agentreceipt AGENTS.md

AgentReceipt repository instructions

Instruction version

This checked-in AGENTS.md is the canonical instruction file for the Agent Learning Loop pilot. AgentReceipt captures its repository-relative path and SHA-256 content hash when each Pilot Task starts. Do not edit this file while collecting a baseline cohort; a change creates a new instruction version and closes the current cohort.

Pilot baseline capture

Before the first substantive edit of an enrolled Pilot Task:

1. Run agentreceipt start to start one non-blocking AgentReceipt capture for the task and keep it active through final validation or abandonment.

2. Add one signed enrollment marker with this exact, privacy-safe shape:

   pilot_phase=baseline task_id=<repo-local-id> task_type=<feature|bug-fix|refactor|test|documentation|ci|maintenance> intended_surface=<production|tests|documentation|configuration-ci|dependencies> planned_complexity=<small|medium|large> delivery_path=<pr|direct-commit> proposal=none
   

3. Run agentreceipt status and require State: active, - git: active, and - filesystem: active before editing.

4. If the status gate fails, do not edit. Repair capture under the same task ID and rerun the gate. If repair requires a new capture, repeat the signed marker with the same task ID; group all resulting receipts into one Pilot Observation. Count one enrollment miss only if the Pilot Task never yields an Eligible Receipt.

5. Use agentreceipt start --watch only as an optional live view in a second terminal. agentreceipt stop imports matching Codex Provider Evidence before finalization; unavailable Provider Evidence remains a gap rather than an automatic exclusion.

6. Do not put task descriptions, prompts, personal data, secrets, absolute paths, or raw provider output in the marker.

7. Stop and finalize the capture only after validation, abandonment, or an explicit rejection.

The pilot ledger records the receipt ID or hash, instruction-file hash, normalized provider and filesystem evidence, and manually adjudicated downstream result. Receipts and replay bundles stay local; only privacy-reviewed IDs, hashes, aggregate measurements, findings, decisions, and reviewed repository changes may be committed or posted.