Sprint 117 — Full Agent Critic Feedback Integration + Code Refactoring

Background

Sprints 114–116 completed establishing the Critic (Codex cross-review) agent and dispatch pipeline integration, but invocation remained as manual trigger only. Even when code-changing agents left commits, Critic review was not automatically applied, making the actual merge-gate role incomplete.

Additionally, frontend/src/lib/api.ts was a single 860-line file mixing 17 domain APIs, selected as a demo refactoring target to demonstrate the Critic feedback loop in practice.

Goals

Phase	Content	Status
A	Build Critic auto-chain workflow after code-changing agent completion	✅ Complete
B	api.ts (860 lines) → api/ 12 domain module separation (demo case)	✅ Complete
C	Receive and judge Critic cross-review	✅ Complete

Key Decisions (D1~D5)

D1. Insert auto-critic chain into Runner script cleanup()

Selected: Add HEAD comparison logic to cleanup() function in oracle-spawn.sh runner template. Alternatives: dispatch post-run, per-agent hooks. Rationale: Preserving HEAD_BEFORE per agent is most accurate. dispatch post-run bundles multiple agent completions making it impossible to pinpoint commit range.

D2. Separate new script oracle-auto-critic.sh

Selected: Separate auto-Critic chain logic into a dedicated script. Rationale: Single responsibility principle. Independent smoke testing possible. Runner template simply makes the call.

D3. Explicitly list 9 code-changing agents in .claude-team.json as single source

Included: conductor, gatekeeper, librarian, architect, postman, curator, herald, palette, sensei Excluded: scribe (documentation), critic (self-reference prevention), scout (verification) Rationale: Single consistency point. Synchronized with CODE_CHANGING_AGENTS in oracle-auto-critic.sh.

D4. Critic self-reference prevention

Selected: Immediately skip in oracle-auto-critic.sh if input agent is critic. Rationale: Prevent infinite loop (Critic → Critic → ...).

D5. Phase B demo target is api.ts domain separation

Rationale: (1) Single 860-line file with the most surface-area debt, (2) 17 domain sections already clearly separated by // ── comments — low risk, (3) Barrel export enables keeping 72 import locations unchanged.

Implementation

Phase A — Infrastructure (outside repo + inside repo)

Outside repo (~/.claude/oracle/bin/):

• oracle-auto-critic.sh new: receives agent name + task_id + base_commit arguments; auto-creates Critic task via oracle-create-task.sh --simple if agent is code-changing and HEAD has changed
• oracle-spawn.sh modified: HEAD_BEFORE=$(git rev-parse HEAD) added to runner template + oracle-auto-critic.sh call inserted in cleanup()

Inside repo:

• .claude-team.json: dispatch.codeChangingAgents[9] + dispatch.autoCritic{enabled,trigger,method} added (commit 17cf39a)
• .claude/commands/agents/_base.md: automatic Critic review rules section added (not committed via gitignore)

Phase B — api.ts 860 lines → api/ 12 files

File	Lines	Domain
client.ts	147	fetchApi, fetchPublicApi, ApiError, StudyRequiredError
types.ts	126	Problem, Submission, Study and 12 shared interfaces
auth.ts	69	authApi, settingsApi
study.ts	133	studyApi, shareLinkApi, StudyStats
problem.ts	26	problemApi
submission.ts	80	submissionApi, draftApi, aiQuotaApi
external.ts	86	solvedacApi, programmersApi
notification.ts	43	notificationApi
review.ts	82	reviewApi, studyNoteApi
public.ts	54	publicApi (no auth required)
feedback.ts	77	feedbackApi, adminApi
index.ts	59	barrel re-export

Compatibility: 72 import (@/lib/api) locations unchanged. tsc --noEmit 0 errors. (commit fff8314)

Phase C — Critic Cross-Review

• Session ID: 019db399-45eb-7343-9295-bc072cdbd085
• Command: codex review --base 78e43d2
• Judgment: ✅ "no actionable regressions were identified"
• Out-of-scope P1 found (immediately corrected): critic.md stdin syntax (--base <ref> - <<< "...") is incompatible with codex 0.122.0 CLI — --base/--commit and [PROMPT] (- stdin included) are mutually exclusive. Corrected to use --uncommitted "prompt" or prompt alone.

Verification

5-point consistency verification (applying Sprint 116 lessons)

#	Verification Point	Status
1	.claude-team.json codeChangingAgents (9)	✅
2	oracle-auto-critic.sh CODE_CHANGING_AGENTS	✅
3	oracle-create-task.sh VALID_AGENTS	✅
4	oracle-spawn.sh VALID_AGENTS + cleanup() auto-critic call	✅
5	oracle-watchdog.sh get_tier()	✅
6	_base.md auto-Critic section	✅

Smoke test (4/4 PASS)

Scenario	Expected	Result
code-changing agent + HEAD same	Skip	✅
Non-code-changing (scribe)	Skip	✅
Self-reference prevention (critic)	Skip	✅
code-changing agent + HEAD changed	Critic task auto-created	✅

Refactoring regression verification

• tsc --noEmit: 0 errors
• Jest: 120 suites / 1259 tests PASS
• Coverage: lines 85.99% / branches 76.95% (above thresholds 83/71)

Key Lessons Learned

1. Critic review scope is diff only: codex review --base <ref> reviews changes only. Full repo audit requires separate design (Sprint 118 target).
2. Watch for Codex output contamination: frontend/coverage/ HTML and .next/ build artifacts mixed into review context causing 1MB+ output explosion. Need rm -rf coverage .next before review.
3. critic.md CLI syntax must be tested: The stdin syntax documented when established in Sprint 114 differs from actual codex 0.122.0. Re-verify on next codex version upgrade.
4. Interactive vs tmux dispatch: Auto-Critic chain fires only in runner script cleanup(). In interactive sessions, manual Oracle invocation required. Sprint 118 E2E verification is tmux-based.

Carried Over

• auto-Critic chain tmux dispatch mode E2E verification (only smoke in interactive — cleanup hook not triggered)
• P0/P1 auto-detection → auto-fix delegation loop (currently Oracle manual mediation)

Commit Map

Commit	Scope	Content
17cf39a	chore(infra)	.claude-team.json autoCritic metadata
fff8314	refactor(frontend)	api.ts 860 lines → api/ 12-file separation
(outside repo)	-	oracle-auto-critic.sh new, oracle-spawn.sh runner template modified
(.claude/ gitignore)	-	_base.md auto-Critic section, critic.md stdin syntax corrected, prompts 12/12 rebuilt