# Project export: Triage This document was generated by HackStack to give an AI agent context about a hackathon project. Sections are labeled with their provenance; content marked as truncated was cut to keep this document small. ## Project metadata - Hackathon: UC Berkeley AI Hackathon 2026 - Tagline: An agent that reproduces your bugs by actually using your app. - Devpost: https://devpost.com/software/triage-y10mpq - GitHub: https://github.com/HansChundekad/Triage - Video: https://www.youtube.com/embed/Q-ArgjSnVzw?enablejsapi=1&hl=en_US&rel=0&start=&version=3&wmode=transparent - Result: winner (Best use of Band for agent to agent communication) - Team: 2 GitHub contributor(s) — Hans (123 commits), Claude Opus 4.8 (93 commits) ## Devpost submission (written by the team) No Devpost description available. ## README (from the GitHub repository) # TRIAGE Autonomous bug-reproduction agent. A developer pastes a GitHub issue; TRIAGE opens a real cloud browser, clicks through a live app to trigger the bug, captures the crash, diagnoses the root cause, and writes a structured report. Built for the UC Berkeley AI Hackathon 2026. ## Architecture (three agents, one Band room) - **ParserAgent** (`triage/parser_agent/`) — GitHub issue → structured repro steps. - **ReproAgent** (`triage/repro_agent/`) — drives a real Browserbase browser. **All** browser work lives here. - **HypothesisAgent** (`triage/hypothesis_agent/`) — diagnoses root cause, can redirect ReproAgent to retry. Integrations: Browserbase (cloud browser), Band (multi-agent coordination), Arize Phoenix (retry-loop tracing), Claude (reasoning). See `docs/TRIAGE_OVERVIEW.md` (source of truth) and `docs/TRIAGE_INTEGRATIONS.md`. ## Status Phase 1 — repo scaffold. No agent logic or integration code yet. ## Setup ```bash pip install -e ".[dev]" # installs python-dotenv + pytest only in Phase 1 cp .env.example .env # then fill in your keys pytest # runs the config-loader tests ``` ## Detected evidence (automated analysis) Indexed codebase: 132 recognized source files, 840 KB. - Anthropic (technology) — detected in the code - CSS (language) — detected in the code - FastAPI (technology) — detected in the code - HTML (language) — detected in the code - Python (language) — detected in the code - React (technology) — detected in the code - TypeScript (language) — detected in the code - AI coding agent: Claude Code — evidence: config files committed to the repository; commit authorship or trailers ## Codebase structure (from repository index) ### Files (120 of 143) ``` .env.example .gitignore backend/__init__.py backend/README.md backend/requirements.txt backend/run_manager.py backend/server.py CLAUDE.md docs/DEMO_RUNBOOK.md docs/STATUS.md docs/superpowers/plans/2026-06-20-phase-1-repo-scaffold.md docs/superpowers/plans/2026-06-20-phase-2-band-module.md docs/superpowers/plans/2026-06-20-phase-3-hypothesis-agent.md docs/superpowers/plans/2026-06-20-phase-3-repro-echo.md docs/superpowers/plans/2026-06-20-phase4-repro-agent-browserbase.md docs/superpowers/plans/2026-06-20-phase5-hypothesis-agent-real.md docs/superpowers/plans/2026-06-20-phase5-parser-agent-real.md docs/superpowers/plans/2026-06-20-phase6-retry-loop.md docs/superpowers/plans/2026-06-20-phase7-frontend.md docs/superpowers/plans/2026-06-20-phase7a-tracing-substrate.md docs/superpowers/plans/2026-06-20-phase7b-phoenix-evaluator.md docs/superpowers/plans/2026-06-20-phase7c-claude-synthesis.md docs/superpowers/plans/2026-06-21-phase7.5-outer-loop.md docs/superpowers/plans/2026-06-21-phoenix-to-ax-migration.md docs/superpowers/specs/2026-06-20-phase7-arize-design.md docs/superpowers/specs/2026-06-20-phase7-frontend-design.md docs/superpowers/specs/2026-06-21-phoenix-to-ax-migration.md docs/superpowers/specs/phase7-report.schema.json docs/TRIAGE_INTEGRATIONS.md docs/TRIAGE_OVERVIEW.md frontend/.env.example frontend/.gitignore frontend/index.html frontend/package.json frontend/README.md frontend/src/api.test.ts frontend/src/api.ts