offtrack

git diff for AI agent runs.

Record golden trajectories. Re-run after any change. See the first step where your agent went off track — and gate CI on it.

---

You bumped a model. Or tweaked a prompt. Or upgraded your agent framework.

Every eval score still looks fine. But your agent now refunds $842 without checking the refund policy — cheaper, faster, and catastrophically wrong:

$ offtrack check

  ✗ refund/over-limit  0/5 aligned
      =  step 0  llm_call     (both runs identical)
      =  step 1  tool_call    lookup_order({"order_id": "TEST-1"})
      ▲ first divergence — expected check_refund_policy, got issue_refund
        baseline  check_refund_policy({"amount_usd": 842.0})
        this run  issue_refund({"order_id": "TEST-1", "amount_usd": 842.0})
      divergence rate rose 0% → 100% (Fisher exact p=0.004, effect +100%)
      Δtokens -46%  Δlatency -30%

  FAIL — exit code 1

Eval scores tell you whether behavior changed. offtrack shows you where: the first divergent step, with cost and latency deltas, from traces stored locally in SQLite. Statistical verdicts (PASS / FAIL / INCONCLUSIVE) gate CI so stochastic LLM variance doesn't cause false alarms — and honest uncertainty doesn't hide behind a green check.

Quickstart

pip install offtrack
offtrack init                 # scaffold offtrack.yaml + baselines/
offtrack record               # capture golden trajectories (N runs per task)
# ...change your model / prompt / framework...
offtrack check                # first divergent step + verdict + exit code

Capturing traces — three ways

Zero-code (OpenAI / Anthropic SDKs) — two lines at your agent's entrypoint:

import offtrack.capture
offtrack.capture.install()   # patches whichever SDKs are importable

Every LLM call is recorded, and tool calls are reconstructed by message delta — real names, args, and results, no changes to your agent loop.

LangGraph / LangChain — pass the callback:

from offtrack.integrations.langgraph import OfftrackCallbackHandler

handler = OfftrackCallbackHandler()
graph.invoke(inputs, config={"callbacks": [handler]})
handler.finish()

Anything else — write capture events (JSONL) to $OFFTRACK_TRACE_DIR, or export OpenTelemetry GenAI traces there: offtrack ingests both semconv generations and the OpenInference flavor. Claude Code sessions import directly with offtrack ingest claude-code <session.jsonl>.

How it works

1. Record — offtrack record runs each task N times (default 5) and stores the trajectories: every LLM call, tool call, argument, token count, cost, latency. Baselines auto-export to baselines/*.json — committed to git, so changing golden behavior is a reviewed act in PRs.
2. Align — new runs are aligned against baselines with Needleman-Wunsch sequence alignment over steps (tool-name gating + structural argument similarity). Reordered parallel calls, retries, and volatile fields (UUIDs, timestamps — masked by default) don't cause false alarms.
3. Localize — the report points at the first divergent step: missing, extra, or changed — with the argument-level diff and whether the trajectories resynced afterward.
4. Verdict — baselines are recorded N times so run-to-run variance is measured, not assumed. FAIL requires an exact Fisher test plus a minimum effect size; PASS requires an exact upper confidence bound. Anything else is INCONCLUSIVE with a prescription: "run 2 more repetitions."

CI

- uses: abhi13-tech/offtrack@v0
  with:
    suite: offtrack.yaml
  env:
    OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}

FAIL fails the check and posts a PR comment with the divergence table. INCONCLUSIVE warns without blocking (strictness is opt-in via inconclusive-as: fail). Exit codes: 0 pass, 1 fail, 3 inconclusive, 4 setup error — infra problems never masquerade as regressions.

This repo gates itself: every PR runs offtrack against its own demo agent (dogfood.yml), and the gate's failure path is itself under test.

pytest

@pytest.mark.offtrack(task="refund/over-limit")
def test_refund_flow(offtrack):
    with offtrack.record():
        run_my_agent("refund order TEST-1")
    offtrack.assert_matches_baseline()
    offtrack.assert_cost_under(usd=0.05)
    offtrack.assert_max_steps(30)

pytest -n auto (xdist) is fully supported — workers spill trajectories to append-only files; the controller merges at session end.

What offtrack is not

These are complements, not competitors — offtrack is the trajectory-diff layer:

	offtrack	promptfoo	LangSmith	EvalView
Unit of comparison	step-level trajectory	prompt/output pairs	hosted traces + evals	eval-run snapshots
First-divergence localization	✓	—	—	—
Semantic alignment (retries, reordering, masks)	✓	—	—	exact-match only
Statistical verdict (variance-aware)	✓	—	—	pass@k
Local-first, no account	✓ SQLite	✓	hosted	✓
Vendor-neutral ingest (OTel GenAI, both semconv generations)	✓	—	—	—

offtrack doesn't score answer quality — pair it with an eval framework for that. It catches the thing eval scores structurally miss: procedural regressions in how the agent got there.

Demo

examples/refund-agent/ is a complete offline demo — a small support agent with a written refund policy and two scripted personas: fake-careful (follows policy) and fake-sloppy (a stand-in for a cheaper model that skips the policy check). No API key, no network, $0:

cd examples/refund-agent
AGENT_MODEL=fake-careful offtrack record    # golden: lookup → policy check → escalate
AGENT_MODEL=fake-sloppy  offtrack check     # FAIL: skips policy, refunds $842 directly

The fake mode is a scripted LLM, not a mocked test — the entire real pipeline (capture, ingest, alignment, stats) executes identically in fake and real mode.

Roadmap

• LangGraph callback + OpenAI/Anthropic SDK capture shims — shipped in 0.2.0 (offtrack.capture.install())
• Claude Code session ingest — shipped in 0.2.0 (offtrack ingest claude-code)
• Semantic matchers — shipped in 0.3.0 (align.final_answer: lexical | embedding; LLM-judge matcher still to come)
• offtrack bisect — shipped in 0.4.0 (offtrack bisect --good <ref>)
• CrewAI / Pydantic-AI adapters · HTML report viewer

Development

git clone https://github.com/abhi13-tech/offtrack && cd offtrack
uv sync --all-extras --group dev
uv run pytest          # 200 tests, all offline
uv run mypy            # strict
just dogfood           # run the gate against the demo locally

MIT licensed.