How long does the integration take?

Change one base URL in an OpenAI-compatible client, or install the Python or TypeScript SDK; the quickstart routes a first request in a few minutes.

Will this lock us into another vendor?

No. The router is Apache-2.0 open source, your provider keys stay yours, and the API is OpenAI-compatible, so leaving is a change of base URL.

What happens when a provider fails?

The router falls back to the next model your policy allows and records the fallback in the trace, so an outage costs a retry, not an incident.

run-reviewers - Prompt - OpenSmartRoute

Name: run-reviewers
Author: koblas

Prompt file imported from koblas/quarry (.claude/commands/run-reviewers.md). Fill in {{arguments}} before use. Copyright stays with the author.

Run reviewers on: {{arguments}}

Step 1: List target files

Paths provided (comma-separated, e.g. src/main, src/test): split on commas, trim, use Glob to list all files under each path:

Glob(pattern="**/*", path="<path1>")
Glob(pattern="**/*", path="<path2>")

Run all globs in parallel (single message).

No paths (pipeline mode): detect changed files via git from <base>, then combine into a deduplicated list. <base> is the start of the range:

First gate — <base> = git merge-base HEAD origin/main: every commit on the feature branch, not only the newest.
Re-gate after a fix pass — <base> = the commit the fix pass started from: the fix commits only, not the whole feature branch. The orchestrator records git rev-parse HEAD before dispatching each fix pass and names that commit in the invoking prompt. Missing on a re-gate → ask for it; never default to the merge-base.

git diff --name-only <base> 2>/dev/null          # committed since <base> + uncommitted tracked
git ls-files --others --exclude-standard 2>/dev/null

Both empty → stop and report; never fall back to reviewing the whole tree (a missing origin/main empties the first line silently).

Pass the range on in Step 5 as <base>..HEAD (plus uncommitted work). A reviewer given a range reads a diff; one given nothing re-reads whole packages, which is where a re-gate's cost actually goes.

Step 2: Discover reviewer agents

Use Grep to find agents with type: reviewer in frontmatter. Both searches in parallel:

Grep(pattern="type: reviewer", path="$HOME/.claude/agents/", glob="**/Agent.md")
Grep(pattern="type: reviewer", path=".claude/agents/", glob="**/Agent.md")

Global path missing → skip silently; project reviewers are enough.

Sanity check: this step yields zero reviewers → stop, report REVIEWER DISCOVERY FAILED. Do not continue, do not report PASS. Every reviewer lives at .claude/agents/<name>/Agent.md with type: reviewer + a triggers: list in frontmatter; zero matches means layout drifted, not that code is clean.

For each matched file, Read only first 10 lines (frontmatter). Check type: reviewer appears inside the YAML frontmatter block (between --- markers), not body text. Discard files where it only appears in body.

From each valid reviewer's frontmatter extract name + triggers. Read all matched files in parallel.

Step 3: Apply project trigger overrides

.claude/review-triggers.json exists in project root → read it, override triggers for matching reviewer names. Missing → skip this step.

Step 4: Filter by relevance

For each reviewer, check whether ANY target file matches ANY of its triggers globs (after overrides). Skip reviewers with no matching files.

Step 5: Launch relevant reviewers in parallel

Spawn all matching reviewers in a single message via Agent:

Agent(subagent_type="<name>", prompt="<pipeline mode only — run: review feature: <slug> unit: -
(CLAUDE.md → *Rules* → Run tags)>
Review <commit range, or the listed paths>. Read
.claude/briefs/review.md first<, and .claude/briefs/proof.md — only for test-reviewer and
correctness-reviewer><, and .claude/briefs/navigation.md — only for arch-reviewer and
correctness-reviewer>. Scope: <the matched files, listed>. Start from the diff
and read only what it touches; widen only when the diff cannot settle a question, and say
which finding forced it. Report every finding you have in this round — a MINOR held back for
a later pass costs a whole extra gate. <On a re-gate: your prior findings were X; confirm each
is closed or still open. Findings STATE.md already records as deferred are out of scope.>")

Name the files. A reviewer told only "focus on internal/" reads the package; one handed six paths reads six diffs.

