fix: MiMo remove unsupported body fields and preserve reasoning content by Meetpatel006 · Pull Request #1253 · Twigpine/openclaude

Meetpatel006 · 2026-05-19T09:22:49Z

Summary

Strip store and stream_options fields from MiMo API requests (both OpenGateway and Xiaomi direct paths) to prevent API 400 errors
Preserve real
easoning_content during conversation replay for MiMo models instead of injecting empty fallback
Add shouldPreserveThinkingBlocksForProviderReplay() to skip thinking-block stripping for mimo-v2* models

Changes

gitlawb-opengateway.ts — add
emoveBodyFields: ['store', 'stream_options'] to gateway config
xiaomi-mimo.ts — replace
easoningContentFallback: '' with
emoveBodyFields, stop synthesizing empty reasoning
**
untimeMetadata.ts** — add fallback inference for mimo-v2* models with correct shim settings
conversationRecovery.ts — preserve thinking blocks for MiMo during replay
Tests — 4 new test cases in openaiShim.test.ts + MiMo assertion block in conversationRecovery.hooks.test.ts

Test plan

OpenGateway MiMo replays real
easoning_content without empty fallback
Xiaomi MiMo replays real
easoning_content without empty fallback
MiMo does not synthesize empty
easoning_content when missing
Streaming requests strip stream_options and store
Conversation recovery preserves thinking blocks for MiMo

Stop the Gitlawb Opengateway preset from forcing synthetic empty reasoning_content onto prior assistant tool-call messages. MiMo rejects that shape after Agent/sub-agent tool calls with upstream Param Incorrect errors. Add an OpenAI shim regression test covering MiMo tool history through https://opengateway.gitlawb.com/v1 while preserving the separate OpenGateway Gemini signature replay behavior. Validation: bun test src/services/api/openaiShim.test.ts; bun run build

…reserve reasoning content

…rip unsupported options

Copilot

Pull request overview

Fixes MiMo (Xiaomi) API integration issues by (1) stripping unsupported store and stream_options fields from request bodies for both OpenGateway and direct Xiaomi paths, (2) removing the empty reasoning_content fallback so real reasoning content is preserved during conversation replay, and (3) preserving thinking blocks for mimo-v2* models when restoring conversations from logs.

Changes:

Replace reasoningContentFallback: '' with removeBodyFields: ['store', 'stream_options'] in both the OpenGateway and Xiaomi MiMo transport configs, and add the same fallback inference for mimo-v2* models in runtimeMetadata.ts.
Add shouldPreserveThinkingBlocksForProviderReplay() in conversationRecovery.ts to skip thinking-block stripping when the active model is mimo-v2*.
Add 4 new test cases in openaiShim.test.ts and a MiMo-specific assertion block in conversationRecovery.hooks.test.ts.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
src/integrations/gateways/gitlawb-opengateway.ts	Drop reasoning-content fallback config; strip `store`/`stream_options` from request body.
src/integrations/vendors/xiaomi-mimo.ts	Same change for the direct Xiaomi MiMo vendor path.
src/integrations/runtimeMetadata.ts	Add `mimo-v2*` inference branch matching the gateway/vendor config.
src/utils/conversationRecovery.ts	Add helper to preserve thinking blocks for `mimo-v2*` models during replay.
src/utils/conversationRecovery.hooks.test.ts	Add MiMo replay assertions confirming thinking blocks are preserved.
src/services/api/openaiShim.test.ts	Add 4 tests covering OpenGateway/Xiaomi reasoning preservation and body-field stripping.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Vasanthdev2004 · 2026-05-19T14:50:33Z

Blockers

None found.

Non-Blocking

338 additions is large for a body-field fix — mostly test coverage.

Looks Good

Removes unsupported body fields (store, stream_options) for MiMo
Preserves reasoning content for mimo-v2* models
Adds conversation recovery for thinking blocks
Good test coverage (4 new tests)
Fixes both OpenGateway and direct Xiaomi paths

Verdict: Approve — clean MiMo compatibility fix.

Vasanthdev2004

Clean MiMo compatibility fix.

kevincodex1

Awesome! Thanks for working on this.

