trusty-review

Fast local PR-review service — LLM-backed code review with search and analysis context.

trusty-review fetches GitHub PR diffs, retrieves code context from trusty-search, queries trusty-analyze for complexity data, then calls an LLM (AWS Bedrock by default) to produce a structured review verdict with actionable findings.

It ships as:

• a one-shot CLI (run / compare subcommands)
• a long-lived HTTP webhook server (serve subcommand, port 7880)
• a JSON-RPC 2.0 / MCP stdio service (serve --stdio) for Claude Code integration

Installation

Install from prebuilt binary

Download the latest prebuilt binary for your platform from the GitHub Releases page. Binaries follow the tag convention trusty-review-v<version>:

Extract and place the trusty-review binary on your PATH.

Install with cargo

cargo install --git https://github.com/bobmatnyc/trusty-tools trusty-review --locked

This builds and installs the trusty-review binary from the latest main branch.

With Homebrew (recommended)

brew tap bobmatnyc/trusty
brew install trusty-review

Or install directly without tapping:

brew install bobmatnyc/trusty/trusty-review

Homebrew provides:

• Automatic updates via brew upgrade trusty-review
• Standard macOS / Linux PATH integration
• Easy dependency management

Prerequisites

Required: A GitHub token (GITHUB_TOKEN) or GitHub App credentials for PR fetching and (optionally) posting review comments. Set PR_INTELLIGENCE_DRY_RUN=false to enable comment posting (default: dry-run).

LLM credentials: AWS Bedrock credentials (env vars, ~/.aws/credentials, IAM role, or SSO) for the default bedrock/ provider, or OPENROUTER_API_KEY for OpenRouter models.

Contributor profiling (trusty-review profile): requires a pre-populated tga SQLite database. Set TRUSTY_TGA_DB or pass --db <path>. Compiled in by default; omit with --no-default-features --features http-server,mcp for a slimmer build without tga/rusqlite compilation.

Sidecar services (optional, degrade gracefully when absent):

• trusty-search on :7878 — code-context hybrid search for richer reviews
• trusty-analyze on :7879 — complexity and quality metrics

cargo install --git https://github.com/bobmatnyc/trusty-tools trusty-search --locked
cargo install --git https://github.com/bobmatnyc/trusty-tools trusty-analyze --locked
trusty-search start
trusty-analyze serve

Quick start — one-shot review

# Review a GitHub PR (Bedrock credentials required)
trusty-review run owner repo 123

# Review a local unified diff
trusty-review run --local-diff /path/to/patch.diff

# Override the reviewer model
trusty-review run owner repo 123 --reviewer-model bedrock/us.anthropic.claude-haiku-4-5

# Compare models
trusty-review compare owner repo 123

HTTP server

# Start the HTTP daemon on port 7880
trusty-review serve

# Custom port / bind address
trusty-review serve --port 8080 --bind 0.0.0.0

Endpoints:

MCP stdio service (Claude Code integration)

# Start the MCP stdio server
trusty-review serve --stdio

Wire into Claude Code via .mcp.json:

{
  "mcpServers": {
    "trusty-review": {
      "command": "trusty-review",
      "args": ["serve", "--stdio"]
    }
  }
}

MCP tools

review_pr

{
  "name": "review_pr",
  "arguments": {
    "owner": "bobmatnyc",
    "repo":  "trusty-tools",
    "pr":    625,
    "reviewer_model": "bedrock/us.anthropic.claude-haiku-4-5"
  }
}

Returns a ReviewResult JSON object with:

• grade (A+ | A | A- | B+ | B | B- | C+ | C | C- | D+ | D | D- | F) — letter grade
• verdict (APPROVE | APPROVE* | REQUEST_CHANGES | BLOCK | UNKNOWN)
• findings (array of findings with severity + confidence)
• input_tokens / output_tokens — LLM token usage
• cost_estimate_usd — estimated API cost

Grade → Verdict mapping

The verdict is derived from the grade per a fixed product decision (APPROVE floor = B-):

The final verdict is max(grade_verdict, severity_floor(findings)) — the grade never produces a verdict weaker than what the severity floor already requires. After verification (Phase 2), the grade is re-clamped to stay consistent with the post-verification verdict.

