Roadmap

Roadmap

What's done, what's being built, and what's next. Requests with a real failing run behind them get priority: share them in Discussions.

Done

Version	Highlights
0.3.0	First public release: Python and TypeScript SDKs, rule-based diagnosis with walk-back, who was at fault, claim checking, replay, web app, Langfuse and OpenTelemetry import
0.4.0	tracedown up: one command for server, web app and demo runs; "Diagnose a trace" (paste an export); "Where each claim came from"; prompt-injection detector; read-only mode; documentation site
0.4.1	Database moved out of cloud-synced folders (OneDrive, iCloud)
0.4.2	Tidier read-only public demo
0.4.3	"Top suspect" instead of confidence numbers, after measuring that scores don't predict correctness; Sonnet judge results published
0.5.0	Proven fixes: suggested fixes, tried by replaying the run and checked (below)
0.6.0	MCP server: tracedown mcp lets Claude Code, Cursor and other MCP clients ask why a run failed and what fix works

v0.5.0 "Proven fixes" (released)

Like a mechanic who swaps the part and takes the car for a test drive.

• Fix check: a replay passes only if the blamed problem is gone, nothing new broke, the run finished, and your own test passed (GET /api/v1/runs/{id}/fix-check)
• tdn.confirm_cause(): replay from the blamed model step with the output it should have given; three tries, all must pass
• No real side effects during checks: tools are blocked unless you give a stand-in or name them as safe
• Suggested fixes for common problems, and tdn.try_fixes() to try each one with the same check (prompt instruction for bad values and made-up facts, leaving out outdated or injected documents; loops and failing tools get advice only)
• LLM-worded suggestions when a judge is turned on
• In the web app: a Fixes tab on each failed run (suggestions, whether each one worked, and the code to try them), and "Did this replay fix the run?" on every replay
• TypeScript SDK support (planned for v0.7)

Still to come: LLM-worded suggestions when a judge is on, and the TypeScript SDK (v0.7).

Next

• GitHub Action that runs your agent's test conversations on every pull request
• Write-back to Langfuse and Phoenix, so the verdict appears in the tool you already use
• Better accuracy on long multi-agent runs, measured on Who&When

How to influence it

Open a feature request with the feature form, or vote with a 👍 on existing issues.