session-chronicle（会议纪事）

Session Chronicle

Excavate Claude Code session logs to capture complete provenance for research findings, ADR decisions, and code contributions. Traces UUID chains across multiple auto-compacted sessions.

CRITICAL PRINCIPLE: Registry entries must be self-contained. Record ALL session UUIDs (main + subagent) at commit time. Future maintainers should not need to run archaeology to understand provenance.

S3 Artifact Sharing: Artifacts can be uploaded to S3 for team access. See S3 Sharing ADR.

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

When to Use This Skill

• User asks "who created this?" or "where did this come from?"
• User says "document this finding" with full session context
• ADR or research finding needs provenance tracking
• Git commit needs session UUID references
• Tracing edits across auto-compacted sessions
• Creating a registry entry for a research session

File Ownership Model

Key Principle: Only findings/ is committed. Research artifacts go to gitignored outputs/ and S3.

Part 0: Preflight Check

Verify session storage, find project sessions, and check required tools (jq, brotli, aws, op).

Full scripts: Preflight Scripts

Summary of steps:

1. Verify Session Storage - Confirm ~/.claude/projects/ exists
2. Find Current Project Sessions - Encode CWD path, enumerate main + subagent .jsonl files
3. Verify Required Tools - Check jq, brotli, aws, op are installed

Step 4 (MANDATORY before any S3 share): Sanitize

Raw Claude Code session JSONL files are dangerous to share. They commonly contain real credentials (AWS keys, GitHub PATs, Telegram bot tokens, Tailscale API keys, 1Password service tokens), internal hostnames, Tailscale CGNAT IPs, emails, and other infrastructure secrets that leak into prompts via screenshots, env dumps, and shell commands.

Before zipping + uploading to S3, run the sanitizer:

SKILL_DIR="$(find $HOME/.claude/plugins/marketplaces/cc-skills -type d -name session-chronicle | head -1)"
"$SKILL_DIR/scripts/sanitize_sessions.py" \
  --input  /path/to/raw/claude-sessions-export-raw \
  --output /path/to/sanitized/claude-sessions-export \
  --report /path/to/redaction_report.txt

The sanitizer is field-aware (does not destroy UUIDs, tool-use IDs, or forex decimals — v1 had a 92% phone-regex false-positive rate that murdered structural identifiers) and covers:

• AWS / GitHub / OpenAI / Anthropic / Slack / Stripe / Google / JWT / Bearer / Authorization
• Tailscale API keys (tskey-*), CGNAT IPs (100.64–127.x.x), tailnet DNS (*.ts.net), tailnet names
• 1Password service tokens (ops_*), op:// URLs, 32-char item IDs after op CLI context
• Cloudflare API tokens + Global API Key + CF_AppSession
• Doppler (dp.*), Docker PAT (dckr_pat_*), npm (npm_*), Supabase (sbp_*), SendGrid (SG.*)
• Telegram bot tokens (<bot_id>:<secret> format) — catches tokens pasted into env dumps
• ClickHouse URLs with embedded credentials
• .internal hostnames, 172.25.x.x private range
• PEM private key blocks (OPENSSH / RSA / EC / DSA / PGP / ED25519)
• Generic password=, api_key=, secret= declarations in JSON/YAML/env format
• Email addresses
• Phone numbers — only when separators present (prevents UUID/decimal destruction)

Output: a redaction report listing per-pattern counts. Review before packaging to confirm nothing important was destroyed (sanity check: UUID integrity should be preserved).

S3 upload sequence: always raw → sanitize → zip → S3 → presigned URL. Never upload -raw/ directly.

Part 1: AskUserQuestion Flows

Flow A: Identify Target for Provenance

When the skill is triggered, first identify what the user wants to trace:

AskUserQuestion:
  question: "What do you want to trace provenance for?"
  header: "Target"
  multiSelect: false
  options:
    - label: "Research finding/session"
      description: "Document a research session with full session context for reproducibility"
    - label: "Specific code/feature"
      description: "Trace who created a specific function, feature, or code block"
    - label: "Configuration/decision"
      description: "Trace when and why a configuration or architectural decision was made"
    - label: "Custom search"
      description: "Search session logs for specific keywords or patterns"

Flow B: Confirm GitHub Attribution

CRITICAL: Every registry entry MUST have GitHub username attribution.

