Use case / Coding agents
Know why the patch failed the test.
Coding agents edit files, run tests and try again. Trace every edit and every test run so a broken patch points back to the step that caused it.
- Edits
- Recorded
- Test runs
- Linked
- Retries
- Counted
Edit and test loop
The step where the patch went wrong.
Each task starts a trace so every later step can be judged against it.
See which files the agent read before it edited anything.
Edits are spans, so a change and its reason sit side by side.
A failing test is attached to the exact patch that produced it.
Retries show what the agent changed after seeing the error, and what it cost.
- 01ok
Task received
“Fix the date bug in the invoice export.”
- 02ok
Repo searched
Six files read, two edited.
- 03ok
Patch written
One change in the export function.
- 04failed
Test failed
One test fails on a leap-year date.
- 05ok
Retry with the error
The error is sent back and the patch is rewritten.
Instrument it
Record every edit and every test run.
Wrap the plan, each edit and the test run in spans, and a failing change points straight back to the step that caused it.
$ npm install tracify-sdk
import { span, traceAgent } from "tracify-sdk";
export const fixBug = traceAgent(async (task: string) => {
const patch = await span("write-patch", () => writePatch(task), { type: "llm", input: task });
const result = await span("run-tests", () => runTests(patch), { type: "tool" });
return result.passed ? patch : span("retry", () => writePatch(task, result.errors), { type: "llm" });
}, { environment: "production" });
The failure to catch
A confident patch that quietly breaks a test it never ran.
Trace planning, file edits, test runs and retries across coding agents so a broken change points back to the step that caused it.
What Tracify watches.
Edit and test loops
Recorded on every run, so it can be searched, compared and alerted on.
Files touched
Recorded on every run, so it can be searched, compared and alerted on.
Test outcome
Recorded on every run, so it can be searched, compared and alerted on.
Tokens per fix
Recorded on every run, so it can be searched, compared and alerted on.
