car-runtime 0.24.0 → 0.24.1

package/index.d.ts

@@ -160,8 +160,14 @@ export class CarRuntime {
   /**
    * Set replan configuration on this runtime.
    * `maxReplans` = 0 disables replanning (default).
+   * `replanOnRejected` (default false): when true, validator/policy/capability
+   * rejections (not just runtime failures) also trigger rollback + replan.
    */
-  setReplanConfig(maxReplans: number, delayMs?: number | null): Promise<void>;
+  setReplanConfig(
+    maxReplans: number,
+    delayMs?: number | null,
+    replanOnRejected?: boolean | null,
+  ): Promise<void>;
 
   // --- State ---
 
@@ -326,9 +332,17 @@ export class CarRuntime {
   /**
    * Generate with full tracking. Returns JSON with `text`, `tool_calls`,
    * `usage`, `model_used`, `latency_ms`, `time_to_first_token_ms`,
-   * `trace_id`. `time_to_first_token_ms` is wall-clock to the first
-   * sampled token (populated by local Candle/MLX paths; `null` for
-   * non-streaming remote calls).
+   * `trace_id`, `stop_reason`. `time_to_first_token_ms` is wall-clock to
+   * the first sampled token (populated by local Candle/MLX paths; `null`
+   * for non-streaming remote calls). `stop_reason` is the raw provider
+   * termination reason (OpenAI `finish_reason`, Anthropic `stop_reason`,
+   * Google `finishReason`); `null` for local backends or providers that
+   * don't report one. A value of `"length"`/`"max_tokens"`/`"MAX_TOKENS"`
+   * means the output was truncated at the token cap. On local Qwen3
+   * hybrid-thinking models it is also set to `"thinking_recovered"` when
+   * reasoning consumed the whole token budget and the runtime retried
+   * with reasoning suppressed to produce a direct answer, or
+   * `"thinking_truncated"` when even that retry was empty (car-releases#60).
    *
    * **Note:** intent is not exposed on the tracked path until the
    * positional argument list is converted to an options object —
@@ -481,7 +495,10 @@ export class CarRuntime {
   /**
    * Unified registry (local + remote). Returns JSON array of
    * `{ id, name, provider, capabilities, param_count, size_mb,
-   *   context_length, available, is_local, public_benchmarks }`.
+   *   context_length, available, is_local, max_output_tokens,
+   *   public_benchmarks }`. `max_output_tokens` is the registry-declared
+   *   per-model output ceiling (`null` when the entry omits it; callers
+   *   then fall back to a fraction of `context_length`).
    * `public_benchmarks` is `[{ name, score, harness?, source_url?,
    * measured_at? }]` with score on a 0.0–1.0 scale; ships empty in
    * the built-in catalog and is populated via curated registry data.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "car-runtime",
-  "version": "0.24.0",
+  "version": "0.24.1",
   "description": "Common Agent Runtime — a deterministic execution layer for AI agents",
   "main": "index.js",
   "types": "index.d.ts",