| name | ai-session-audit |
| description | Audit retained coding-agent sessions, tools, workers, tokens, and routing. Use for retrospective AI-tool efficacy checks. |
AI Session Audit
Extract deterministic facts before making semantic judgments about agent
quality. The bundled command reads T3 canonical provider events from existing
logs. It does not inspect T3's private database or copy prompt and tool content
into its output.
Commands
Run a bounded summary:
python3 <skill-dir>/scripts/ai_session_audit.py summary \
--since 2026-08-20 --format markdown
Export scrubbed events for DuckDB or another analytics backend:
python3 <skill-dir>/scripts/ai_session_audit.py events \
--since 2026-08-20 --output audit.ndjson
Use --provider, --thread, and --until to narrow either command. Output
goes to stdout unless --output is explicit. Existing output requires
--force.
Workflow
- Run
summary against the smallest useful date range.
- Treat tool order, turn count, provider-reported tokens, skill observations,
worker tasks, and model routes as deterministic evidence.
- Read references/schema.md before comparing providers
or interpreting token and task fields.
- Use
events when longitudinal queries or visualization justify a database.
Read references/backends.md before selecting one.
- Separate observed behavior from inferred policy compliance. A transcript can
prove that a skill ran. Deciding whether one should have run still requires a
frozen routing corpus or a reviewed rubric.
Boundaries
- Keep source logs read-only.
- Do not emit prompts, commands, tool arguments, tool results, reasoning, or
assistant prose.
- Do not present processed-token observations as billed usage or cost.
- Do not compare policy revisions unless the report records a revision label or
another verified activation boundary.
- Do not treat more tools, workers, tokens, or turns as inherently better.