Run the coverage gate once, before spawning, and paste its output into every prompt: .claude/scripts/uncovered-diff.py <base>. Uncovered added lines are an untested-change finding no reviewer needs to rediscover by reading; handing every reviewer the same list stops three of them paying to find it separately. Its rows are grouped per run with the enclosing function, which is also the cheapest form to paste. If it exits 1 (anything outside the "declared unreachable" section, (no coverage block) file rows included), send it back to the developer before spending a review round at all. Exit 2 is a failing test — also back to the developer; never a PASS. Also paste .claude/scripts/test-stats.py --base <base> --changed so reviewers read the test-count deltas instead of recounting.

Then run the mutation sample once, and paste its output into test-reviewer's prompt only: .claude/scripts/mutation-sample.py --base <base> (no --profile needed: the coverage gate above already bounced any uncovered line). It mutates up to 20 changed lines (guards first) in an isolated $TMPDIR copy — never the worktree — and lists each SURVIVED mutant. Costs no model tokens. Exit 2 (tool failure, or unmutated tests already failing) → back to the developer, like the coverage gate. Record its summary line in the round's METRICS.md row. Worst case is --max × --timeout (20 × 120s): run it with run_in_background and wait for it. Skip it when test-reviewer is not in this round or no production .go file changed; on a re-gate run it only if test-reviewer re-runs.

Do NOT review code yourself — only orchestrate.

Spawn fails with Agent type '<name>' not found → agent was added after session start, roster not re-enumerated. Report REVIEWER SPAWN FAILED: <name> — restart the session. Do not report PASS for a reviewer that never ran.

Step 6: Report

Reviewers return findings in shared format [BLOCKER|MAJOR|MINOR|NIT] <file>:<line> — <defect> with Failure: and Fix: lines. MINOR/NIT carry Why: instead of Failure: — a smell has no constructible failure by definition. Accept either; missing Failure: on a MINOR is not malformed.

Merge into one severity-ranked list, deduplicating findings two reviewers raised on same file:line. Keep reviewer name as prefix so developer knows who to re-run.

## Review Report

### Target
<path or "changed files">

### Triggered reviewers
- <name>: triggered by <matched files>

### Skipped reviewers
- <name>: no files matched triggers

### BLOCKER
<findings, prefixed with reviewer name>

### MAJOR
<findings, prefixed with reviewer name>

### MINOR
<findings, prefixed with reviewer name>

### NIT
<findings, prefixed with reviewer name>

### Strengths
<positive notes worth another author copying — omit if none>

### Verdict: PASS | PASS WITH FOLLOW-UPS | BLOCKED
- **BLOCKED** — any BLOCKER or MAJOR exists. Must be fixed before feature is done.
- **PASS WITH FOLLOW-UPS** — only MINOR/NIT findings. Feature is done; findings go to
  developer as fix-if-cheap, not a mandatory round trip.
- **PASS** — no findings.

Pipeline mode: persist the report. Write it to docs/specifications/<feature-slug>/REVIEW-<NN>.md (next unused two-digit round) before dispatching any fix pass. Fix-pass prompt cites that file instead of pasting findings, the METRICS.md round row counts from it, and pipeline-reviewer's retro reads it; report left only in conversation is gone.

Style preferences never fail the gate. A refactor-advisor suggestion alone is PASS WITH FOLLOW-UPS, not BLOCKED.

Install run-reviewers and it becomes one more candidate the router can pick - when it fits.

A listing is a routing target with a manifest: what it does, which domains it covers, what it costs and who publishes it. Once installed it sits beside your own models and tools, is scored like any other candidate for each request, and shows up in the trace when it wins. Ratings come from workspaces that installed it, one per account.

