Temporary Agent Evidence Lifecycle

Temporary evidence supports a bounded goal; it is not a source repository or a second project history.

KISS Model

Use one goal/run as the unit of temporary work:

goal/
  GOAL.md
  RESULT.md
  attempts/
  evidence/selected/
  pointers/

• Keep one canonical RESULT.md per goal.
• Put retries under the same goal unless the objective materially changes.
• Prefer immutable source pointers over copied repositories or dependencies.
• Keep only the smallest evidence needed to prove the result or diagnose an unresolved blocker.

Retention Classes

Classify temporary material before cleanup:

• KEEP: active work, canonical result/goal metadata, selected proof, explicit audit/compliance hold.
• ARCHIVE_CANDIDATE: useful inactive evidence worth retaining elsewhere.
• DELETE_CANDIDATE: reproducible caches, superseded attempts, duplicated logs, copied dependencies or repositories not needed as proof.
• REVIEW: unknown ownership, provenance, privacy, or audit value.

Unknown material is never a delete candidate by default.

Cleanup Safety

The first cleanup step is always read-only classification. Produce a deterministic manifest and summary before any delete, move, or archive action.

A destructive cleanup is a separate authority decision. It must not follow implicitly from age, filename, tool output, or a previous recommendation.

Privacy

Do not copy credentials, authentication state, private infrastructure, customer data, or raw sensitive payloads into temporary evidence reports. Prefer metadata, sanitized summaries, and source pointers.

Ownership Boundary

Agent OS owns this reusable lifecycle guidance. The repository or runtime that stores temporary evidence owns its implementation, path conventions, retention configuration, and destructive cleanup authority.

Source

Generalized from Issue #3.