fix(dashboard-agent): summarise what the transcript recorded about a watch

A watch can expire or be cancelled with nothing written back into the
transcript, so asking the summariser for "any watch that is running" preserved
an old confirmation as current state.

Author: kathiekiwi

internal-packages/dashboard-agent/src/compaction.test.ts

@@ -19,6 +19,7 @@ import {
   safeTail,
   shouldCompactConversation,
   STATIC_PREFIX_TOKENS,
+  SUMMARY_INSTRUCTION,
   withDurableState,
 } from "./compaction";
 
@@ -389,6 +390,31 @@ describe("the summariser's input", () => {
   });
 });
 
+/**
+ * A watch's lifecycle is server-side: it can expire or be cancelled with nothing written back
+ * into the transcript. So the summary can only ever say what the transcript RECORDED — asking
+ * for what is running turns an old confirmation into a claim that it still is, and the next
+ * answer tells the user a watch is on that ended hours ago. The property, not the sentence:
+ * the watch line asks for a record and never for present state.
+ */
+describe("the summary instruction never asks for present state", () => {
+  const watchLine = SUMMARY_INSTRUCTION.split("\n").find((line) => /watch/i.test(line));
+
+  it("has a line about watches at all", () => {
+    expect(watchLine).toBeDefined();
+  });
+
+  it("asks what the transcript recorded, not what is true now", () => {
+    expect(watchLine).toMatch(/record/i);
+    // The transcript cannot know, so the instruction has to say why.
+    expect(watchLine).toMatch(/expire|cancel/i);
+  });
+
+  it("never asks for a watch that is running, scheduled or active", () => {
+    expect(watchLine).not.toMatch(/(?:that is|still|currently)\s+(?:running|active|scheduled)/i);
+  });
+});
+
 /** Records what each model call was actually given, and summarises predictably. */
 function capturingModel(prompts: string[], summarized: string[] = []) {
   return new MockLanguageModelV3({

internal-packages/dashboard-agent/src/compaction.ts

@@ -61,16 +61,16 @@ const SUMMARY_MODEL = "anthropic:claude-haiku-4-5" as const;
  */
 const SUMMARY_MAX_OUTPUT_TOKENS = 1_000;
 
-const SUMMARY_INSTRUCTION = `You are compacting a support conversation between a user and an agent that reads a Trigger.dev dashboard, so the agent can keep going with a shorter history.
+export const SUMMARY_INSTRUCTION = `You are compacting a support conversation between a user and an agent that reads a Trigger.dev dashboard, so the agent can keep going with a shorter history.
 
 Write a summary in under 400 words, as notes rather than prose. Keep, in this order:
 1. What the user is trying to do, in their own terms, and anything they asked to be remembered.
 2. Facts already established, with the run ids, queue names, task identifiers, error fingerprints and numbers they rest on. Never restate a number you cannot see.
 3. Any investigation that is open: its investigationId, its title and its current outcome.
-4. Any watch that is running or has reported, and what it said.
+4. Any watch the transcript records — what it was set up to watch, and what it said if it reported. Write it as what the transcript recorded, never as what is true now: a watch can expire or be cancelled without saying so here, so never present one as current.
 5. What was asked most recently and what is still unanswered.
 
-Drop tool mechanics, retries, and anything already superseded. Do not add advice, and do not invent anything that is not in the transcript.`;
+Drop tool mechanics, retries, and anything already superseded. Do not add advice, and do not invent anything that is not in the transcript. Everything you write is a record of what the transcript said, not a claim about the present.`;
 
 /** A summary that reads as a summary, and never as the user's next question. */
 export function summaryMessage(summary: string, durableState?: string): ModelMessage {