fix(agent): seed hydrated transcripts with estimated context usage

packages/agent/src/adapters/claude/session/jsonl-hydration.test.ts

@@ -496,6 +496,29 @@ describe("conversationTurnsToJsonlEntries", () => {
     expect(conv[1].parentUuid).toBe(conv[0].uuid);
   });
 
+  it("seeds only the last assistant message with the estimated context size", () => {
+    const lines = conversationTurnsToJsonlEntries(
+      [
+        { role: "user", content: [{ type: "text", text: "q".repeat(400) }] },
+        {
+          role: "assistant",
+          content: [{ type: "text", text: "a".repeat(400) }],
+        },
+        { role: "user", content: [{ type: "text", text: "more" }] },
+        { role: "assistant", content: [{ type: "text", text: "final" }] },
+      ],
+      config,
+    );
+
+    const assistants = parseConversationEntries(lines).filter(
+      (entry) => entry.type === "assistant",
+    );
+    expect(assistants.at(-1)?.message.usage.input_tokens).toBeGreaterThan(0);
+    for (const entry of assistants.slice(0, -1)) {
+      expect(entry.message.usage.input_tokens).toBe(0);
+    }
+  });
+
   it("emits one line per assistant block with shared message id", () => {
     const lines = conversationTurnsToJsonlEntries(
       [

packages/agent/src/adapters/claude/session/jsonl-hydration.ts

@@ -421,6 +421,13 @@ export function conversationTurnsToJsonlEntries(
   const permissionMode = config.permissionMode ?? "default";
   const baseTime = Date.now() - turns.length * 3000;
   let turnIndex = 0;
+  // Seed the last assistant message with the estimated context size so the
+  // SDK's auto-compact accounting isn't blind until the first API response.
+  const estimatedTokens = turns.reduce(
+    (sum, turn) => sum + estimateTurnTokens(turn),
+    0,
+  );
+  const lastAssistantTurn = turns.findLast((turn) => turn.role === "assistant");
 
   for (const turn of turns) {
     const timestamp = new Date(baseTime + turnIndex * 3000).toISOString();
@@ -530,7 +537,8 @@ export function conversationTurnsToJsonlEntries(
               stop_reason: isLast ? lastStopReason : null,
               stop_sequence: null,
               usage: {
-                input_tokens: 0,
+                input_tokens:
+                  turn === lastAssistantTurn && isLast ? estimatedTokens : 0,
                 cache_creation_input_tokens: 0,
                 cache_read_input_tokens: 0,
                 output_tokens: 0,