allenai Review Plan
Independent review of a plan produced by plan-work. The reviewer reads only what is on disk; it does not see the planner's reasoning, by design.
Input
work/<slug>/README.md with status: pending-review and a populated # Instructions section.
Output
Same README, updated:
- On approval — frontmatter
status: ready. Optionally append a short# Review Notessection summarizing why the plan was approved (one or two sentences). - On rejection — frontmatter
status: pending-plan. Append or overwrite a# Review Feedbacksection with specific, actionable issues for the planner to address.
Checks
For each step in the plan, verify:
- Concreteness. Does the step name a specific tool/skill/command, or is it hand-wavy ("analyze the data", "look up papers")? Hand-wavy steps are rejected.
- Tool fit. Is the chosen tool/skill the right one for the job? Cross-reference available skills (e.g.
asta-tools:find-literaturefor literature search,asta-tools:experimentfor running experiments). Flag misuse. - Dependencies. Are inputs from prior steps/work referenced by path? Do later steps actually use earlier outputs?
- Artifacts. Does each step that produces data say where the artifact lands (ideally
work/<slug>/data/...)? - Scope. Does the plan actually accomplish the Goal? Or does it overshoot / undershoot?
- Realism. Are runtimes, dependencies, and credentials plausible given the environment?
Verdict rules
- Approve only if all six checks pass.
- Otherwise reject and write
# Review Feedbacklisting the failed checks with one bullet per issue. Reference the specific step number.
Out of scope
- Rewriting the plan. The reviewer only flags issues; plan-work revises.
- Assessing executed work. That is review-work.
- Modifying anything outside the work README.