Homework — Session Mistake Extractor

On session cleanup ("세션 정리") or /homework invocation, analyze the current and linked previous sessions to extract 찐빠 (mistakes: rule violations, scope-creep, hallucinations, premature hypotheses, missed conventions, etc.), then report findings via omcustom-feedback with a mandatory user confirmation gate.

This skill is the dedicated entry point for R011's "Session-End Retrospective Feedback (Model-Drafted)" pattern. It formalizes the retrospective workflow that produced issue #1266.

Usage

/homework                          # Analyze current session, report findings
/homework --dry-run                # Analyze only, no omcustom-feedback invocation
/homework --days 3                 # Include linked sessions from last N days
/homework --severity high          # Filter to critical/high findings only

Trigger Detection

Default: OFF for auto-activation. This skill does NOT auto-run on session cleanup or session-end signals unless explicitly enabled. The default is false — silence is a non-trigger.

Activate ONLY when:

• Explicit /homework invocation (always runs)
• Explicit opt-in: the user/project has explicitly directed homework to run on session cleanup — e.g., a CLAUDE.md directive such as "run homework on every session cleanup", or a settings flag. A one-off explicit request ("회고 돌려줘", "homework 실행") also counts.

If no explicit opt-in exists, treat session-cleanup / session-end phrases ("세션 정리", "숙제", "회고", "끝", "종료", "마무리", "done", "wrap up", "session cleanup", "end session") as NON-triggers — do NOT auto-run. You MAY briefly remind the user that /homework is available, then proceed with normal session-end handling.

When auto-activation IS explicitly enabled and fires as a session-end signal, this skill runs BEFORE sys-memory-keeper's MEMORY.md update (R011 session-end self-check order: homework → memory save).

Workflow

Phase 1: Trigger Parsing

Parse arguments:

• --dry-run: analyze only, skip Phase 5 (omcustom-feedback invocation)
• --days <n>: extend linked-session search to last N days (default: 0 = current session only)
• --severity <level>: filter output to this severity and above (default: all)

Phase 2: Session Gathering

2a. Current Session Transcript

Attempt deterministic transcript extraction. Preferred sources (in order):

1. Grep for rule-violation markers in the current conversation context:

   • Safety classifier trip signals: [Safety], [R001], [Warning], classifier denial messages
   • Self-corrections: "sorry", "I made an error", "let me correct", "이전 답변 수정", "죄송합니다"
   • Premature hypothesis signals: "I assume", "probably", "should be" followed by a contradiction in a later turn
   • Interrupt + re-plan events: user corrections, "no", "wrong", "다시", "아니"

2. Transcript files (CC v2.1.x session JSONL):

   find ~/.claude/projects -name "session-*.jsonl" -newer "$(date -v-1d +%Y-%m-%dT00:00:00)" 2>/dev/null | head -5
   

   Parse type: "error", type: "correction", type: "feedback" events.

3. Fallback: Rely on the model's recall of the current conversation (impressionistic, lower confidence — mark findings as [recall] not [transcript]).

2b. Linked Previous Sessions (when --days > 0)

Use the episodic-memory plugin's search-conversations / episodic-memory:remembering-conversations to retrieve sessions from the last N days linked to this project. Pass project directory as context.

If episodic-memory is unavailable, scan JSONL files:

find ~/.claude/projects -name "session-*.jsonl" \
  -newer "$(date -v-${DAYS}d +%Y-%m-%dT00:00:00)" 2>/dev/null

R020 read-before-characterize: Do NOT characterize a session's mistakes before reading it. Read the session transcript (or a representative sample) first, then characterize.

Phase 3: Mistake (찐빠) Analysis

Categorize each finding with the following structure (mirror #1266 format):

찐빠 #N — [{severity}] {short title}
├── 증상: {what was observed — cite evidence: session line ref or commit SHA}
├── 근거: {transcript evidence or recall note — always read first per R020}
├── 원인: {root cause — why did this happen?}
├── 영향 규칙: {R0xx, R0yy — affected rule IDs}
└── 제안: {concrete corrective action or harness change}

Severity scale:

Level	Criteria	Examples
Critical	Safety classifier trip, credential exposure, scope-creep into privileged domains, working-tree loss	R001 violation, secret dump, unauthorized infra action
High	Rule violation with downstream impact, hallucinated fact acted upon, premature hypothesis causing permanent change	R020 Parallel Read+Change, wrong root cause → wrong fix
Medium	Process gap, missed convention, advisory rule ignored	R007 header missing, bypassPermissions omitted, count sync missed
Low	Minor style drift, non-impactful oversight	honorific regression, ecomode token waste

Mistake categories to look for:

Category	Signals
Rule violations (R0xx)	Header missing (R007), tool prefix absent (R008), file write by orchestrator (R010), sequential when parallel required (R009)
Scope-creep	Subagent task expanding beyond its named scope (R010 Subagent Scope-Creep STOP Protocol)
Hallucinated facts	External UI fields stated as fact (R003 Unverifiable External Product UI), in-cluster hostnames, unverified URLs
Premature hypotheses	Diagnosis before reading evidence (R020 Read-Before-Characterize), parallel Read+permanent-change dispatch (R020 Variant)
Missed conventions	Count sync drift (3-way sync), template mirror omitted, bypassPermissions missing
Over-claim completion	[Done] without verification (R020), test-skip masking failures

Do NOT over-claim. If evidence for a finding is weak or based on recall only, mark it [recall, low-confidence] and note what would be needed to confirm it. R020 read-before-characterize applies to this analysis itself.