frontend/src/App.test.tsx frontend/src/App.tsx frontend/src/components/BrowserView.test.tsx frontend/src/components/BrowserView.tsx frontend/src/components/LiveLog.test.tsx frontend/src/components/LiveLog.tsx frontend/src/components/ReportCard.test.tsx frontend/src/components/ReportCard.tsx frontend/src/components/UrlInput.test.tsx frontend/src/components/UrlInput.tsx frontend/src/fixtures/recorded-run.json frontend/src/main.tsx frontend/src/replay.test.ts frontend/src/replay.ts frontend/src/styles/app.css frontend/src/styles/tokens.css frontend/src/test-setup.ts frontend/src/types.test.ts frontend/src/types.ts frontend/tsconfig.json frontend/vercel.json frontend/vite.config.ts pyproject.toml README.md scripts/demo.sh scripts/handshake.py scripts/hypothesis_echo_demo.py scripts/phase6_live_run.py scripts/phase7_eval.py scripts/phase7_traced_run.py scripts/test_echo_chain.py scripts/three_way_smoke.py tests/__init__.py tests/_tracing_helpers.py tests/test_backend_server.py tests/test_band_module.py tests/test_claude_call_spans.py tests/test_config.py tests/test_eval_ax_logging.py tests/test_eval_code_checks.py tests/test_eval_ground_truth.py tests/test_eval_judges.py tests/test_eval_run_eval.py tests/test_hypothesis_agent.py tests/test_hypothesis_reasoning.py tests/test_memory_backend_ax.py tests/test_memory_backend_phoenix.py tests/test_memory_distill.py tests/test_memory_facade.py tests/test_memory_history.py tests/test_parser_agent.py tests/test_parser_claude.py tests/test_parser_github.py tests/test_parser_prior_context.py tests/test_phase7_capture.py tests/test_phase75_booth_injection.py tests/test_repro_browser_tracing.py tests/test_repro_browser.py tests/test_repro_echo_tracing.py tests/test_repro_echo.py tests/test_repro_loop.py tests/test_run_artifacts.py tests/test_run_trace.py tests/test_synthesis_schema.py tests/test_synthesize.py tests/test_tracing_setup.py triage/__init__.py triage/config.py triage/eval/__init__.py triage/eval/code_checks.py triage/eval/ground_truth.py triage/eval/judges.py triage/eval/run_eval.py triage/hypothesis_agent/__init__.py triage/hypothesis_agent/__main__.py triage/hypothesis_agent/agent.py triage/hypothesis_agent/README.md triage/hypothesis_agent/reasoning.py triage/memory/__init__.py triage/memory/backends/__init__.py triage/memory/backends/ax.py triage/memory/backends/phoenix.py triage/memory/distill.py [23 more files omitted for size] ``` ### Dependencies - backend/requirements.txt: fastapi@>=0.115, sse-starlette@>=2.1, uvicorn[standard]@>=0.30 - frontend/package.json: @testing-library/jest-dom@^6.4.8, @testing-library/react@^16.0.1, @types/react@^18.3.5, @types/react-dom@^18.3.0, @vitejs/plugin-react@^4.3.1, jsdom@^25.0.0, react@^18.3.1, react-dom@^18.3.1, typescript@^5.5.4, vite@^5.4.3, vitest@^2.0.5 - pyproject.toml: anthropic, arize@>=8.35, arize-ax-cli@>=0.25, arize-otel@>=0.13, arize-phoenix@>=17.9, arize-phoenix-client@>=2.9, arize-phoenix-evals@>=3.1, band-sdk, httpx, openinference-instrumentation-anthropic@>=1.0.6, opentelemetry-sdk@>=1.42, pandas@>=2.0, playwright, pytest@>=8.0, python-dotenv@>=1.0, stagehand ### Recent commits (newest first) - fix(frontend): point demo fixture replay links at real Browserbase sessions - fix(demo): RUN SUMMARY trace hint reflects actual attempt count - docs: add demo-day runbook (commands, surfaces, AX deep link, outer-loop story) - feat(demo): opt-in auto-open of Browserbase live view per attempt - fix(demo): RUN SUMMARY reads persisted attempts (true fail→succeed count) - feat(demo): one clean entrypoint + live-surface visibility (Phase 8 demo-prep) - fix(trace): key span ids on run-unique attempt id (redirect_parser hero trace) - docs(ax): correct read-latency claim — by-trace-id reads ~2s, not hours - chore(ax): gitignore the ax CLI export temp dir - docs(ax): integrations + status + deps for Phoenix→AX migration - feat(ax): log evaluators to AX via spans.update_evaluations (eval.