Identified issue

The problem turned out to be prompt-history serialization in SessionPrompt.prompt and MessageV2.toModelMessages.

SessionPrompt.prompt assembles the next model call from prior session history. For older assistant turns, MessageV2.toModelMessages was replaying historical assistant text too aggressively. That meant intermediate progress text from a previous task was still present in the prompt for a later unrelated task.

The symptom was not “wrong parent ID” or “provider resumed the wrong response”. In the broken cases, the new user message and parentID were correct, but the model was still being shown stale prior-task assistant text and would re-anchor to it.

The relevant code path is:

• packages/opencode/src/session/prompt.ts: SessionPrompt.prompt
• packages/opencode/src/session/message-v2.ts: MessageV2.toModelMessages

The local fix narrowed historical assistant turns to their last step before serializing them into later prompts:

// packages/opencode/src/session/message-v2.ts
const historic = !!options?.active && msg.info.parentID !== options.active;

for (const part of historic ? laststep(msg.parts) : msg.parts) {
  if (part.type === "text") {
    assistantMessage.parts.push({
      type: "text",
      text: part.text,
    });
  }
  if (historic) continue;
  // keep full step/tool/reasoning continuity only for the active turn
}

That change preserves older assistant turns as history, but stops replaying all of their intermediate progress chatter into future prompts.

Why this matches the user-visible bug

The stale content was typically progress text from the previous task, such as “I’m checkpointing...”, “I’m staging...”, or “The reasoning-core checkpoint is committed.” When that text was still in the serialized prompt for a new unrelated request, the assistant could start on the new request and then slide back into the earlier task.

That also explains why /undo is a strong reproducer without being the root cause: undo makes it easier to put the session back into a state where the most recent surviving prior assistant turn is exactly the stale one that gets replayed.

How this was verified

Verification used three separate checks.

Likely related upstream issue: #16612 Sometimes replies to the previous turn (stale context): keeps repeating an earlier command/answer instead of responding to the latest input.
Probably not the same bug: #10287 Critical bug revert undo messages, which appears to be about incorrect file-state restoration on undo rather than stale prompt history.

1. Session export inspection

The exported session data showed assistant messages correctly parented to the latest user turn while still emitting content from an older task. That ruled out a simple parent-link bug.

2. Instrumented runtime logging

Session.revert cleanup completed correctly, but the next SessionPrompt.prompt call still serialized model history containing stale assistant text from the previous task. That ruled out the initial undo-race hypothesis.

3. Deterministic tests

Two local tests made the issue concrete:

• packages/opencode/test/session/stale-history.test.ts

  • session.prompt stale history reproduction > keeps only the last step from prior assistant turns
  • This seeds one completed “reasoning core” turn, then sends a different prompt like “Sketch the finite-state-machine core instead.”
  • The model call should include only the final assistant step from the old turn, not the earlier progress text.

• packages/opencode/test/session/message-v2.test.ts

  • trims historic assistant turns to their last step
  • This verifies MessageV2.toModelMessages directly.

The prompt-level repro looked like this:

// packages/opencode/test/session/stale-history.test.ts
expect(calls[0]?.messages).toEqual(
  expect.arrayContaining([
    {
      role: "assistant",
      content: [
        { type: "text", text: "The reasoning-core checkpoint is committed." },
      ],
    },
  ]),
);

expect(calls[0]?.messages).not.toEqual(
  expect.arrayContaining([
    {
      role: "assistant",
      content: expect.arrayContaining([
        { type: "text", text: "I’m checkpointing the reasoning core first." },
      ]),
    },
  ]),
);

That is the behavioral distinction that fixed the bug locally.

Local resolution

The local resolution was:

1. Keep full assistant continuity only for the active turn.
2. For older assistant turns, serialize only the last step instead of replaying all intermediate assistant progress text.
3. Add regression coverage in MessageV2.toModelMessages and in the end-to-end prompt assembly path.

References:

• packages/opencode/src/session/prompt.ts: SessionPrompt.prompt
• packages/opencode/src/session/message-v2.ts: MessageV2.toModelMessages
• packages/opencode/test/session/stale-history.test.ts
• packages/opencode/test/session/message-v2.test.ts