Run reviewers on: **{{arguments}}** ## Step 1: List target files **Paths provided** (comma-separated, e.g. `src/main, src/test`): split on commas, trim, use `Glob` to list all files under each path: ``` Glob(pattern="**/*", path="<path1>") Glob(pattern="**/*", path="<path2>") ``` Run all globs in parallel (single message). **No paths** (pipeline mode): detect changed files via git from `<base>`, then combine into a deduplicated list. `<base>` is the start of the range: - **First gate** — `<base>` = `git merge-base HEAD origin/main`: every commit on the feature branch, not only the newest. - **Re-gate after a fix pass** — `<base>` = the commit the fix pass started from: the fix commits only, not the whole feature branch. The orchestrator records `git rev-parse HEAD` before dispatching each fix pass and names that commit in the invoking prompt. Missing on a re-gate → ask for it; never default to the merge-base. ```bash git diff --name-only <base> 2>/dev/null # committed since <base> + uncommitted tracked git ls-files --others --exclude-standard 2>/dev/null ``` Both empty → stop and report; never fall back to reviewing the whole tree (a missing `origin/main` empties the first line silently). **Pass the range on** in Step 5 as `<base>..HEAD` (plus uncommitted work). A reviewer given a range reads a diff; one given nothing re-reads whole packages, which is where a re-gate's cost actually goes. ## Step 2: Discover reviewer agents Use `Grep` to find agents with `type: reviewer` in frontmatter. Both searches in parallel: ``` Grep(pattern="type: reviewer", path="$HOME/.claude/agents/", glob="**/Agent.md") Grep(pattern="type: reviewer", path=".claude/agents/", glob="**/Agent.md") ``` Global path missing → skip silently; project reviewers are enough. **Sanity check:** this step yields **zero** reviewers → stop, report `REVIEWER DISCOVERY FAILED`. Do not continue, do not report PASS. Every reviewer lives at `.claude/agents/<name>/Agent.md` with `type: reviewer` + a `triggers:` list in frontmatter; zero matches means layout drifted, not that code is clean. For each matched file, `Read` only first 10 lines (frontmatter). Check `type: reviewer` appears **inside the YAML frontmatter block** (between `---` markers), not body text. Discard files where it only appears in body. From each valid reviewer's frontmatter extract `name` + `triggers`. Read all matched files in parallel. ## Step 3: Apply project trigger overrides `.claude/review-triggers.json` exists in project root → read it, override triggers for matching reviewer names. Missing → skip this step. ## Step 4: Filter by relevance For each reviewer, check whether ANY target file matches ANY of its `triggers` globs (after overrides). Skip reviewers with no matching files. ## Step 5: Launch relevant reviewers in parallel Spawn all matching reviewers in a **single message** via `Agent`: ``` Agent(subagent_type="<name>", prompt="<pipeline mode only — run: review feature: <slug> unit: - (CLAUDE.md → *Rules* → Run tags)> Review <commit range, or the listed paths>. Read .claude/briefs/review.md first<, and .claude/briefs/proof.md — only for test-reviewer and correctness-reviewer><, and .claude/briefs/navigation.md — only for arch-reviewer and correctness-reviewer>. Scope: <the matched files, listed>. Start from the diff and read only what it touches; widen only when the diff cannot settle a question, and say which finding forced it. Report every finding you have in this round — a MINOR held back for a later pass costs a whole extra gate. <On a re-gate: your prior findings were X; confirm each is closed or still open. Findings STATE.md already records as deferred are out of scope.