Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

Bug: reasoning_content is replayed in the request history, causing tool-call markup leakage

Open Beginner friendly
  • Is something wrong?
#56,352 2 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
2/5
Estimated time
1-3 hours
Newbie friendliness
66/100
Issue type
Bug
Clarity
Clearly specified
Activity status
Active
Tech stack
typescript
Domain
api, backend

Research direction

Start by tracing how session history is assembled into provider request payloads, then find where assistant messages retain reasoning_content. Check whether that field is sent in subsequent requests and whether the request-building path has tests for reasoning models. Done means history can still retain reasoning for display while requests omit it, with a regression test covering the behavior.

Written by the indexing model from the issue text.

Description

opencode persists the model's reasoning_content (chain-of-thought) in the session history and sends it back in every subsequent request. DeepSeek's API states that reasoning_content is the current turn's CoT and must not be passed back.

Impact: with reasoning models (e.g. deepseek-v4.1-flash), the model imitates patterns from its own past reasoning - including malformed tool-call markup - so tool calls are silently lost (tool_calls == 0, finish_reason == "stop"). It is self-reinforcing: the more it leaks, the more that markup appears in reasoning.

A/B (one captured opencode request, 1011 messages / 4.78 MB, replayed to the same endpoint with a "write a ~120-line script" instruction appended):

variant leak rate
with reasoning_content 4 / 5
without reasoning_content 0 / 5
last 300 msgs, with 2 / 5
last 300 msgs, without 0 / 5

"Leak" is defined strictly as: markup inside delta.content and tool_calls == 0 and finish_reason == "stop".

Observed message structure in the captured request: assistant messages 480, of which reasoning_content present in 480.

Excluded by A/B: max_tokens (256k vs 128k -> 4/5 vs 4/5), system prompt content, temperature, tool_choice, parallel_tool_calls, thinking/reasoning_effort fields, and endpoint (official api.deepseek.com vs the opencode gateway). The dose-response (100 msgs 0/5 -> 300 msgs 2-3/5 -> 1009 msgs 4/5) is fully explained by the replayed CoT.

A different client (CodeBuddy) with the same model and the same real leaked-text corpus in a 1.13 MB history does not reproduce it (0/5).

Suggested fix: do not send reasoning_content back when building the request payload (keep it for UI display only); optionally gate behind a provider/model capability flag.

Reproduction: run a long session with a reasoning model, then ask it to write a long file - occasionally the tool call is emitted as text; re-sending the same request with reasoning_content stripped removes the failure (0/5 vs 4/5).

Environment: opencode 2.0.30 (Windows), providers opencode-go and official deepseek, model deepseek-v4.1-flash (variant max).

Dominant language
TypeScript
Stars
215k
Forks
28.9k
Avg merge
11h 19m
Merged PRs (30d)
320

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from anomalyco/opencode

All issues in anomalyco/opencode

Similar issues

More TypeScript issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.