Repository navigation
core: standalone signed-empty reasoning turns are replayed into later requests #46881
Description
Description
V2 replays completed assistant turns whose only content is empty reasoning plus provider metadata.
In packages/core/src/session/runner/to-llm-message.ts, same-model reasoning retains provider metadata, and the meaningful filter considers metadata sufficient even when part.text === "". The existing omits empty assistant turns test explicitly preserves a standalone { text: "", state: { signature } } turn.
The Amazon Bedrock adapter serializes each retained part as:
{
"reasoningContent": {
"reasoningText": {
"text": "",
"signature": "..."
}
}
}
Controlled forks isolated the effect:
- The exact prompt with the clean full history (~620k provider tokens) produced visible text and tools.
- The same prompt with one or two standalone signed-empty turns still worked.
- The same prompt with five accumulated signed-empty turns deterministically returned another signature-only end_turn with no visible output.
- Manual compaction repaired the session because compaction rendering drops empty reasoning parts.
Expected: signed reasoning state is retained when it accompanies semantic assistant output—text, readable reasoning, or a tool call—but a standalone metadata-only assistant turn is omitted from future model history or marked failed by the no-output handling described separately.
Actual: state-only turns accumulate and can poison subsequent same-model requests.
Plugins
No response
OpenCode version
opencode2 v0.0.0-dev-18900; current V2 source af332a2 remains affected.
Steps to reproduce
- Create a V2 session using a provider that returns signed reasoning metadata.
- Persist several completed assistant turns containing only reasoning.text = "" and a provider signature.
- Submit a prompt that works in a fork before those turns.
- Observe the provider request replay each empty reasoningText plus signature.
- Observe the later response repeat the signature-only pattern.
- Compact the same session and retry; the prompt works because the empty reasoning turns are no longer replayed.
Screenshot and/or share link
No response
Operating System
No response
Terminal
No response
Activity
github-actions commentedon Sep 2, 2026
This issue might be related to existing issues about empty or incomplete assistant turns being replayed to providers. Please check:
Your issue focuses specifically on signed empty reasoning turns (Bedrock provider metadata/signature accumulating across turns), which appears to be a distinct variant of the broader empty-assistant-turn replay problem. Linking these in case the fix should be coordinated.
Rocklis commentedon Sep 4, 2026
Confirmed on current
dev. The existingtoLLMMessagestest actually preserves a standalone empty reasoning part when it has a provider signature, so the reported history shape is reproducible without a live Bedrock call.I have a small replay-side fix: provider metadata is kept when the same assistant turn also contains real text or a tool call, but a turn containing only empty reasoning metadata is omitted. The full
packages/coresuite passes (1,097 tests); I'll open the PR next.bbartels commentedon Sep 4, 2026
not sure what the relation between dev/v2 are, but I was using v2
Rocklis commentedon Sep 4, 2026
Yep, that is the path this fix targets.
devis just the repository branch I tested against; the affected replay code lives in the V2packages/corepath. The fix is in #47357.Rocklis commentedon Sep 7, 2026
I've moved #47357 onto
v2(3079fb4) and rerun the 24 replay tests there. My earlier full-suite result was fromdev, notv2; the PR description now makes that distinction.@bbartels, could you try this head with the session that accumulated signature-only turns? The change drops standalone empty reasoning turns, but keeps the signature when the same turn has text or a tool call.
opencodejr commentedon Oct 28, 2026
Thanks for the detailed write-up and the forked-session comparison.
The replay is confirmed in code on the latest
dev(35a4a68). Inpackages/core/src/session/runner/to-llm-message.ts, themeaningfulfilter keeps a reasoning part with empty text when it has provider metadata, so a turn that holds only an empty reasoning part with a signature is sent again on later same-model requests. The existing testomits empty assistant turnsexpects that turn to be kept, so the replay is deliberate today.What would make it a bug is how Bedrock reacts once several of these turns pile up: signature-only replies with no visible output. That needs a real Bedrock account and a long session, and the sandbox I reproduce in has no provider credentials. A mock can't show how the real model responds, so a maintainer with Bedrock access will need to confirm it.
Related: