cngx 0.2.0

cngx 0.2.0

Your AI coding agent says "done, all tests pass." cngx runs what it claimed and blocks the merge when that is not true.

This release repositions cngx around one honest, non-gameable idea: bind an agent's claim to a real run.

New: cngx verify

cngx verify --output-file agent_message.md -- pytest

cngx runs the command, parses the real result, reads what the agent claimed, and blocks (exit 1) when the agent claimed success but the checks fail, or when its reported counts do not match the real run. The verdict comes from the command's real exit code, so a fabricated "all tests passed" cannot pass.

Verified against the live OpenAI API: gpt-4o-mini wrote "I ran the test suite, and all tests passed successfully. The code is now ready to merge." It ran nothing. cngx ran the real tests and returned BLOCKED, Real result: FAILED (failures=2), exit 1.

• Result parsers for pytest, unittest, jest/vitest, go test, cargo test, and generic exit-code commands.
• Sources for the claim: --output-file, --stdin, --claim.
• Offline --evidence-file to gate an existing CI/test log without executing.
• --require-claim, --timeout, --json.

Changed

• cngx quickstart now runs real tests in a throwaway project (no API keys) and shows a false claim blocked, then a real fix verified.
• README, docs, and the GitHub Action lead with verify. The Action gains a command input.
• cngx check (heuristic text policy) is demoted to an advanced lint and points to verify, because scoring prose can be gamed by a fabricated claim.
• Community tracker seed records removed so charts reflect only real submits.

Install

pipx install cngx
cngx quickstart

Full changelog: https://github.com/aadi-joshi/cngx/blob/main/CHANGELOG.md

cngx 0.1.10

Summary

• Tracker no longer plots the same fingerprint twice (vertical spikes fixed)
• Submit + Lambda reject duplicate fingerprint shapes
• Honest tagline: CI gate for merge-ready agents that never showed tests
• Explicitly no fake token-savings claim until a circuit breaker exists

cngx 0.1.9

Summary

• Public tracker no longer shows harness tabs (cngx-e2e-test, cngx-cli-live)
• submit rejects test/harness model names and probe baselines
• submit only attaches same-model fingerprints to a baseline

Notes

Charts stay sparse until more real multi-day community submits land. That is expected.

cngx 0.1.8

Fixes

• Claude adapter: stop sending temperature+top_p together (broke live cngx check -a claude on current Haiku/Sonnet)
• Windows: cngx diff no longer crashes on cp1252 consoles
• Diff recommendations no longer claim no critical issues when major shifts exist

Live verified before release

• Claude Haiku policy check PASS
• Anthropic proxy stream + non-stream fingerprints persisted
• OpenAI wrap child process OK

cngx 0.1.7

Fix

Windows standalone binaries failed because PowerShell did not expand dist/*.whl, so PyInstaller built without typer/pydantic installed. Install step now uses bash and asserts imports.

Install

pipx install cngx
pip install cngx==0.1.7

cngx 0.1.6

Highlights

• Windows standalone binaries actually run (pydantic_core collected; lazy package import)
• Anthropic /v1/messages fingerprinted like OpenAI (stream + non-stream)
• cngx wrap warns when Gemini is the only configured path
• Docs/SECURITY honesty pass

Install

pipx install cngx
pip install cngx==0.1.6

Verified locally: Windows cngx.exe --help, version, quickstart, and offline check (exit 1 on unverified patch).

cngx 0.1.5

Highlights

• Fix Windows PyInstaller binaries (bundle pydantic and related imports)
• Offline --evidence-file injects a concrete result line into gated text when the CI log is real
• Remove unwired cngx enforce and unused Postgres backend
• Honest docs: proxy latency, Gemini wrap limits, tracker privacy vs edge logs
• PyPI Documentation URL points at the GitHub docs tree
• mkdocs build --strict in CI

Install

pipx install cngx
# or
pip install cngx==0.1.5

See CHANGELOG.md for the full list.

v0.1.4

Highlights

• Fix cngx watch crash (Typer OptionInfo leak)
• Accept GEMINI_API_KEY as well as GOOGLE_API_KEY
• Default Gemini model: gemini-flash-latest
• --evidence-file for offline gates (require real pytest/CI result lines)
• PyPI: Alpha classifier, drop unused cloud/platform/postgres extras
• Absolute README links for PyPI rendering
• Stronger coding-agent policy (result-shaped evidence patterns)

Install

pipx install cngx==0.1.4
cngx quickstart

v0.1.3

Offline policy gate

Gate existing agent output in CI with no LLM calls:

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

• --output-file, --stdin, --response-file
• coding_agent_verification.yaml + lenient variant
• GitHub Action output-file input
• Gate a coding agent in CI

Fix

• Semantic drift PC1 projection (vt[0] instead of wrong u[:, 0] axis)

v0.1.2

OIDC trusted publishing pipeline test. No functional changes.