* columns) - feat(ax): implement AX trace-query read backend + flip seam default to ax - feat(ax): setup_tracing registers arize.otel tracer (batch=False); phoenix fallback - feat(ax): add ARIZE_* config + TRIAGE_TRACE_BACKEND selector (default ax) - docs(ax): migration spec + plan (Phoenix→Arize AX), verified APIs - docs(phase7.5): precise harness cut-path (elif→if restore) per final review - docs(phase7.5): record outer-loop completion, cut-path, and Phoenix→AX seam - refactor(phase7.5): isolate trace read-back behind a swappable backend seam - feat(phase7.5): harness injects learned context (memory ON supersedes force-retry) - feat(phase7.5): booth path posts learned context into room + shapes attempt 1 ## Key source files (fetched from GitHub, selected and truncated for size) ### CLAUDE.md ```markdown # TRIAGE Autonomous bug-reproduction agent (UC Berkeley AI Hackathon 2026). Reproduces a reported bug by driving a real cloud browser through a live app, then diagnoses and reports root cause. **Read first:** `docs/TRIAGE_OVERVIEW.md` (architecture, source of truth) and `docs/TRIAGE_INTEGRATIONS.md` (how to connect Browserbase/Band/Arize). **Current state / what's next:** `docs/STATUS.md`. ## Architecture Three agents coordinate in one Band room: - **ParserAgent** (`triage/parser_agent/`) — GitHub issue → structured repro steps. - **ReproAgent** (`triage/repro_agent/`) — drives the Browserbase browser. - **HypothesisAgent** (`triage/hypothesis_agent/`) — diagnoses, can redirect for retry. - Shared Band layer: `triage/shared/band.py`. Config loader: `triage/config.py`. ## Non-negotiable rules 1. All browser/Stagehand work lives in **ReproAgent only**. 2. Every cross-agent Band message must **@mention** a recipient (no @mention = no one sees it). 3. Agent names are exact: `ParserAgent` / `ReproAgent` / `HypothesisAgent` — never generic. 4. Three **distinct Band identities** (`BAND_PARSER_*`, `BAND_REPRO_*`, `BAND_HYPOTHESIS_*`). 5. **New Browserbase session per retry** (no sessionId). 6. Band **messages** = directed @mention talk; **events** = logs. Don't mix. 7. **Verify SDK details against live docs before writing integration code**; flag drift. 8. Arize `bug.detected` must be honest — the fail→succeed flip must be real. ## Commands Always use the repo venv (`.venv/`): - Install: `.venv/bin/pip install -e ".[dev]"` - Test: `.venv/bin/pytest` - Config is loaded via `from triage.config import load_config` (fails loud on missing env vars). ## Conventions - All-Python; target Python 3.11+ (dev on 3.14). - TDD for real logic (write the failing test first). - Per-task commits, scoped messages. Never commit secrets — `.env` is gitignored; keep `.env.example` in sync when adding a config var. - When adding a config var: update `.env.example`, `triage/config.py`, and its test together. ``` ### docs/DEMO_RUNBOOK.md ```markdown # TRIAGE — Demo Runbook Everything for the booth. Test-script demo, native surfaces, no in-app UI. --- ## 0. Pre-demo checklist (do once, before judges arrive) - [ ] Terminal open at `/Users/hanschundekad/Triage`. - [ ] Browser logged into **browserbase.com**. - [ ] **Band web app** open at app.band.ai → room **`741d69f0-006d-436a-9d45-cc43b88cbc75`** (pinned; all 3 agents are participants). All demo runs stream into this one room. - [ ] **Arize AX** open (project `triage-bug-repro`). Have the deep link (§4) ready. - [ ] (Optional) `cd frontend && npm run dev` running, for the report-card surface. - [ ] **Pre-seed the outer loop:** do one `./scripts/demo.sh` run ~10+ min before showing the "Arize makes it smarter" story, so AX has the history indexed. (Memory already returns a real hint from 4 prior runs — verified — so this is belt-and-suspenders.) Sanity check the memory feed is live: ```bash TRIAGE_OUTER_LOOP=1 .venv/bin/python -c "from dotenv import load_dotenv; load_dotenv(); from triage.config import load_config; from triage.memory import load_learned_context; print(load_learned_context(load_config()))" ``` Expect a `Prior-run memory: across N past run(s)…` line. --- ## 1. Commands | # | Purpose | Command | |---|---|---| | **A** | **Main demo** — deterministic fail→succeed (hero) | `./scripts/demo.sh` | | **B** | **"Arize makes it smarter"** — reads memory, first-try success | `TRIAGE_OUTER_LOOP=1 .venv/bin/python scripts/phase7_traced_run.py` | | **C** | **Report card** surface | `cd frontend && npm run dev` → open URL → click **Demo** | Both A and B use the hardcoded GitHub issue (`…/StrideAI/issues/1`). No UI input. --- ## 2. Run of show (Demo A — the hero) 1. **Run `./scripts/demo.sh`.