perf-merged-fix-not-deployed-to-fleet-daemons
I diagnosed the problem class and produced a verified solution. Rather than treating this as a code bug, I confirmed it is a deploy-lag defect: the fix is correct and merged, but no daemon in the fleet contains it.
github.com/deployBunker/bunker) and located the actual fix: 651c4d8 (internal/server/diskusage.go, TTL snapshot cache).git merge-base --is-ancestor 651c4d8 16fff6d → exit 1 (deployed sandbox daemon)a6a52a1, 33c9e00 → exit 1; 098a087, d5e457b (HEAD) → exit 0main (/tmp/bunkerd-head, commit d5e457b) and ran the PERF-001 cache tests — all 4 pass.scripts/deploy-gate.sh, a fail-closed gate that resolves each host's running bunkerd revision, checks ancestry, emits a P1 board row, and runs an 8× TTL-plateau probe. Verified against pre/post-fix binaries, a simulated SSH remote path, and fake slow-vs-cached CLIs./workspace/solution.md — title, root-cause analysis, exact deploy commands, the full deploy-gate script, an optional wire-exposed commit hardening patch, and a verification/acceptance section./workspace/evidence.md — raw command transcripts backing every claim.~/bunker/scripts/deploy-gate.sh — the tested, ready-to-use gate.Key rule captured: after merging any perf fix to a daemon, re-measure against LIVE deployments before declaring the win; every deploy-lag row is P1/P2 with git merge-base --is-ancestor <fix> <deployed-revision> as its acceptance test.
# Evidence - Problem class: perf-merged-fix-not-deployed-to-fleet-daemons - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-26T05:10:30.853Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "Perf finding measured twice in one hunt (bunker fleet 2026-09-26): a merged perf fix shows ZERO live improvement because deployed daemons predate the fix commit. Dedication: PERF-001 TTL disk-usage cache merged 2026-09-24 (651c4d8), but 3/3 reachable bunker nodes ran pre-fix daemons (dedi-2 Sep 23 binary, mvp Sep 24 restart with Sep 23 binary, las-02 Sep 24 binary) so `bunker list` still paid the full synchronous walk server-side (dedi-2: 1.4s of 1.8s total, 114,774 files; mvp: ~340ms). Method that worked: (1) 8x repeat calls to detect a TTL plateau (absent => fix not live), (2) contrast host with smaller file count to prove linearity in files, (3) RTT + CLI-floor + on-box loopback decomposition to split network vs handler, (4) binary mtime / daemon ActiveEnterTimestamp vs fix-merge date, with git merge-base --is-ancestor as the deploy gate. Lesson: after merging any perf fix to a daemon, the perf lane must re-measure against LIVE deployments before declaring the win, and deploy-lag rows belong on the board as P1/P2 with the ancestor-check gate.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "perf-merged-fix-not-deployed-to-fleet-daemons", "provider": "openrouter", "solved_at": "2026-09-26T05:10:30.901Z", "version": ""}I diagnosed the problem class and produced a verified solution. Rather than treating this as a code bug, I confirmed it is a deploy-lag defect: the fix is correct and merged, but no daemon in the fleet contains it.
github.com/deployBunker/bunker) and located the actual fix: 651c4d8 (internal/server/diskusage.go, TTL snapshot cache).git merge-base --is-ancestor 651c4d8 16fff6d → exit 1 (deployed sandbox daemon)a6a52a1, 33c9e00 → exit 1; 098a087, d5e457b (HEAD) → exit 0main (/tmp/bunkerd-head, commit d5e457b) and ran the PERF-001 cache tests — all 4 pass.scripts/deploy-gate.sh, a fail-closed gate that resolves each host's running bunkerd revision, checks ancestry, emits a P1 board row, and runs an 8× TTL-plateau probe. Verified against pre/post-fix binaries, a simulated SSH remote path, and fake slow-vs-cached CLIs./workspace/solution.md — title, root-cause analysis, exact deploy commands, the full deploy-gate script, an optional wire-exposed commit hardening patch, and a verification/acceptance section./workspace/evidence.md — raw command transcripts backing every claim.~/bunker/scripts/deploy-gate.sh — the tested, ready-to-use gate.Key rule captured: after merging any perf fix to a daemon, re-measure against LIVE deployments before declaring the win; every deploy-lag row is P1/P2 with git merge-base --is-ancestor <fix> <deployed-revision> as its acceptance test.
# Evidence - Problem class: perf-merged-fix-not-deployed-to-fleet-daemons - Model: openrouter/deepseek/deepseek-v4.1-flash - Solved: 2026-09-26T05:10:30.853Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "Perf finding measured twice in one hunt (bunker fleet 2026-09-26): a merged perf fix shows ZERO live improvement because deployed daemons predate the fix commit. Dedication: PERF-001 TTL disk-usage cache merged 2026-09-24 (651c4d8), but 3/3 reachable bunker nodes ran pre-fix daemons (dedi-2 Sep 23 binary, mvp Sep 24 restart with Sep 23 binary, las-02 Sep 24 binary) so `bunker list` still paid the full synchronous walk server-side (dedi-2: 1.4s of 1.8s total, 114,774 files; mvp: ~340ms). Method that worked: (1) 8x repeat calls to detect a TTL plateau (absent => fix not live), (2) contrast host with smaller file count to prove linearity in files, (3) RTT + CLI-floor + on-box loopback decomposition to split network vs handler, (4) binary mtime / daemon ActiveEnterTimestamp vs fix-merge date, with git merge-base --is-ancestor as the deploy gate. Lesson: after merging any perf fix to a daemon, the perf lane must re-measure against LIVE deployments before declaring the win, and deploy-lag rows belong on the board as P1/P2 with the ancestor-check gate.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4.1-flash", "problem_class": "perf-merged-fix-not-deployed-to-fleet-daemons", "provider": "openrouter", "solved_at": "2026-09-26T05:10:30.901Z", "version": ""}