kalibra 0.1.1

Regression detection and CI quality gates for AI agents.

Kalibra
Regression detection and CI quality gates for AI agents.

---

You change a prompt, swap a model, or refactor a tool — did the agent get better or worse?

Kalibra compares two populations of traces and tells you, with statistical rigor, what changed — success rate, cost, latency, tokens, per-task regressions, per-span breakdowns. Two dependencies. One command.

pip install kalibra
kalibra demo

  Kalibra Compare
  ──────────────────────────────────────────────────────────
  Baseline       100 traces   (baseline.jsonl)
  Current        100 traces   (current.jsonl)
  Direction ~ MIXED

  Trace metrics

  ▲ Success rate      50.0% → 75.0%  +25.0 pp   (p=0.000)
  ▲ Cost              $0.0358 → $0.0213 median  -40.5%
  ▼ Duration          7.6s → 15.2s median  +99.1%
  ≈ Steps             4 → 4 steps/trace (median)  +0.0%
  ▼ Error rate        0.2% → 4.3%  +4.1 pp
  ≈ Token usage       7,746 → 7,738 tokens/trace (median)  -0.1%
  ≈ Token efficiency  8,443 → 7,090 tokens/success (median)  -16.0%
  ▲ Cost / quality    $0.0385 → $0.0189 per success (median)  -51.0%

  Trace breakdown

  ~ Per trace         20 matched — ✓ 10 improved, ✗ 5 regressed

  Span breakdown

  ▼ Per span          5 matched — ✗ 1 regressed, ~ 4 mixed

  ──────────────────────────────────────────────────────────
  ~ MIXED — no quality gates configured

Add -v for per-task outcome changes, per-span breakdowns, and confidence intervals.

Why Kalibra

Aggregate metrics hide task-level regressions. Your success rate went from 80% to 82% — great. Except five tasks that used to pass now fail, masked by eight new easy ones that pass. Kalibra catches this.

• 10 metrics — success rate, cost, duration, steps, error rate, tokens, token efficiency, cost/quality, per-task breakdown, per-span breakdown
• Statistical rigor — bootstrap 95% CIs on continuous metrics, two-proportion z-test on rates, noise thresholds to ignore jitter
• Quality gates — require: success_rate_delta >= -5 fails your CI pipeline (exit 1) when thresholds are violated
• Any JSONL — flat traces, nested spans, non-standard field names. Use --suggest to auto-detect field mappings
• Three output formats — terminal (human), markdown (PR comments), JSON (automation)
• Two dependencies — click + pyyaml. No ML frameworks, no API keys

Quickstart

1. Install

pip install kalibra

2. Try the demo

kalibra demo

This creates a kalibra-demo/ directory with sample traces and runs a comparison. Afterwards:

kalibra compare kalibra-demo/baseline.jsonl kalibra-demo/current.jsonl -v

3. Compare your own data

If your fields don't match the defaults, let Kalibra figure it out:

kalibra inspect your-traces.jsonl --suggest

This scans your data and prints a copy-pasteable compare command with the right --outcome, --cost, --trace-id flags.

Quality gates for CI

# kalibra.yml
baseline:
  path: ./baselines/production.jsonl
current:
  path: ./eval-output/canary.jsonl

require:
  - success_rate_delta >= -2     # max 2pp success rate drop
  - regressions <= 5             # max 5 tasks regressed
  - cost_delta_pct <= 20         # max 20% cost increase
  - span_regressions <= 3        # max 3 span types regressed

kalibra compare        # reads kalibra.yml, exits 1 on failure

Field mapping

Kalibra works with any JSONL shape. Map your fields in config or on the command line:

# kalibra.yml
fields:
  outcome: metadata.result
  cost: agent_cost.total_cost
  task_id: metadata.task_name

kalibra compare a.jsonl b.jsonl --outcome metadata.result --cost usage.total_cost

Comparing files with different schemas? Override fields per source:

baseline:
  path: ./langfuse.jsonl
  fields: { outcome: metadata.result, cost: usage.total_cost }
current:
  path: ./braintrust.jsonl
  fields: { outcome: scores.correctness, cost: metrics.cost }

Python API

from kalibra.loader import load_traces
from kalibra.engine import compare
from kalibra.renderers import render

baseline = load_traces("baseline.jsonl")
current = load_traces("current.jsonl")

result = compare(baseline, current, require=["success_rate_delta >= -5"])
print(render(result, "terminal", verbose=True))
print("passed:", result.passed)

Commands

kalibra compare [a.jsonl b.jsonl]     Compare traces — flags, config, or positional args
kalibra compare -v                    Verbose — CIs, per-task/per-span detail
kalibra compare --format markdown     Markdown for PR comments
kalibra compare --format json         Machine-readable JSON
kalibra compare --metrics             List all threshold fields
kalibra inspect traces.jsonl          Show data coverage and fields
kalibra inspect traces.jsonl --suggest  Auto-detect field mappings
kalibra init                          Create kalibra.yml interactively
kalibra demo                          Run comparison on built-in sample data

Development

git clone https://github.com/khan5v/kalibra.git
cd kalibra
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest