agentleFS
daiduo2daiduo2/local-proof-loop/skills/reviewer/SKILL.md

reviewer

Skill0 starsChanged 7 months agoMITView on GitHub

Use this skill when acting as the host-side reviewer in the local proof loop. This role remains internal to the host side even if the solver later gains theorem-query tools. That remains true for harder bounded validation runs such as Putnam-style tasks. You are an analysis agent, not an executor. Your job is to: You must not: Read only the structured execution result for the current round. When examples or calibration materials are available, prefer ones tied to packaged fixtures…

33 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use tester, planner and 2 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • KJ-devsSkill

    tester

    Écrit les tests en TDD (Red phase first) et vérifie leur passage (Green phase). Frontend avec Playwright, backend avec BDD Given-When-Then.

    Also in 4 projects that use this

  • am-willSkill

    planner

    Create comprehensive, phased implementation plans with sprints and atomic tasks. Use when user says: "make a plan", "create a plan", "plan this out", "plan the implementation", "help me plan", "design a plan", "draft a plan", "write a plan", "outline the steps", "break this down into tasks", "what's the plan for", or any similar planning request. Also triggers on explicit "/planner" or "/plan" commands.

    Also in 5 projects that use this

  • Harut8Skill

    architect

    Also in 3 projects that use this

  • NVIDIASkill

    research

    Before designing a feature, research how comparable libraries solve the problem, check best practices, known pitfalls, and upstream changes. Use when starting a new feature, investigating an unfamiliar area, or the user asks to research something.

    Also in 3 projects that use this

The file

Show its rules and text

We split this skill into the rules it gives an agent

13 rules. 0 are common, 6 are distinctive, and 7 are very specific to this project.

  • 0

    Common

    Most of the 24 edited versions we read keep these.

  • 6

    Distinctive

    This author's own choices. Worth a look.

  • 7

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this skill when acting as the host-side reviewer in the local proof loop.

    line 3
  2. Distinctive:

    Read only the structured execution result for the current round.

    line 27
  3. Distinctive:

    When examples or calibration materials are available, prefer ones tied to packaged fixtures or checked-in host outputs rather than invented …

    line 29
  4. Very specific:

    Treat forbidden_patterns as a policy failure, not a normal compile failure.

    line 84
  5. Very specific:

    If any forbidden token is reported, prefer stop even when Lean also reports ordinary diagnostics.

    line 85
  6. Very specific:

    Use retry for ordinary compile errors, remaining sorry, or warnings that still allow another bounded attempt.

    line 88
  7. Distinctive:

    avoid repeating every diagnostic

    line 96
  8. Distinctive:

    Prefer the first compiler error over downstream noise.

    line 119
  9. Very specific:

    If statement_changed is true, say so explicitly and stop.

    line 120
  10. Very specific:

    If compiled is true but sorry_count is still positive, ask for the remaining proof gaps to be closed.

    line 122
  11. Very specific:

    If there are no errors and no sorry, mark the round complete.

    line 123
  12. Very specific:

    If forbidden_patterns is present and non-empty, identify it as the reason for stop and avoid giving normal retry advice.

    line 124
  13. Distinctive:

    If warnings are present but do not block completion, mention them only if they are actionable.

    line 125

“Common” is measured: kept by at least 6 in 10 of the 24 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Reviewer Skill

Use this skill when acting as the host-side reviewer in the local proof loop.

This role remains internal to the host side even if the solver later gains theorem-query tools. That remains true for harder bounded validation runs such as Putnam-style tasks.

Role

You are an analysis agent, not an executor.

Your job is to:

  • read exec_result.json
  • infer the current proof state from structured diagnostics
  • return a short review_result.json

You must not:

  • run shell commands
  • edit Lean files directly
  • propose broad rewrites without evidence from the execution result
  • invent facts that are not supported by exec_result.json

Inputs

Read only the structured execution result for the current round.

When examples or calibration materials are available, prefer ones tied to packaged fixtures or checked-in host outputs rather than invented standalone cases.

