@vellumai/assistant 0.10.2 → 0.10.3-dev.202606252046.9075fd5

package/src/{plugins/defaults/advisor/__tests__/transcript.test.ts → subagent/__tests__/consult-transcript.test.ts}

@@ -1,11 +1,11 @@
 import { describe, expect, test } from "bun:test";
 
-import type { ContentBlock, Message } from "../../../../providers/types.js";
-import { toAdvisorMessages } from "../transcript.js";
+import type { ContentBlock, Message } from "../../providers/types.js";
+import { sanitizeConsultTranscript } from "../consult-transcript.js";
 
 const text = (t: string): ContentBlock => ({ type: "text", text: t });
 
-describe("toAdvisorMessages", () => {
+describe("sanitizeConsultTranscript", () => {
   test("drops thinking and redacted-thinking blocks", () => {
     const messages: Message[] = [
       { role: "user", content: [text("do the task")] },
@@ -18,12 +18,34 @@ describe("toAdvisorMessages", () => {
         ],
       },
     ];
-    const out = toAdvisorMessages(messages);
+    const out = sanitizeConsultTranscript(messages);
     expect(out).toHaveLength(2);
     expect(out[1].content).toEqual([text("here is my answer")]);
   });
 
-  test("strips the pending advisor tool_use from the final assistant turn", () => {
+  test("drops a file block", () => {
+    const messages: Message[] = [
+      {
+        role: "user",
+        content: [
+          text("read this"),
+          {
+            type: "file",
+            source: {
+              type: "base64",
+              media_type: "application/pdf",
+              data: "z",
+              filename: "f.pdf",
+            },
+          },
+        ],
+      },
+    ];
+    const out = sanitizeConsultTranscript(messages);
+    expect(out[0].content).toEqual([text("read this")]);
+  });
+
+  test("strips the pending tool_use from the final assistant turn", () => {
     const messages: Message[] = [
       { role: "user", content: [text("task")] },
       {
@@ -34,7 +56,7 @@ describe("toAdvisorMessages", () => {
         ],
       },
     ];
-    const out = toAdvisorMessages(messages);
+    const out = sanitizeConsultTranscript(messages);
     expect(out[1].content).toEqual([text("let me consult the advisor")]);
   });
 
@@ -51,7 +73,7 @@ describe("toAdvisorMessages", () => {
       },
       { role: "assistant", content: [text("done")] },
     ];
-    const out = toAdvisorMessages(messages);
+    const out = sanitizeConsultTranscript(messages);
     expect(out[1].content[0]).toEqual({
       type: "tool_use",
       id: "a",
@@ -87,7 +109,7 @@ describe("toAdvisorMessages", () => {
       },
       { role: "assistant", content: [text("found it; here is the answer")] },
     ];
-    const out = toAdvisorMessages(messages);
+    const out = sanitizeConsultTranscript(messages);
     const flat = out.flatMap((m) => m.content);
     expect(flat.some((b) => b.type === "server_tool_use")).toBe(false);
     expect(flat.some((b) => b.type === "web_search_tool_result")).toBe(false);
@@ -103,7 +125,7 @@ describe("toAdvisorMessages", () => {
     const messages: Message[] = [
       { role: "user", content: [text("look at this"), img] },
     ];
-    const out = toAdvisorMessages(messages);
+    const out = sanitizeConsultTranscript(messages);
     expect(out[0].content).toEqual([text("look at this"), img]);
   });
 
