cngx documentation

Local CLI that fingerprints coding-agent output and gates merges when verification is missing. Also proxies long agent sessions for baseline-relative drift. No cloud account required.

overview

Day-one use: cngx check --output-file gates agent text in CI with no API keys. Longer runs: wrap the agent, pin a baseline, and watch for mid-session verification collapse.

Install with pipx, gate offline or run agents through cngx wrap, optionally submit opt-in metrics to this tracker.

install

Recommended: pipx puts cngx on your PATH in an isolated environment. No virtualenv to manage.

pipx install cngx
cngx version

Alternative: pip install cngx inside a project virtualenv.

No Python: standalone binaries on GitHub Releases (PyInstaller builds per platform).

Docker is optional for containerizing the proxy on a server. It is not required for normal CLI use.

quickstart

Zero API keys. Mock adapter demo in under 30 seconds.

cngx quickstart

gate agent output

Gate existing agent text in CI with no API keys: cngx check --output-file output.txt -c policy.yaml. Full walkthrough: docs/guides/gate-coding-agent.md

wrap

Zero-code instrumentation for existing agent CLIs. Starts the local proxy if needed and injects SDK base URL environment variables into the child process.

cngx wrap -- aider
cngx wrap --session-id long-run -- python my_agent.py

Variable	Set to	Used by
`OPENAI_BASE_URL`	`http://127.0.0.1:8642/v1`	OpenAI SDK, many compatible tools
`OPENAI_API_BASE`	same	legacy alias in some agent wrappers
`ANTHROPIC_BASE_URL`	`http://127.0.0.1:8642`	Anthropic SDK, Claude Code-style CLIs

Manual base URL configuration remains for tools that ignore environment overrides.

watch and sessions

cngx watch runs the proxy with a live terminal dashboard. Each turn gets a session id and turn number.

After 20+ turns, cngx can raise a session stability warning when verification-step variance collapses (distinct from per-turn structural drift). See session thresholds in the repository docs under docs/concepts/sessions.md.

cngx report --session my-long-run

drift detection

Structural drift compares heuristic fingerprint metrics (depth, verification steps, hedging, corrections) to your pinned baseline. Live proxy uses KSWIN and MDDM streaming tests; batch paths use Mann-Whitney U with Benjamini-Hochberg FDR and the Cauchy Combination Test (CCT).

Semantic drift (optional, watch --semantic) uses local embeddings. Neither structural nor semantic drift alone proves the model got worse.

Alerts require corroboration across multiple metrics. Length-only shifts do not alert alone.

cli reference

Command	Purpose
`cngx init --yes`	Create `.cngx/` local database
`cngx quickstart`	Mock demo, no API keys
`cngx wrap -- <cmd>`	Run agent through proxy (recommended)
`cngx watch`	Proxy + live TUI dashboard
`cngx pin --label NAME`	Pin baseline fingerprint
`cngx diff --baseline NAME`	Compare recent traffic to baseline
`cngx check -c POLICY.yaml "prompt"`	CI policy check (exit 0/1/2)
`cngx report --session ID`	Session trajectory summary
`cngx submit --baseline NAME`	Opt-in tracker submission (numeric metrics only)

submit and privacy

By default nothing leaves your machine. cngx submit is opt-in, shows the exact JSON before sending, and posts only after you confirm.

Submitted payloads contain only: model name, timestamp, numeric metrics, drift score, and your baseline label. No prompts, outputs, trace IDs, or task names. Application storage keeps payload fields only; CloudFront/API Gateway access logs are separate infrastructure concerns.

No GitHub account or pull request is required.

cngx submit --baseline my-baseline --dry-run
cngx submit --baseline my-baseline

tracker schema

Each community record is one JSON object (schema version 1):

Field	Type	Description
`schema_version`	int	Always `1`
`record_id`	string	UUID
`timestamp`	string	ISO-8601 UTC
`model`	string	Model name
`baseline_label`	string	Submitter's baseline label
`drift_score`	float	0 to 1 vs baseline
`depth`	int	Reasoning depth
`verification_steps`	int	Verification step count
`hedging_ratio`	float	Uncertainty language ratio