review-work

Review Work

Independent assessment of executed work. Reads only what is on disk; it does not see the executor's reasoning, by design.

Input

work/<slug>/README.md with status: pending-assessment, plus the artifacts under work/<slug>/data/.

Output

Same README, updated:

# Assessment populated with a verdict and explanation (see template below).
Frontmatter status: done (if goal was accomplished) or status: needs-input (if it was not and human direction is needed) or status: pending-plan (if the assessment recommends a fresh plan attempt).

Checks

Goal vs Results. Re-read the # Goal section. Compare against # Results. Does the work actually answer the goal, or only adjacent questions?
Artifacts exist. For every artifact named in # Results, verify it is present under work/<slug>/data/.
Plan adherence. Did execution follow the plan? Documented deviations are acceptable; undocumented substitution is a red flag.
Evidence quality. Are claims in # Results supported by the artifacts? Spot-check by reading a sample. For prose claims backed by .ev evidence entries, load the asta-tools:check-claims skill and apply its rubric rather than a second opinion of your own.

Assessment template

# Assessment

## Verdict
<accomplished | partial | not accomplished>

## Reasoning
<one or two paragraphs explaining the verdict against the Goal>

## Root cause (if not fully accomplished)
<one of: incorrect assumption | technical blocker | missing data | scope drift | other — with a sentence of detail>

## Recommended next status
<done | needs-input | pending-plan>

Verdict rules

accomplished → set status: done.
partial with a clear next step the agent can take → set status: pending-plan and write a short note in the Assessment recommending what to plan next.
partial or not accomplished requiring user input → set status: needs-input.

Out of scope

Re-running the work. That is do-work.
Drafting a new plan. That is plan-work.
Committing or indexing. That is save-work.