| name | artifact-inline-no-tools-plan-review |
| version | 1.0.0 |
| description | Recover adversarial plan reviews on large or sandbox-hostile repos by embedding grounded facts + full plan text directly in a compact prompt and explicitly forbidding reviewer tool use. |
| category | software-development |
| tags | ["review","adversarial-review","prompt-packaging","codex","gemini","plan-review"] |
Artifact-inline no-tools plan review
Use when adversarial plan reviews keep failing because providers try to search the repo, read ignored files, or hit sandbox/path/tooling problems.
When this fits
- Codex/Gemini review runs fail due file access, ignored-path restrictions, sandbox errors, or repo-scale grep/search timeouts
- You already have the key grounded facts and the full plan text locally
- You want a fast rerun review focused on the plan artifact itself, not exploratory repo browsing
Core pattern
- Build a compact prompt file in the real workspace, not
/tmp/.
- Put these into the prompt explicitly:
- adversarial reviewer instructions
- grounded facts already verified by you
- the exact review questions
- the full current plan text
- Add an explicit restriction:
- "Do NOT use tools, repo search, file reads, or web access. Review ONLY the grounded facts and plan text included below."
- Dispatch Codex/Gemini with the compact prompt.
- Save canonical review artifacts under
scripts/review/results/.
- If Gemini still fails due capacity, save an
UNAVAILABLE artifact instead of leaving the slot blank.
Why this works
On large repos, providers may waste the run on:
- ignored-path read failures
- repo grep timeouts
- sandboxed shell/file-access failures
- startup agent noise that derails actual review
Inlining the artifact and forbidding tool use turns the run into pure document review. This worked well for worldenergydata plan reruns on issues #343 and #344 after earlier attempts failed or produced noisy/non-substantive runs.
Minimal prompt skeleton
Adversarial review request for Issue #NNN.
Rules:
- You are adversarial. Assume defects until proven otherwise.
- Do not praise. Do not restate the plan.
- Focus only on what is wrong, missing, risky, ambiguous, or insufficiently grounded.
- IMPORTANT: Do NOT use tools, repo search, file reads, or web access. Review ONLY the grounded facts and plan text included below.
- Return APPROVE only if the draft is truly implementation-ready.
- Every finding must cite a specific plan section or quoted claim.
Grounded facts:
- ...
- ...
Review questions:
1. ...
2. ...
Plan under review:
[full plan text]
Required output format:
Verdict: APPROVE | MINOR | MAJOR
Retrieval adequacy: adequate | inadequate
Key findings:
- <severity> <finding>
Main blockers to fix:
1. <item>
2. <item>
Implementation notes
- Write prompt files with
terminal/real filesystem paths under .planning/quick/
- Prefer absolute paths in
$(cat ...) when dispatching to avoid cwd surprises
- Save raw compact outputs separately from canonical review artifacts
- If the review returns duplicated output, keep only the canonical distilled artifact in
scripts/review/results/
Good use cases
- plan review reruns after first-wave provider/tool failures
- large repos with ignore rules that block provider tool access
- architecture/planning reviews where the artifact text is the real object under review
Not for
- code review where the diff itself is too large and must be inspected from the repo
- cases where the grounded facts are weak or unverified
- implementation validation that truly requires runtime inspection