You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fix the push_repo_memory patch-size cap breach in Metrics Collector — the run wrote a 131 KB memory patch against a 10 KB configured limit (12 KB with overhead), a 13x overshoot that hard-failed the job.
audit on the failed run surfaces the exact gate error: Patch additions size (131 KB, 133495 bytes) exceeds maximum allowed size (12 KB, 12288 bytes, configured limit: 10 KB with 20% overhead allowance).
audit-diff between the failed run and the prior success shows core_consumed jumped +272% (557 → 2072 GitHub API points) with the same 12 total API calls — the agent's per-call data volume grew sharply, not the call count.
The daily metrics write ballooned far past its historical size — likely a larger-than-usual metrics/daily/YYYY-MM-DD.json payload (possibly from an added metrics field, a backfill, or a fetch-loop pulling more source data — consistent with the 272% API-usage spike) that blew through the 10 KB patch cap before the existing 30-day pruning (metrics-collector.md:211) had a chance to offset it.
Separately, audit why run 30782232562 emitted a payload ~13x the norm: diff metrics/daily/2026-08-03.json size against the prior week's files to confirm whether this is a one-off spike or a durable growth trend before just raising the cap.
If the growth is durable, add active pruning/summarization of metrics/daily/*.json (not just the 30-day delete) so the cap doesn't need repeated bumps.
Success criteria
Metrics Collector's push_repo_memory step completes without a patch-size gate failure on the next 3 consecutive scheduled runs.
max_patch_size change is reflected in both .github/workflows/metrics-collector.md and its compiled .lock.yml.
No regression in the 30-day metrics/daily/ pruning behavior.
Parent: #49835 · Analyzed runs: 30782232562 (failed), 30730908285 (comparator, success)
Related to #49835
Problem
Fix the
push_repo_memorypatch-size cap breach in Metrics Collector — the run wrote a 131 KB memory patch against a 10 KB configured limit (12 KB with overhead), a 13x overshoot that hard-failed the job.Affected workflow: Metrics Collector (
.github/workflows/metrics-collector.lock.yml, jobpush_repo_memory)Run analyzed: 30782232562 (2026-08-03T03:36Z, failed)
Comparator run: 30730908285 (2026-08-02T03:36Z, succeeded)
Evidence
auditon the failed run surfaces the exact gate error:Patch additions size (131 KB, 133495 bytes) exceeds maximum allowed size (12 KB, 12288 bytes, configured limit: 10 KB with 20% overhead allowance).audit-diffbetween the failed run and the prior success showscore_consumedjumped +272% (557 → 2072 GitHub API points) with the same 12 total API calls — the agent's per-call data volume grew sharply, not the call count.max_patch_size: 10240is set in.github/workflows/metrics-collector.lock.yml:646andMAX_PATCH_SIZE: 10240env at line 1751 — unchanged since prior fixes for other workflows (Prevent Agent Performance Analyzer repo-memory push failures by raising patch-size cap #48012, fix(daily-code-metrics): raise max-file-size to 256KB and enforce 90-day history.jsonl pruning #43870 raised caps elsewhere but Metrics Collector's own cap was never revisited).Probable root cause
The daily metrics write ballooned far past its historical size — likely a larger-than-usual
metrics/daily/YYYY-MM-DD.jsonpayload (possibly from an added metrics field, a backfill, or a fetch-loop pulling more source data — consistent with the 272% API-usage spike) that blew through the 10 KB patch cap before the existing 30-day pruning (metrics-collector.md:211) had a chance to offset it.Proposed remediation
max_patch_sizefor Metrics Collector'spush_repo_memoryconfig (metrics-collector.lock.yml:646, and the corresponding.mdsource) from 10 KB to a value sized to the actual daily payload plus headroom (mirror the Prevent Agent Performance Analyzer repo-memory push failures by raising patch-size cap #48012/fix(daily-code-metrics): raise max-file-size to 256KB and enforce 90-day history.jsonl pruning #43870 pattern of cap increases) — e.g. 32-64 KB.metrics/daily/2026-08-03.jsonsize against the prior week's files to confirm whether this is a one-off spike or a durable growth trend before just raising the cap.metrics/daily/*.json(not just the 30-day delete) so the cap doesn't need repeated bumps.Success criteria
push_repo_memorystep completes without a patch-size gate failure on the next 3 consecutive scheduled runs.max_patch_sizechange is reflected in both.github/workflows/metrics-collector.mdand its compiled.lock.yml.metrics/daily/pruning behavior.Parent: #49835 · Analyzed runs: 30782232562 (failed), 30730908285 (comparator, success)
Related to #49835