…nt (Twigpine#1253) * Fix OpenGateway MiMo agent tool history Stop the Gitlawb Opengateway preset from forcing synthetic empty reasoning_content onto prior assistant tool-call messages. MiMo rejects that shape after Agent/sub-agent tool calls with upstream Param Incorrect errors. Add an OpenAI shim regression test covering MiMo tool history through https://opengateway.gitlawb.com/v1 while preserving the separate OpenGateway Gemini signature replay behavior. Validation: bun test src/services/api/openaiShim.test.ts; bun run build * fix(MiMo): update integration to remove unnecessary body fields and preserve reasoning content * test(MiMo): enhance tests to verify reasoning_content handling and strip unsupported options --------- Co-authored-by: jatmn <the@jat.mn>

…tinuity DeepSeek V4 / Moonshot providers reject multi-turn conversations where a redacted_thinking block is silently dropped from the assistant message. The shim only matched b.type === 'thinking' when selecting the source for reasoning_content and when filtering redacted content out of the visible text stream, so any Anthropic-side redacted_thinking block never made it onto the OpenAI-shaped outgoing message and the provider 400'd on the next turn. This patch widens the filter at openaiShim.ts:585 to also match redacted_thinking for both the reasoning_content source selection and the textContent exclusion, so the block is preserved in the chain-of-thought stream (it carries no visible text, but the signature is what the provider needs to validate continuity). Adds a focused test in openaiShim.test.ts covering the redacted_thinking path end-to-end: the assistant message that carries a tool_call now has reasoning_content: '' attached even when the only thinking-shaped block is redacted_thinking, which is the field DeepSeek/Kimi strict-validate. Rebased against current upstream/main (c634fd8 9e942da). Noting the rest of the original PR Twigpine#1228 patch series was already absorbed by other merged work and is intentionally not re-applied here, to avoid duplication: - stripThinkingBlocks provider-conditional gating: PR Twigpine#1248 (0d3c157, 2026-05-23) — fix(recovery): keep thinking blocks on resume for reasoning-echo providers. This covers the recovery-path half of the original 1d333b9 patch with a more accurate shouldPreserveThinkingBlocksForProviderReplay helper instead of unconditional inject. - reasoning_content injection gating: PR Twigpine#828 (67de6bd, 2026-04-22) — fix(openai-shim): echo reasoning_content on assistant tool-call messages for Moonshot. This threaded preserveReasoningContent through convertMessages and _doOpenAIRequest with isMoonshotBaseUrl gating, then extended by PR Twigpine#1253 (aab2fbc, 2026-05-19) — fix: MiMo remove unsupported body fields and preserve reasoning content — which split the reasoningContentFallback flag ('omit' for MiMo, '' for DeepSeek/Moonshot/Z.AI). That 3-flag scheme is what the test suite in upstream/main encodes and differs from the unconditional inject in the original 1d333b9 patch. - reasoning_content heuristic tightening: PR Twigpine#1201 (1fc5116, 2026-06-03) — fix(api): tighten reasoning_content heuristic to prevent false-positive provider configuration leaks. This replaced the loose includes() check with a segment-boundary-aware inference and added 10 regression tests in runtimeMetadata.test.ts. - Test coverage carried in the original 64ad59b (4 edge-case tests for DeepSeek V4 reasoning_content continuity): three of the four were not carried forward because their scenarios are already covered upstream: * 'array-content assistant without thinking block still gets reasoning_content: ""' — covered by the tool-call branch in openaiShim.test.ts merged via Twigpine#828. * 'string-content assistant message gets reasoning_content: ""' — covered by the same tool-call branch in Twigpine#828. * 'synthetic interruption assistant message carries reasoning_content: ""' — covered by the filterOrphanedThinkingOnlyMessages path tested in conversationRecovery.test.ts via Twigpine#1248. The fourth, 'redacted_thinking block preserves continuity with reasoning_content: ""', is the only scenario that exercises a code path this patch actually changes (the b.type === 'redacted_thinking' match), so it is retained as the targeted regression test for this commit. The README.md note about DeepSeek V4 thinking mode is not part of this commit — the wider rebase is in upstream and a documentation follow-up will be filed separately if needed. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