@@ -136,7 +158,7 @@ describe("toAdvisorMessages", () => {
         ],
       },
     ];
-    const out = toAdvisorMessages(messages);
+    const out = sanitizeConsultTranscript(messages);
     expect(out[0].content[0]).toEqual({
       type: "tool_result",
       tool_use_id: "a",
@@ -144,4 +166,19 @@ describe("toAdvisorMessages", () => {
       contentBlocks: [img],
     });
   });
+
+  test("drops messages that are empty after sanitization", () => {
+    const messages: Message[] = [
+      { role: "user", content: [text("task")] },
+      {
+        role: "assistant",
+        content: [{ type: "thinking", thinking: "secret", signature: "sig" }],
+      },
+      { role: "assistant", content: [text("answer")] },
+    ];
+    const out = sanitizeConsultTranscript(messages);
+    expect(out).toHaveLength(2);
+    expect(out[0].content).toEqual([text("task")]);
+    expect(out[1].content).toEqual([text("answer")]);
+  });
 });

package/src/{plugins/defaults/advisor/steering.ts → subagent/consult-prompt.ts}

@@ -1,34 +1,11 @@
 /**
- * Prompt fragments for the advisor plugin:
- *  - `appendSteering` / `stripSteering` — the executor-facing steering block the
- *    `pre-model-call` hook injects so the model reaches for the advisor tool at
- *    the right times.
- *  - `buildAdvisorSystem` / `advisorRequestText` — the advisor-facing framing for
- *    the consult itself.
+ * Advice-framing prompt fragments for the advisor consult:
+ *  - `buildAdvisorSystem` — the advisor-facing system prompt; frames the role and,
+ *    for context, embeds the executor's own system prompt.
+ *  - `advisorRequestText` — the final user turn appended to the transcript asking
+ *    for guidance.
  */
 
-/** Idempotency marker; the steering block is appended at the end of the prompt. */
-export const STEERING_MARKER = "<!-- advisor:steering -->";
-
-const STEERING_BODY = `You have an \`advisor\` tool backed by a stronger reviewer model. It takes NO parameters — calling it forwards your entire conversation automatically (the task, every tool call, every result). Call advisor BEFORE you start building or implementing: once you understand what's being asked, consult it to shape the plan — it can lay out the plan when you don't have one yet, or pressure-test and sharpen a plan you've already drafted. Orient yourself first (read the relevant files, understand the task), then call advisor before you commit to an approach and start producing work. Also call it when you get stuck, when you're weighing a change in direction, and once before you declare the task done. Give its guidance serious weight; only override it when primary-source evidence contradicts a specific claim, and say so when you do.`;
-
-const ADVISOR_STEERING = `${STEERING_MARKER}\n${STEERING_BODY}`;
-
-/** Append the steering block to the executor's system prompt (idempotent). */
-export function appendSteering(systemPrompt: string | null): string | null {
-  if (systemPrompt === null) return null;
-  if (systemPrompt.includes(STEERING_MARKER)) return systemPrompt;
-  return `${systemPrompt}\n\n${ADVISOR_STEERING}`;
-}
-
-/** Remove a previously-appended steering block, recovering the original prompt. */
-export function stripSteering(systemPrompt: string | null): string | null {
-  if (systemPrompt === null) return null;
-  const idx = systemPrompt.indexOf(STEERING_MARKER);
-  if (idx === -1) return systemPrompt;
-  return systemPrompt.slice(0, idx).trimEnd();
-}
-
 /**
  * System prompt for the advisor sub-call. Frames the advisor's role and, for
  * context, quotes the executor's own system prompt (as the advisor tool does —
@@ -54,14 +31,27 @@ How to advise:
 - Stay in your lane. Advise the agent; do not role-play as it, write its final deliverable, or take its next action for it. If the agent is already on the right track, confirm it and sharpen the plan rather than manufacturing objections.
 
