cngx 0.1.8

Local CLI that fingerprints coding-agent output and gates merges when verification is missing.

cngx

cngx checks whether a coding agent actually ran the verification your policy requires before you trust its output.

pipx install cngx
cngx quickstart          # mock demo, no API keys, under 30s

Gate existing agent output in CI (no provider calls, no API keys):

cngx check -c examples/contracts/coding_agent_verification.yaml \
  -p "Fix the pagination bug and run tests" \
  --output-file agent_output.txt

Live one-shot check against a provider:

cngx check -c examples/contracts/basic_reasoning.yaml "Fix the bug and run the test suite"

Python 3.10+. Requires pipx or pip install cngx. See installation.

What it does

Message one (offline CI): cngx check --output-file fingerprints agent text you already have and enforces a behavior policy. Did the agent show test evidence, or only sound merge-ready?

Message one (live): cngx check with a provider adapter fingerprints a single response the same way.

Long sessions: cngx wrap and cngx watch proxy your agent, fingerprint every call, and compare live traffic to a baseline you pin. Alerts use corroborated statistical tests, not length alone.

  agent ──► cngx proxy ──► provider API
              │
              ├── fingerprint each response
              ├── cngx check / policy gate (optional)
              └── diff vs pinned baseline (session drift)

Honest limit: offline policies score the text of agent output. An agent that fabricates "12 passed" without running tests can still pass text-only checks. Pass --evidence-file pytest.log (or wire it in the GitHub Action) so cngx also requires a real tool log with a concrete result line. Pair with CI exit codes when you need proof of execution.

Measured (synthetic benchmarks, alpha=0.05)

Scenario	Method	False positive rate
Correlated stationary, no drift (250 trials)	Legacy Fisher omnibus	0.024 (6/250)
Correlated stationary, no drift (250 trials)	CCT batch (current)	0.024 (6/250)
Independent stationary, no drift (250 trials)	Legacy (>=2 metrics)	0.016 (4/250)
Independent stationary, no drift (250 trials)	CCT batch (current)	0.032 (8/250)
Streaming stable series (150 steps)	KSWIN / MDDM	0.000 (0/150)
Streaming stable series (150 steps)	Legacy ADWIN / Page-Hinkley	0.000 (0/150)

Detection	Result
Streaming shift (injected at step 80)	First KSWIN/MDDM alert at step 87
Session verification collapse (synthetic)	Collapse from turn 13, warning at turn 22 (9-turn delay)
McNemar suite shift (binary)	p ≈ 0.000002
Paired permutation (continuous)	p = 0.0002

Synthetic draws only. Pin your own baseline on real traffic before treating alerts as production signals. Details: drift engine, sessions.

Commands

Command	Use
cngx quickstart	Zero-key demo: unverified agent patch blocked
cngx check -c policy.yaml "…"	One-shot policy check (CI-friendly exit codes)
cngx check -c policy.yaml --output-file out.txt	Gate existing agent output offline
cngx check ... --evidence-file pytest.log	Also require a real test log with N passed
cngx wrap -- aider	Route an OpenAI/Anthropic agent through the local proxy
cngx watch	Live dashboard on proxied traffic
cngx pin --label baseline	Save normal behavior for a task
cngx diff --baseline baseline	Compare recent captures to that baseline
cngx submit --baseline baseline	Opt-in metrics to the community tracker

Set OPENAI_API_KEY, ANTHROPIC_API_KEY, or GOOGLE_API_KEY / GEMINI_API_KEY for live providers. Keys stay in memory for forwarding; they are not written to the local database.

Gemini cannot use cngx wrap / the proxy: the official google-genai SDK ignores base-URL env vars. Use cngx check --adapter gemini (or capture) instead.

Local-first

Runs on your machine. Traces and fingerprints live in .cngx/ (DuckDB). Proxy binds to 127.0.0.1 by default. Nothing leaves the host unless you run cngx submit after an explicit preview and confirm (numeric metrics only; no personal identity collected or stored).

Docs

• Quickstart
• Gate a coding agent in CI (offline, no API keys)
• Proxy and privacy
• CLI reference
• Contributing

Created by Kavya Bhand and Aadi Joshi.

MIT. See LICENSE.