cinch ai

cinch ai is the explicit AI boundary in Cinch. It does not run during copy, send, pull, desktop sync, or cinch mcp.

v1 has one workflow:

cat error.log | cinch ai fix
cinch ai fix latest
cinch ai fix 01HXABCD

Use --no-send to generate the prompt only:

cargo test 2>&1 | cinch ai fix --no-send
cinch ai fix latest --no-send

Input priority:

1. stdin, when piped
2. latest, from the local store
3. a clip ID prefix, from the local store

--copy copies the prompt or provider answer back to your system clipboard.

Provider boundary

If no provider is configured, cinch ai fix behaves like --no-send: it prints the AI-ready prompt and does not call a provider.

Supported provider names:

Provider	Use case
hosted-bedrock	Cinch-operated managed provider for hosted distributions. Requires operator endpoint config.
bedrock-byok	BYOK AWS Bedrock boundary for self-hosters and developers. The v1 CLI documents the boundary; AWS request signing is not wired into this binary yet.
openai-compatible	Ollama, LM Studio, llama.cpp servers, or OpenAI-compatible APIs.

OpenAI-compatible local example:

export CINCH_AI_PROVIDER=openai-compatible
export CINCH_AI_BASE_URL=http://localhost:11434/v1
export CINCH_AI_MODEL=llama3.1

cargo test 2>&1 | cinch ai fix

Hosted provider example:

export CINCH_AI_PROVIDER=hosted-bedrock
export CINCH_AI_HOSTED_URL=https://api.example.com/ai/fix
export CINCH_AI_HOSTED_TOKEN=...

cinch ai fix latest

Privacy model

• AI provider calls happen only under cinch ai.
• --no-send does not call an AI provider.
• cinch mcp is read-only, and local-only by default; it contacts the relay only for fleet-scoped reads (CINCH_MCP_FLEET=1).
• cinch copy, cinch send, cinch pull, and desktop sync do not call AI providers.
• Provider errors print the assembled prompt so the context is not lost.