v0.8.13

Added

• Two opt-in, provider-agnostic reliability options for generate that make local-model tool loops survivable. Both default to off, so existing behavior is unchanged.
  • tool_call_repair_attempts (default 0): when a step returns text with no structured tool calls, salvage a call the model emitted as assistant text in Hermes (<tool_call>{...}</tool_call>), bare/json-fenced, or Qwen3-Coder XML form. A candidate runs only if its name resolves and its arguments validate against the tool's input_schema, so plain JSON-in-prose never false-positives. Salvaged calls run the normal execute path and are marked salvaged (with salvaged_format) on their ToolCallRecord, plus a tool_call_salvaged trajectory event. The budget is shared across the whole call, including schema-repair passes.
  • max_tool_calls_per_step (default unlimited, must be >= 1): execute only the first N tool calls of a step and drop the rest for that turn (the model can re-issue them next turn). Dropped calls surface as ToolCallRecords with error.message: 'dropped_max_tool_calls_per_step' and a tool_calls_dropped event. Set to 1 for runtimes that mishandle parallel tool calls.
• Both options are threaded through EngineDefaults (same per-call-wins merge rule) and model_call. The subprocess claude_cli provider records option_ignored for each, since it does not run the shared tool loop.