Review Policy File
Atlas Review lets you customize review behavior per project with an optional policy file. This is the fastest way to tell Atlas which paths deserve deep review, which files should be collapsed or skipped, and how strict the review should be overall.
Where To Find It
Section titled “Where To Find It”Open:
- Atlas Settings
- Review
- Policy File
From there, Atlas shows:
- a project selector
- a JSON editor
- starter examples you can insert
- buttons to reload, load the starter, save, or reset the file
If the project does not have a policy file yet, that is normal. Atlas keeps using its built-in review behavior until you save one.
What Atlas Saves
Section titled “What Atlas Saves”When you save a policy from the app, Atlas writes it to:
.atlas/review-policy.json
This file is repo-owned, so it can live with the project and be shared with the rest of the team.
What The Policy Changes
Section titled “What The Policy Changes”The policy file affects how Atlas Review shapes and prioritizes review work for that project.
defaultschanges global behavior like evidence requirements and recommendation caps.linterscontrols the deterministic linter pre-pass Atlas can run before the main review.promptAppendadds focused instructions for Atlas as a simple list of short prompts, such as asking it to treat billing or auth paths as high risk.pathsadds per-path shaping rules, for example making tests shallow, lockfiles collapsed, or generated output skipped.
Trusted Vs Local Preview
Section titled “Trusted Vs Local Preview”Atlas supports two ways to read the policy file for manual reviews:
Trusted base branch: use the policy from the trusted branch or PR base. This is the default.Local working tree preview: preview policy edits that only exist in your local checkout.
Use local preview when you want to test policy changes before committing them.
Important:
- manual reviews started from Atlas can preview the local working-tree policy
- automatic and other authoritative review runs still use the trusted base-branch policy
That means you can safely experiment locally without changing the behavior of the review runs your team depends on.
Recommended First Edits
Section titled “Recommended First Edits”Most teams only need a few changes:
- keep the linter pre-pass enabled when the repo already has supported tools configured
- make tests
shallow - mark generated output like
distorbuildasskip - collapse generated protobuf outputs like
*.pb.go - collapse lockfiles and snapshots
- push high-risk areas like
auth,billing, orpaymentstodeep - add a few short
promptAppendentries for your most sensitive flows
The built-in starter examples in the app already cover these common cases:
- Generated outputs
- Go and protobuf-generated outputs
- Tests and snapshots
- Auth and payments
- Migrations and config
- Stack-aware defaults for Go modules and protobuf or Buf schemas when those files are present
Starter Example
Section titled “Starter Example”{ "version": 1, "defaults": { "requireEvidence": true, "allowPraise": false, "maxRecommendations": 8 }, "promptAppend": [], "paths": [ { "glob": "**/*.test.*", "fileClass": "shallow", "promptAppend": "Review tests for correctness and regression coverage." }, { "glob": "**/dist/**", "fileClass": "skip" }, { "glob": "**/package-lock.json", "fileClass": "collapsed" }, { "glob": "**/auth/**", "fileClass": "deep" } ]}Field Guide
Section titled “Field Guide”defaults
Section titled “defaults”Use this section to tune review behavior across the whole project.
requireEvidence: keep this on if you want findings grounded in evidenceallowPraise: usually keep this off unless you want positive callouts in reviewsmaxRecommendations: caps how many recommendations Atlas should aim to return
promptAppend
Section titled “promptAppend”Use this for short, high-signal guidance. Good examples:
- “Treat billing and refund flows as high risk.”
- “Call out missing permission checks and rollback gaps.”
- “Focus on user-visible regressions, data loss, and broken migrations.”
Do not use this section for style preferences or team formatting rules. Prefer one short instruction per array item.
linters
Section titled “linters”This section controls the deterministic linter pre-pass that Atlas can run before the main review.
Use it to:
- enable or disable linter pre-pass for the whole project
- enable or disable specific tools
- allow Atlas-managed fallback when a supported tool is missing from the repo
The supported linter keys are:
eslinttscdotnetdotnetFormatsemgrep
For full setup and behavior, see Linter Pre-Pass.
This is the most useful part of the file for most teams.
Each path rule can include:
glob: which files it applies tofileClass: how much review effort Atlas should spend therepromptAppend: extra guidance for that path
Supported fileClass values:
deepstructuralshallowcollapsedskip
A practical mental model:
deep: high-risk source code that deserves full attentionstructural: migrations, schema, and configuration where rollout or compatibility matters Protobuf contracts and Buf config usually belong here because backward compatibility matters more than implementation style.shallow: lower-risk paths like tests where correctness matters more than stylecollapsed: files Atlas should acknowledge without spending much effort onskip: generated output and other files that should not consume review budget
Save, Reset, And Reload
Section titled “Save, Reset, And Reload”Inside the app:
Load starterloads the default starter into the editor without saving itCreate starter fileorReset to starterwrites the starter policy into the repoSave policynormalizes and saves the JSON to.atlas/review-policy.jsonReloadpulls the latest saved file back into the editor
Current Limits
Section titled “Current Limits”- JSON only in this release
- Atlas normalizes the file when you save it
- The file is optional until you decide to create it
When To Use It
Section titled “When To Use It”Use a policy file when:
- Atlas spends too much time on generated files or lockfiles
- you want protobuf or RPC schema changes reviewed for compatibility while generated outputs stay collapsed
- you want stricter review on auth, billing, payments, migrations, or infrastructure
- you want tests reviewed for correctness without expanding into style feedback
- you want a tighter cap on noisy recommendations