Phase 4: Draft Feedback Issue

Assemble a feedback issue in Korean using the #1266 format:

**제목**: 세션 회고: {date} 세션 찐빠 {N}건 — {top finding title}

**카테고리**: improvement

**본문**:
## 세션 회고 — {YYYY-MM-DD}

### 개요
총 {N}건의 찐빠가 발견되었습니다 (Critical: {c}, High: {h}, Medium: {m}, Low: {l}).

### 찐빠 목록

{찐빠 #1 ~ #N — structured format from Phase 3}

### 하네스 제안 (있는 경우)
{Concrete skill/rule/hook changes that would prevent recurrence}

---
*Generated by `/homework` skill (v0.1.0)*

If --severity filter is active, include only findings at or above the threshold. Note the filter in the issue body.

If no findings are discovered, output:

[homework] 이번 세션에서 찐빠를 발견하지 못했습니다. (세션 정상 종료)

and skip Phase 5.

Phase 5: Report via omcustom-feedback (Phase 4A gate)

MUST go through the omcustom-feedback skill's Phase 4A preview + confirmation gate. NEVER auto-submit. User approval is always required before any GitHub issue is created.

Invoke the omcustom-feedback skill with the drafted issue content. The user will see a preview and must confirm before any GitHub issue is created.

[homework] 피드백 이슈 초안을 omcustom-feedback으로 전달합니다.
아래 미리보기를 확인하고 제출 여부를 결정해 주세요.

If --dry-run is active, skip this phase and output the draft directly to the conversation instead.

Phase 6: Output Summary

[homework] 완료
├── 분석: {N}건 찐빠 발견 (Critical: {c}, High: {h}, Medium: {m}, Low: {l})
├── 제출: {이슈 URL | dry-run (미제출) | 사용자 취소}
└── 다음 액션: {harness 제안 있으면 표시, 없으면 "없음"}

Options Reference

Option	Default	Description
--dry-run	off	Analyze only, no omcustom-feedback invocation
--days <n>	0	Include linked sessions from last N days (0 = current session only)
--severity <level>	all	Filter findings to this level and above

Rules & Cross-References

Rule	Relevance
R011 (SHOULD-memory-integration)	This skill is the dedicated entry point for "Session-End Retrospective Feedback (Model-Drafted)". Runs before sys-memory-keeper MEMORY.md update.
R020 (MUST-completion-verification)	Analysis MUST read transcript evidence before characterizing a mistake. Do NOT characterize before reading (Read-Before-Characterize). Parallel Read + Permanent-Change Dispatch anti-pattern applies here too.
R016 (MUST-continuous-improvement)	Genuine defects/process gaps → feedback issue. The /homework output feeds R016's continuous improvement loop.
R010 (MUST-orchestrator-coordination)	Any file writes in this workflow must be delegated to subagents. This skill orchestrates, never directly writes except to .claude/outputs/.
R001 (MUST-safety)	Analysis must not dump credential values, secrets, or PII. Reference sensitive items by name only.

Related Skills

Skill	Relationship
omcustom-feedback	Reporting channel (Phase 5). Model-invocable with Phase 4A confirmation gate.
instinct-extractor	Cross-session failure-pattern mining (complements /homework's single-session focus). For multi-session patterns, run instinct-extractor after homework.
episodic-memory:search-conversations	Cross-session retrieval for --days mode.
sys-memory-keeper	Runs after /homework at session end (R011 order: homework → memory save).

Artifact Output

.claude/outputs/sessions/{YYYY-MM-DD}/homework-{HHmmss}.md

If Phase 5 is skipped (--dry-run), the draft issue body is written to this artifact path for reference.

Permission Mode Note

This skill does not spawn subagents directly. If future versions delegate analysis to subagents, ALL Agent tool calls MUST include mode: "bypassPermissions" per R010 Universal bypassPermissions.

Context Fork Note

This skill does NOT use context: fork. The fork cap is at 10/12; homework is a single-agent orchestration skill and does not require a forked context.

Limitations

• Transcript availability: CC session-*.jsonl schema may change; Phase 2a source (1) grep is more resilient than JSONL parsing.
• Recall accuracy: When transcript files are unavailable, findings are marked [recall] and confidence is lower.
• episodic-memory dependency: --days mode degrades gracefully when the plugin is unavailable (falls back to JSONL scan).
• Scope: This skill analyzes session behavior, not code quality. For code-quality retrospectives, use dev-review or adversarial-review.