Session log audit

Skill krzemienski/crucible/skills/session-log-audit

Locate and inspect the JSONL session log for a robust trial; emit per-trial line citations proving the required behaviors fired. Use this skill whenever a trial completes, whenever an Oracle requests behavior verification, whenever VG-11 of a Crucible run executes, or whenever a reviewer needs to confirm "did the hook actually fire?" Reads from the canonical Claude Code session log path (~/.claude/projects/ENCODED-CWD/SESSION-ID.jsonl). Produces evidence/session-logs/TRIAL-ID/INDEX.md with cited line numbers per behavior. Never edits session logs — fix the plugin and re-run instead.From its SKILL.md

Install

npx -y skills add krzemienski/crucible --skill session-log-audit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.5 KB, as published. Nobody here has run it

Session Log Audit

Scope

This skill handles location, copying, and per-trial behavior auditing of Claude Code session logs (PRD §1.17 FR-LOG, MSC-13, MSC-14).

Does NOT handle: producing the trials themselves (that's planning/validation), modifying session-log content (sealed audit trail), or evaluating reviewer/Oracle outputs.

Security

Read-only access to ~/.claude/projects/.
Refuse to write to or modify any session log file.
Sanitize paths before logging — encoded-cwd patterns may include user-identifiable strings.
If a session log contains secret values (API keys mistakenly typed), cite line ranges only — never reproduce the secret in INDEX.md.

Workflow

Determine <encoded-cwd> for $TARGET_REPO: replace every non-alphanumeric character with -.
List candidate session files: ls -lt ~/.claude/projects/<encoded-cwd>/*.jsonl.
Match <session-id> to the trial by: a. SDK-driven trials: read init.json from the SDK transcript and extract session_id. b. Interactive trials: take the most recently modified .jsonl matching the trial's wall-clock window.
Copy the session log into evidence/session-logs/<trial>/session.jsonl (NEVER edit it).
Run audit.py (provided in scripts/) over the copy. Output INDEX.md citing line numbers for each required behavior:
- PreToolUse hook fired
- PostToolUse hook fired
- Stop / completion-attempt hook fired
- The named skill from MODE.txt was invoked (crucible:<skill-name>)
- For validation-mode trials: zero Write/Edit tool uses
If any required behavior is missing, REFUSE to mark the audit complete; surface the missing behavior to the planner for plugin fix + trial re-run.

Produced artifacts

evidence/session-logs/<trial>/session.jsonl — RAW session log copy (sealed, never edited)
evidence/session-logs/<trial>/INDEX.md — cited line numbers per behavior with counts
evidence/session-logs/SUMMARY.md — aggregated index across all trials
scripts/audit.py — the parser (single source of truth)

Forbidden actions

Editing session.jsonl content.
Synthesizing line citations not present in the actual log.
Marking a behavior verified without an actual line number citation.
Hand-authoring INDEX.md without running audit.py.

Example

Trial-01 ran /crucible:planning "add header to README":

encoded-cwd = -Users-nick-Desktop-crucible
Latest matching .jsonl: ~/.claude/projects/-Users-nick-Desktop-crucible/abc-123.jsonl
Copy → evidence/session-logs/trial-01/session.jsonl
Run audit.py:
- PreToolUse: lines [12, 47, 89, 134] (count=4)
- PostToolUse: lines [13, 48, 90, 135] (count=4)
- Stop: lines [201] (count=1)
- crucible:planning skill invocation: lines [5, 23] (count=2)
- Writes/Edits: lines [45, 132] (count=2 — planning trial, edits expected)
INDEX.md captures all citations. PASS.