The vlang code-review tooling, moved out of per-checkout .claude/ directories
into one git repository so it runs identically on Windows (PowerShell / git-bash),
WSL and macOS. Git is the sync mechanism; nothing is copied by hand.
| Path | Contents |
|---|---|
skills/vreview/ |
The /vreview review process (SKILL.md) |
skills/code-review/ |
run_detectors.sh (thin bash runner kept for WSL/macOS habit; same manifest), H2/QUIC conformance matrices, rfc-texts/. The shell detectors were retired once their V ports passed the same fixtures |
skills/phase-r/, skills/review-respond/, skills/adversarial-review/ |
Companion skills (see below) |
detectors/ |
The detectors, one V program each (detect_*.v, same <file> <base_copy_or_empty> contract), and run_detectors.v, the batch runner: computes targets and base revision with git, compiles the detectors once into a hash-keyed cache under the OS temp dir, dispatches by a basename manifest. No bash needed |
probes/ |
vprobe.v, the executed probes /vreview hard rule 4 requires: matrix (base build against head build per place, kind of object and account), race (several copies of a command at once), mutate (one test run per one-line mutant, generated for the functions that changed), trace (which inputs each caller of a function is tested with) and siblings (which sites of one situation lack a call another site gained). The file's header describes the spec |
hooks/ |
hooks.json (plugin hook wiring), require_code_review.sh (commit gate), check_powershell_parse.ps1, check_ps1_file_parse.ps1 (reference copies, wired globally) |
reference/ |
code-review-reference.html quick reference, its code-review-misses-data.js table data, and the code-review-misses.md narrative corpus |
tools/ |
review_cost.v: what /vreview and /adversarial-review runs cost, read from Claude Code session transcripts (see below) |
| Skill | Use it for |
|---|---|
/review-tools:vreview |
Reviewing your own change before every commit. Writes the marker the commit gate checks |
/review-tools:phase-r |
A bug someone else found: reproduce it on the unfixed code before fixing it |
/review-tools:review-respond <PR#> [review or comment] [owner/repo] |
One round of review on a PR, from a maintainer, another person or a bot such as Codex: fetch the findings (review bodies, inline comments and plain PR comments), check each against the code, reproduce each real bug with /phase-r, fix, run /vreview, commit, push to your fork, and reply where each finding was posted. Replaced /codex-respond on 2026-11-14 |
/review-tools:adversarial-review |
A multi-agent pass that tries to break a finished change; run it only with the user's go-ahead |
/luna-respond was removed on 2026-11-14: it was a placeholder for a review service that was
never connected.
v run tools/review_cost.v [projects-dir] [--since YYYY-MM-DD] [--list] [--selftest] reads
the session transcripts under ~/.claude/projects (or the folder given) and prints one table. A
/vreview run lasts from the command or Skill call to the end of that turn, with the subagents
it started. A stop that a Stop hook blocks does not end the turn, and a command typed after an
interrupted run starts a new run. An /adversarial-review run is one Workflow run whose name,
summary or the first 400 characters of its script mention "adversarial", counted from the
subagent transcripts of that run; the main session's work around it is left out. Weighted cost
counts input 1, cache writes 1.25 (5 min) or 2 (1 h), cache reads 0.1 and output 5. --list
prints each adversarial run, and --selftest checks the counting on made-up transcripts.
On this machine, 2026-11-15 (all projects):
| vreview | adversarial-review | |
|---|---|---|
| Runs found | 228 | 80 (70 since 2026-10-15) |
| Median model calls | 14 | 237 |
| Median output tokens | 17K | 159K |
| Median new input | 97K | 1.17M |
| Median weighted cost | 900K | 4.54M |
| Recent runs (2026-10-15+) | 655K median (140K to 4.31M) | 3.82M median (134K to 44.95M) |
A /vreview run that goes on to fix or commit in the same turn counts that work too, so its
figures run high.
See CONTRIBUTING.md: a rev is the rule, its reference page entry and its corpus rows in one commit, and the commit gate checks all three.
See MIGRATION.md. Short form: clone, claude plugin marketplace add <clone>,
claude plugin install review-tools@review-tools-local --scope user; update with git pull.
Run the detector fixture suite with bash tests/run_fixtures.sh (see tests/README.md). Run the
probe self-test with bash tests/vprobe/selftest.sh [path/to/v] on Linux, WSL or macOS.
For OpenCode (skills + /vreview slash commands, no Claude Code needed),
see OPENCODE.md.
The first commit is a verbatim copy of the canonical files as of 2026-10-18:
C:\Users\john3\.claude\skills\{vreview,code-review,phase-r,codex-respond,luna-respond,adversarial-review},
S:\repo\vlang\.claude\hooks\, C:\Users\john3\.claude\hooks\check_*.ps1 and
S:\Code Review\. The HTML reference was at rev 66.
Later commits (in order): fixture suite (tests/), Claude Code plugin packaging
(.claude-plugin/), retirement of sync_skills.sh, runner-owns-git rewrite of
run_detectors.sh, and the port of detectors to V (detectors/*.v).