When posted to GitHub, the review comment includes a footer:

Grade: B+ · 🤖 Reviewed by Trusty-Review (`us.anthropic.claude-sonnet-4-6`) · tokens ↑1234 ↓567 · est. $0.01

(↑ = input tokens, ↓ = output tokens). The footer appears identically in dry-run output.

review_diff

{
  "name": "review_diff",
  "arguments": {
    "diff": "diff --git a/src/lib.rs ...",
    "context": "Refactoring the auth module",
    "reviewer_model": "bedrock/us.anthropic.claude-sonnet-4-6"
  }
}

review_health

{ "name": "review_health", "arguments": {} }

Returns a health status object:

{
  "status": "ok",
  "version": "0.3.2",
  "dry_run": true,
  "reviewer_model": "us.anthropic.claude-sonnet-4-6",
  "inference": "ok",
  "deps": {
    "trusty_search": {
      "required": true,
      "reachable": true
    },
    "trusty_analyze": {
      "required": false,
      "reachable": true
    }
  }
}

Status values:

• ok — all dependencies healthy and inference reachable.
• degraded — a required dependency (trusty-search) or inference is unreachable.
• unknown — cannot determine health state.

Inference field values:

• ok — AWS Bedrock and/or OpenRouter accessible.
• unreachable — both inference providers unreachable (network/DNS error).
• auth_error — inference provider reachable but auth failed (bad API key).
• unknown — inference probe could not determine status.

Environment variables

AWS credentials can also be supplied via ~/.aws/credentials, IAM roles, or SSO. The full AWS credential chain is supported.

Context sources & the conformance gate

trusty-review enriches a review with external context sources (JIRA, Confluence, GitHub Issues, and the intent/method-conformance back gate). Each source resolves independently from layered config — an env var beats a TOML key beats the built-in default (the ContextSourcesConfig struct):

Enable values are lenient (true/1/yes/on); mode is live (only mode supported today) or semantic (not yet implemented for these sources).

The conformance back gate (DOC-15)

The conformance source is the BACK gate of the intent/method-conformance capability (docs/specs/intent-conformance.md). During review it resolves "what method did the ticket/spec prescribe?" via the shared intent-source resolver (ISR) and surfaces it so the reviewer LLM can flag a diff that explicitly contradicts that method (matrix M5). A gap or an unresolved intent surfaces nothing — the gate is conservative and fail-open, so it never manufactures a false-positive finding.

It is default-DISABLED (unlike the other sources, it does not auto-enable on mere credential presence) because it issues a GitHub ticket fetch and is opt-in. Turn it on explicitly:

# Env (one-shot)
TRUSTY_REVIEW_CONTEXT_CONFORMANCE_ENABLED=true trusty-review serve

# …or in $XDG_CONFIG_HOME/trusty-review/config.toml
[context.sources.conformance]
enabled = true
mode = "live"

The gate is backed by the intent_source module in trusty-common, gated behind that crate's intent-source Cargo feature (which trusty-review already enables). See the Cargo features note below.

Reviewer model

The default reviewer model is us.anthropic.claude-sonnet-4-6 on AWS Bedrock.

Override via:

• CLI flag: --reviewer-model bedrock/us.anthropic.claude-haiku-4-5
• Env var: PR_INTELLIGENCE_REVIEWER_MODEL=bedrock/us.anthropic.claude-haiku-4-5
• Config file: $XDG_CONFIG_HOME/trusty-review/config.toml

Provider prefix convention:

• bedrock/<id> — AWS Bedrock Converse API (no API key needed, uses AWS credential chain)
• openrouter/<id> — OpenRouter (requires OPENROUTER_API_KEY)
• Bare id — uses the configured default provider

Cargo features

The conformance back gate (above) is backed by trusty-common's intent-source feature, which trusty-review enables unconditionally (it is part of the dependency declaration, not a trusty-review feature). That feature gates the shared intent-source resolver (ISR) so the other trusty-common consumers that do not need it pay nothing.

Slim build (no contributor profiling, no tga/rusqlite compilation):

cargo install --git https://github.com/bobmatnyc/trusty-tools trusty-review \
  --locked --no-default-features --features http-server,mcp

License

MIT — see LICENSE.