Permit0 AI Permit0 Permit0

Reviewing Plan Docs for permit0

When to Use

Use this skill when the user asks you to review, audit, critique, or validate planning documents under docs/plans/<feature>/. The output is a set of review docs written to docs/plan-reviews/<feature>/.

Workflow

Phase 1: Read the Plan

Read every file in docs/plans/<feature>/ in numbered order. For each doc, note:

• What it claims (the design decisions, schemas, code changes).
• What assumptions it makes.
• What it explicitly defers or omits.

Phase 2: Codebase Walkthrough

Do a thorough read of the actual source code that the plan references. This is not optional -- you must verify claims against real code.

For integration plans, always read:

1. The reference implementation the plan says it is based on. For example, if the plan says "replicate the Claude Code integration," read crates/permit0-cli/src/cmd/hook.rs end to end.
2. The crate(s) the plan says it will modify. Read every file in those crates, not just the ones the plan mentions.
3. The crates the plan says are NOT modified. Spot-check at least the public API surface to verify the plan's claim that no changes are needed.
4. Packs and risk rules if the feature touches normalization or scoring (read packs/permit0/email/ as the reference pack).
5. The daemon/serve code if the feature involves --remote mode (crates/permit0-cli/src/cmd/serve.rs).
6. Tests -- read existing test files to understand current coverage and what the plan's new tests need to complement.

Phase 3: Write the Review

Write review docs to docs/plan-reviews/<feature>/. One review file per plan doc, plus a summary.

Output Structure

docs/plan-reviews/<feature>/
  00-summary.md               # Overall verdict and key findings
  01-review-overview.md       # Review of 00-overview.md
  02-review-protocol.md       # Review of 01-protocol.md
  03-review-implementation.md # Review of 02-implementation.md
  04-review-configuration.md  # Review of 03-configuration.md
  05-review-testing.md        # Review of 04-testing.md
  06-review-limitations.md    # Review of 05-limitations.md

Skip review files for plan docs that don't exist (e.g. if there is no 01-protocol.md, skip 02-review-protocol.md).

Review Doc Format

Every review doc must follow this structure:

# Review: <plan doc title>

**Reviewer:** Cursor Agent (session <timestamp or ID>)
**Plan doc:** `docs/plans/<feature>/<filename>`
**Review date:** <date>

## Verdict

One of: APPROVE | APPROVE WITH COMMENTS | REQUEST CHANGES | REJECT

## Summary

2-3 sentences: what the plan doc gets right, what it gets wrong.

## Detailed Findings

### Finding 1: <short title>

**Severity:** Critical | Major | Minor | Nit
**Location:** <section or line reference in the plan doc>
**Claim:** <what the plan says>
**Reality:** <what the code actually does>
**Recommendation:** <what to change>

### Finding 2: ...

## Verified Claims

List the specific claims you checked against the codebase and confirmed
are correct. This is as important as the findings -- it shows the review
was thorough.

## Questions for the Author

Numbered list of open questions that the reviewer could not resolve by
reading the code.

Summary Doc Format (00-summary.md)

# Plan Review Summary: <feature>

**Reviewer:** Cursor Agent (session <timestamp or ID>)
**Review date:** <date>
**Plan location:** `docs/plans/<feature>/`

## Overall Verdict

One of: APPROVE | APPROVE WITH COMMENTS | REQUEST CHANGES | REJECT

## Key Findings

Bulleted list of the most important findings across all review docs,
ordered by severity.

## Statistics

| Metric | Count |
|--------|-------|
| Plan docs reviewed | N |
| Critical findings | N |
| Major findings | N |
| Minor findings | N |
| Nits | N |
| Verified claims | N |
| Open questions | N |

## Recommendation

1-2 paragraphs: should the team proceed with implementation as-is,
or are changes needed first?

Distinguishing Your Review

Multiple agents may review the same plan. To make your review distinguishable:

1. Always include the reviewer line: **Reviewer:** Cursor Agent (session <unique identifier>) -- use the current timestamp or conversation ID.
2. Cite specific file paths and line numbers from the codebase when verifying or disputing claims. Other reviewers may make the same point but your evidence trail will differ.
3. Include a "Verified Claims" section -- this is your proof-of-work showing which parts of the codebase you actually read.

What Makes a Good Review

DO

• Verify every code snippet in the plan against the actual source. Plans often show idealized code that doesn't match current function signatures or module structure.
• Check for missing error handling. Plans often describe the happy path and skip what happens on failure.
• Validate the "files NOT changed" list. If the plan claims the engine is untouched, confirm no engine changes are actually needed.
• Look for wire-format mismatches. If the plan describes a JSON schema for communication between two components, verify both sides agree. This is where bugs hide (e.g. one side sends "human", the other expects "humanintheloop").
• Check that test coverage matches the claimed behavior. If the plan says "HITL maps to deny," there should be a test for that.
• Flag implicit assumptions. If the plan assumes network access in a sandboxed environment, call it out.

DON'T

• Don't rubber-stamp. If you only found nits, you probably didn't read deeply enough.
• Don't rewrite the plan. Findings should be specific and actionable, not alternative designs.
• Don't skip the codebase walkthrough. A review based only on reading the plan docs is not a review -- it's proofreading.

Severity Definitions

Source: permit0-ai/permit0 — distributed by TomeVault.