Templates
PR safety gate
Decide whether a pull request is safe to merge without a human reviewer.
Decide whether a pull request is safe to merge without a human reviewer. Pattern: composite_scoring. Model: jev-1.13.0. Outage rule: review.
When to use it
- Bots or agents open many small pull requests (dependency bumps, generated fixes, copy changes) and you want the safe ones merged without waiting.
- Your team can name the risks that always need a person: sensitive code, weakened tests, destructive migrations, changes that mix several concerns.
When not to use it
- You want a code review with comments. That is text generation; keep an LLM or a person for it.
- Your branch protection already requires a human review on every pull request. Change that policy first.
- The only rule is size. Diff size is a number, so compare it in code (this template does that in a check).
Questions
| Id | Type | Asks | Answers | Gating |
|---|---|---|---|---|
touches_sensitive_code | noul | Does the change in pr.diff (files pr.files) modify authentication, authorization, billing, payments, secrets handling, or permission checks? | true or false | yes |
tests_weakened | noul | Does pr.diff delete tests, skip tests, loosen assertions, or lower a coverage or lint threshold? | true or false | yes |
risky_migration | noul | Does pr.diff include a database or schema migration that drops, renames, or rewrites existing columns, tables, or data? | true or false | yes |
mixed_concerns | noul | Judged by pr.title, pr.description and pr.diff, does this pull request mix several unrelated changes that a reviewer would want split up? | true or false | yes |
Reading the result
- Merge only when
overallActionisautoandrouteisauto_merge. Anything else waits for a person. large_diffis a code check onpr.lines_changed, not a model question. Change the 400 line limit to suit your repo.- Replace the sensitive areas in
touches_sensitive_codewith your own (auth, billing, permissions, secrets, infrastructure). - CI still has to pass. This set decides on review, not on correctness.
Spec
Save it as bandwise/sets/pr-safety-gate.json and edit it for your data.
{
"schemaVersion": 1,
"model": "jev-1.13.0",
"input": {
"schema": {
"type": "object",
"required": [
"pr"
],
"properties": {
"pr": {
"type": "object",
"required": [
"title",
"files",
"diff",
"lines_changed"
],
"properties": {
"title": {
"type": "string"
},
"description": {
"type": "string"
},
"files": {
"type": "array",
"items": {
"type": "string"
}
},
"diff": {
"type": "string"
},
"lines_changed": {
"type": "integer",
"minimum": 0
}
}
}
}
}
},
"checks": [
{
"id": "large_diff",
"when": {
"input": "pr.lines_changed",
"gte": 400
}
}
],
"stages": [
{
"id": "risks",
"questions": {
"touches_sensitive_code": {
"type": "noul",
"instructions": "Does the change in `pr.diff` (files `pr.files`) modify authentication, authorization, billing, payments, secrets handling, or permission checks?",
"criteria": {
"true": "At least one changed line sits in code that decides who can do what, handles money, or handles credentials.",
"false": "The changed lines are outside those areas, such as copy, styling, docs, or unrelated features."
},
"meta": {
"label": "Touches sensitive code"
}
},
"tests_weakened": {
"type": "noul",
"instructions": "Does `pr.diff` delete tests, skip tests, loosen assertions, or lower a coverage or lint threshold?",
"criteria": {
"true": "Existing checks were removed, skipped, or made easier to pass.",
"false": "Tests were added, kept, or made stricter, or the change touches no tests."
},
"meta": {
"label": "Tests weakened"
}
},
"risky_migration": {
"type": "noul",
"instructions": "Does `pr.diff` include a database or schema migration that drops, renames, or rewrites existing columns, tables, or data?",
"criteria": {
"true": "A migration changes or removes existing data or structure in a way that is hard to undo.",
"false": "Migrations only add new tables, columns, or indexes, or the change has no migration."
},
"meta": {
"label": "Risky migration"
}
},
"mixed_concerns": {
"type": "noul",
"instructions": "Judged by `pr.title`, `pr.description` and `pr.diff`, does this pull request mix several unrelated changes that a reviewer would want split up?",
"criteria": {
"true": "Several unrelated changes are bundled, so a reviewer could miss one.",
"false": "The change has one clear purpose."
