@poncho-ai/harness 0.50.1 → 0.50.3

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.50.1 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.50.3 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            530.75 KB
+ESM dist/index.js            530.79 KB
 ESM dist/isolate-BNQ6P3HI.js 51.41 KB
-ESM ⚡️ Build success in 224ms
+ESM ⚡️ Build success in 229ms
 DTS Build start
-DTS ⚡️ Build success in 6878ms
-DTS dist/index.d.ts 89.28 KB
+DTS ⚡️ Build success in 7430ms
+DTS dist/index.d.ts 89.60 KB

package/CHANGELOG.md

@@ -1,5 +1,36 @@
 # @poncho-ai/harness
 
+## 0.50.3
+
+### Patch Changes
+
+- [`a67fb45`](https://github.com/cesr/poncho-ai/commit/a67fb45162823d832296ae9af137eb566d9f2f97) Thanks [@cesr](https://github.com/cesr)! - harness: forward `tenantId` through `continueFromToolResult`. Resumed runs (after an approval checkpoint) ran tools with `ctx.tenantId` undefined, so tenant-scoped stores (memory, VFS, todos) resolved the default `"__default__"` tenant instead of the caller's — surfacing as `memory_main_get` returning empty after an approval resume.
+
+## 0.50.2
+
+### Patch Changes
+
+- [#133](https://github.com/cesr/poncho-ai/pull/133) [`60e21c8`](https://github.com/cesr/poncho-ai/commit/60e21c8c18054d2824cf7364b71dabfd07574b48) Thanks [@cesr](https://github.com/cesr)! - harness: don't inject the "interrupted by a time limit" bridge when resuming from tool results
+
+  A taskless `run()` (continuation checkpoint, or `continueFromToolResult` —
+  e.g. resuming after an approval gate) injected a synthetic user message
+  telling the model its turn was "interrupted by a time limit ... continue
+  EXACTLY from where you left off, do not re-summarize." That bridge is only
+  appropriate when the model was actually cut off mid-response (the last
+  message is an `assistant` turn, which some providers also reject as a
+  conversation-ending message).
+
+  When the last message is a `tool` result — which is the case for every
+  `continueFromToolResult` resume — the conversation already ends in a
+  provider-valid user `tool_result` block and the model continues from the
+  results naturally. Injecting the bridge there was a bug: after a normal
+  approval resolution the model was told its turn had been killed by a time
+  limit, causing it to distrust and re-derive context (re-reading the VFS,
+  concluding it had "hallucinated" data it had legitimately loaded). The
+  bridge is now only added when the last message is an `assistant` turn, and
+  its wording no longer hard-codes "time limit" (max-steps checkpoints use
+  the same path).
+
 ## 0.50.1
 
 ### Patch Changes

package/dist/index.d.ts

@@ -1434,6 +1434,11 @@ declare class AgentHarness {
             error?: string;
         }>;
         conversationId?: string;
+        /** Must be forwarded for the continuation run, otherwise tenant-scoped
+         *  tool stores (memory, VFS, todos) resolve the default "__default__"
+         *  tenant on resume instead of the caller's — e.g. memory_main_get
+         *  returns empty after an approval checkpoint. */
+        tenantId?: string;
         parameters?: Record<string, unknown>;
         abortSignal?: AbortSignal;
     }): AsyncGenerator<AgentEvent>;

package/dist/index.js

@@ -10441,10 +10441,10 @@ ${this.skillFingerprint}`;
         }
       } else {
         const lastMsg = messages[messages.length - 1];
-        if (lastMsg && lastMsg.role !== "user") {
+        if (lastMsg && lastMsg.role === "assistant") {
           messages.push({
             role: "user",
-            content: "[System: Your previous turn was interrupted by a time limit. Your partial response above is already visible to the user. Continue EXACTLY from where you left off \u2014 do NOT restart, re-summarize, or repeat any content you already produced. If you were mid-sentence or mid-table, continue that sentence or table. Proceed directly with the next action or output.]",
+            content: "[System: Your previous turn was interrupted before you finished. Your partial response above is already visible to the user. Continue EXACTLY from where you left off \u2014 do NOT restart, re-summarize, or repeat any content you already produced. If you were mid-sentence or mid-table, continue that sentence or table. Proceed directly with the next action or output.]",
             metadata: { timestamp: now(), id: randomUUID5() }
           });
         }
@@ -11568,6 +11568,7 @@ ${this.skillFingerprint}`;
     yield* this.runWithTelemetry({
       messages,
       conversationId: input.conversationId,
+      tenantId: input.tenantId,
       parameters: input.parameters,
       abortSignal: input.abortSignal
     });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.50.1",
+  "version": "0.50.3",
   "description": "Agent execution runtime - conversation loop, tool dispatch, streaming",
   "repository": {
     "type": "git",

package/src/harness.ts

@@ -2388,14 +2388,25 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
         });
       }
     } else {
-      // Continuation run (no explicit task). Some providers (Anthropic) require
-      // the conversation to end with a user message. Inject a transient signal
-      // that is sent to the LLM but never persisted to the conversation store.
+      // Taskless run: either a genuine continuation checkpoint (the model
+      // was cut off mid-response by a soft-deadline / max-steps boundary)
+      // or a tool-result resume — `continueFromToolResult`, e.g. after an
+      // approval gate — whose last message is the tool results.
+      //
+      // Only the first case needs a bridge: ending on an assistant message
+      // is invalid for some providers (Anthropic), and the model must be
+      // told to continue rather than restart. A `tool` last message converts
+      // to a user-role tool_result block, so the conversation already ends
+      // "with a user message" and the model continues from the results
+      // naturally. Injecting the bridge after tool results was a bug: it
+      // told the model — after a normal approval resolution — that its turn
+      // had been "interrupted by a time limit" and to not re-summarize,
+      // which made it distrust and re-derive context it already had.
       const lastMsg = messages[messages.length - 1];
-      if (lastMsg && lastMsg.role !== "user") {
+      if (lastMsg && lastMsg.role === "assistant") {
         messages.push({
           role: "user",
-          content: "[System: Your previous turn was interrupted by a time limit. Your partial response above is already visible to the user. Continue EXACTLY from where you left off — do NOT restart, re-summarize, or repeat any content you already produced. If you were mid-sentence or mid-table, continue that sentence or table. Proceed directly with the next action or output.]",
+          content: "[System: Your previous turn was interrupted before you finished. Your partial response above is already visible to the user. Continue EXACTLY from where you left off — do NOT restart, re-summarize, or repeat any content you already produced. If you were mid-sentence or mid-table, continue that sentence or table. Proceed directly with the next action or output.]",
           metadata: { timestamp: now(), id: randomUUID() },
         });
       }
@@ -3716,6 +3727,11 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
     messages: Message[];
     toolResults: Array<{ callId: string; toolName: string; result?: unknown; error?: string }>;
     conversationId?: string;
+    /** Must be forwarded for the continuation run, otherwise tenant-scoped
+     *  tool stores (memory, VFS, todos) resolve the default "__default__"
+     *  tenant on resume instead of the caller's — e.g. memory_main_get
+     *  returns empty after an approval checkpoint. */
+    tenantId?: string;
     parameters?: Record<string, unknown>;
     abortSignal?: AbortSignal;
   }): AsyncGenerator<AgentEvent> {
@@ -3773,6 +3789,7 @@ Code is wrapped in an async IIFE — use \`return\` to return a value to the too
     yield* this.runWithTelemetry({
       messages,
       conversationId: input.conversationId,
+      tenantId: input.tenantId,
       parameters: input.parameters,
       abortSignal: input.abortSignal,
     });