@poncho-ai/sdk 1.15.0 → 1.15.2

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/sdk@1.15.0 build /home/runner/work/poncho-ai/poncho-ai/packages/sdk
+> @poncho-ai/sdk@1.15.2 build /home/runner/work/poncho-ai/poncho-ai/packages/sdk
 > tsup src/index.ts --format esm --dts
 
 CLI Building entry: src/index.ts
@@ -8,7 +8,7 @@
 CLI Target: es2022
 ESM Build start
 ESM dist/index.js 17.24 KB
-ESM ⚡️ Build success in 19ms
+ESM ⚡️ Build success in 21ms
 DTS Build start
-DTS ⚡️ Build success in 1324ms
-DTS dist/index.d.ts 31.27 KB
+DTS ⚡️ Build success in 1365ms
+DTS dist/index.d.ts 32.36 KB

package/CHANGELOG.md

@@ -1,5 +1,30 @@
 # @poncho-ai/sdk
 
+## 1.15.2
+
+### Patch Changes
+
+- [`3a25676`](https://github.com/cesr/poncho-ai/commit/3a2567666e1bc8d6650818db76d07765c0250264) Thanks [@cesr](https://github.com/cesr)! - Add a per-run `model` override to `RunInput` (and forward it from `runConversationTurn`'s opts). The override is captured once at run start and wins over the agent definition's `model.name` for every step of the run.
+
+  Previously the only way to vary the model per run kind on a shared harness was mutating `parsedAgent.frontmatter.model.name` before each run — but the harness re-reads that field at the start of every step, so a concurrent run's mutation flipped an in-flight run's model mid-turn. Besides being the wrong model, the switch invalidated the run's entire Anthropic prompt cache (caches are per-model), observed in production as the same ~104k-token prefix being cache-written twice back-to-back, once per model. Callers should pass `model` in the run input instead of mutating frontmatter.
+
+## 1.15.1
+
+### Patch Changes
+
+- [`299f574`](https://github.com/cesr/poncho-ai/commit/299f574a2f2f0d4873f42bbcffdf604e9cc4c29c) Thanks [@cesr](https://github.com/cesr)! - Mark in-flight assistant drafts with `metadata.incomplete = true`.
+
+  The orchestrator's per-step draft persist (`persistDraft`) and the
+  approval/device checkpoint and continuation writes now stamp the trailing
+  assistant message `metadata.incomplete = true`; the three terminal writes
+  (normal finalize, cancelled, errored) clear it. This lets a consumer that
+  reconciles a persisted snapshot against a live event stream (e.g. a
+  WebSocket layer) strip the in-flight draft from the authoritative snapshot
+  and rebuild that turn from the event log instead — so the snapshot and the
+  replayed events never both carry the in-flight turn, eliminating
+  reconnect-time duplication. Additive + backwards-compatible: consumers that
+  ignore the flag are unaffected.
+
 ## 1.15.0
 
 ### Minor Changes

package/dist/index.d.ts

@@ -680,6 +680,14 @@ interface Message {
             content: string | string[];
         }>;
         isCompactionSummary?: boolean;
+        /** True while this assistant message is an in-flight DRAFT (the turn
+         *  hasn't finished). Set by the orchestrator's per-step draft persist and
+         *  cleared at finalize. Consumers that reconcile a persisted snapshot with
+         *  a live event stream (e.g. PonchOS's WS layer) strip `incomplete`
+         *  messages from the snapshot and rebuild the in-flight turn from the
+         *  event log instead — so the two never both carry it (no reconnect
+         *  duplication). */
+        incomplete?: boolean;
     };
 }
 /** Extract the text content from a message, regardless of content format. */
@@ -763,6 +771,17 @@ interface RunInput {
      * will hit the cache before the 5-min TTL expires.
      */
     disablePromptCache?: boolean;
+    /**
+     * Model name override for this run, captured once at run start. Takes
+     * precedence over the agent definition's `model.name` for every step of
+     * the run. Use this instead of mutating the parsed agent's frontmatter
+     * when one harness instance serves runs that need different models
+     * (e.g. user turns vs cron jobs) — a frontmatter mutation made while
+     * another run is in flight changes that run's model mid-turn, and the
+     * model switch invalidates its entire Anthropic prompt cache (caches
+     * are per-model).
+     */
+    model?: string;
     /** Scope this run to a specific tenant. */
     tenantId?: string;
     /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/sdk",
-  "version": "1.15.0",
+  "version": "1.15.2",
   "description": "Core types and utilities for building Poncho skills",
   "repository": {
     "type": "git",

package/src/index.ts

@@ -38,6 +38,14 @@ export interface Message {
     toolActivity?: string[];
     sections?: Array<{ type: "text" | "tools"; content: string | string[] }>;
     isCompactionSummary?: boolean;
+    /** True while this assistant message is an in-flight DRAFT (the turn
+     *  hasn't finished). Set by the orchestrator's per-step draft persist and
+     *  cleared at finalize. Consumers that reconcile a persisted snapshot with
+     *  a live event stream (e.g. PonchOS's WS layer) strip `incomplete`
+     *  messages from the snapshot and rebuild the in-flight turn from the
+     *  event log instead — so the two never both carry it (no reconnect
+     *  duplication). */
+    incomplete?: boolean;
   };
 }
 
@@ -144,6 +152,17 @@ export interface RunInput {
    * will hit the cache before the 5-min TTL expires.
    */
   disablePromptCache?: boolean;
+  /**
+   * Model name override for this run, captured once at run start. Takes
+   * precedence over the agent definition's `model.name` for every step of
+   * the run. Use this instead of mutating the parsed agent's frontmatter
+   * when one harness instance serves runs that need different models
+   * (e.g. user turns vs cron jobs) — a frontmatter mutation made while
+   * another run is in flight changes that run's model mid-turn, and the
+   * model switch invalidates its entire Anthropic prompt cache (caches
+   * are per-model).
+   */
+  model?: string;
   /** Scope this run to a specific tenant. */
   tenantId?: string;
   /**