>") ``` Name the files. A reviewer told only "focus on `internal/`" reads the package; one handed six paths reads six diffs. **Run the coverage gate once, before spawning, and paste its output into every prompt:** `.claude/scripts/uncovered-diff.py <base>`. Uncovered added lines are an untested-change finding no reviewer needs to rediscover by reading; handing every reviewer the same list stops three of them paying to find it separately. Its rows are grouped per run with the enclosing function, which is also the cheapest form to paste. If it exits 1 (anything outside the "declared unreachable" section, `(no coverage block)` file rows included), send it back to the developer before spending a review round at all. Exit 2 is a failing test — also back to the developer; never a PASS. Also paste `.claude/scripts/test-stats.py --base <base> --changed` so reviewers read the test-count deltas instead of recounting. **Then run the mutation sample once, and paste its output into `test-reviewer`'s prompt only:** `.claude/scripts/mutation-sample.py --base <base>` (no `--profile` needed: the coverage gate above already bounced any uncovered line). It mutates up to 20 changed lines (guards first) in an isolated `$TMPDIR` copy — never the worktree — and lists each `SURVIVED` mutant. Costs no model tokens. Exit 2 (tool failure, or unmutated tests already failing) → back to the developer, like the coverage gate. Record its summary line in the round's `METRICS.md` row. Worst case is `--max` × `--timeout` (20 × 120s): run it with `run_in_background` and wait for it. Skip it when `test-reviewer` is not in this round or no production `.go` file changed; on a re-gate run it only if `test-reviewer` re-runs. Do NOT review code yourself — only orchestrate. Spawn fails with `Agent type '<name>' not found` → agent was added after session start, roster not re-enumerated. Report `REVIEWER SPAWN FAILED: <name> — restart the session`. Do **not** report PASS for a reviewer that never ran. ## Step 6: Report Reviewers return findings in shared format `[BLOCKER|MAJOR|MINOR|NIT] <file>:<line> — <defect>` with `Failure:` and `Fix:` lines. MINOR/NIT carry `Why:` instead of `Failure:` — a smell has no constructible failure by definition. Accept either; missing `Failure:` on a MINOR is not malformed. Merge into one severity-ranked list, deduplicating findings two reviewers raised on same `file:line`. Keep reviewer name as prefix so developer knows who to re-run. ``` ## Review Report ### Target <path or "changed files"> ### Triggered reviewers - <name>: triggered by <matched files> ### Skipped reviewers - <name>: no files matched triggers ### BLOCKER <findings, prefixed with reviewer name> ### MAJOR <findings, prefixed with reviewer name> ### MINOR <findings, prefixed with reviewer name> ### NIT <findings, prefixed with reviewer name> ### Strengths <positive notes worth another author copying — omit if none> ### Verdict: PASS | PASS WITH FOLLOW-UPS | BLOCKED - **BLOCKED** — any BLOCKER or MAJOR exists. Must be fixed before feature is done. - **PASS WITH FOLLOW-UPS** — only MINOR/NIT findings. Feature is done; findings go to developer as fix-if-cheap, not a mandatory round trip. - **PASS** — no findings. ``` **Pipeline mode: persist the report.** Write it to `docs/specifications/<feature-slug>/REVIEW-<NN>.md` (next unused two-digit round) before dispatching any fix pass. Fix-pass prompt cites that file instead of pasting findings, the `METRICS.md` round row counts from it, and `pipeline-reviewer`'s retro reads it; report left only in conversation is gone. Style preferences never fail the gate. A `refactor-advisor` suggestion alone is **PASS WITH FOLLOW-UPS**, not BLOCKED.

