Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
70 changes: 70 additions & 0 deletions .agents/skills/release-review/SKILL.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,70 @@
---
name: release-review
description: Final gate before an lspkit release. A static, adversarial comparison of a clean main commit against the last release that hunts regressions, leaks, API breaks and weakened gates, and ends in a SAFE or BLOCK verdict. Use when the user says "release review", "regression review", "review against the release", or before tagging a release.
---

# Release Review

This is the last step before `v*` is tagged. **SAFE** means the commit can ship as is. You own every verdict.

## Preconditions (any failure → BLOCK, stop)

- `git fetch --tags origin`. The current branch is `main`, and `HEAD` equals `origin/main`.
- `git status --porcelain` is empty: no staged, unstaged or untracked files, and no stashes kept for this commit.
- The PR that produced `HEAD` passed every check: `gh pr list --state merged --search <sha>`, then `gh pr checks`.
- Baseline: `T=$(gh release view --json tagName -q .tagName)`. Review `git diff $T..HEAD`.

## Static comparison

Read every changed file in full, the old version beside the new (`git show $T:<path>`). Start with the highest blast radius:

1. 🛑 Memory lifetimes ([MEM-LIFETIME](../../../docs/specs/lspkit-spec.md#mem-lifetime-memory-lifetimes)):
- a new ownership cycle;
- a registration with no RAII owner;
- an unowned `tokio::spawn`;
- a keyed map with no eviction path;
- a fix or long-lived structure with no leak regression.
2. Dispatch, cancellation and JSON-RPC framing (`crates/lspkit-server`), sidecar and correlator (`crates/lspkit`).
3. Live scheduler, watcher, VFS and position encoding.
4. Profiler service, sessions and sampler (`crates/lsp-profiler*`), the .NET engine (`dotnet/`), and the profiler UI webviews (`packages/lsp-profiler-ui`).
5. Release plumbing: `release.yml` and `scripts/release-assets.sh`.

Every removed guard, error path, bound or default is a suspect. Trace each change to its callers.

## Must-check list (each hit is a finding)

- **Public API:** list removed or changed `pub` items and exported TS symbols. Any break means the next version needs a minor bump, and the break must be listed in the release notes.
- **Gates weakened:**
- in any of `coverage-thresholds.json`, `.github/workflows/`, `Makefile`, `deny.toml`, `rust-toolchain.toml`, lint configs (`[lints]`, `.oxlintrc.json`, `Directory.Build.props`): a lowered threshold, a raised timeout, a relaxed `if:`, or a lint turned off;
- anywhere in source: a new `#[allow]`, `oxlint-disable` or `#pragma warning disable`.
- **Tests weakened:**
- `git diff $T..HEAD --diff-filter=D --stat`;
- removed `assert`/`expect` lines;
- fewer permutations;
- new `sleep` or timing probes;
- tests that re-implement the code they test.
- **Dependencies:** check the `Cargo.lock` and npm lockfile diffs. A new crate or package needs a reason and a clean history, and versions must stay in sync across manifests and workflows.
- **Forbidden in production code:** `unwrap`/`expect`/`panic!`/`unsafe`, unbounded channels or queues, and `[GROUP-…]` citations that resolve to nothing.

## Legwork

- You may use at most two `pi` sessions (`zai/glm-5.3`). Give each one a single file set, the diff range and concrete questions. Register each on TMC.
- Their output is leads. Confirm every claim in code yourself. Kill a pi that stalls.
- ❌ Subagents, ❌ worktrees, ❌ edits or commits during the review.

## Tests

- This is static: ❌ `make test`/`make ci`, ❌ full suites, ❌ triggering Actions.
- You may run one targeted test (`cargo test -p <crate> --test <file> <name>`) only to confirm a specific suspect.

## Verdict

- A finding is **confirmed** only when you have traced the mechanism and can state input → wrong output. Drop anything you can't confirm.
- **BLOCK** on any confirmed regression, new leak, weakened gate or test, or unlisted API break. Otherwise **SAFE**, with the version bump the API diff calls for.
- File a GitHub issue **only** for a bug that is already in the released tag `$T`. Confirm it at `git show $T:<path>`, search `gh issue list --search` for a duplicate, then run `gh issue create --label bug` with the tag, file:line, repro, and expected vs actual.
- A bug new since `$T` gets **no issue**. It blocks the release: fix it through [fix-bug](../fix-bug/SKILL.md) on a branch and a PR, then re-run this review on the new `main`.
- This review never tags or releases.

## Reply

Give the verdict first, then confirmed findings ordered by severity, then exactly what you reviewed and what you did not.
1 change: 1 addition & 0 deletions .claude/skills/release-review
Loading