Aet investigate

/aet-investigate

Answer one declared investigation question with a portable evidence handoff.

Boundary

• Keep the target workspace read-only.
• Do not execute arbitrary commands, modify code, commit, push, merge, publish, or repair.
• Require one primary hypothesis, at least one competing hypothesis, and a disconfirming search.
• Treat Run Records as historical behavior. They may create Observations and Evidence Candidates, but only explicit proof.inspect and freshness.check results may create Verified Evidence.
• Preserve unknown, counter-evidence, diagnostics, Freshness limitations, and every does_not_prove statement.
• Use policy.json and tool-contract.json without expanding their authority.

Workflow

1. Normalize one declared run when a normalized directory is not already available:

   aet run normalize --source <codex|claude-code> \
     --input <session.jsonl> --output <normalized-run>
   

2. Create an investigation-request/1.0 JSON object. Bind its run_sources entry to the normalized Manifest's source_type and run_group_id. Copy the budgets and privacy boundary from policy.json; narrow them when possible.

3. Run the investigator. Without a Proof it remains observation-only:

   aet investigate --request <request.json> \
     --run <normalized-run> --output <investigation.json>
   

   To inspect an existing deterministic Proof, the Policy must allow proof.inspect and freshness.check, and their budgets must be non-zero:

   aet investigate --request <request.json> \
     --run <normalized-run> --workspace <workspace> \
     --proof <proof.json> --output <investigation.json>
   

   The Proof path and its recorded workspace must remain inside and match the declared read-only workspace. A matching recorded command Candidate may be marked verified; stale Proof remains historical Evidence and cannot answer a current-workspace Claim.

4. Compile and validate the handoff:

   aet bundle create --investigation <investigation.json> \
     --output <evidence-bundle>
   aet bundle validate <evidence-bundle>
   

5. Return only the bounded status, unresolved questions, Bundle path, and the smallest authorized next action. A reviewer does not need AET installed to read manifest.json, index.json, core/*.jsonl, and report.md.

Completion check

Claim completion only when aet bundle validate <evidence-bundle> returns PASS. Confirm the Investigation preserves the original Policy and Ledger, does not promote Agent self-report, and exports real Run Group, Record identity, Proof, Freshness, Source type, and Content Hash bindings. A non-unknown finding requires current deterministic Evidence.

Stop after the validated Bundle. Do not continue into repair, new command execution, Learn, Stage, Adopt, or release work.