pi-vision-handoff 0.9.1 → 0.9.3

package/README.md

@@ -255,7 +255,7 @@ Because pi runs parallel `read` tool calls via `Promise.all` and fires each read
 | `context` (all messages) | `{ type: "input_image", image_url: "data:…" }` (OpenAI Responses) | detected by shape → replaced with `{ type: "input_text", text }` |
 | `context` (all messages) | `{ type: "image", source: { type: "base64", media_type, data } }` (Anthropic Messages) | detected by shape → replaced with `{ type: "text", text }` |
 
-The describer call itself goes through pi's normal model machinery (`completeSimple()`), **not** the agent event loop — so it never re-triggers `context` (no recursion). The `read` tool result keeps its image block untouched (kitty inline + `/resume`); only the `context`-cloned LLM-bound payload has images swapped for text.
+The describer call itself goes through pi's normal model machinery (`completeSimple()`), **not** the agent event loop — so it never re-triggers `context` (no recursion). Every real describer request receives a dedicated helper session ID. For Neuralwatt, the request also receives a matching isolated `X-NW-Conversation-ID` instead of inheriting the main Pi session ID; the unrelated image prompt therefore cannot replace the main agent's server-side cache lineage. The `read` tool result keeps its image block untouched (kitty inline + `/resume`); only the `context`-cloned LLM-bound payload has images swapped for text.
 
 ### Usage reporting
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-vision-handoff",
-  "version": "0.9.1",
+  "version": "0.9.3",
   "description": "Give text-only pi models vision — describe images with a vision model you pick via an interactive picker, then hand off the text description to non-vision models",
   "type": "module",
   "author": "Tom X Nguyen",
@@ -31,9 +31,9 @@
     "README.md"
   ],
   "devDependencies": {
-    "@earendil-works/pi-ai": "0.80.10",
-    "@earendil-works/pi-coding-agent": "0.80.10",
-    "@earendil-works/pi-tui": "0.80.10",
+    "@earendil-works/pi-ai": "0.83.0",
+    "@earendil-works/pi-coding-agent": "0.83.0",
+    "@earendil-works/pi-tui": "0.83.0",
     "@types/node": "25.9.1",
     "@vitest/coverage-v8": "4.1.7",
     "knip": "6.14.1",

package/src/describer.ts

@@ -60,6 +60,7 @@ import {
   type VisionHandoffUsageRecord,
 } from "./usage.js";
 import { imageHash } from "./image.js";
+import { randomUUID } from "node:crypto";
 import { abortWireGuard, fetchInterceptorGuard, timeoutGuard, type AbortWire } from "./dispose.js";
 
 /** Tokens reserved for the describer's input so the requested output budget
@@ -133,17 +134,47 @@ function configSnapshot(cfg: VisionHandoffConfig): {
  * completeSimple() directly cannot resolve custom API ids such as `makora`.
  * The optional access preserves compatibility with older ModelRegistry versions,
  * which registered custom streams globally and do not expose the config getter. */
+const NEURALWATT_CONVERSATION_HEADER = "x-nw-conversation-id";
+
+function isolateVisionRequest(
+  model: Model<Api>,
+  options: SimpleStreamOptions,
+): SimpleStreamOptions {
+  const sessionId = `pi-vision-handoff:${randomUUID()}`;
+  const headers = { ...(options.headers ?? {}) };
+  const inheritedConversationHeader = Object.keys(headers).find(
+    (name) => name.toLowerCase() === NEURALWATT_CONVERSATION_HEADER,
+  );
+
+  // Provider auth can carry the active Pi conversation id. A describer call is
+  // an independent conversation: reusing that id makes an unrelated image
+  // prompt the newest cache lineage for the main agent. Give every real vision
+  // request its own identity while preserving all non-session auth headers.
+  if (inheritedConversationHeader) {
+    headers[inheritedConversationHeader] = sessionId;
+  } else if (model.provider.toLowerCase() === "neuralwatt") {
+    headers["X-NW-Conversation-ID"] = sessionId;
+  }
+
+  return {
+    ...options,
+    sessionId,
+    ...(Object.keys(headers).length > 0 ? { headers } : {}),
+  };
+}
+
 export async function completeVisionModel(
   model: Model<Api>,
   modelRegistry: ModelRegistry,
   context: Context,
   options: SimpleStreamOptions,
 ): Promise<AssistantMessage> {
+  const isolatedOptions = isolateVisionRequest(model, options);
   const provider = modelRegistry.getRegisteredProviderConfig?.(model.provider);
   if (provider?.streamSimple && provider.api === model.api) {
-    return provider.streamSimple(model, context, options).result();
+    return provider.streamSimple(model, context, isolatedOptions).result();
   }
-  return completeSimple(model, context, options);
+  return completeSimple(model, context, isolatedOptions);
 }
 
 /** Dependencies the describer can't own itself (held by the engine). */