Quick answer: A review task package exports a Markdown handoff that carries the context and revision anchors for the change, so another agent can work from the same evidence instead of a loose description. Modification is optional, and the receiving agent assesses every suspicion independently rather than treating the export as a confirmed list of defects.

What the package contains

The export is a structured Markdown document, not a screenshot or a free-form message. It bundles the information an agent needs to understand the change.

  • Context: the bounded related code and the resolved definitions, callers, and imports behind the review.
  • Revision anchors: pointers to the specific revisions and lines, so the agent reads the same version rather than a moving target.
  • Findings: the notes the review raised, each with its evidence class and recommendation.

Anchoring the revision matters because code keeps changing. Without it, an agent can reason about a newer or older version than the one you reviewed and reach conclusions that do not apply.

Markdown is the right container because it stays readable by both people and agents, travels without a proprietary tool, and keeps the evidence as text the next agent can quote and check. A binary report or a loose chat thread would lose the anchors and make it harder to trace a finding back to its source line.

Why a handoff instead of an auto-edit

The review does not rewrite the code by itself. It returns priority, evidence, and a recommendation, and the task package preserves that separation. The receiving agent gets a clear picture of what looks suspicious and why, but it is not handed permission to change anything.

Element Purpose
Context Show the code around the change
Revision anchors Pin the exact version under review
Evidence class Say whether a note is static, dynamic, contract, or hypothesis
Recommendation Suggest a next step, not command an edit

This keeps the handoff useful without turning it into an instruction to apply fixes blindly.

The receiving agent stays independent

The receiving agent assesses each suspicion on its own. It does not inherit the findings as facts, and it should confirm the evidence before acting.

  1. Read the context and confirm the revision anchors match the current code.
  2. Rerun the reproducible commands behind the findings.
  3. Verify model-only hypotheses, which are not confirmed defects.
  4. Watch for paths the static parser cannot resolve, such as dynamic loading.
  5. Decide independently which suspicions hold.

This mirrors how a person reads the review. A model-only claim is always a hypothesis, and a clean review does not prove correctness, so the next agent cannot treat the export as a certificate.

Limits of the package

The handoff is only as complete as the resolved context. It can miss dynamic loading, framework entry points, and external consumers, and a file with no incoming references is not automatically safe to delete. When the agent meets a finding that depends on an unresolved path, it should run a dynamic test or smoke check rather than accepting or dismissing the note on the export alone.

If a project script is missing, the evidence stays unverified rather than passed, and an empty result remains inconclusive. The package carries that status honestly instead of implying success.

Powerduck exports the Markdown handoff while keeping the evidence and revision anchors attached, so the receiving agent has the basis to verify each suspicion rather than trusting a summary.

Frequently asked questions

Does the task package tell the agent to apply the fixes? No. It carries context, revision anchors, and findings with recommendations. Modification is optional, and the agent assesses each suspicion independently.

Why are revision anchors included? They pin the exact version under review so the agent does not reason about a different revision and reach conclusions that no longer apply.

Can the receiving agent trust the findings as confirmed? No. It reruns the reproducible commands and verifies model-only hypotheses, since model-only claims are hypotheses and the export is not a correctness certificate.

The bottom line

A review task package is a Markdown handoff that gives another agent the context and revision anchors it needs, while keeping every finding open to independent verification. Its strength is that the next agent works from the same evidence and version, not that it inherits a list of approved defects. Keep modification optional, rerun the evidence, and treat model-only notes as hypotheses. The package transfers context; it does not transfer judgment. More in the AI code review guide.