fix: preserve assistant reasoning_content when translating history#5
Open
meumlpontocom wants to merge 3 commits intoLordymine:masterfrom
Open
fix: preserve assistant reasoning_content when translating history#5meumlpontocom wants to merge 3 commits intoLordymine:masterfrom
meumlpontocom wants to merge 3 commits intoLordymine:masterfrom
Conversation
Moonshot-backed models (e.g. kimi-k2.6 via the OpenCode Zen proxy) reject multi-turn requests that have `thinking` enabled but carry assistant messages without `reasoning_content`: thinking is enabled but reasoning_content is missing in assistant tool call message at index 8 The Anthropic request that Claude Code replays on each turn includes `thinking` blocks inside prior assistant `content` arrays. The Anthropic → OpenAI translator in `request-conversion.ts` was dropping those blocks, so by the time the history reached the upstream provider, reasoning context was gone and the provider refused the request. This change accumulates `thinking` and `redacted_thinking` blocks found on assistant messages and emits their text as `reasoning_content` on the translated OpenAI-style message. `redacted_thinking` uses a `[redacted]` placeholder so the field is non-empty when the provider insists on its presence. Tests cover single thinking blocks, thinking + tool_use turns, multiple thinking blocks joined with newlines, `redacted_thinking`, and the no-thinking path (field stays absent). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The initial fix only propagated reasoning_content when Anthropic had
actually emitted a `thinking` block for that specific turn. In practice
Claude Code's replayed history contains assistant messages that carry
only `tool_use` blocks — no thinking block at all — and Moonshot still
rejects those with:
thinking is enabled but reasoning_content is missing
in assistant tool call message at index 5
When the incoming request has `thinking: { type: "enabled" }` (or the
equivalent) and an assistant message in the history has tool_calls but
no extractable reasoning, emit a placeholder `reasoning_content` so the
upstream accepts the request. Real thinking blocks, when present, still
win — the placeholder only fills in the gap.
Covered by 3 new tests: placeholder injection, no-placeholder when there
are no tool_calls, and real thinking blocks overriding the placeholder.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Some upstream providers (notably Moonshot AI routing kimi-k2.6 via the
OpenCode Zen proxy → OpenRouter) enable extended thinking server-side
regardless of whether the client requested it. In that configuration
every assistant message with tool_calls in the replayed history must
carry a non-empty `reasoning_content`, even when the original Anthropic
request never set `thinking: { ... }`.
Repro logged against a fresh session: Claude Code sent no `thinking`
field, the proxy forwarded the history untouched, and Moonshot rejected
with:
thinking is enabled but reasoning_content is missing
in assistant tool call message at index 2 / 5 / 8
Drop the `thinkingEnabled` gate and emit `reasoning_content` whenever an
assistant message has tool_calls. Real thinking blocks still win; when
none are present, a `[no reasoning provided]` placeholder fills the gap.
Providers that ignore the field are unaffected — it's a known OpenAI-
compatible extension used by DeepSeek, Moonshot, Qwen, etc.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Author
|
Update: after testing end-to-end against a real Third commit in this PR ( Test suite: 101 pass, 0 fail. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes a runtime proxy failure seen when running Claude Code against a
Moonshot-backed model through the OpenCode Zen proxy (e.g.
kimi-k2.6):The Anthropic request that Claude Code replays on each turn embeds
thinkingblocks inside prior assistantcontentarrays. TheAnthropic → OpenAI translator in
src/proxy/request-conversion.tsonly handled
textandtool_useblocks, silently dropping anythinking/redacted_thinkingentries. By the time the historyreached the upstream provider the reasoning context was gone, and
Moonshot rejects such requests whenever extended thinking is on.
What changed
src/proxy/request-conversion.ts— accumulate thinking text fromassistant
thinkingblocks and emit it asreasoning_contentonthe OpenAI-shaped message.
redacted_thinkingblocks (provider-opaque) become a
[redacted]placeholder so the field is stillnon-empty when the provider requires it.
tests/request-conversion.test.ts— 5 new tests:thinking+ text contentthinking+tool_usein the same assistant turnthinkingblocks joined with\nredacted_thinkingplaceholderreasoning_contentstays absentNo behavior changes for messages that don't include thinking blocks,
so requests to providers that ignore
reasoning_content(or rejectunknown fields) are unaffected — the field is only emitted when the
incoming history actually carries thinking.
Test plan
bun run typecheck— cleanbun test— 84 pass (was 79; 5 new), 0 failbun run build— cleankimi-k2.6via Claude Code: tool-useturn 8 that previously returned HTTP 400 now round-trips
successfully.
🤖 Generated with Claude Code