kms-pr-review

# KMS PR Review — two-pass diff gate Diff-scoped review gate, following `.github/skills/shared/orchestrator-contract.md` (phase order, report schema, sub-agent routing). It is lighter than `/rust-review-all` and `/meta-security` and **does not replace them** for release or security-audit requests. ## Step 0 — Anti-hallucination Read `.github/skills/shared/anti-hallucination.md` before analysing anything. ## Step 1 — Mode and scope - First run `git fetch origin` so the base ref is current. - Argument is a PR number or URL → **PR mode** (report-only). Scope: `gh pr diff <N> --repo Cosmian/kms`. - No argument → **self-review mode**. `<base>`: the PR base if a PR exists for the branch (`gh pr view --json baseRefName -q .baseRefName`), else the upstream branch of a stack, else `develop`. Scope is the union of: - `git diff --name-only origin/<base>...HEAD` - `git diff --name-only HEAD` - `git diff --name-only --cached` - `git ls-files --others --exclude-standard` Never treat an empty or failed diff as clean: report it and stop. ## Step 2 — Phase 1: Structural Scan Run as a sub-agent with cheap/fast routing (contract §6), or inline when the harness has no sub-agent tool. Scan file scope, imports, naming, obvious anti-patterns, hardcoded values, missing error handling, surface risks. If `/kms-caveman` already ran this session on overlapping scope, feed its findings in instead of re-scanning. Run a domain skill only when the diff touches its area; skip it otherwise (token cost). They are report-only. Add their findings to `review-YYYY-MM-DD/01-structural.md` as `S<n>` claims tagged with the source skill, so Phase 2 still disposes of each one. Phase 2 must give a confirmed finding from `/cryptography-review`, `/security-review`, or `/kmip-compliance` the disposition `needs-human`, never `confirmed`/`adjusted`, so no crypto, security, or KMIP change is auto-applied without a human: | Diff touches | Run | |---|---| | `crate/crypto/`, `openssl_providers.rs`, algorithm or key-size choice | `/cryptography-review <paths>` | | `crate/server/src/middlewares/`, `routes/`, `core/operations/`, `crate/access/`, cloud-provider routes, parsing of untrusted input | `/security-review <paths>` | | `crate/kmip/`, `crate/server/src/core/operations/` | `/kmip-compliance <operation>` | In PR mode, also run `/kms-sync-rules` on the PR diff (report-only) and add each unmet rule as an `S<n>` claim. In self-review mode it already ran (`AGENTS.md` §5 step 4). Never run `/meta-security` here (an audit orchestrator, not a diff gate). Write `review-YYYY-MM-DD/01-structural.md`. Each claim has an ID `S<n>`, file:line, evidence, and a recommended fix. ## Step 3 — Phase 2: Semantic Verification Run as a sub-agent with reasoning routing (contract §6), or inline when the harness has no sub-agent tool. Inputs: `review-YYYY-MM-DD/01-structural.md`, the live diff and source, `AGENTS.md`, and the applicable `.github/instructions/*.instructions.md`. - Give every `S<n>` exactly one disposition: `confirmed`, `rejected` (with reason), `adjusted` (with corrected severity/line/fix), `[UNVERIFIED]`, or `needs-human`. Use `needs-human` when the fix needs a design decision, changes public API or behavior outside the task, or needs external access. - Add missed findings as `M<n>` with the same fields (correctness, concurrency, edge cases, cross-file impact, contract compliance). - Write `review-YYYY-MM-DD/02-semantic.md` with the disposition table, the findings, and a **GO / NO-GO** verdict. If the disposition table is missing, the report is invalid: redo Step 3. ## Step 4 — Final report Write `review-YYYY-MM-DD/final.md` using the Unified Report Schema (contract §4). It adds no new findings, links `review-YYYY-MM-DD/01-structural.md` and `review-YYYY-MM-DD/02-semantic.md`, and records `Cycle: <k>` (0 = first review). ## Step 5 — Auto-fix loop (self-review mode only) 1. Verdict GO and no `confirmed`/`adjusted` findings left unapplied → done. 2. Apply the recommended fix of every `confirmed` and `adjusted` finding, at any severity. Never apply `rejected`, `[UNVERIFIED]`, or `needs-human` findings. 3. Re-run `AGENTS.md` §5 step 4 (`/kms-sync-rules`) on the files the fixes touched and apply its rows (for example CLI doc regeneration). When Rust files changed, also re-run step 1 (targeted tests for the touched code) and step 3 (`cargo clippy-all`, `cargo fmt --all`). A newly failing test, warning, or unmet sync rule is a `confirmed` finding of the next cycle. 4. Re-run Steps 1–4 with `Cycle: k+1`, overwriting the `review-YYYY-MM-DD/*.md` files. 5. Stop after **2 fix cycles**: the review at `Cycle: 2` is the last. If it is NO-GO, or any `needs-human`, `confirmed` or `adjusted` finding is still unapplied, stop, report those findings from `review-YYYY-MM-DD/final.md` to the user, and do not start another cycle. ## Step 6 — Output rules - PR mode: never edit files. - Never post to GitHub, push, or resolve review threads. - Attribute each finding to its phase; list blockers first. `/kms-adversarial-test` is the complementary live adversarial test against a running KMS.