Audit what a session actually did and turn each cause of waste or drift into one concrete action. The goal is the most token-efficient run that still follows the canonical flow for the issue being worked on. Adherence comes first: never propose saving tokens by removing, weakening or reordering a served step, a check, a verifier or the approval gate.
Run it in a plain session, never inside one of the installed roles. Audit one session by default, and several only when the user names several ids. Never choose sessions yourself.
Rules
- Sessions are evidence. Never change one, and never run a runtime operation against one: an operation can write to the session or the response cache. Read with file tools.
- Read a client transcript only after the user confirms it (step 2).
- Keep the audit itself cheap: start from the digest, and open a file only to confirm a specific finding, by search or line range. Never read a whole transcript,
design.jsonor session. - Change nothing without the user's approval of that specific finding. Approving one finding approves no other.
- Measured facts come from the scripts, judged ones from you. Label each. Report usage a client did not record as unavailable, never as zero.
Procedure
<skill> below is this skill's folder.
1. Check the installation and the sessions
The target needs answers.json in its maintenance folder. If it has none, there is nothing to audit: offer bootstrap-orchestrator-system. The digest stops and names any session id it cannot find.
2. Find the transcripts
node <skill>/scripts/digest.mjs locate --repo <target> --session <id> [--session <id>]...
It lists, per session, the Claude Code and Copilot CLI transcripts whose runtime calls name that session, and prints no content. A file can hold more than one session, and a resumed session can span several files. Show the list and ask the user to confirm the files to read, in one question for all the sessions. For another client, ask for the path; if there is no parser for it, continue with the session store only and say so.
3. Build the digest
node <skill>/scripts/digest.mjs digest --repo <target> --session <id>... --transcript <id>=<file>...
4. Audit
Load contracts/audit.md. Read the issue once per distinct issue. Produce one finding per cause, with several sessions merged into one finding.
5. Report, then act finding by finding
Present the report in the format of contracts/audit.md, most costly first. Then load contracts/actions.md and take each finding through its action, asking the user per finding. End with what was applied, what was written, and what was left.