#1228) * fix(openai-shim): preserve redacted_thinking blocks for reasoning continuity DeepSeek V4 / Moonshot providers reject multi-turn conversations where a redacted_thinking block is silently dropped from the assistant message. The shim only matched b.type === 'thinking' when selecting the source for reasoning_content and when filtering redacted content out of the visible text stream, so any Anthropic-side redacted_thinking block never made it onto the OpenAI-shaped outgoing message and the provider 400'd on the next turn. This patch widens the filter at openaiShim.ts:585 to also match redacted_thinking for both the reasoning_content source selection and the textContent exclusion, so the block is preserved in the chain-of-thought stream (it carries no visible text, but the signature is what the provider needs to validate continuity). Adds a focused test in openaiShim.test.ts covering the redacted_thinking path end-to-end: the assistant message that carries a tool_call now has reasoning_content: '' attached even when the only thinking-shaped block is redacted_thinking, which is the field DeepSeek/Kimi strict-validate. Rebased against current upstream/main (c634fd8 9e942da). Noting the rest of the original PR #1228 patch series was already absorbed by other merged work and is intentionally not re-applied here, to avoid duplication: - stripThinkingBlocks provider-conditional gating: PR #1248 (0d3c157, 2026-05-23) — fix(recovery): keep thinking blocks on resume for reasoning-echo providers. This covers the recovery-path half of the original 1d333b9 patch with a more accurate shouldPreserveThinkingBlocksForProviderReplay helper instead of unconditional inject. - reasoning_content injection gating: PR #828 (67de6bd, 2026-04-22) — fix(openai-shim): echo reasoning_content on assistant tool-call messages for Moonshot. This threaded preserveReasoningContent through convertMessages and _doOpenAIRequest with isMoonshotBaseUrl gating, then extended by PR #1253 (aab2fbc, 2026-05-19) — fix: MiMo remove unsupported body fields and preserve reasoning content — which split the reasoningContentFallback flag ('omit' for MiMo, '' for DeepSeek/Moonshot/Z.AI). That 3-flag scheme is what the test suite in upstream/main encodes and differs from the unconditional inject in the original 1d333b9 patch. - reasoning_content heuristic tightening: PR #1201 (1fc5116, 2026-06-03) — fix(api): tighten reasoning_content heuristic to prevent false-positive provider configuration leaks. This replaced the loose includes() check with a segment-boundary-aware inference and added 10 regression tests in runtimeMetadata.test.ts. - Test coverage carried in the original 64ad59b (4 edge-case tests for DeepSeek V4 reasoning_content continuity): three of the four were not carried forward because their scenarios are already covered upstream: * 'array-content assistant without thinking block still gets reasoning_content: ""' — covered by the tool-call branch in openaiShim.test.ts merged via #828. * 'string-content assistant message gets reasoning_content: ""' — covered by the same tool-call branch in #828. * 'synthetic interruption assistant message carries reasoning_content: ""' — covered by the filterOrphanedThinkingOnlyMessages path tested in conversationRecovery.test.ts via #1248. The fourth, 'redacted_thinking block preserves continuity with reasoning_content: ""', is the only scenario that exercises a code path this patch actually changes (the b.type === 'redacted_thinking' match), so it is retained as the targeted regression test for this commit. The README.md note about DeepSeek V4 thinking mode is not part of this commit — the wider rebase is in upstream and a documentation follow-up will be filed separately if needed. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(openai-shim): propagate redacted_thinking.data into reasoning_content --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Twigpine#1228) * fix(openai-shim): preserve redacted_thinking blocks for reasoning continuity DeepSeek V4 / Moonshot providers reject multi-turn conversations where a redacted_thinking block is silently dropped from the assistant message. The shim only matched b.type === 'thinking' when selecting the source for reasoning_content and when filtering redacted content out of the visible text stream, so any Anthropic-side redacted_thinking block never made it onto the OpenAI-shaped outgoing message and the provider 400'd on the next turn. This patch widens the filter at openaiShim.ts:585 to also match redacted_thinking for both the reasoning_content source selection and the textContent exclusion, so the block is preserved in the chain-of-thought stream (it carries no visible text, but the signature is what the provider needs to validate continuity). Adds a focused test in openaiShim.test.ts covering the redacted_thinking path end-to-end: the assistant message that carries a tool_call now has reasoning_content: '' attached even when the only thinking-shaped block is redacted_thinking, which is the field DeepSeek/Kimi strict-validate. Rebased against current upstream/main (c634fd8 9e942da). Noting the rest of the original PR Twigpine#1228 patch series was already absorbed by other merged work and is intentionally not re-applied here, to avoid duplication: - stripThinkingBlocks provider-conditional gating: PR Twigpine#1248 (0d3c157, 2026-05-23) — fix(recovery): keep thinking blocks on resume for reasoning-echo providers. This covers the recovery-path half of the original 1d333b9 patch with a more accurate shouldPreserveThinkingBlocksForProviderReplay helper instead of unconditional inject. - reasoning_content injection gating: PR Twigpine#828 (67de6bd, 2026-04-22) — fix(openai-shim): echo reasoning_content on assistant tool-call messages for Moonshot. This threaded preserveReasoningContent through convertMessages and _doOpenAIRequest with isMoonshotBaseUrl gating, then extended by PR Twigpine#1253 (aab2fbc, 2026-05-19) — fix: MiMo remove unsupported body fields and preserve reasoning content — which split the reasoningContentFallback flag ('omit' for MiMo, '' for DeepSeek/Moonshot/Z.AI). That 3-flag scheme is what the test suite in upstream/main encodes and differs from the unconditional inject in the original 1d333b9 patch. - reasoning_content heuristic tightening: PR Twigpine#1201 (1fc5116, 2026-06-03) — fix(api): tighten reasoning_content heuristic to prevent false-positive provider configuration leaks. This replaced the loose includes() check with a segment-boundary-aware inference and added 10 regression tests in runtimeMetadata.test.ts. - Test coverage carried in the original 64ad59b (4 edge-case tests for DeepSeek V4 reasoning_content continuity): three of the four were not carried forward because their scenarios are already covered upstream: * 'array-content assistant without thinking block still gets reasoning_content: ""' — covered by the tool-call branch in openaiShim.test.ts merged via Twigpine#828. * 'string-content assistant message gets reasoning_content: ""' — covered by the same tool-call branch in Twigpine#828. * 'synthetic interruption assistant message carries reasoning_content: ""' — covered by the filterOrphanedThinkingOnlyMessages path tested in conversationRecovery.test.ts via Twigpine#1248. The fourth, 'redacted_thinking block preserves continuity with reasoning_content: ""', is the only scenario that exercises a code path this patch actually changes (the b.type === 'redacted_thinking' match), so it is retained as the targeted regression test for this commit. The README.md note about DeepSeek V4 thinking mode is not part of this commit — the wider rebase is in upstream and a documentation follow-up will be filed separately if needed. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(openai-shim): propagate redacted_thinking.data into reasoning_content --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

