ARE Agent Residue Evidence

Task-scoped residue evidence

Tests finish. Residue remains.

ARE shows the files, directories, processes, and listening ports that appeared inside one Agent task—before anyone decides what should be cleaned.

Download v0.3.0 Run the reproducible demo
A test task flows through scoped observation, human review, authorized cleanup, and verification.
ARE supplies evidence. The Agent explains it. The user owns the cleanup decision.

The checkpoint

Observe before cleanup.

  1. 01Begin

    Declare the exact workspace and task-owned temporary roots before tests or builds.

  2. 02Review

    Receive grouped, path-aliased evidence without a machine-wide scan.

  3. 03Decide

    The user authorizes cleanup; the Agent uses its ordinary host tools.

  4. 04Verify

    ARE rechecks the same stable candidates and records a new revision.

Narrow by design

Evidence, never intervention.

Local and offline

No network access, telemetry, upload path, or remote service.

No cleanup tool

ARE cannot delete files, terminate processes, close ports, or decide that a candidate is safe.

Task isolation

Opaque owner capabilities isolate evidence. Public IDs never authorize report access.

Honest scope

NO_CANDIDATES_OBSERVED is scoped to declared roots. It is never a host-wide cleanliness claim.

Agent workflow

Add one evidence checkpoint.

Install the verified MCPB or native package. Compatible Agents receive the same eight-tool, evidence-only contract.

begin_task_observation
  ↓ run tests or builds
end_task_observation
  ↓ explain candidates and ask
user authorizes cleanup
  ↓ Agent cleans with host tools
verify_task_residue
Open the complete quickstart →

Trust boundary

A checkpoint, not a certification.

ARE complements sandboxing, test isolation, build provenance, and host policy. It answers a narrower question: what task-scoped residue did this observation actually see?

Threat model Report contract Security policy