AI: BYOK + local, and an agent built for weak models

Status: accepted

Decision

1. BYOK first. Providers at launch: Anthropic, OpenRouter, OpenAI, Gemini, plus
   any OpenAI-compatible endpoint (which covers Ollama/LM Studio/koboldcpp for local).
   No Vaudeville-operated inference, no account, no proxy server. Keys in the local
   vault (OS keychain, else encrypted file).
2. AI is optional. The Converter, inspection, validation, deterministic Doctor
   checks, and all project management work with zero keys. The first key ask happens
   at the first AI moment, never at install.
3. The agent must survive weak models. Design constraints inherited from
   RoleCall's Orison program (proven in production there):
   • Tiny always-on tool surface + deferred tool schemas revealed on demand.
   • Resource verbs (ls/read/write/search over vaud:// URIs) instead of dozens
     of noun-tools.
   • Prompted-XML tool-calling fallback for models without native tool use; one tool
     per iteration in that mode.
   • Spill store: big tool results go to disk handles, never into context.
   • Explicit samplers on every request; max_tokens leaves room for tool args.
   • Staged edits with validate-before-commit; the agent never writes user files
     directly.
4. Model roles are user-mappable. Interview, treatment rewrites, test-stage
   inference, and bulk audits can each point at a different configured model
   (e.g. cheap local for audits, frontier for rewrites). Sane single-model default.
5. Eval harness at M2, before feature work on the agent. Golden tasks +
   replayable transcripts + a model matrix; agent changes must move
   success-rate/tokens-per-task numbers.

Why

• Local-first + BYOK is the only stance consistent with the privacy pitch and with
  an audience that already owns keys and local rigs.
• The weak-model constraints aren't charity: they make frontier-model behavior
  cheaper and more reliable too (Orison's measured lesson).