\n\nMultiple LLMs review your code in parallel, debate conflicting opinions, then a head agent delivers the final verdict. Different models catch different bugs — consensus filters the noise.\n\n\n\n\n---\n\n## Quick Start\n\n```bash\nnpm i -g codeagora\nagora init\ngit diff | agora review\n```\n\n`agora init` auto-detects your API keys and CLI tools, then generates a config.\n\n---\n\n## Supported Providers (Tier 1)\n\n| Provider | Type | Cost |\n|----------|------|------|\n| Groq | API | Free |\n| Anthropic | API | Paid |\n| Claude Code | CLI | Subscription |\n| Gemini CLI | CLI | Free |\n| Codex CLI | CLI | Subscription |\n\n[Full provider list (24+ API, 12 CLI) ->](docs/PROVIDERS.md)\n\n---\n\n## How It Works\n\n```\ngit diff | agora review\n\n Pre --- Semantic Diff Classification\n --- TypeScript Diagnostics\n --- Change Impact Analysis\n |\n L1 --- Reviewer A (security) --+\n --- Reviewer B (logic) --+-- parallel specialist reviews\n --- Reviewer C (general) --+\n |\n Filter -- Hallucination Check (file/line validation)\n --- Self-contradiction Filter\n --- Evidence Dedup\n |\n L2 --- Adversarial Discussion (supporters must disprove)\n --- Static analysis evidence in debate\n |\n L3 --- Head Agent --> ACCEPT / REJECT / NEEDS_HUMAN\n |\n Output -- Triage: N must-fix / N verify / N ignore\n```\n\n---\n\n## Web Dashboard\n\nReal-time web UI for monitoring reviews, browsing sessions, and managing configuration.\n\n```bash\nagora dashboard # Start on http://localhost:6274\nagora dashboard -p 8080 # Custom port\n```\n\nFeatures:\n- **9 pages** — Dashboard, Sessions, Models, Costs, Discussions, Config, Pipeline, Compare, Review Detail\n- **Live pipeline** — WebSocket-powered real-time stage progression and discussion updates\n- **Model intelligence** — Leaderboard, quality trends, selection frequency charts\n- **httpOnly cookie auth** — Secure token exchange via `POST /api/auth`\n- **Server-side pagination** — Filterable by status, search, date range\n\nThe dashboard token is printed on startup and persisted to `.ca/dashboard-token`.\n\n---\n\n## Interactive TUI\n\nTerminal UI for running reviews without leaving the terminal.\n\n```bash\nagora tui\n```\n\n8 screens: Review Setup, Pipeline Progress, Results, Diff Viewer, Debate, Config, Model Selector, Provider Status. Navigate with arrow keys, `Enter` to select, `q` to quit.\n\n---\n\n## MCP Server (Claude Code / Cursor)\n\n9-tool MCP server for AI IDE integration.\n\n```json\n// claude_desktop_config.json or .cursor/mcp.json\n{\n \"mcpServers\": {\n \"codeagora\": {\n \"command\": \"npx\",\n \"args\": [\"-y\", \"@codeagora/mcp\"]\n }\n }\n}\n```\n\nTools: `review_diff`, `review_pr`, `review_staged`, `session_list`, `session_detail`, `explain_session`, `config_get`, `config_set`, `health_check`.\n\n---\n\n## Notifications\n\n```bash\nagora notify 2026-03-27/001 # Send notification for a past session\n```\n\nSupported channels:\n- **Discord** — Real-time thread updates + summary (webhook URL in config)\n- **Slack** — Summary notification (webhook URL in config)\n- **Generic webhook** — HMAC-SHA256 signed payloads over HTTPS\n\nConfigure in `.ca/config.json` under `notifications`.\n\n---\n\n## Extensions\n\nAll extensions are optional — install only what you need.\n\n| Package | Install | What it does |\n|---------|---------|-------------|\n| [@codeagora/web](https://www.npmjs.com/package/@codeagora/web) | `npm i -g @codeagora/web` | Web dashboard — 9-page SPA with real-time pipeline monitoring, session history, model leaderboard, cost tracking |\n| [@codeagora/tui](https://www.npmjs.com/package/@codeagora/tui) | `npm i -g @codeagora/tui` | Interactive terminal UI — run reviews, browse sessions, edit config, watch debates in real-time |\n| [@codeagora/mcp](https://www.npmjs.com/package/@codeagora/mcp) | `npm i -g @codeagora/mcp` | MCP server (9 tools) — integrates with Claude Code, Cursor, and any MCP-compatible IDE |\n| [@codeagora/notifications](https://www.npmjs.com/package/@codeagora/notifications) | `npm i -g @codeagora/notifications` | Webhooks — Discord (real-time threads + summary), Slack (summary), generic (HMAC-SHA256 signed) |\n\nEach extension works standalone or together. The core `codeagora` CLI includes everything needed for command-line reviews and GitHub Actions.