@poncho-ai/harness 0.59.4 → 0.59.6

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.59.4 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.59.6 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            556.98 KB
 ESM dist/isolate-F2PPSUL6.js 53.82 KB
-ESM ⚡️ Build success in 270ms
+ESM dist/index.js            557.73 KB
+ESM ⚡️ Build success in 256ms
 DTS Build start
-DTS ⚡️ Build success in 8326ms
-DTS dist/index.d.ts 101.34 KB
+DTS ⚡️ Build success in 8276ms
+DTS dist/index.d.ts 101.66 KB

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # @poncho-ai/harness
 
+## 0.59.6
+
+### Patch Changes
+
+- [`e573f72`](https://github.com/cesr/poncho-ai/commit/e573f72ca31627e48dbdbf296946a72c59a488db) Thanks [@cesr](https://github.com/cesr)! - Preserve the LLM transcript when a turn dies. The errored branch of
+  runConversationTurn persisted only the display draft — `_harnessMessages`
+  was never updated, so the model's next turn had no memory of the entire
+  failed interaction (its user message included), even though the user could
+  see it on screen. Both the errored branch and the cancelled-without-
+  `run:cancelled.messages` fallback now append a faithful plain-text
+  reconstruction (user message + assistant text-so-far + tool activity + an
+  interruption note) to the transcript. Plain text on purpose: replaying real
+  tool_use blocks would need paired results or the next API call rejects the
+  dangling pair.
+
+## 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

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>;
 }

package/dist/index.js

@@ -11921,7 +11921,8 @@ ${this.skillFingerprint}`;
       tenantId: input.tenantId,
       parameters: input.parameters,
       abortSignal: input.abortSignal,
-      suppressTelemetry: input.suppressTelemetry
+      suppressTelemetry: input.suppressTelemetry,
+      model: input.model
     });
   }
   async runToCompletion(input) {
@@ -14497,6 +14498,20 @@ var runConversationTurn = async (opts) => {
     };
   } catch (error) {
     flushTurnDraft(draft);
+    const reconstructTranscriptTail = (reason) => {
+      const parts = [];
+      if (draft.assistantResponse.length > 0) parts.push(draft.assistantResponse);
+      if (draft.toolTimeline.length > 0) {
+        parts.push(`Tool activity before interruption:
+${draft.toolTimeline.join("\n")}`);
+      }
+      parts.push(`[This turn was interrupted: ${reason}. The work above may be incomplete.]`);
+      return [
+        ...conversation._harnessMessages ?? [],
+        userMessage,
+        { role: "assistant", content: parts.join("\n\n") }
+      ];
+    };
     const aborted = opts.abortSignal?.aborted === true;
     if (aborted || runCancelled) {
       if (draft.assistantResponse.length > 0 || draft.toolTimeline.length > 0 || draft.sections.length > 0) {
@@ -14507,7 +14522,7 @@ var runConversationTurn = async (opts) => {
             latestRunId,
             contextTokens: 0,
             contextWindow: 0,
-            harnessMessages: cancelHarnessMessages,
+            harnessMessages: cancelHarnessMessages ?? reconstructTranscriptTail("cancelled"),
             toolResultArchive: opts.harness.getToolResultArchive(opts.conversationId)
           },
           { shouldRebuildCanonical: true }
@@ -14550,6 +14565,9 @@ var runConversationTurn = async (opts) => {
     }
     if (draft.assistantResponse.length > 0 || draft.toolTimeline.length > 0 || draft.sections.length > 0) {
       conversation.messages = buildMessages(false);
+      conversation._harnessMessages = reconstructTranscriptTail(
+        error instanceof Error ? `error \u2014 ${error.message}` : "error"
+      );
       conversation.updatedAt = Date.now();
       await opts.conversationStore.update(conversation);
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.59.4",
+  "version": "0.59.6",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/harness.ts

@@ -3841,6 +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];
@@ -3900,6 +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

@@ -420,6 +420,31 @@ export const runConversationTurn = async (
     };
   } catch (error) {
     flushTurnDraft(draft);
+
+    // The LLM transcript (`_harnessMessages`) is normally only written at a
+    // clean finalize / a cancel that delivered `run:cancelled.messages`. A
+    // turn that dies any other way (in-process error, abort that never
+    // surfaced the cancel event) would leave the transcript WITHOUT this
+    // turn at all — the display shows the partial work but the model has
+    // amnesia about the whole interaction on the next turn. Reconstruct a
+    // faithful plain-text record from the draft instead: the user message
+    // plus an assistant message carrying the text-so-far + tool activity.
+    // Plain text on purpose — replaying real tool_use blocks would need
+    // paired results or the next API call 400s on the dangling pair.
+    const reconstructTranscriptTail = (reason: string): Message[] => {
+      const parts: string[] = [];
+      if (draft.assistantResponse.length > 0) parts.push(draft.assistantResponse);
+      if (draft.toolTimeline.length > 0) {
+        parts.push(`Tool activity before interruption:\n${draft.toolTimeline.join("\n")}`);
+      }
+      parts.push(`[This turn was interrupted: ${reason}. The work above may be incomplete.]`);
+      return [
+        ...(conversation._harnessMessages ?? []),
+        userMessage,
+        { role: "assistant" as const, content: parts.join("\n\n") },
+      ];
+    };
+
     const aborted = opts.abortSignal?.aborted === true;
     if (aborted || runCancelled) {
       if (
@@ -434,7 +459,8 @@ export const runConversationTurn = async (
             latestRunId,
             contextTokens: 0,
             contextWindow: 0,
-            harnessMessages: cancelHarnessMessages,
+            harnessMessages:
+              cancelHarnessMessages ?? reconstructTranscriptTail("cancelled"),
             toolResultArchive: opts.harness.getToolResultArchive(opts.conversationId),
           },
           { shouldRebuildCanonical: true },
@@ -484,6 +510,12 @@ export const runConversationTurn = async (
       draft.sections.length > 0
     ) {
       conversation.messages = buildMessages(false); // terminal: errored
+      // Keep the LLM transcript faithful too (see reconstructTranscriptTail
+      // above) — without this, the next turn's model context skipped the
+      // whole errored interaction.
+      conversation._harnessMessages = reconstructTranscriptTail(
+        error instanceof Error ? `error — ${error.message}` : "error",
+      );
       conversation.updatedAt = Date.now();
       await opts.conversationStore.update(conversation);
     }