},
"meta": {
"label": "Mixes unrelated changes"
}
}
}
}
],
"policies": {
"touches_sensitive_code": {
"type": "noul",
"gating": true,
"noul": {
"trueAt": 0.8,
"falseAt": 0.2,
"reviewMargin": 0.1
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
},
"tests_weakened": {
"type": "noul",
"gating": true,
"noul": {
"trueAt": 0.8,
"falseAt": 0.2,
"reviewMargin": 0.1
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
},
"risky_migration": {
"type": "noul",
"gating": true,
"noul": {
"trueAt": 0.8,
"falseAt": 0.2,
"reviewMargin": 0.1
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
},
"mixed_concerns": {
"type": "noul",
"gating": true,
"noul": {
"trueAt": 0.8,
"falseAt": 0.2,
"reviewMargin": 0.1
},
"actions": {
"high": {
"kind": "auto"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "review"
}
}
}
},
"composites": [
{
"id": "risk",
"kind": "weighted",
"terms": [
{
"q": "touches_sensitive_code",
"weight": 0.3
},
{
"q": "tests_weakened",
"weight": 0.25
},
{
"q": "risky_migration",
"weight": 0.25
},
{
"q": "mixed_concerns",
"weight": 0.1
},
{
"check": "large_diff",
"weight": 0.1
}
],
"policy": {
"type": "composite",
"gating": true,
"levelThresholds": {
"high": 0.5,
"medium": 0.2
},
"actions": {
"high": {
"kind": "review"
},
"medium": {
"kind": "review"
},
"low": {
"kind": "auto"
}
}
}
}
],
"routes": [
{
"when": {
"any": [
{
"q": "touches_sensitive_code",
"eq": true
},
{
"q": "tests_weakened",
"eq": true
},
{
"q": "risky_migration",
"eq": true
},
{
"check": "large_diff"
}
]
},
"output": "needs_human"
},
{
"when": {
"composite": "risk",
"gte": 0.2
},
"output": "needs_human"
}
],
"defaultRoute": "auto_merge",
"savings": {
"comparatorModel": "claude-haiku-4-5",
"estOutputTokensPerQuestion": 60,
"kind": "decision"
},
"onUnavailable": "review"
}Example states
The expected outcome is what a person would decide. It is not a recorded model answer.
Dependency patch bump
Expected: auto_merge: one purpose, no sensitive code, no tests touched.
{
"pr": {
"title": "Bump date-fns from 3.6.0 to 3.6.1",
"description": "Patch release with a timezone parsing fix.",
"files": [
"package.json",
"pnpm-lock.yaml"
],
"diff": "- \"date-fns\": \"3.6.0\"\n+ \"date-fns\": \"3.6.1\"",
"lines_changed": 14
}
}Permission check change
Expected: needs_human: it changes who can delete a project.
{
"pr": {
"title": "Let editors delete projects",
"description": "Product asked for editors to be able to clean up old projects.",
"files": [
"src/auth/permissions.ts",
"src/auth/permissions.test.ts"
],
"diff": "- delete: [\"owner\", \"admin\"],\n+ delete: [\"owner\", \"admin\", \"editor\"],",
"lines_changed": 6
}
}Skipped flaky test
Expected: needs_human: a test was skipped to get the build green.
{
"pr": {
"title": "Fix CI",
"description": "The checkout test is flaky, skipping it for now.",
"files": [
"tests/checkout.spec.ts"
],
"diff": "- it(\"charges the saved card\", async () => {\n+ it.skip(\"charges the saved card\", async () => {",
"lines_changed": 2
}
}Borderline cases
One case near the line for each question. Use them to test your wording before you trust the thresholds.
touches_sensitive_code
A logging change inside the auth module that does not change any decision.
{
"pr": {
"title": "Log failed logins at info level",
"files": [
"src/auth/login.ts"
],
"diff": "- logger.debug(\"login failed\", { userId })\n+ logger.info(\"login failed\", { userId })",
"lines_changed": 2
}
}tests_weakened
A snapshot was regenerated: the assertion still exists but now expects new output.
{
"pr": {
"title": "Update invoice snapshot",
"description": "The invoice footer copy changed.",
"files": [
"tests/__snapshots__/invoice.test.ts.snap"
],
"diff": "- \"Thanks for your business\"\n+ \"Thank you for your business\"",
"lines_changed": 2
}
}risky_migration
Adds a NOT NULL column with a default, which rewrites a large table on some databases.
{
"pr": {
"title": "Add status to orders",
"files": [
"migrations/0042_order_status.sql"
],
"diff": "+ALTER TABLE orders ADD COLUMN status text NOT NULL DEFAULT 'open';",
"lines_changed": 1
}
}mixed_concerns
A feature plus a small drive-by rename in a file the feature also touches.
{
"pr": {
"title": "Add CSV export",
"description": "Adds CSV export to reports. Also renames fmt to formatCell while I was there.",
"files": [
"src/reports/export.ts",
"src/reports/format.ts"
],
"diff": "+export function toCsv(rows) { ... }\n-function fmt(cell) {\n+function formatCell(cell) {",
"lines_changed": 58
}
}Try it
Save an example state as state.json, then run the spec locally. Local mode makes no network call and needs no key; answers are synthetic unless a recorded fixture matches.
pnpm bandwise run --local bandwise/sets/pr-safety-gate.json state.jsonBandwise is an independent product built on TypeSafe's System One models. It is not TypeSafe's documentation. For the System One models themselves, see docs.typesafe.ai.