Paste a raw stack trace → see the failure as a path through your codebase graph. Post-mortem, local-first, zero setup. No debugger session, no Docker, no API key, works offline.
$ cd my-repo
$ npx crashpath # paste a traceback into the browser
$ pytest 2>&1 | npx crashpath # or pipe it straight in
$ npx crashpath demo # zero-config demo, right nowA stack trace is a principled filter over your codebase — it selects the ~10 functions that matter right now. crashpath parses only the files the trace touches, builds a ground-truth graph with tree-sitter, and overlays the runtime failure path:
- ▬ red — the runtime trace: what actually executed. Exact, always.
- ▬ grey — static call/import edges parsed by tree-sitter. Best-effort, honestly labeled.
- ⤳ violet, dashed — ghost edges: runtime hops with no static edge. That's not a bug — that's your decorator/registry/framework dispatch made visible, labeled with the mechanism.
- Blast radius: 1-hop callers (via
git grep) and callees around the failure path. - Click any node → the source, crash line highlighted.
←/→walks the spine.
Also: multi-trace picker for messy logs · sourcemap resolution for minified JS
(dist/bundle.js:1:4823 → src/pricing.ts:19) · chained exceptions as stacked spines ·
--ref v1.4.2 to analyze the exact version that crashed (git worktree) ·
crashpath export -t trace.txt -o failure.html for a self-contained artifact you can attach to an issue.
| Documented pain elsewhere | crashpath's answer |
|---|---|
| Large repos break whole-repo indexers (timeouts, context limits) | Trace-as-filter: parse only trace-reachable files — typically <50, even in monorepos |
| AI-generated diagrams aren't ground truth | Three layers, visually distinct: runtime fact / parsed evidence / badged inference. AI can never draw nodes or edges |
| Mermaid rendering ceiling | Custom SVG renderer, deterministic layout, full interactivity |
| Setup friction: Docker, keys, DBs | npx crashpath. Nothing else. Fully offline core |
| Per-language native indexers killed Sourcetrail | tree-sitter WASM grammars + ~200-line frame parsers; add a language |
| Static analysis can't see runtime dispatch | Ghost edges — the trace is runtime truth |
| Line numbers drift between prod and HEAD | --ref checks out the crashed version in a throwaway worktree |
Python and JavaScript/TypeScript (including sourcemaps) in v1 — done well, on a 32-fixture corpus of real, ugly traces (k8s/JSON/CI log wrapping, chained exceptions, pytest formats, minified stacks). Frame-parse rate: 32/32. More languages via the plugin interface.
$ crashpath --ai ollama --model qwen3:4b # fully local, no key
$ ANTHROPIC_API_KEY=… crashpath --ai anthropic
$ OPENAI_API_KEY=… crashpath --ai openaiThe payload is ~3–5k tokens (frames + snippets), small enough for a local 3B model. Output is a schema-validated annotation — root-cause hypothesis, ghost-edge explanations, suggested fix — rendered in a clearly-badged "AI inference" panel. AI never draws graph structure.
$ claude mcp add crashpath -- npx crashpath mcpTools: map_trace(trace_text, repo_path?, ref?) → structured failure summary + a local UI URL for
the human; export_trace_map(trace_text, format) → standalone artifact path.
The server binds 127.0.0.1 only. Source reads are sandboxed to the repo root. Zero network
calls unless you explicitly pass --ai (and with --ai ollama, still none beyond localhost).
- Static edges are name-resolution, not type inference — deliberately (that treadmill killed better-funded tools). Ambiguity is skipped, not guessed; ghost edges mark what static analysis can't see.
- Post-mortem only: no live debugger integration in v1.
- Two languages in v1. Ruby/Go/JVM/… belong to contributors.
- pytest's default (non-native) traceback format is parsed to the failing test frame only.
$ npm ci && npm --prefix ui ci
$ npm test # vitest, incl. the 32-fixture corpus gate
$ npm run corpus # parse-rate table
$ npx playwright testArchitecture notes: docs/architecture.md. Fixture corpus rules: CLAUDE.md — goldens are evidence, never regenerate silently.
MIT © Zhao Han Qing