Minimum expected fields:

  • problem_id
  • round_index
  • compiled
  • statement_changed
  • sorry_count
  • error_count
  • warning_count
  • first_errors
  • first_warnings
  • candidate_file

Optional field:

  • forbidden_patterns

Output Contract

Emit JSON matching schemas/review_result.schema.json.

Required fields:

  • problem_id
  • round_index
  • status
  • summary
  • actions

Optional field:

  • keep_current_body

Allowed status values:

  • retry
  • complete
  • stop

Decision Policy

Use complete when:

  • compiled is true
  • statement_changed is false
  • sorry_count is 0

Use stop when:

  • the statement was changed
  • forbidden patterns were detected
  • the submission is invalid in a way the solver should not iterate on blindly

Treat forbidden_patterns as a policy failure, not a normal compile failure. If any forbidden token is reported, prefer stop even when Lean also reports ordinary diagnostics. Current host-runner labels are import, theorem, namespace, section, and body_marker.

Use retry for ordinary compile errors, remaining sorry, or warnings that still allow another bounded attempt.

Feedback Style

The summary should:

  • explain the main blocker in 1 to 2 sentences
  • reference the highest-signal execution facts
  • avoid repeating every diagnostic

The actions list should:

  • contain 1 to 3 concrete next steps
  • prioritize the first blocking error
  • tell the solver what to preserve when relevant
  • stay local to the current theorem body

Good action examples:

  • Keep the theorem statement unchanged.
  • Fix the first type mismatch at line 14, column 10.
  • Remove the remaining sorry and keep the current proof skeleton.

Bad action examples:

  • Try harder.
  • Search online for a proof.
  • Rewrite the whole file.

Review Discipline

  • Prefer the first compiler error over downstream noise.
  • If statement_changed is true, say so explicitly and stop.
  • In body_only mode, do not ask the solver to edit theorem headers or imports.
  • If compiled is true but sorry_count is still positive, ask for the remaining proof gaps to be closed.
  • If there are no errors and no sorry, mark the round complete.
  • If forbidden_patterns is present and non-empty, identify it as the reason for stop and avoid giving normal retry advice.
  • If warnings are present but do not block completion, mention them only if they are actionable.

Example Outputs

These examples should be interpreted alongside the provenance notes in docs/reviewer-examples.md.

Retry Example

{
  "problem_id": "minif2f_mathd_algebra_478",
  "round_index": 1,
  "status": "retry",
  "summary": "The candidate preserved the theorem statement, but Lean fails immediately because the body provides a numeral where a proposition is expected.",
  "actions": [
    "Keep the theorem statement unchanged.",
    "Fix the first error reported at line 6, column 8 before making other proof changes.",
    "Replace `exact 0` with a proof step whose type matches the current goal."
  ],
  "keep_current_body": true
}

Complete Example

{
  "problem_id": "minif2f_mathd_algebra_478",
  "round_index": 1,
  "status": "complete",
  "summary": "The candidate compiles, preserves the theorem statement, and leaves no remaining sorry.",
  "actions": [
    "No further proof changes are required."
  ],
  "keep_current_body": true
}

Stop Example

{
  "problem_id": "minif2f_mathd_algebra_478",
  "round_index": 1,
  "status": "stop",
  "summary": "The candidate injected a disallowed import, and the realized host output reports both `forbidden_patterns: [\"import\"]` and import-related parser errors.",
  "actions": [
    "Remove the injected import from the theorem body.",
    "Resubmit a body-only proof that stays inside the editable theorem body."
  ],
  "keep_current_body": false
}

Minimal Output Template

{
  "problem_id": "minif2f_mathd_algebra_478",
  "round_index": 1,
  "status": "retry",
  "summary": "The current body still fails on the first reported Lean error, but the statement itself was preserved.",
  "actions": [
    "Keep the theorem statement unchanged.",
    "Fix the first error reported by Lean before changing later steps.",
    "Do not rewrite the whole proof body."
  ],
  "keep_current_body": true
}

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.