Reviewer Overview

The reviewer runs in GitHub Actions (and Azure DevOps summary-only) and posts a structured review comment on PRs. Azure DevOps summary-only uses the PR-level changes endpoint (cumulative diff). It can use:

OpenAI/ChatGPT (native transport) with a ChatGPT login bundle.
Claude (Anthropic Messages API) with an ANTHROPIC_API_KEY .
Copilot (via Copilot CLI) for teams already using GitHub Copilot.
Copilot direct HTTP transport (experimental) for custom gateways.

Recommended onboarding

CLI wizard: intelligencex setup wizard
Local web UI (preview): intelligencex setup web

Related docs:

Runtime Flow

flowchart LR
  classDef trigger fill:#BAE6FD,stroke:#0369A1,color:#082F49,stroke-width:2px;
  classDef engine fill:#DDD6FE,stroke:#5B21B6,color:#2E1065,stroke-width:2px;
  classDef provider fill:#FDE68A,stroke:#B45309,color:#451A03,stroke-width:2px;
  classDef output fill:#A7F3D0,stroke:#047857,color:#052E2B,stroke-width:2px;

  A["Pull request event"] --> B["GitHub Actions job"]
  B --> C["IntelligenceX review pipeline"]
  C --> D["Context builder<br/>diff selection, chunking, redaction"]
  D --> E["Provider call<br/>OpenAI, Claude, or Copilot"]
  E --> F["Findings parser and formatter"]
  F --> G["PR summary comment"]
  F --> H["Inline comments (if enabled)"]
  C --> I["Thread triage and optional auto-resolve"]

  class A,B trigger;
  class C,D,F,I engine;
  class E provider;
  class G,H output;

Trust model (short version)

BYO GitHub App is supported for branded bot identity.
Secrets are stored in GitHub Actions (you control access).
Web UI binds to localhost only; tokens never leave your machine.

Engine Scope

Review pipeline: resolve inputs, build context, assemble prompt, call provider, parse inline comments, post summary/inline output.
Providers and transports: OpenAI (native/appserver), Claude (Anthropic Messages API), OpenAI-compatible HTTP endpoints (Ollama/OpenRouter/etc.), and Copilot (CLI/direct).
Context builder: diff-range selection, file filtering, chunking, redaction, language hints, related PRs.
Formatter/output: summary templates, inline comment formatting, structured findings block.
Thread triage/auto-resolve: load threads, require evidence, summarize/append optional replies.

Success Metrics

Review latency (p50/p95) from job start to posted comment.
Failure rate for review runs (auth, preflight, provider errors).
Reviewer usefulness score (maintainer feedback or “kept” findings rate).
Inline quality (false-positive rate based on fixes/confirmations).

Default Mode + Model Policy

Default review mode: hybrid (summary + inline when supported; falls back to summary-only).
Default provider/model: OpenAI with gpt-5.4 unless configured otherwise; Claude and Copilot are opt-in.
Safe defaults: skip drafts; skip workflow changes unless allowed; no secrets/writes on untrusted PRs; core reviewer defaults fail-open only for transient errors, while the bundled GitHub workflow exports fail-open env defaults for provider/runtime failures so auth outages do not block CI; budget summary enabled; auto-resolve limited to bot threads with evidence; secrets audit on.

Reusable workflow (quick start)

jobs:
  review:
    uses: EvotecIT/IntelligenceX/.github/workflows/review-intelligencex-core.yml@<pinned-sha>
    with:
      reviewer_source: source
      openai_transport: native
      review_config_path: .intelligencex/reviewer.json
      mode: hybrid
      length: medium
    secrets:
      INTELLIGENCEX_AUTH_B64: ${{ secrets.INTELLIGENCEX_AUTH_B64 }}
      INTELLIGENCEX_GITHUB_APP_ID: ${{ secrets.INTELLIGENCEX_GITHUB_APP_ID }}
      INTELLIGENCEX_GITHUB_APP_PRIVATE_KEY: ${{ secrets.INTELLIGENCEX_GITHUB_APP_PRIVATE_KEY }}

The long value after @ is a pinned workflow commit SHA. Keep it pinned for security; update it intentionally when you upgrade.

reviewer_source: source is best when you want the latest workflow/reviewer behavior from source. Use reviewer_source: release when you prefer a packaged release artifact for tighter version control. Use style (review tone/style profile) and output_style (rendering preset) as optional inputs when needed.

Inputs → environment mapping (short)

The reusable workflow maps with: inputs to environment variables the reviewer reads.

Workflow input	Environment variable
`repo`	`INPUT_REPO`
`pr_number`	`INPUT_PR_NUMBER`
`reviewer_token`	`INTELLIGENCEX_GITHUB_TOKEN`
`reviewer_source`	`REVIEWER_SOURCE`
`reviewer_release_repo`	`REVIEWER_RELEASE_REPO`
`reviewer_release_tag`	`REVIEWER_RELEASE_TAG`
`reviewer_release_asset`	`REVIEWER_RELEASE_ASSET`
`reviewer_release_url`	`REVIEWER_RELEASE_URL`

Minimal config (native OpenAI)

{
  "review": {
    "provider": "openai",
    "openaiTransport": "native",
    "model": "gpt-5.4",
    "mode": "inline",
    "length": "long",
    "reviewUsageSummary": true
  }
}

Minimal config (Claude)

{
  "review": {
    "provider": "claude",
    "model": "claude-opus-4-1",
    "mode": "inline",
    "length": "long",
    "reviewUsageSummary": true
  }
}

Quick flow (end-to-end)

# 1) Auth login (stores tokens locally)
intelligencex auth login

# 2) Setup reviewer (creates PR)
intelligencex setup wizard

What to configure next

Model/provider + output style
Review length and strictness
Auto-resolve/triage behavior for bot threads
Triage-only mode (skip full review, only triage threads)
Usage summary line (optional)

How to interpret the review comment

By default:

Todo List ✅ and Critical Issues ⚠️ are treated as merge blockers.
Other Issues 🧯 are suggestions and should not block merges by default.

Usage and credits line

Enable reviewUsageSummary to append limits/credits. OpenAI uses the ChatGPT account snapshot, while Claude uses the live Anthropic provider-limit snapshot. See Configuration .