Sprint 104 — CI Improvement: Full Expansion + AI Coverage Integration + Measurement

Background

Continuing from Sprint 103's "pilot then expand" decision, this sprint's core tasks were expanding the setup-node-service composite action from github-worker only to all Node services, and integrating ai-analysis (FastAPI/pytest) into the global coverage-gate. Simultaneously, the Oracle __AGENT_DONE__ marker bug carried over from Sprint 102 was resolved through Oracle self-fix, restoring agent chain automation reliability.

This sprint also completes the expansion phase of Channeltalk CI refactoring Principle ⑤ (validate in a small service first, then expand).

Goals

1. Expand setup-node-service composite action to all Node services (quality-nestjs / audit-npm / test-node matrix integration)
2. Integrate ai-analysis (Python/pytest) coverage into the global coverage-gate
3. Restore automation reliability by fixing the Oracle agent chain __AGENT_DONE__ marker missing bug
4. Obtain Sprint 103 pilot actual measurement data (github-worker pre/post timing comparison)

Work Summary

Task	Agent	Output
Composite action full-service expansion	Architect	PR #113 (00650bf)
ai-analysis coverage-gate integration	Architect	PR #113 (00650bf)
Oracle runner trap fix	Oracle (direct)	~/.claude/oracle/bin/oracle-spawn.sh
github-worker timing measurement	Sensei	✅ Sprint 105 retrospective completion
ADR update	Scribe	This document + sprint-103.md retrospective

Changes

1. Composite Action Expansion (ci.yml)

• Removed matrix.service != 'github-worker' condition from quality-nestjs / audit-npm / test-node 3 jobs
• Inline 3 steps (Setup Node + Cache + Install) × 3 jobs = 67 lines deleted
• All Node services (gateway, identity, submission, problem, github-worker) now go through composite
• audit-npm retains install-command: 'npm ci --ignore-scripts' (security scan policy)

2. AI Coverage-Gate Integration (ci.yml)

• test-ai-analysis: added --cov-report=lcov:coverage/lcov.info to pytest
• artifact path expanded to multi-line (xml + lcov)
• test-ai-analysis added to coverage-gate needs, if condition expanded
• scripts/check-coverage.mjs: no changes required (recursive discovery automatically recognizes ai-analysis)

3. Oracle Runner __AGENT_DONE__ Marker (Local Infrastructure)

• ~/.claude/oracle/bin/oracle-spawn.sh:127-149 runner heredoc reconstructed
• set -euo pipefail → set -uo pipefail (errexit removed)
• cleanup() function + trap cleanup EXIT: guarantees marker echo + lock removal + reap + dispatch on any exit path
• Covers all 4 failure modes: tee pipe SIGPIPE / empty output / tmux detached
• Reference memory: memory/issue-oracle-pipeline-agent-done-marker.md

4. Measurement Attempt (Sensei)

Sensei attempted Pre/Post sample collection via gh run list + jobs API. Pre 4-run average (Quality 22.2s / Audit 19.8s / Test 19.2s) successfully obtained, but both PR #111/#113 modified only workflow files — detect-changes skipped github-worker → Post sample count: 0. No runs observed where composite action executed through actual install/test path. Full report: ~/.claude/oracle/inbox/sensei-sprint-104-timing.md

Sprint 105 [B] retrospective actual measurement result (task-20260421-sp105-b3-finalize):

Job	Pre-change avg (n=4)	Post-change avg (n=3)²	Difference	Practical verdict
Quality — github-worker	22.2s (σ 5.8s)	22.3s (σ 2.5s)	+0.1s (+0.4%)	✅ No change
Audit — github-worker	19.8s (σ 3.7s)	18.0s (σ 3.0s)	-1.8s (-8.9%)	✅ No change
Test GitHub Worker	19.2s (σ 1.9s)	20.0s (σ 1.0s)	+0.8s (+3.9%)	✅ No change

² Post n=3: 2 natural runs (run 24702740418·24702828670) + 1 synthetic (run 24703075569 rebuild_all). Detail: ~/.claude/oracle/inbox/sensei-sprint-105-timing.md

Lessons Learned

1. CI infrastructure PRs cannot generate their own actual measurement data due to detect-changes skip

Both PR #111 (Sprint 103) and PR #113 (Sprint 104) modified only .github/workflows/ci.yml. detect-changes did not detect services/github-worker/** changes — entire matrix skipped. Response: Standardize workflow_dispatch(rebuild_all=true) once alongside composite/pipeline change PRs, or use a dummy touch commit. Deferred to Sprint 105.

2. "Pilot then expand" decisions are justified by code duplication removal alone

67-line deletion (25% compression) secures maintainability. Actual performance measurement is a follow-up validation, not the basis for the expansion decision. Reconfirmed that Sprint 103's original plan ("measure from 2nd run after merge") was an asynchronous structure.

3. trap EXIT covers all 4 shell runner failure modes in one guard

set -e + tee pipeline fails to execute subsequent cleanup if even one of normal exit / claude failure / SIGPIPE / signal receipt is missing. trap EXIT intercepts all exit paths — single guard handles all 4 cases. Borrowable for other dispatch system designs.

4. Script-zero-change is the best approach for AI coverage integration

scripts/check-coverage.mjs already has recursive discovery structure implying extensibility. Simply adding the lcov reporter to pytest enables automatic recognition. Maintains the simplicity of handling Python/Node coverage in a single script. Sprint 103 lesson "test 0-artifact scenario for infrastructure PRs" is also applicable here.

Carried Over (Sprint 105)

• github-worker actual measurement completion — re-measurement trigger: when 5 real github-worker change PRs are available
• workflow_dispatch(rebuild_all=true) guardrail introduction (Sensei recommendation, Sprint 104 Lesson 1 response)
• L2 cache layer (build output) — scope definition needed
• Frontend build optimization — scope definition needed
• Global coverage threshold upgrade to 70% — after actual measurement data obtained
• commitlint scope-enum management automation (pre-commit hook consideration)

References

• Sprint 103 ADR: docs/adr/sprints/sprint-103.md
• PR #113: composite expansion + AI coverage (00650bf)
• Channeltalk CI refactoring: https://channel.io/ko/team/blog/articles/backend-ci-refactoring-73fca77d
• Sensei measurement report: ~/.claude/oracle/inbox/sensei-sprint-104-timing.md
• Oracle issue memory: ~/.claude/projects/-Users-leokim-Desktop-leo-kim-AlgoSu/memory/issue-oracle-pipeline-agent-done-marker.md