Tracing for agents

Trace answers that fit in context.

Your coding agent pays for every token it reads. Provon answers in structured JSON sized for a prompt — start with a summary, open a span only when the question demands it.

Explore the tracing CLI Instrument traces
provon traces summarize ok

$ provon traces summarize tr_8f9a2b

"status":

"error · run_tests"

"spans":

12 (1 failed)

"latency":

1.42s

"nextActions":

provon traces span tr_8f9a2b sp_4d21

Context budget per answer
summarize ~0.8k
overview ~3k
raw spans ~12k

Small by default, so the agent spends its window on reasoning.

Progressive disclosure

Ask the smallest useful question.

  1. Search

    01

    Narrow to candidates by error text, span name, or user.

    $ provon traces search "permission denied" --level error

  2. Summarize

    02

    Read timing, tokens, cost, and failed spans first.

    $ provon traces summarize <trace-id> --top 5

  3. Inspect

    03

    Open one span only when the question needs it.

    $ provon traces span <trace-id> <span-id>

  4. Export

    04

    Bundle evidence with Git context for a handoff.

    $ provon traces export <trace-id> --out ./evidence

Built for tools

Evidence without the noise.

Parseable envelopes

Every command returns ok, data, warnings, and nextActions.

Bounded responses

Overview before spans. Raw detail stays an explicit choice.

Local diagnosis

provon diagnose runs offline with no API key and no upload.

Repo-aware handoff

Exports record branch, commit, and dirty paths, not code.

Start with one production trace.

Connect an agent, run a task, and read the evidence.

Open quickstart