polar-code-review

# Polar Code Review Review the diff against the things only Polar knows: its conventions, its Accepted ADRs, its shared helpers, its API contract, its deploy shape, its billing domain. Not a bug hunt and not a security review — `/code-review` and `/security-review` do those better, and they verify findings and can apply fixes. This adds what they cannot know. This is only a router. Every lens is a skill you can also run on its own. ## 1. Diff ```bash git fetch origin main git diff origin/main --stat git diff origin/main ``` That is the branch plus the working tree. If fetch fails, use `main` instead of `origin/main`. Keep the full diff — every agent gets it, not just file names. ## 2. Route Run the first four always. Add the rest only when the trigger matches. | Skill | Runs when | |---|---| | `conventions-check` | always | | `adr-check` | always | | `reuse-check` | always | | `slop-check` | always | | `api-surface-review` | `**/schemas.py`, `**/endpoints.py`, `polar/openapi.py`, `docs/openapi.json`, `sdk/`, `clients/packages/client/` | | `ship-safety` | `migrations/versions/`, `**/tasks.py`, `polar/models/`, `server/scripts/`; an endpoint is removed; or the diff spans `server/` and `clients/` with a dependency between them | | `billing-review` | the billing paths listed in that skill's Scope section | ## 3. Launch Launch one independent subagent per selected skill in a single turn. They must not wait on each other. Use the host's general-purpose agent type (`general-purpose` in Claude Code, `generalPurpose` in Cursor, `general` if that is all the host exposes). Same prompt for every lens: ``` Read `.agents/skills/<skill name>/SKILL.md` and follow it exactly. Review this diff. Only changed lines are in scope — never the rest of the repo. [INSERT FULL DIFF] Use the output format the skill defines. High-confidence findings only. If you are unsure whether something is a defect, put it under Question instead of asserting it. ``` Each skill declares what it does **not** own, so the lenses do not overlap by construction. `adr-check` predates this skill and has its own terser format: either `No violations`, or a list of ADR id, `file:line`, what breaks, and the fix. Map its findings to 🔴. ## 4. Merge - **Deduplicate.** One line, one finding. Precedence when two lenses hit the same line: `billing-review` → `api-surface-review` → `ship-safety` → `adr-check` → `conventions-check` → `reuse-check` → `slop-check`. - **Surface conflicts.** If two lenses disagree, say so rather than picking silently. - **Cut the padding.** A short report that is all true beats a long one that is half true. ## 5. Report ```markdown ## Polar Review ### 🔴 Blocking - **[lens]** `file:line` — <what breaks>. Fix: <fix> ### 🟠 Should fix - **[lens]** `file:line` — <claim>. Fix: <fix> ### 🟡 Questions - **[lens]** `file:line` — <question> ### 🧹 Delete - `file:line-line` — <why> ### Notes <deploy notes from ship-safety, surface delta from api-surface-review. Omit if neither ran.> ### Coverage Ran: <...>. Skipped: <...> (no matching paths). ### Verdict ✅ APPROVED | ❌ CHANGES REQUESTED — n blocking, n should-fix ``` Any 🔴 means CHANGES REQUESTED. 🟠 alone is a judgement call: say which way you lean and why. The 🟡 section is expected to have content — Polar reviews are mostly questions.