** Say: *"A real GitHub bug report is handed to three coordinating agents."* 2. **💬 Band room** (or terminal `💬` stream): ParserAgent → ReproAgent → "attempt 1 failed" → HypothesisAgent diagnoses → `redirect_parser` → ParserAgent re-parses → retry. *"Three distinct agents coordinating over @mentions in a real room."* 3. **🌐 Browserbase**: two tabs auto-open — attempt 1 (browser blanks/fails), attempt 2 (browser succeeds). *"It drives a real cloud browser, not a mock."* 4. **Terminal RUN SUMMARY**: `verdict: REPRODUCED ✅`, honest `bug.detected flip: False → True`, both session URLs. 5. **📋 Report card** (frontend Demo button): synthesized root cause — verdict, per-step ok/fail/crash, mechanism, eval scores. 6. **📊 Arize** (§3–4): walk the trace. Run takes ~1–3 min. --- ## 3. What's in the Arize trace (the walkthrough) - **One `triage_run` → two `repro_attempt` spans** (the inner-loop retry, captured). - Expand each → `browser_execution` → `stagehand_action` per step (4 on attempt 1, 8 on attempt 2). - **Evaluations: `repro_fidelity` = not_reproduced (0)** on attempt 1 vs **reproduced (1)** on attempt 2 — LLM-judge scores, honestly different. Plus `root_cause_correctness`, `honesty`. --- ## 4. Arize deep link (use this — the trace LIST [truncated — 4696 more characters] ``` ### pyproject.toml ``` [build-system] requires = ["hatchling"] build-backend = "hatchling.build" [project] name = "triage" version = "0.1.0" description = "Autonomous bug-reproduction agent — UC Berkeley AI Hackathon 2026" requires-python = ">=3.11" dependencies = [ # --- Installed in Phase 1 (config loader + tests need these) --- "python-dotenv>=1.0", # --- Listed for later phases; NOT imported yet. Versions verified # against live docs in the phase that first uses each. --- "anthropic", # Claude reasoning (all agents) "httpx", # GitHub issue fetch (ParserAgent) "band-sdk", # Band coordination (Phase: Band) "stagehand", # Browserbase driver (ReproAgent) "playwright", # CDP screenshots + console errors (ReproAgent) # --- Arize AX (PRIMARY trace backend) — verified 2026-06-21 --- "arize-otel>=0.13", # arize.otel.register → otlp.arize.com (WRITE path) "arize>=8.35", # ArizeClient: spans.update_evaluations (eval logging) "arize-ax-cli>=0.25", # `ax` CLI: spans export (READ-back) + verification # --- Arize Phoenix (FALLBACK backend + LOCAL judge engine) — verified 2026-06-20 --- "arize-phoenix>=17.9", # tracing fallback (TRIAGE_TRACE_BACKEND=phoenix) "openinference-instrumentation-anthropic>=1.0.6", # auto-instrument Claude calls (both backends) "opentelemetry-sdk>=1.42", # manual retry spans "arize-phoenix-evals>=3.1", # LOCAL LLM-judge engine (backend-agnostic, Phase 7B) "arize-phoenix-client>=2.9", # Phoenix annotation fallback (Phase 7B) "pandas>=2.0", # eval dataframes (Phase 7B) ] [project.optional-dependencies] dev = ["pytest>=8.0"] [tool.hatch.build.targets.wheel] packages = ["triage"] [tool.pytest.ini_options] testpaths = ["tests"] ``` ### backend/requirements.txt ``` fastapi>=0.115 uvicorn[standard]>=0.30 sse-starlette>=2.1 ``` ### frontend/package.json ``` { "name": "triage-frontend", "private": true, "version": "0.1.0", "type": "module", "scripts": { "dev": "vite", "build": "tsc -b && vite build", "preview": "vite preview", "test": "vitest run" }, "dependencies": { "react": "^18.3.1", "react-dom": "^18.3.1" }, "devDependencies": { "@testing-library/jest-dom": "^6.4.8", "@testing-library/react": "^16.0.1", "@types/react": "^18.3.5", "@types/react-dom": "^18.3.0", "@vitejs/plugin-react": "^4.3.1", "jsdom": "^25.0.0", "typescript": "^5.5.4", "vite": "^5.4.3", "vitest": "^2.0.5" } } ``` ### backend/server.py ```python from __future__ import annotations import json import re from fastapi import FastAPI, HTTPException from fastapi.middleware.cors import CORSMiddleware from pydantic import BaseModel from sse_starlette.sse import EventSourceResponse from backend.run_manager import RunRegistry ISSUE_RE = re.compile(r"^https://github\.com/[^/]+/[^/]+/issues/\d+/?