review_code
Adversarial multi-model code review. Submit a diff, a module, or a spec+implementation and get back a structured pass/fail verdict with each issue's type, severity, location, explanation, and suggested fix.
Why call this instead of reviewing your own output: a single model shares its blind spots with itself. This routes your code through a panel of different models plus a set of deterministic detectors, catching what self-review misses — path/contract violations, module incoherence (dangling imports, broken cross-references), syntax and call-arity regressions in a diff's post-image, and 'prose instead of tool calls' (output that describes an action rather than emitting it). The panel adds semantic judgment on top and never overrides a deterministic finding.
Call it before shipping or merging, as a second opinion on a risky change, or as a gate in an autonomous build loop. Choose depth='fast' (one model, low latency) or 'deep' (full panel, higher recall). Deep review audits files of any size in milestone chunks so every panel model contributes; the price (shown in the 402) and the payment window scale with file size. Per-call limit ~1,600 lines — larger inputs return 413, so split by file/module and call once per file. Paid per call via x402 (USDC on Base); the price is announced in the 402 response before any charge.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| depth | No | "fast" = single-model, low latency; "deep" = full panel, higher recall (default "deep"). | |
| context | No | optional spec/intent, related interfaces, constraints, or what the change should do. | |
| payload | Yes | a unified diff, a complete module/file, or a spec plus its implementation (include enough context lines for diffs). | |
| language | Yes | primary language of the payload (e.g. "typescript", "python"). |