{ "prompt": "Run reviewers on: **{{arguments}}**\n\n## Step 1: List target files\n\n**Paths provided** (comma-separated, e.g. `src/main, src/test`): split on commas, trim, use\n`Glob` to list all files under each path:\n\n```\nGlob(pattern=\"**/*\", path=\"<path1>\")\nGlob(pattern=\"**/*\", path=\"<path2>\")\n```\n\nRun all globs in parallel (single message).\n\n**No paths** (pipeline mode): detect changed files via git from `<base>`, then combine into a\ndeduplicated list. `<base>` is the start of the range:\n\n- **First gate** — `<base>` = `git merge-base HEAD origin/main`: every commit on the feature\n branch, not only the newest.\n- **Re-gate after a fix pass** — `<base>` = the commit the fix pass started from: the fix\n commits only, not the whole feature branch. The orchestrator records `git rev-parse HEAD`\n before dispatching each fix pass and names that commit in the invoking prompt. Missing on a\n re-gate → ask for it; never default to the merge-base.\n\n```bash\ngit diff --name-only <base> 2>/dev/null # committed since <base> + uncommitted tracked\ngit ls-files --others --exclude-standard 2>/dev/null\n```\n\nBoth empty → stop and report; never fall back to reviewing the whole tree (a missing\n`origin/main` empties the first line silently).\n\n**Pass the range on** in Step 5 as `<base>..HEAD` (plus uncommitted work). A reviewer given a\nrange reads a diff; one given nothing re-reads whole packages, which is where a re-gate's cost\nactually goes.\n\n## Step 2: Discover reviewer agents\n\nUse `Grep` to find agents with `type: reviewer` in frontmatter. Both searches in parallel:\n\n```\nGrep(pattern=\"type: reviewer\", path=\"$HOME/.claude/agents/\", glob=\"**/Agent.md\")\nGrep(pattern=\"type: reviewer\", path=\".claude/agents/\", glob=\"**/Agent.md\")\n```\n\nGlobal path missing → skip silently; project reviewers are enough.\n\n**Sanity check:** this step yields **zero** reviewers → stop, report\n`REVIEWER DISCOVERY FAILED`. Do not continue, do not report PASS. Every reviewer lives at\n`.claude/agents/<name>/Agent.md` with `type: reviewer` + a `triggers:` list in frontmatter;\nzero matches means layout drifted, not that code is clean.\n\nFor each matched file, `Read` only first 10 lines (frontmatter). Check `type: reviewer`\nappears **inside the YAML frontmatter block** (between `---` markers), not body text. Discard\nfiles where it only appears in body.\n\nFrom each valid reviewer's frontmatter extract `name` + `triggers`. Read all matched files in\nparallel.\n\n## Step 3: Apply project trigger overrides\n\n`.claude/review-triggers.json` exists in project root → read it, override triggers for matching\nreviewer names. Missing → skip this step.\n\n## Step 4: Filter by relevance\n\nFor each reviewer, check whether ANY target file matches ANY of its `triggers` globs (after\noverrides). Skip reviewers with no matching files.\n\n## Step 5: Launch relevant reviewers in parallel\n\nSpawn all matching reviewers in a **single message** via `Agent`:\n\n```\nAgent(subagent_type=\"<name>\", prompt=\"<pipeline mode only — run: review feature: <slug> unit: -\n(CLAUDE.md → *Rules* → Run tags)>\nReview <commit range, or the listed paths>. Read\n.claude/briefs/review.md first<, and .claude/briefs/proof.md — only for test-reviewer and\ncorrectness-reviewer><, and .claude/briefs/navigation.md — only for arch-reviewer and\ncorrectness-reviewer>. Scope: <the matched files, listed>. Start from the diff\nand read only what it touches; widen only when the diff cannot settle a question, and say\nwhich finding forced it. Report every finding you have in this round — a MINOR held back for\na later pass costs a whole extra gate. <On a re-gate: your prior findings were X; confirm each\nis closed or still open. Findings STATE.md already records as deferred are out of scope.>\")\n```\n\nName the files. A reviewer told only \"focus on `internal/`\" reads the package; one handed six\npaths reads six diffs.\n\n**Run the coverage gate once, before spawning, and paste its output into every prompt:**\n`.claude/scripts/uncovered-diff.py <base>`. Uncovered added lines are an untested-change\nfinding no reviewer needs to rediscover by reading; handing every reviewer the same list stops\nthree of them paying to find it separately. Its rows are grouped per run with the enclosing\nfunction, which is also the cheapest form to paste. If it exits 1 (anything outside the\n\"declared unreachable\" section, `(no coverage block)` file rows included), send it back to\nthe developer before spending a review round at all. Exit 2 is a failing test — also back to\nthe developer; never a PASS. Also paste `.claude/scripts/test-stats.py --base <base> --changed` so\nreviewers read the test-count deltas instead of recounting.