Structure

Layers are inferred from where files sit in the tree, not from reading the code. A project that names its directories unconventionally will read oddly here — open the file browser to check anything the diagram implies.

Languages

Share of indexed source by file size. Binary and vendored files are excluded.

Dependencies

Declared in the repository’s manifests at the indexed commit. A declared package is not proof it is used, and runtime dependencies are listed first.

Feature verification

• Arize (AX/Phoenix) tracing of the retry loopVerified

  Arize Phoenix tracing of the retry loop

  Claimed on readmehigh confidence

  • triage/tracing/setup.py:17 — setup_tracing registers either Arize AX (arize.otel.register) or Phoenix (phoenix.otel.register) as the OpenTelemetry backend, selected by cfg.trace_backend
  • triage/repro_agent/browser.py:122 — run_repro accepts a run_trace and wraps browser execution in trace spans
• Bounded retry loop (loop-safety cap)Verified

  Autonomous retry loop until bug is reproduced or diagnosis is confirmed

  Claimed on readmehigh confidence

  • triage/repro_agent/loop.py:30 — MAX_REPRO_ATTEMPTS = 3 hard-caps total browser attempts per repro cycle
• Bug detection via dual signal (blank page + console crash match)Verified

  TRIAGE captures the crash to confirm the bug actually reproduced

  Claimed on readmehigh confidence

  • triage/repro_agent/browser.py:52 — detect_bug requires both blank_body and a matching console error substring before flagging bug_detected
• Claude used for reasoning across all three agentsVerified

  Built with claude

  Claimed on Devposthigh confidence

  • triage/parser_agent/agent.py:59 — extract_steps is called with an injected anthropic_client to parse the GitHub issue into steps
  • triage/hypothesis_agent/reasoning.py:16 — diagnose() calls the Anthropic client (model claude-sonnet-4-6) for root-cause reasoning
• Demo replay mode with a recorded fixture runVerified

  (not explicit in README, but present as a demo feature; included since BUILT WITH lists these integrations for a working demo)

  Claimed on readmemedium confidence

  • frontend/src/replay.ts:6 — startReplay drip-feeds a StreamEvent[] to the UI on a timer, replaying a recorded run
  • frontend/src/fixtures/recorded-run.json — Bundled fixture data used as the replay source
• FastAPI backend serving run status and SSE event streamVerified

  Built with fastapi

  Claimed on Devposthigh confidence

  • backend/server.py:16 — FastAPI app exposes POST /api/runs, GET /api/runs/{id}, and an SSE GET /api/runs/{id}/stream backed by RunRegistry
• Frontend web UI: paste a GitHub issue URL, live run view, report cardVerified

  A developer pastes a GitHub issue; TRIAGE... writes a structured report (implies a UI)

  Claimed on readmehigh confidence

  • frontend/src/App.tsx:12 — React app wires UrlInput, LiveLog, BrowserView, and ReportCard components to a live/replay run stream
  • frontend/src/api.ts:1 — startLiveRun/startReplayRun connect the UI to the backend run stream
• GitHub issue parsing into structured repro stepsVerified

  A developer pastes a GitHub issue; ParserAgent converts it into structured repro steps

  Claimed on readmehigh confidence

  • triage/parser_agent/github.py:41 — fetch_issue performs a real GET against the GitHub REST API for a given issue URL
  • triage/parser_agent/agent.py:47 — post_initial_steps fetches the issue, calls extract_steps (Claude), formats and sends steps @ReproAgent
• Honest fail-to-succeed bug.detected tracing flipVerified

  Arize Phoenix tracing of the retry loop must be honest about the fail-to-succeed flip (project rule)

  Claimed on readmemedium confidence

  • triage/memory/backends/ax.py:17 — Docstring and code identify attributes['bug.detected'] on each repro_attempt span as 'the honest reproduction signal (rule 8's fail->succeed flip)', read back for prior-run context
• Multi-agent coordination via Microsoft Band platformVerified

  Built with microsoft-band

  Claimed on Devposthigh confidence

  • triage/shared/band.py:105 — BandAgent.connect uses band.platform.link.BandLink and band.client.rest to connect, create/join a chat room, and subscribe over websocket
• New Browserbase session per retry attempt (no session reuse)Verified

  Rule: new Browserbase session per retry, never reused

  Claimed on readmehigh confidence

  • triage/repro_agent/browser.py:128 — run_repro always calls client.sessions.start() fresh with no sessionId parameter, and the module docstring states a new session is created every call
• Real cloud browser bug reproduction via Browserbase/StagehandVerified

  TRIAGE opens a real cloud browser and clicks through a live app to trigger the bug

  Claimed on readmehigh confidence

  • triage/repro_agent/browser.py:128 — run_repro starts a real Browserbase/Stagehand session (client.sessions.start) and connects Playwright over CDP
  • triage/repro_agent/browser.py:187 — Each repro step is executed via session.observe(...) then session.act(...), matching the claimed observe->act->screenshot loop
• Root cause diagnosis via Claude reasoningVerified

  TRIAGE diagnoses the root cause and writes a structured report

  Claimed on readmehigh confidence

  • triage/hypothesis_agent/reasoning.py:24 — diagnose() builds a Diagnosis dataclass (decision, root_cause, redirect_instruction) via a Claude call with a detailed system prompt
  • triage/hypothesis_agent/agent.py:92 — on_message callback invokes diagnose() with the real anthropic client and posts the diagnosis back over Band
• Three-agent Band-coordinated architecture (ParserAgent, ReproAgent, HypothesisAgent) with retry redirectVerified

  Three agents coordinate in one Band room; HypothesisAgent can redirect ReproAgent to retry

  Claimed on readmehigh confidence

  • triage/shared/band.py:20 — Defines three distinct AgentName identities and payload contracts (ReproStepsPayload, ReproResultPayload, HypothesisPayload) shared across agents
  • triage/repro_agent/loop.py:66 — classify_message routes by sender identity into steps/confirm/redirect, and extract_tweak parses HypothesisAgent's retry instruction
  • triage/hypothesis_agent/agent.py:40 — route_diagnosis maps a Diagnosis to @ReproAgent or @ParserAgent redirect targets
• LLM-as-judge evaluation of repro fidelity and root-cause correctnessCode-supported

  (implied by eval module, not explicit in README/Devpost, included for completeness)

  Claimed on readmemedium confidence

  • triage/eval/judges.py:47 — make_fidelity_judge/make_root_cause_judge build phoenix.evals classifiers, but this is an internal dev-eval harness rather than a user-facing feature explicitly claimed in the README

An AI agent derived these features from the project’s Devpost page and readme, then searched the code for each one. Verified features are backed by cited code; claimed-only features had no supporting code, which is not by itself proof a feature is missing.