@poncho-ai/harness 0.59.3 → 0.59.5

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.59.3 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.59.5 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
 > node scripts/embed-docs.js && tsup src/index.ts --format esm --dts
 
 [embed-docs] Generated poncho-docs.ts with 4 topics
@@ -8,9 +8,9 @@
 CLI tsup v8.5.1
 CLI Target: es2022
 ESM Build start
-ESM dist/index.js            556.92 KB
+ESM dist/index.js            557.00 KB
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM ⚡️ Build success in 268ms
+ESM ⚡️ Build success in 252ms
 DTS Build start
-DTS ⚡️ Build success in 8081ms
-DTS dist/index.d.ts 101.11 KB
+DTS ⚡️ Build success in 7698ms
+DTS dist/index.d.ts 101.66 KB

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # @poncho-ai/harness
 
+## 0.59.5
+
+### Patch Changes
+
+- [`d14c390`](https://github.com/cesr/poncho-ai/commit/d14c390ce6830f7446ea7a4e934d2cb76833c455) Thanks [@cesr](https://github.com/cesr)! - `continueFromToolResult` accepts and forwards the per-run `model` override, so approval-checkpoint continuations run on the same model as the checkpointed run instead of re-reading the (possibly concurrently-mutated) agent frontmatter.
+
+## 0.59.4
+
+### 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.
+
+- Updated dependencies [[`3a25676`](https://github.com/cesr/poncho-ai/commit/3a2567666e1bc8d6650818db76d07765c0250264)]:
+  - @poncho-ai/sdk@1.15.2
+
 ## 0.59.3
 
 ### Patch Changes

package/dist/index.d.ts

@@ -1582,6 +1582,11 @@ declare class AgentHarness {
         /** Emit no telemetry for the continuation run (e.g. resuming an
          *  incognito turn after an approval). */
         suppressTelemetry?: boolean;
+        /** Per-run model override for the continuation run — same semantics as
+         *  `RunInput.model`. Forward the model the checkpointed run was using,
+         *  otherwise the continuation falls back to the agent definition's
+         *  (possibly concurrently-mutated) frontmatter model. */
+        model?: string;
     }): AsyncGenerator<AgentEvent>;
     runToCompletion(input: RunInput): Promise<HarnessRunOutput>;
 }
@@ -2334,6 +2339,12 @@ interface RunConversationTurnOpts {
      * built with an OTLP exporter attached.
      */
     suppressTelemetry?: boolean;
+    /**
+     * Forwarded to `RunInput.model`. Per-run model override, captured once at
+     * run start — safe under concurrent runs on a shared harness, unlike
+     * mutating the parsed agent's frontmatter.
+     */
+    model?: string;
     /** Per-event hook — called for every AgentEvent yielded by the run, in order. */
     onEvent?: (event: AgentEvent) => void | Promise<void>;
 }

package/dist/index.js

@@ -10719,7 +10719,7 @@ ${this.skillFingerprint}`;
       }
       return pushEvent({ type: "run:cancelled", runId, messages: trimToValidPrefix(snapshot) });
     };
-    const resolvedModelName = agent.frontmatter.model?.name ?? "claude-opus-4-5";
+    const resolvedModelName = input.model ?? agent.frontmatter.model?.name ?? "claude-opus-4-5";
     const contextWindow = agent.frontmatter.model?.contextWindow ?? getModelContextWindow(resolvedModelName);
     yield pushEvent({
       type: "run:started",
@@ -11086,7 +11086,7 @@ ${textContent}` };
             }
             return [];
           };
-          const modelName = agent.frontmatter.model?.name ?? "claude-opus-4-5";
+          const modelName = input.model ?? agent.frontmatter.model?.name ?? "claude-opus-4-5";
           if (step === 1) {
             modelLog.item(`${modelName} (provider=${agent.frontmatter.model?.provider ?? "anthropic"})`);
           }
@@ -11921,7 +11921,8 @@ ${this.skillFingerprint}`;
       tenantId: input.tenantId,
       parameters: input.parameters,
       abortSignal: input.abortSignal,
-      suppressTelemetry: input.suppressTelemetry
+      suppressTelemetry: input.suppressTelemetry,
+      model: input.model
     });
   }
   async runToCompletion(input) {
@@ -14323,7 +14324,8 @@ var runConversationTurn = async (opts) => {
         files: opts.files && opts.files.length > 0 ? opts.files : void 0,
         abortSignal: opts.abortSignal,
         disablePromptCache: opts.disablePromptCache,
-        suppressTelemetry: opts.suppressTelemetry
+        suppressTelemetry: opts.suppressTelemetry,
+        model: opts.model
       },
       initialContextTokens: conversation.contextTokens ?? 0,
       initialContextWindow: conversation.contextWindow ?? 0,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.59.3",
+  "version": "0.59.5",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",
@@ -34,7 +34,7 @@
     "mustache": "^4.2.0",
     "yaml": "^2.4.0",
     "zod": "^3.22.0",
-    "@poncho-ai/sdk": "1.15.1"
+    "@poncho-ai/sdk": "1.15.2"
   },
   "peerDependencies": {
     "esbuild": ">=0.17.0",

