Repository navigation
Each layer's output ceiling comes from its own model
Two fixes.
Output ceilings and reasoning effort (#504)
Per-layer models let every layer run a different model — but the output ceiling was still derived from a shared default rather than from the model actually assigned to that layer. A layer pointed at a model with a smaller ceiling could be asked for more than it could return; one pointed at a larger model was capped below what it could do.
Each layer now sizes its ceiling from its own model, and reasoning effort is requested explicitly rather than left to the provider's default.
Touches the shared LLM libraries (anthropicwire, openaicompat, a new reasoning package) and every provider — Claude, Bedrock, Azure Foundry and LiteLLM.
Qdrant readiness in tests (#509)
wait.ForListeningPort("6334/tcp") proves only that the TCP port accepts a connection. Qdrant's listener accepts before the gRPC service is serving, so the first call raced it and died in transport:
rpc error: code = Unavailable desc = connection error:
desc = "error reading server preface: connection reset by peer"
The probe now waits for Qdrant to be serving. Test-only — two lines per file.
Images are cosign-signed; the tag is GPG-signed. See docs/verifying-artifacts.md.