AlgoSu TechKO

Clearing Carryover Work (Critic SIGTERM root cause · seed representativeness · 256~259 ADR gap)

SummaryCleared three carryover items accumulated across 255~263 in one pass. (1) Identified the root cause of the Critic gate's repeated ACP SIGTERM interruptions — codex was launched *inside* the ACP session process tree, so the child was reaped (SIGTERM) when the turn ended; this was already structurally resolved by Sprint 262's detached-cron redesign (0 SIGTERM across 8 post-redesign PRs #507·#509~515; the observed signal 13 was benign SIGPIPE from codex's internal grep). Added ensure-critic-crons.sh for idempotent re-registration against cron loss. (2) Re-reviewed BOJ recommendation seed representativeness → #521 `10e4c377` (12→23); the Critic caught 3 tier mislabels vs solved.ac ground truth (15649 GOLD→SILVER, 1005·2098 PLATINUM→GOLD), exposing 0 real PLATINUM → added 3 verified PLATINUM seeds (3653·11375·1761) to meet item2's goal, plus a Trivy brace-expansion 5.0.7→5.0.8 override. (3) 256~259 retrospective ADR gap → #522 `eb338f75` (KR+EN, index 201). A parallel Oracle session repeatedly pre-applied the same fixes — discriminated by measurement only, avoiding rework.

Date
ImpactCritical

Date: 2026-07-28

Goal

Clear three items carried over across 255~263 in a single Sprint 264 pass. The scope is accumulated-debt cleanup, not feature development.

Target (user-confirmed — all three)

  • item1 — Identify the root cause of the Critic gate's repeated ACP SIGTERM interruptions (carried 261→263)
  • item2 — Re-review the actual representativeness of the BOJ/Programmers recommendation seed list (carried 255~261)
  • item3 — Close the 256~259 retrospective ADR gap (write 4)

Decisions

D1. Critic SIGTERM root cause — in-session launch → turn reap (item1, already structurally resolved)

The root cause of the Critic gate's past repeated SIGTERM interruptions was that codex was launched inside the ACP session process tree, so the child was reaped (signal 15) together when the session turn ended. This was already structurally resolved by Sprint 262's detached-cron redesign (launch-critic.sh → queue marker → critic-runner cron launches codex detached, outside the ACP session). Measuring the 8 post-redesign PR logs (#507·#509~515) showed 0 SIGTERM; the observed signal 13 was benign SIGPIPE from codex's internal grep pipe (review completed normally). In other words, item1 was not "unresolved-and-open" but "resolved-yet-undocumented", and this sprint confirmed it empirically.

D2. Idempotent re-registration against cron loss — ensure-critic-crons.sh (item1 residual risk)

If any of the three detached crons (critic-runner·critic-gate-watchdog·merge-watch) becomes unregistered or lost, the gate stops silently (more dangerous than SIGTERM — no alert at all). To eliminate this single point of failure, ensure-critic-crons.sh (idempotent re-registration) was added as an Oracle-profile operations script. It lives outside this repo (profile scripts) and is not a PR target.

D3. Seed representativeness — correct to solved.ac ground-truth tiers + secure real PLATINUM (item2, #521)

BOJ recommendation seeds were expanded 12→23 (to reinforce GOLD/PLATINUM fallback), but the Critic gate caught 3 tier mislabels (P2) vs solved.ac ground truth: 15649 (N and M (1)) actually SILVER but labeled GOLD; 1005 (ACM Craft)·2098 (Traveling Salesman) actually GOLD but labeled PLATINUM. These mislabels left 0 real PLATINUM, so item2's goal (securing high-tier fallback) itself was unmet. → Corrected to ground-truth tiers and added 3 verified PLATINUM seeds (3653·11375·1761). As a side fix, the Trivy-flagged brace-expansion 5.0.7 CVE-2026-14257 (transitive HIGH) was resolved via a 5.0.8 override.

D4. Closing the 256~259 retrospective ADR gap (item3, #522)

The 256~259 retrospective ADR gap, previously under a parallel session's jurisdiction, was handled within this sprint's scope. The Scribe delegation generated only 256 and then stalled (ACP background stall), so the missing pieces (KR 257·258·259 + EN 256·257·258·259) were completed by hand after gathering commit facts. The docs/adr/README.md retrospective-ADR count was corrected 197→201 (passing CI Quality — docs --strict).

Implementation

  • item1: root-cause identification (doc/log analysis) + ensure-critic-crons.sh (profile scripts, outside repo). No code change.
  • item2 (#521 10e4c377): services/problem/src/problem/recommendation-seeds.ts — seeds 12→23, 3 ground-truth tier corrections + 3 PLATINUM additions; package.json brace-expansion 5.0.8 override. tsc exit 0 · jest 19 suites/266 tests pass (Oracle re-verified directly).
  • item3 (#522 eb338f75): docs/adr/sprints/sprint-{256,257,258,259}.md (KR) + docs/adr-en/sprints/… (EN) + docs/adr/README.md count. EN 213/213 · doc-refs intact.

Verification (physical facts): both #521 and #522 Critic-verdict CLEAN (confirmed by Oracle direct log read), CI green, squash-merged after mergeState CLEAN — #521 10e4c377, #522 eb338f75 confirmed on origin/main.

Incidents

  1. Parallel Oracle session collision (recurring): (a) While trying to push+PR the 263 retrospective ADR, discovered a parallel session had already merged PR #520 — cleaned up the revived stale branch via the GitHub API. (b) Before I started, a parallel session had already pre-applied and pushed all of #521's 3 Critic fixes (tier corrections, PLATINUM additions, Trivy) and #522's README fix — discriminated by measurement (remote commit history) and avoided rework (my local branch lagged remote, so push was withheld). Exactly the parallel-session PR-collision pattern in memory.
  2. Auto-verdict false positive (reconfirmed): the .verdict auto-parser labeled #521·#522 "CLEAN-hint" and the #521 re-review "REVIEW-NEEDED", but both codex verdict sections were CLEAN. The auto-verdict is a hint only — reconfirmed that the gate is Oracle's direct log read (seed-investigation keywords in the log body misled the scanner).
  3. #522 BEHIND: #521 and Dependabot #517 merged into main first, leaving #522's branch BEHIND so auto-merge did not fire → gh pr update-branch, then auto-merged.
  4. EN ADR written by hand: ANTHROPIC_API_KEY unset, so translate-adr.mjs (exit 2) could not run → EN ADR written by hand (same as Sprint 261~263). Key re-rotation is deprecated per user instruction.