mra review

Context-aware code review that auto-selects strategy based on diff size.

Strategies

Strategy	When	How
Light	< 50 lines, ≤ 3 files	Single pass, 2 turns (~15s)
Standard	< 300 lines	Single pass, 3 turns (~30s)
Debate	Large diffs or API changes	2 analysts + mailbox voting (~3 min)

mra review my-api                     # Codex by default
mra review my-api --pr 123            # post inline comments
mra review my-api --provider claude --strategy debate   # force Claude debate mode
mra review my-api --base development  # compare against a specific branch

Providers

Review defaults to Codex. Admins can switch the default:

mra config review.providerMode codex
mra config review.providerMode claude
mra config review.providerMode fallback
mra config review.providerMode dual

CLI --provider overrides are blocked unless review.allowUserOverride is enabled or MRA_REVIEW_ADMIN_OVERRIDE=1 is set. fallback tries primary then secondary; dual runs both providers and merges their standard single-pass findings. In this phase Codex uses single-pass review; debate and personas remain Claude-only.

New installations use Codex plus the standard strategy. Unversioned legacy configs preserve Claude behavior until explicitly migrated. Codex runs from a trusted MRA cwd against a sanitized read-only snapshot; repository AGENTS, Claude rules, and skills are supplied only as untrusted review context.

For machine integrations, use mra integration describe|doctor|review. Protocol v1 is analysis-only, emits a SHA-bound JSON artifact, and never receives GitHub credentials or approval intent. Protocol v1 advertises only Codex because it is the provider with enforced sanitized execution. Claude, fallback, and dual remain available for ordinary reviews, but their output is not approval-eligible evidence.

--personas (opt-in)

Swap the two generic debate agents for 5 named domain experts:

mra review my-api --personas

Persona	Focus
security-auditor	Secrets, injection, auth (Troy Hunt)
api-contract-guardian	Cross-repo signature drift
performance-hawk	N+1, hot-path I/O, bundle bloat
refactoring-sage	Code smells, naming, cohesion (Fowler)
test-architect	Kent Beck 11 principles

See Personas for the full design.

Read-only guarantee

All review agents run with --disallowedTools "Write,Edit,NotebookEdit". They cannot modify files — only read and report.