NoWreck：比对 AI 修改声明与代码实际结构，定性「幻觉」

A CLI tool that verifies AI coding assistant claims against actual structural changes — catching hallucinated functions, fake calls, and missed modifications before they ship.

Deterministic AI Verifier — v0.11.0

NoWreck is a deterministic structural verifier for AI-generated code-change claims. When an AI describes a code change, NoWreck compares the claims against structural evidence derived by its own scanners — the verifier never asks another AI for an opinion. Where the evidence comes from depends on the mode: in Pre/Post and Claims modes, from actual before/after repository snapshots; in Prompt Mode, from the model's own proposed diff.

$ nowreck fix "Add email validation to auth.py"

  Summary
  ────────────────────
  ● 3 claims total
  ● 2 confirmed
  ● 1 contradicted

  CONFIRMED
  ─────────
  ✓ ADD_FUNCTION validate_email → auth.py  (conf: 100%)
    Evidence: Function 'validate_email' was added in auth.py

  CONTRADICTED
  ────────────
  ✗ CALLS_FUNCTION validate_email → auth.py  (conf: 100%)
    Evidence: Function 'validate_email' was added in auth.py

Hallucinated functions or classes — a claim that something was added when it isn't there

Fake internal API calls — a claim that a function calls another function when the call doesn't exist

What's New In v0.11.0

NoWreck v0.11.0 — Multi-Provider Support (Anthropic + Gemini)

Release date: August 2026 Previous release: v0.10.0 (Independent Verification Architecture) Focus: Extend NoWreck beyond the OpenAI-compatible format to two major providers with different API shapes — Anthropic (Claude) and Google Gemini — via isolated provider adapters, with zero changes for existing OpenAI-compatible users.

What's new in v0.11.0

Provider adapters ✅

Non-OpenAI providers get adapters, not rewrites. The existing ModelProvider core logic (retries, parsing, failure saving) is untouched. Adapters translate between the provider's native format and the internal format:

ModelProvider
├── _detect_adapter(base_url) → adapter
├── OpenAIAdapter — passthrough (OpenAI, Groq, OpenRouter, Grok,
│   Kimi, Ollama, LM Studio, any compatible endpoint)
├── AnthropicAdapter — /v1/messages, x-api-key auth, top-level system
└── GeminiAdapter — /v1beta/models/{model}:generateContent,
    x-goog-api-key auth, systemInstruction

Anthropic

Endpoint: POST /v1/messages
Auth: x-api-key header (+ required anthropic-version header)
System messages extracted from the messages array into the top-level
system field
Response parsed from the content block array

nowreck config set base_url https://api.anthropic.com
nowreck config set api_key sk-ant-...
nowreck config set model claude-sonnet-4-20250514
nowreck fix "Add email validation to auth.py"

Gemini

Endpoint: POST /v1beta/models/{model}:generateContent (model in URL path)
Auth: x-goog-api-key header
Messages translated to contents with parts; system messages go to
systemInstruction; assistant role maps to model
Response parsed from candidates[0].content.parts[*].text

nowreck config set base_url https://generativelanguage.googleapis.com
nowreck config set api_key AIza...
nowreck config set model gemini-2.0-flash
nowreck fix "Add email validation to auth.py"

Auto-detection + explicit override ✅

The adapter is inferred from base_url (api.anthropic.com → Anthropic, generativelanguage.googleapis.com → Gemini, anything else → OpenAI). An optional provider config key overrides auto-detection when needed:

nowreck config set provider anthropic  # or: gemini, openai

No new CLI flags. The same commands work with every provider.

Phase 4 integration tests ✅

Six behavioral tests drive the real _default_http_call HTTP path (urlopen patched) and prove per-provider request shape, auth headers, and response round-trip through the correct adapter.