@poncho-ai/harness 0.50.0 → 0.50.2

package/.turbo/turbo-build.log

@@ -1,5 +1,5 @@
 
-> @poncho-ai/harness@0.50.0 build /home/runner/work/poncho-ai/poncho-ai/packages/harness
+> @poncho-ai/harness@0.50.2 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
@@ -9,8 +9,8 @@
 CLI Target: es2022
 ESM Build start
 ESM dist/isolate-BNQ6P3HI.js 51.41 KB
-ESM dist/index.js            530.56 KB
-ESM ⚡️ Build success in 259ms
+ESM dist/index.js            530.76 KB
+ESM ⚡️ Build success in 181ms
 DTS Build start
-DTS ⚡️ Build success in 6896ms
+DTS ⚡️ Build success in 5658ms
 DTS dist/index.d.ts 89.28 KB

package/CHANGELOG.md

@@ -1,5 +1,55 @@
 # @poncho-ai/harness
 
+## 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
+
+- [#131](https://github.com/cesr/poncho-ai/pull/131) [`9e0ccd8`](https://github.com/cesr/poncho-ai/commit/9e0ccd8307ab056f94b7efc3e25a0cb71ee75625) Thanks [@cesr](https://github.com/cesr)! - harness: strip `poncho-upload://` scheme in `S3UploadStore.get` / `.delete`
+
+  `createUploadStore({ provider: "s3" })` wraps `S3UploadStore` in
+  `CachedUploadStore`, whose `put` returns a `poncho-upload://<key>` ref
+  and stores the underlying S3 object at the bare `<key>`. On read,
+  `CachedUploadStore.get` checks an in-memory cache (10-minute TTL); on
+  miss it falls through to `S3UploadStore.get(<ref>)`. Pre-fix, the S3
+  store treated the scheme-prefixed ref as a literal S3 key and hit the
+  backend with `poncho-upload://<key>` — guaranteed `NoSuchKey`.
+
+  In practice this meant a chat message with an attached image worked on
+  the turn it was uploaded (cache hit) and then started showing as
+  "[Attached file: … — file is no longer available]" on every follow-up
+  turn ~10 minutes later (cache miss → S3 NoSuchKey → outer catch in the
+  harness resolver). The same path worked for the local-fs store, which
+  strips the scheme in both `get` and `delete`.
+
+  `S3UploadStore.get` now strips the scheme before issuing
+  `GetObjectCommand`. `S3UploadStore.delete` already stripped `https://`
+  and now strips `poncho-upload://` too.
+
 ## 0.50.0
 
 ### Minor Changes

package/dist/index.js

@@ -2612,16 +2612,17 @@ var S3UploadStore = class {
       }
       return Buffer.from(await response.arrayBuffer());
     }
+    const key = urlOrKey.startsWith(PONCHO_UPLOAD_SCHEME) ? urlOrKey.slice(PONCHO_UPLOAD_SCHEME.length) : urlOrKey;
     await this.ensureClient();
     const result = await this.client.send(
-      new this.s3Sdk.GetObjectCommand({ Bucket: this.bucket, Key: urlOrKey })
+      new this.s3Sdk.GetObjectCommand({ Bucket: this.bucket, Key: key })
     );
-    if (!result.Body) throw new Error(`uploads: empty body for S3 key ${urlOrKey}`);
+    if (!result.Body) throw new Error(`uploads: empty body for S3 key ${key}`);
     return Buffer.from(await result.Body.transformToByteArray());
   }
   async delete(urlOrKey) {
     await this.ensureClient();
-    const key = urlOrKey.startsWith("https://") ? new URL(urlOrKey).pathname.slice(1) : urlOrKey;
+    const key = urlOrKey.startsWith("https://") ? new URL(urlOrKey).pathname.slice(1) : urlOrKey.startsWith(PONCHO_UPLOAD_SCHEME) ? urlOrKey.slice(PONCHO_UPLOAD_SCHEME.length) : urlOrKey;
     await this.client.send(
       new this.s3Sdk.DeleteObjectCommand({ Bucket: this.bucket, Key: key })
     );
@@ -10440,10 +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() }
           });
         }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@poncho-ai/harness",
-  "version": "0.50.0",
+  "version": "0.50.2",
   "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() },
         });
       }

package/src/upload-store.ts

@@ -349,11 +349,19 @@ export class S3UploadStore implements UploadStore {
       }
       return Buffer.from(await response.arrayBuffer());
     }
+    // Strip the `poncho-upload://` scheme — LocalUploadStore does this for
+    // both get/delete but S3 used to treat the full URI as a literal S3
+    // key, so resolving a persisted file part (data: "poncho-upload://…")
+    // on a follow-up turn 404'd against R2 and the harness emitted a
+    // "file is no longer available" placeholder.
+    const key = urlOrKey.startsWith(PONCHO_UPLOAD_SCHEME)
+      ? urlOrKey.slice(PONCHO_UPLOAD_SCHEME.length)
+      : urlOrKey;
     await this.ensureClient();
     const result = await this.client.send(
-      new this.s3Sdk.GetObjectCommand({ Bucket: this.bucket, Key: urlOrKey }),
+      new this.s3Sdk.GetObjectCommand({ Bucket: this.bucket, Key: key }),
     );
-    if (!result.Body) throw new Error(`uploads: empty body for S3 key ${urlOrKey}`);
+    if (!result.Body) throw new Error(`uploads: empty body for S3 key ${key}`);
     return Buffer.from(await result.Body.transformToByteArray());
   }
 
@@ -361,7 +369,9 @@ export class S3UploadStore implements UploadStore {
     await this.ensureClient();
     const key = urlOrKey.startsWith("https://")
       ? new URL(urlOrKey).pathname.slice(1)
-      : urlOrKey;
+      : urlOrKey.startsWith(PONCHO_UPLOAD_SCHEME)
+        ? urlOrKey.slice(PONCHO_UPLOAD_SCHEME.length)
+        : urlOrKey;
     await this.client.send(
       new this.s3Sdk.DeleteObjectCommand({ Bucket: this.bucket, Key: key }),
     );