\n\n[Extension guide ->](docs/EXTENSIONS.md)\n\n---\n\n## GitHub Actions\n\nAdd CodeAgora to any repo in 2 steps:\n\n**1. Create `.ca/config.json`** (or run `agora init`):\n\n```json\n{\n \"mode\": \"pragmatic\",\n \"reviewers\": [\n { \"id\": \"r1\", \"model\": \"llama-3.3-70b-versatile\", \"backend\": \"api\", \"provider\": \"groq\", \"enabled\": true, \"timeout\": 120 },\n { \"id\": \"r2\", \"model\": \"qwen/qwen3-32b\", \"backend\": \"api\", \"provider\": \"groq\", \"enabled\": true, \"timeout\": 120 },\n { \"id\": \"r3\", \"model\": \"meta-llama/llama-4-scout-17b-16e-instruct\", \"backend\": \"api\", \"provider\": \"groq\", \"enabled\": true, \"timeout\": 120 }\n ]\n}\n```\n\n**2. Add the workflow** (`.github/workflows/codeagora-review.yml`):\n\n```yaml\nname: CodeAgora Review\non:\n pull_request:\n types: [opened, synchronize, reopened]\n\npermissions:\n contents: read\n pull-requests: write\n statuses: write\n\njobs:\n review:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v4\n - uses: bssm-oss/CodeAgora@v2\n with:\n github-token: ${{ secrets.GITHUB_TOKEN }}\n env:\n GROQ_API_KEY: ${{ secrets.GROQ_API_KEY }}\n```\n\n**3. Add `GROQ_API_KEY`** to your repo's Settings > Secrets > Actions.\n\nEvery PR gets inline review comments, a summary verdict, and a commit status check. Add `review:skip` label to any PR to bypass.\n\n---\n\n## Documentation\n\n| Doc | Content |\n|-----|---------|\n| [CLI Reference](docs/CLI_REFERENCE.md) | All commands and options |\n| [Configuration](docs/CONFIGURATION.md) | Config file guide |\n| [Providers](docs/PROVIDERS.md) | Full provider list with tiers |\n| [Architecture](docs/ARCHITECTURE.md) | Pipeline design and project structure |\n| [Extensions](docs/EXTENSIONS.md) | Web, TUI, MCP, Notifications |\n| [Troubleshooting](docs/TROUBLESHOOTING.md) | Common errors and fixes, exit codes |\n| [FAQ](docs/FAQ.md) | Frequently asked questions |\n\n---\n\n## Development\n\n```bash\npnpm install && pnpm build\npnpm test # 3386 tests\npnpm test:coverage # with coverage report\npnpm typecheck\npnpm cli review path/to/diff.patch\n```\n\n---\n\n## Benchmarks\n\nGolden-bug fixtures under `benchmarks/golden-bugs/` drive the false-negative measurement framework (see #472).\n\n**Score pre-computed results** (fast, no API calls):\n\n```bash\npnpm bench:fn -- --validate-only # schema-check fixtures\npnpm bench:fn -- --results path/to/results-dir # score against pre-computed review output\npnpm bench:fn -- --results path/to/results-dir --json # CI-friendly JSON report\n```\n\n**Run the live pipeline against every fixture** (produces the results dir above):\n\n```bash\nexport OPENROUTER_API_KEY=...\npnpm bench:fn:run -- --results ./bench-out\npnpm bench:fn -- --results ./bench-out\n```\n\nThe driver uses `benchmarks/.ca/config.json` — a lean 3-reviewer OpenRouter setup. A full run over the 4 seed fixtures costs roughly $0.04–$0.10 depending on discussion rounds. Add `--fixtures id1,id2` to restrict, `--skip-head` to skip the L3 verdict stage.\n\nTwo fixture kinds live side by side:\n\n- **Recall cases** (`expectedFindings` non-empty) — review must surface each listed bug. Misses count as FN.\n- **FP regression cases** (`expectedFindings` is `[]`) — review must report nothing. Any finding is a regression.\n\nCurrent seed fixtures: 3 recall cases (off-by-one, null-deref, SQL injection) + 1 FP regression (PR #490 moderator regex). See `benchmarks/golden-bugs/README.md` for fixture format.\n\n### Baseline (n=3, 2026-04-20)\n\nThree live runs with the default 3-reviewer OpenRouter config ([#24666562754](https://github.com/bssm-oss/CodeAgora/actions/runs/24666562754), [#24667305646](https://github.com/bssm-oss/CodeAgora/actions/runs/24667305646), [#24667897271](https://github.com/bssm-oss/CodeAgora/actions/runs/24667897271)):\n\n| Metric | Mean | Min | Max |\n|---|---|---|---|\n| recall@3 | 100.0% | 100.0% | 100.0% |\n| recall@5 | 100.0% | 100.0% | 100.0% |\n| recall@10 | 100.0% | 100.0% | 100.0% |\n| FPs per fp-regression fixture | 2.3 | 2 | 3 |\n| fp-regression triggered | 3/3 runs |\n\n**Recall stable** — all three recall cases (off-by-one, null-deref, SQL injection) caught in top-3 on every run.\n\n**FP regression triggered on every run** — but the *content* of the phantom findings shifts between runs: CRITICAL×3 about unhandled `JSON.parse` on run 1, WARNING×2 about regex DoS + input size on run 2, WARNING + CRITICAL about unbounded string + missing type import on run 3. Each individual claim is a plausible-sounding, code-level assertion that the review would make against a real diff, which is exactly why the current calibration stack does not filter them. This confirms the \"high-confidence corroborated FP\" blind spot documented in `project_calibration_stack.md`. This fixture is the regression gate for future calibration work (see #468).\n\n---\n\n## License\n\nMIT\n","readmeFilename":"README.md"}
![\"CodeAgora](\"assets/logo.svg\"){kind=logo}
\n![\"Version\"](\"https://img.shields.io/npm/v/codeagora?color=%2305A6B9\")
![\"Tests\"](\"https://img.shields.io/badge/tests-3168%20passing-%23191A51\"){kind=icon}
![\"Node\"](\"https://img.shields.io/badge/node-%3E%3D20-%2305A6B9\"){kind=icon}
![\"License\"](\"https://img.shields.io/badge/license-MIT-%23191A51\"){kind=icon}