AskUserQuestion:
  question: "Who should be attributed as the creator?"
  header: "Attribution"
  multiSelect: false
  options:
    - label: "Use git config user (Recommended)"
      description: "Attribute to $(git config user.name) / $(git config user.email)"
    - label: "Specify GitHub username"
      description: "I'll provide the GitHub username manually"
    - label: "Team attribution"
      description: "Multiple contributors - list all GitHub usernames"

Flow C: Confirm Session Scope

CRITICAL: Default to ALL sessions. Registry must be self-contained.

AskUserQuestion:
  question: "Which sessions should be recorded in the registry?"
  header: "Sessions"
  multiSelect: false
  options:
    - label: "ALL sessions (main + subagent) (Recommended)"
      description: "Record every session file - complete provenance for future maintainers"
    - label: "Main sessions only"
      description: "Exclude agent-* subagent sessions (loses context)"
    - label: "Manual selection"
      description: "I'll specify which sessions to include"

IMPORTANT: Always default to recording ALL sessions. Subagent sessions (agent-*) contain critical context from Explore, Plan, and specialized agents. Omitting them forces future maintainers to re-run archaeology.

Flow D: Preview Session Contexts Array

Before writing, show the user the full session_contexts array, then confirm:

AskUserQuestion:
  question: "Review the session_contexts array that will be recorded:"
  header: "Review"
  multiSelect: false
  options:
    - label: "Looks correct - proceed"
      description: "Write this to the registry"
    - label: "Add descriptions"
      description: "Let me add descriptions to some sessions"
    - label: "Filter some sessions"
      description: "Remove sessions that aren't relevant"
    - label: "Cancel"
      description: "Don't write to registry yet"

Flow E: Choose Output Format

AskUserQuestion:
  question: "What outputs should be generated?"
  header: "Outputs"
  multiSelect: true
  options:
    - label: "registry.jsonl entry (Recommended)"
      description: "Master index entry with ALL session UUIDs and GitHub attribution"
    - label: "iterations.jsonl entries"
      description: "Detailed iteration records in sessions/<id>/"
    - label: "Full session chain archive (.jsonl.br)"
      description: "Compress sessions with Brotli for archival"
    - label: "Markdown finding document"
      description: "findings/<name>.md with embedded provenance table"
    - label: "Git commit with provenance"
      description: "Structured commit message with session references"
    - label: "Upload to S3 for team sharing"
      description: "Upload artifacts to S3 with retrieval command in commit"

Flow F: Link to Existing ADR

AskUserQuestion:
  question: "Link this to an existing ADR or design spec?"
  header: "ADR Link"
  multiSelect: false
  options:
    - label: "No ADR link"
      description: "This is standalone or ADR doesn't exist yet"
    - label: "Specify ADR slug"
      description: "Link to an existing ADR (e.g., 2025-12-15-feature-name)"
    - label: "Create new ADR"
      description: "This finding warrants a new ADR"

Part 2: Session Archaeology Process

Scan ALL session files, build the session_contexts array, and optionally trace UUID chains.

Full scripts: Archaeology Scripts

Summary of steps:

1. Full Project Scan - Enumerate all main + subagent sessions with line counts and timestamps
2. Build session_contexts Array - Create the array with ALL sessions (session_uuid, type, entries, description)
3. Trace UUID Chain (optional) - Follow parent UUID references across sessions for detailed provenance

Part 3: Registry Schema

Two NDJSON files track provenance:

• findings/registry.jsonl - Master index, one self-contained JSON object per line
• findings/sessions/<id>/iterations.jsonl - Iteration-level tracking per session

Full schema, examples, and field reference: Registry Schema Reference

Required Fields (registry.jsonl)

Part 4: Output Generation

Brotli compression for session archival and structured git commit messages with provenance.

