[gsmachado]

Summary

Hermes could not use tools on E2EE models. Venice already emitted the call: a short preamble in delta.content, then delta.tool_calls, then finish_reason: "tool_calls". This proxy kept only the text and closed the stream with finish_reason: "stop". Hermes stored that cut-off message and never ran a tool. The follow-up was then rejected, because an assistant tool-call message is content: null and ChatMessageDto required content. toPlainMessages also stripped tool_calls, tool_call_id, and name before encryption.

After this change:

• Streaming responses emit delta.tool_calls chunks in upstream order. The final chunk uses the real finish reason (tool_calls, stop, or length).
• Non-streaming responses assemble those deltas by index onto message.tool_calls. A tool-only turn has content: null.
• Follow-up messages may set content: null and include tool_calls, tool_call_id, and name.
• Tool-result bodies are passed through to encrypt, so they stay ciphertext. Tool names, JSON schemas, call ids, and arguments stay plaintext. That is the Venice E2EE rule: the enclave chat template reads those fields as JSON.

Library pin

@axlabs/venice-e2ee is pinned to the merged library commit 7b4e6528f7f870402224bc887de06f0952f7f575 (AxLabs/venice-e2ee#1) via git+https, not the registry. 0.2.0 still drops tool calls, and 0.3.0 is not on npm. This environment has no npm credentials, and the library CI workflow does not publish.

Follow-up, from a clean library main:

npm run release -- minor   # 0.2.0 -> 0.3.0, publishes @axlabs/venice-e2ee

Then in this repo, replace the git pin with "@axlabs/venice-e2ee": "^0.3.0", refresh the lockfile, and release the proxy (npm run release -- patch). decryptStream now yields structured events instead of strings, so do not point this proxy back at 0.2.0.

Tests

• npm test — 34 passed
• npm run lint — clean
• npm run build — nest build succeeded
• npm ci — installs the pinned library commit over HTTPS