Imported from eerwitt/prout (AGENTS.md). Install upstream with npx skills add eerwitt/prout. Copyright stays with the author.

AGENTS.md

Agent guidance for this repository lives in CLAUDE.md. Read it before making any change — it covers the repo layout, the two-step build (Bazel for the vendored LiteRT-LM C ABI, CMake for Prout), coding conventions, and verified LiteRT-LM facts you should not re-derive.

Non-negotiable security invariants

Even if you read nothing else:

1. Credentials never leave locked memory except inside an authorized request — never into stdout/stderr, logs, audit records, error messages, or agent-visible responses.
2. The audit log is append-only. Never rewrite, reorder, or truncate existing records; each machine writes only to its own log file.
3. The model recommends; code enforces. No arbiter verdict authorizes credential decryption until schema-validated against the service policy.
4. Never modify third_party/litert-lm/ — it is a vendored submodule.