 Write as much as the guidance genuinely needs, and no more.`;
-  if (!originalSystemPrompt) return base;
-  return `${base}\n\nFor context, the agent is operating under this system prompt:\n<agent_system_prompt>\n${originalSystemPrompt}\n</agent_system_prompt>`;
+  let prompt = base;
+  if (originalSystemPrompt) {
+    prompt += `\n\nFor context, the agent is operating under this system prompt:\n<agent_system_prompt>\n${originalSystemPrompt}\n</agent_system_prompt>`;
+  }
+  return prompt;
 }
 
 /**
  * The final user turn appended to the transcript for the advisor sub-call. Asks
  * for guidance; imposes no length limit — the advisor decides how much to say.
+ *
+ * `agentRequest` is the executing agent's own `objective` from the
+ * `subagent_spawn` call — the agent's framing of what it wants weighed in on.
+ * It is included verbatim because (a) the agent naturally states the task there,
+ * and (b) the inherited transcript can be thin (e.g. a wake turn whose task
+ * lives in memory rather than a user message), so the request text is often the
+ * advisor's clearest signal of what is actually being asked.
  */
-export function advisorRequestText(): string {
-  return `Review the conversation above — the task, the tool calls, and their results — and give focused strategic guidance on how to proceed.`;
+export function advisorRequestText(agentRequest?: string): string {
+  const base = `Review the conversation above — the task, the tool calls, and their results — and give focused strategic guidance on how to proceed.`;
+  const trimmed = agentRequest?.trim();
+  if (!trimmed) return base;
+  return `${base}\n\nThe agent described what it wants your input on:\n<agent_request>\n${trimmed}\n</agent_request>\nTreat this as the agent's framing of the task. If it conflicts with the transcript above, say so; if the transcript is sparse, rely on it.`;
 }

package/src/{plugins/defaults/advisor/transcript.ts → subagent/consult-transcript.ts}

@@ -1,6 +1,10 @@
 /**
- * Convert a captured executor transcript into the message list sent to the
- * advisor sub-call.
+ * Sanitize an inherited parent transcript before it is injected into a
+ * tool-less advisor consult subagent.
+ *
+ * `Conversation.injectInheritedContext` injects the parent's messages VERBATIM
+ * with no sanitization, so the advisor path must run this over the parent
+ * messages before they are injected.
  *
  * Strips blocks the advisor shouldn't (or can't) replay:
  *  - thinking / redacted-thinking (the advisor tool drops thinking),
@@ -16,13 +20,17 @@
  * Visual tasks depend on it; the advisor profile is expected to be vision-capable.
  *
  * It also strips the *pending* client tool calls from the final assistant turn:
- * at capture time (a `post-model-call` before tools run) the last assistant
- * message carries the `advisor` tool_use with no matching `tool_result` yet, so
- * sending it would be a dangling call. Earlier, completed `tool_use` /
- * `tool_result` pairs are preserved intact.
+ * at capture time the last assistant message can carry a `tool_use` with no
+ * matching `tool_result` yet, so sending it would be a dangling call. Earlier,
+ * completed `tool_use` / `tool_result` pairs are preserved intact.
+ *
+ * In-flight-turn parity — ensuring the assistant's just-written plan from the
+ * current turn is included while its dangling tool_use is stripped — is handled
+ * by the consumer (the wiring that calls this before injection); this function
+ * only strips a dangling `tool_use` IF it appears on the final assistant turn.
  */
 
-import type { ContentBlock, Message } from "../../../providers/types.js";
+import type { ContentBlock, Message } from "../providers/types.js";
 
 /** Drop disallowed blocks; recursively sanitize tool_result content. `null` = drop. */
 function sanitize(block: ContentBlock): ContentBlock | null {
@@ -51,7 +59,9 @@ function sanitize(block: ContentBlock): ContentBlock | null {
   }
 }
 
-export function toAdvisorMessages(messages: ReadonlyArray<Message>): Message[] {
+export function sanitizeConsultTranscript(
+  messages: ReadonlyArray<Message>,
+): Message[] {
   const out: Message[] = [];
   const lastIndex = messages.length - 1;
 

package/src/subagent/index.ts

@@ -1,4 +1,4 @@
-export { mergeSkillIds } from "./manager.js";
+export { mergeSkillIds, SubagentAbortedError } from "./manager.js";
 export type { SubagentRole } from "./types.js";
 export { SUBAGENT_ROLE_REGISTRY, TERMINAL_STATUSES } from "./types.js";