Project Configuration

syntaxis/config.json holds workflow policy that belongs to you, in a format the skills themselves can read. Every skill checks its relevant settings before it acts.

The line it can’t cross

Configuration is free to make review and verification stricter, and it can tune branch names and how far automation is allowed to go on its own.

It can never grant permission to commit, merge, push, deploy, publish, send anything, delete data, wave off a failing check, or accept a finding. Those approvals sit outside what configuration is allowed to touch, by design.

What ships by default

{
  "schemaVersion": 1,
  "workflow": {
    "stepReview": "feature",
    "checkpointCommits": "disabled"
  },
  "git": {
    "featureBranchPrefix": "feature/",
    "fixBranchPrefix": "fix/",
    "rollbackBranchPrefix": "rollback/"
  },
  "verification": {
    "logicTests": "when-configured",
    "uiEvidence": "when-available"
  },
  "review": {
    "independentExecution": "automatic"
  },
  "qualityGates": {
    "regular":    { "audit": "manual", "independentReview": "when-sensitive", "check": "manual", "tryGuide": "manual" },
    "continuous": { "audit": "manual", "independentReview": "when-sensitive", "check": "manual", "tryGuide": "manual" }
  },
  "continuous": {
    "maxFeatures": null,
    "maxRepairAttempts": 2,
    "finalIntegrationAudit": false
  }
}

No file at all just means the built-in defaults apply. An invalid file still lets read-only status fall back to those defaults, but anything that would actually mutate the workflow stops and points you at /doctor instead of guessing what you meant.

How often review actually happens

stepReview	What that looks like
feature (default)	One review packet, once every step is done
every	A pause for review after each individual step

Onboarding presents this choice as Efficient versus Guided, but all it’s actually storing is these two low-level settings. Pair every with checkpointCommits: enabled for the most hands-on version of the loop.

Cadence only governs when you’re asked to approve something — it has nothing to do with whether you can ask for an explanation afterward. Either style ends with an optional, read-only code walkthrough.

The quality gates

qualityGates.regular governs the normal workflow plus Autopilot. qualityGates.continuous governs Continuous Mode specifically.

Mode	What triggers it
manual	Nothing automatic — only when you ask
always	Every single work item in that workflow
when-sensitive	Sensitive or unusually broad work (audit, independent review)
when-behavioral	Work with an observable behavior change (check)
when-user-facing	Anything user-facing (guides)

Independent review ships set to when-sensitive, so it kicks in on its own for sensitive work while leaving small, ordinary features alone. Switch it to manual to turn off that automatic trigger — /audit independent current still works as an explicit request either way.

How independent review actually executes

review.independentExecution decides how a selected review runs — not whether one gets selected in the first place.

Value	What happens
automatic (default)	A fresh, isolated reviewer child, whenever the adapter can prove isolation, exact reviewer identity, model, and completion
manual	Always hand off through a fresh session instead

When automatic execution isn’t actually supported, it drops back to the manual handoff rather than letting review get skipped outright. That automatic reviewer is a generic child of the current runtime, told what to do only by the project’s own installed Audit skill and review contract — no global agent role or outside framework gets required or discovered along the way.

Verification settings

Setting	Values
verification.logicTests	when-configured, required
verification.uiEvidence	when-available, required

Limits inside Continuous Mode

Setting	What it controls
continuous.maxFeatures	Caps how many items run in one pass; null means run to the end of the plan
continuous.maxRepairAttempts	Caps how many repair attempts one feature gets
continuous.finalIntegrationAudit	Runs one integration audit at the very end

After you change something

Run /doctor. It flags schema problems and any drift between what’s configured and what the project is actually doing.