Repository navigation
Choose the respondent model and the pre-authorized spend #120
Description
Question
Which model answers the survey when the user says "run it", and what spend is pre-authorized?
Candidates: Claude Haiku 4.5 on Bedrock through the credential the agent-dev stack already uses (no new vendor, but a bearer token the wrapper does not hold yet, and no boto3 in its venv); OpenAI or Gemini keys from Doppler beca (Plan #109 already names them for small pilots); a direct Anthropic key (Plan #109 rules it out). Decide the default model, the per-run spend the agent may incur without asking (recommendation: run without asking when estimate_cost is under one dollar, ask above), and whether the free test model is still run first as a mechanics check or skipped for the user's run. Use grilling and domain-modeling.
Activity
beca-oc commentedon Nov 2, 2026
Resolution (owner, 2026-11-01)
us.anthropic.claude-haiku-4-5-20251001-v1:0, us-east-1, the provider the agent-dev Onyx stack already uses. No new vendor and no direct Anthropic API key.estimate_costis under one dollar runs without asking; above that, Beca shows the estimate and asks. The owner approved the model with this recommended limit attached.Unblocks #123. #127 still waits on #121 for what the wrapper needs to call Bedrock.