Full scripts and templates: Output Generation

Summary:

• Compression - Brotli-9 compress each session to outputs/research_sessions/<id>/*.jsonl.br (gitignored)
• Manifest - Auto-generated manifest.json with target_id, count, timestamp
• Commit message - Template includes registry_id, attribution, session counts, S3 retrieval commands

Part 5: Confirmation Workflow

Final Confirmation Before Write

ALWAYS show the user what will be written before appending:

AskUserQuestion:
  question: "Ready to write to registry. Confirm the entry:"
  header: "Confirm"
  multiSelect: false
  options:
    - label: "Write to registry"
      description: "Append this entry to findings/registry.jsonl"
    - label: "Edit first"
      description: "Let me modify some fields before writing"
    - label: "Cancel"
      description: "Don't write anything"

Before this question, display:

1. Full JSON entry (pretty-printed)
2. Count of session_contexts entries
3. GitHub username attribution
4. Target file path

Post-Write Verification

After writing, verify:

# Validate NDJSON format
tail -1 findings/registry.jsonl | jq . > /dev/null && echo "Valid JSON"

# Show what was written
echo "Entry added:"
tail -1 findings/registry.jsonl | jq '.id, .created_by.github_username, (.session_contexts | length)'

Part 6: Workflow Summary

1. PREFLIGHT
   ├── Verify session storage location
   ├── Find ALL sessions (main + subagent)
   └── Check required tools (jq, brotli)

2. ASK: TARGET TYPE
   └── AskUserQuestion: What to trace?

3. ASK: GITHUB ATTRIBUTION
   └── AskUserQuestion: Who created this?

4. ASK: SESSION SCOPE
   └── AskUserQuestion: Which sessions? (Default: ALL)

5. BUILD session_contexts ARRAY
   ├── Enumerate ALL main sessions
   ├── Enumerate ALL subagent sessions
   └── Collect metadata (entries, timestamps)

6. ASK: PREVIEW session_contexts
   └── AskUserQuestion: Review before writing

7. ASK: OUTPUT FORMAT
   └── AskUserQuestion: What to generate?

8. ASK: ADR LINK
   └── AskUserQuestion: Link to ADR?

9. GENERATE OUTPUTS
   ├── Build registry.jsonl entry (with iterations_path, iterations_count)
   ├── Build iterations.jsonl entries (if applicable)
   └── Prepare commit message

10. ASK: FINAL CONFIRMATION
    └── AskUserQuestion: Ready to write?

11. WRITE & VERIFY
    ├── Append to registry.jsonl
    ├── Append to sessions/<id>/iterations.jsonl
    └── Validate NDJSON format

12. SANITIZE (MANDATORY before any S3 share)
    ├── Run scripts/sanitize_sessions.py on the staging directory
    ├── Review redaction_report.txt (verify no structural destruction)
    └── Produce sanitized/ directory for downstream packaging

13. (OPTIONAL) S3 UPLOAD
    ├── Package sanitized/ directory (zip or brotli per-file)
    └── Upload compressed archives, generate presigned URL if sharing externally

Success Criteria

1. Complete session enumeration - ALL main + subagent sessions recorded
2. GitHub attribution - created_by.github_username always present
3. Self-contained registry - Future maintainers don't need archaeology
4. User confirmation - Every step has AskUserQuestion confirmation
5. Valid NDJSON - All entries pass jq validation
6. Reproducible - Session UUIDs enable full context retrieval

References

• S3 Sharing ADR
• S3 Retrieval Guide
• Registry Schema Reference
• Preflight Scripts
• Archaeology Scripts
• Output Generation
• NDJSON Specification
• jq Manual
• Brotli Compression

Troubleshooting

Post-Execution Reflection

After this skill completes, check before closing:

1. Did the command succeed? — If not, fix the instruction or error table that caused the failure.
2. Did parameters or output change? — If the underlying tool's interface drifted, update Usage examples and Parameters table to match.
3. Was a workaround needed? — If you had to improvise (different flags, extra steps), update this SKILL.md so the next invocation doesn't need the same workaround.

Only update if the issue is real and reproducible — not speculative.