package/src/harness.ts

@@ -2376,7 +2376,7 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
       return pushEvent({ type: "run:cancelled", runId, messages: trimToValidPrefix(snapshot) });
     };
 
-    const resolvedModelName = agent.frontmatter.model?.name ?? "claude-opus-4-5";
+    const resolvedModelName = input.model ?? agent.frontmatter.model?.name ?? "claude-opus-4-5";
     const contextWindow =
       agent.frontmatter.model?.contextWindow ?? getModelContextWindow(resolvedModelName);
 
@@ -2836,7 +2836,12 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
           return [];
         };
 
-        const modelName = agent.frontmatter.model?.name ?? "claude-opus-4-5";
+        // Per-run override wins over frontmatter. Reading frontmatter here is
+        // what made model selection racy: the field is re-read every step, so
+        // a concurrent run's setHarnessModel-style mutation flipped this
+        // run's model mid-turn (and a model switch drops the whole per-model
+        // Anthropic prompt cache).
+        const modelName = input.model ?? agent.frontmatter.model?.name ?? "claude-opus-4-5";
         if (step === 1) {
           modelLog.item(`${modelName} (provider=${agent.frontmatter.model?.provider ?? "anthropic"})`);
         }
@@ -3836,6 +3841,11 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
     /** Emit no telemetry for the continuation run (e.g. resuming an
      *  incognito turn after an approval). */
     suppressTelemetry?: boolean;
+    /** Per-run model override for the continuation run — same semantics as
+     *  `RunInput.model`. Forward the model the checkpointed run was using,
+     *  otherwise the continuation falls back to the agent definition's
+     *  (possibly concurrently-mutated) frontmatter model. */
+    model?: string;
   }): AsyncGenerator<AgentEvent> {
     const messages = [...input.messages];
     const lastMsg = messages[messages.length - 1];
@@ -3895,6 +3905,7 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
       parameters: input.parameters,
       abortSignal: input.abortSignal,
       suppressTelemetry: input.suppressTelemetry,
+      model: input.model,
     });
   }
 

package/src/orchestrator/run-conversation-turn.ts

@@ -74,6 +74,12 @@ export interface RunConversationTurnOpts {
    * built with an OTLP exporter attached.
    */
   suppressTelemetry?: boolean;
+  /**
+   * Forwarded to `RunInput.model`. Per-run model override, captured once at
+   * run start — safe under concurrent runs on a shared harness, unlike
+   * mutating the parsed agent's frontmatter.
+   */
+  model?: string;
   /** Per-event hook — called for every AgentEvent yielded by the run, in order. */
   onEvent?: (event: AgentEvent) => void | Promise<void>;
 }
@@ -230,6 +236,7 @@ export const runConversationTurn = async (
         abortSignal: opts.abortSignal,
         disablePromptCache: opts.disablePromptCache,
         suppressTelemetry: opts.suppressTelemetry,
+        model: opts.model,
       },
       initialContextTokens: conversation.contextTokens ?? 0,
       initialContextWindow: conversation.contextWindow ?? 0,