$") app = FastAPI(title="TRIAGE frontend backend") app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"], ) _registry = RunRegistry() class RunRequest(BaseModel): issueUrl: str @app.post("/api/runs") async def start_run(req: RunRequest): if not ISSUE_RE.match(req.issueUrl.strip()): raise HTTPException(422, "not a GitHub issue URL") return {"runId": _registry.create(req.issueUrl.strip())} @app.get("/api/runs/{run_id}") def snapshot(run_id: str): try: return _registry.snapshot(run_id) except KeyError: raise HTTPException(404, "unknown run") @app.get("/api/runs/{run_id}/stream") async def stream(run_id: str): if not _registry.has(run_id): raise HTTPException(404, "unknown run") async def gen(): async for name, data in _registry.stream(run_id): yield {"event": name, "data": json.dumps(data)} return EventSourceResponse(gen()) @app.get("/api/replays") def replays(): # The frontend ships its own bundled fixture; this lists server-side ones (none yet). return [] ``` ### frontend/src/main.tsx ```typescript import React from "react"; import { createRoot } from "react-dom/client"; import App from "./App"; import "./styles/tokens.css"; import "./styles/app.css"; createRoot(document.getElementById("root")!).render( ); ``` ### frontend/src/App.tsx ```typescript import { useRef, useState } from "react"; import UrlInput from "./components/UrlInput"; import LiveLog from "./components/LiveLog"; import BrowserView from "./components/BrowserView"; import ReportCard from "./components/ReportCard"; import { startLiveRun, startReplayRun } from "./api"; import type { ReproReport, RunPhase, StreamEvent, SessionInfo } from "./types"; const API_BASE = (import.meta.env.VITE_API_BASE as string | undefined) ?? ""; type Status = "idle" | "running" | "report" | "error"; export default function App() { const [status, setStatus] = useState("idle"); const [phase, setPhase] = useState(null); const [events, setEvents] = useState([]); const [report, setReport] = useState(null); const [errorMsg, setErrorMsg] = useState(""); const [session, setSession] = useState(null); const [mode, setMode] = useState<"live" | "replay">("replay"); const cancelRef = useRef void)>(null); function onEvent(e: StreamEvent) { setEvents((prev) => [...prev, e]); if (e.type === "session") setSession(e.session); if (e.type === "status") setPhase(e.phase); if (e.type === "report") { setReport(e.report); setStatus("report"); } if (e.type === "error") { setErrorMsg(e.message); setStatus("error"); } } function onRun(runMode: "live" | "replay", url: string) { cancelRef.current?.(); setEvents([]); setReport(null); setErrorMsg(""); setPhase(null); setStatus("running"); setSession(null); setMode(runMode); cancelRef.current = runMode === "replay" ? startReplayRun(onEvent) : startLiveRun(API_BASE, url, onEvent); } return (
Triage reproduces your bugs by using your app
{status === "error" &&
{errorMsg}
} {status !== "idle" && (
{/* LiveLog (Task 6) */}
{phase &&
phase: {phase}
}
{report && }
)}
); } ``` ### triage/__init__.py ```python """TRIAGE — autonomous bug-reproduction agent.""" ``` ### frontend/vite.config.ts ```typescript import { defineConfig } from "vite"; import react from "@vitejs/plugin-react"; export default defineConfig({ plugins: [react()], test: { environment: "jsdom", globals: true, setupFiles: ["./src/test-setup.ts"], }, }); ``` [116 more indexed source files omitted to keep this export small. The full file list is in the Codebase structure section above.]