v1.6.0 - True Streaming with Tools and Model Registry Refresh

Changed

• CLI built-in tools are now disabled by default (--tools= is appended to the CLI invocation). The upstream CLI is a full agent: with its tools enabled, a plain chat request could turn into minutes of multi-turn file/shell loops inside the proxy's own working directory — easily outliving QODERCN_TIMEOUT_MS and surfacing in agent clients (e.g. Claude Code) as empty responses ("No response requested.") or upstream_timeout errors. Measured latency for the same model/prompt dropped from ~220s to ~3s. Set QODERCN_CLI_TOOLS=1 to keep the CLI's built-in tools.

Fixed

• Requests could hang forever when the CLI exited but its stdio pipes stayed open (e.g. a descendant process inheriting stdout): close never fires, so the request promise never settled and even the timeout kill could not end it. Both CLI runners now settle from the exit event after a short flush grace period, and force-settle shortly after a timeout kill.
• Duplicated text in longer streaming replies: the CLI's stream-json assistant events carry cumulative snapshots of the growing message (same message id, text blocks growing from the start — verified against qodercli 1.1.41), not incremental deltas. Forwarding them directly replayed earlier text whenever a reply spanned multiple snapshots. A snapshot tracker now aligns text blocks per message and emits only newly grown suffixes.
• .env.example suggested QODERCN_TIMEOUT_MS=120000, below the 300s code default and far below what queued, tool-heavy agent requests need; it now suggests 600000.

Added

• True streaming for tool-declared requests: stream: true requests that carry tools no longer wait for the fully buffered reply. The text part streams live through a safety gate (findStreamingSafePrefixLength) that withholds any tail which could still be or grow into the tool-call JSON payload — unclosed or closed ```json fences, unbalanced or "tool_calls"-bearing JSON objects, trailing partial backticks. Once the stream completes, the withheld tail is parsed and tool calls are appended as structured delta.tool_calls chunks (OpenAI) or tool_use blocks with input_json_delta (Anthropic). Agent clients like Claude Code now show live progress and receive tool confirmations in real time, while structured tool calling is unchanged. Skipped when SERVER_TOOL_EXECUTION=1 — the multi-round server-side tool loop stays buffered.
• Model registry update (matching qodercli --list-models, qodercli 1.1.41):
  • New routing tiers: ultimate (Ultimate), performance (Performance), efficient (Efficient), lite (Lite), cantus (Cantus).
  • New models: qwen3.8-flash (Qwen3.8-Flash), kimi-k3 (Kimi-K3), glm-5.3-flash (GLM-5.3-Flash).
  • Renamed to their current CLI offerings: qwen3.8-max-preview → qwen3.8-max (Qwen3.8-Max, effort aliases renamed accordingly), glm-5.2 → glm-5.3 (GLM-5.3), minimax-m2.7 → minimax-m3 (MiniMax-M3).
  • Removed qwen3.6-flash, which is no longer offered by the CLI. Requests for removed model IDs fall back to auto.
  • Synchronized model keys in opencode.json and both Chinese/English READMEs.