\n\n**Then run the mutation sample once, and paste its output into `test-reviewer`'s prompt only:**\n`.claude/scripts/mutation-sample.py --base <base>` (no `--profile` needed: the coverage gate\nabove already bounced any uncovered line). It\nmutates up to 20 changed lines (guards first) in an isolated `$TMPDIR` copy — never the\nworktree — and lists each `SURVIVED` mutant. Costs no model tokens. Exit 2 (tool failure, or\nunmutated tests already failing) → back to the developer, like the coverage gate. Record its\nsummary line in the round's `METRICS.md` row. Worst case is `--max` × `--timeout` (20 × 120s):\nrun it with `run_in_background` and wait for it. Skip it when `test-reviewer` is not in this\nround or no production `.go` file changed; on a re-gate run it only if `test-reviewer` re-runs.\n\nDo NOT review code yourself — only orchestrate.\n\nSpawn fails with `Agent type '<name>' not found` → agent was added after session start, roster\nnot re-enumerated. Report `REVIEWER SPAWN FAILED: <name> — restart the session`. Do **not**\nreport PASS for a reviewer that never ran.\n\n## Step 6: Report\n\nReviewers return findings in shared format\n`[BLOCKER|MAJOR|MINOR|NIT] <file>:<line> — <defect>` with `Failure:` and `Fix:` lines.\nMINOR/NIT carry `Why:` instead of `Failure:` — a smell has no constructible failure by\ndefinition. Accept either; missing `Failure:` on a MINOR is not malformed.\n\nMerge into one severity-ranked list, deduplicating findings two reviewers raised on same\n`file:line`. Keep reviewer name as prefix so developer knows who to re-run.\n\n```\n## Review Report\n\n### Target\n<path or \"changed files\">\n\n### Triggered reviewers\n- <name>: triggered by <matched files>\n\n### Skipped reviewers\n- <name>: no files matched triggers\n\n### BLOCKER\n<findings, prefixed with reviewer name>\n\n### MAJOR\n<findings, prefixed with reviewer name>\n\n### MINOR\n<findings, prefixed with reviewer name>\n\n### NIT\n<findings, prefixed with reviewer name>\n\n### Strengths\n<positive notes worth another author copying — omit if none>\n\n### Verdict: PASS | PASS WITH FOLLOW-UPS | BLOCKED\n- **BLOCKED** — any BLOCKER or MAJOR exists. Must be fixed before feature is done.\n- **PASS WITH FOLLOW-UPS** — only MINOR/NIT findings. Feature is done; findings go to\n developer as fix-if-cheap, not a mandatory round trip.\n- **PASS** — no findings.\n```\n\n**Pipeline mode: persist the report.** Write it to\n`docs/specifications/<feature-slug>/REVIEW-<NN>.md` (next unused two-digit round) before\ndispatching any fix pass. Fix-pass prompt cites that file instead of pasting findings, the\n`METRICS.md` round row counts from it, and `pipeline-reviewer`'s retro reads it; report left\nonly in conversation is gone.\n\nStyle preferences never fail the gate. A `refactor-advisor` suggestion alone is\n**PASS WITH FOLLOW-UPS**, not BLOCKED.", "variables": [ "arguments" ], "notes": "Run reviewers on specific paths or on changed files. Used by the pipeline (no args = git diff) and for ad-hoc reviews (with paths). Tools: Read, Glob, Grep, Bash, Agent. Arguments: <optional paths, e.g. src/main, src/test>.", "metadata": { "source": { "provider": "github-claude-commands", "repository": "https://github.com/koblas/quarry", "path": ".claude/commands/run-reviewers.md", "ref": "6d16a5b22976d641cca9c8108fc658879bad90b5", "url": "https://github.com/koblas/quarry/blob/6d16a5b22976d641cca9c8108fc658879bad90b5/.claude/commands/run-reviewers.md", "key": "koblas/quarry/.claude/commands/run-reviewers.md" } } }

run-reviewers

About

Step 1: List target files

Step 2: Discover reviewer agents

Step 3: Apply project trigger overrides

Step 4: Filter by relevance

Step 5: Launch relevant reviewers in parallel

Step 6: Report

Install run-reviewers and it becomes one more candidate the router can pick - when it fits.

What it costs to adopt - in engineering time.

One base URL

No vendor lock-in

Failover built in

Use it

Manifest

Reviews