08-23: the Workflows source runs its local reusable autofix workflow and uses the repository's CI concurrency/context contract; the consumer template calls the published reusable workflow with SHA-pinned actions, SERVICE_BOT_PAT, and workflow-run PR context recovery. Do not align wholesale because the consumer is a deployable wrapper rather than the source implementation.
+main_sha256 = 96c8131203431b7da850474e06fa169a2580fa5fa4759674be712ea408f17d16
+template_sha256 = 73536de641ae21b89b5de24848a850e2753f5bc4d608f90498b44007fc2b8cfe
+divergence = Timeout handling 2026-09-14: four identical conclusion-filter edits applied to source and consumer autofix; both still exclude cancellations. Verified the exact same transformation against origin/main on both sides. Fingerprints refreshed; existing divergence and its review date are unchanged. Prior divergence: Intentional divergence reviewed 2026-08-23: the Workflows source runs its local reusable autofix workflow and uses the repository's CI concurrency/context contract; the consumer template calls the published reusable workflow with SHA-pinned actions, SERVICE_BOT_PAT, and workflow-run PR context recovery. Do not align wholesale because the consumer is a deployable wrapper rather than the source implementation.
 divergence_reviewed = 2026-08-23
-fingerprint_refreshed = 2026-08-23
+fingerprint_refreshed = 2026-09-14
 
 [pair.18]
 main = .github/workflows/maint-coverage-guard.yml

docs/WORKFLOW_GUIDE.md

@@ -142,6 +142,13 @@ _Inline Gate helper_
 - **`agents-autofix-rebase.yml`** — See Agents section; merges the base branch when Gate flags stale branches, preventing common “needs rebase” failures from blocking progress.
 - **`agents-bot-comment-autolabel.yml`** / **`agents-bot-comment-handler.yml`** — Automatically harvest trusted bot review comments and apply them via the autofix pathways without manual labelling.
 
+Autofix budgets count only same-head Gate runs concluded `failure` or `timed_out`
+with at least one job carrying either conclusion. Cancelled, skipped, successful,
+and jobless runs do not consume the budget or trigger `needs-human`. The heavy
+loop reports the number of Gate failures examined and the attempts with failing
+jobs; a current Gate without a failed job cannot escalate even after earlier
+failures. Consumer Gate Followups uses the same rule.
+
 #### Autofix & Lint Coordination
 1. **Gate emits signals** — `pr-00-gate.yml` attaches artifacts describing the failing job plus `mergeable_state`. When lint/format/typecheck/test jobs fail it dispatches `autofix_gate_failure`; when the PR is dirty/behind it dispatches `autofix_rebase_needed`.
 2. **CI autofix first pass (`autofix.yml`)** — Runs Ruff/formatters/tests where possible and pushes fixes directly to the branch so same-run Gate retries can pass without escalation.