jatmn and others added 3 commits May 18, 2026 21:12

fix(MiMo): update integration to remove unnecessary body fields and p…

b07814f

…reserve reasoning content

test(MiMo): enhance tests to verify reasoning_content handling and st…

92bb3bc

…rip unsupported options

Copilot AI review requested due to automatic review settings May 19, 2026 09:22

Copilot started reviewing on behalf of Meetpatel006 May 19, 2026 09:23 View session

Copilot AI reviewed May 19, 2026

View reviewed changes

Meetpatel006 mentioned this pull request May 19, 2026

Fix OpenGateway MiMo agent tool history #1243

Closed

Vasanthdev2004 approved these changes May 19, 2026

View reviewed changes

Vasanthdev2004 requested a review from kevincodex1 May 19, 2026 15:31

kevincodex1 approved these changes May 19, 2026

View reviewed changes

kevincodex1 merged commit aab2fbc into Twigpine:main May 19, 2026
5 of 6 checks passed

github-actions Bot mentioned this pull request May 19, 2026

chore(main): release 0.14.0 #1217

Merged

ajsai47 mentioned this pull request May 30, 2026

fix(status-line): show N/A instead of 0% for providers that strip stream_options (#448) #1444

Closed

WesleySouzaSilva mentioned this pull request Jun 9, 2026

fix(openai-shim): guarantee reasoning_content continuity for DeepSeek… #1228

Merged

4 tasks done

jatmn mentioned this pull request Jun 16, 2026

fix(openai-shim): restore reasoning support for proxy and localhost providers #1155

Closed

Navigation Menu

Repository navigation

fix: MiMo remove unsupported body fields and preserve reasoning content - #1253

fix: MiMo remove unsupported body fields and preserve reasoning content#1253
kevincodex1 merged 3 commits into
Twigpine:mainfrom
Meetpatel006:issue-1218-mimo-subagents

Meetpatel006 commented May 19, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Vasanthdev2004 commented May 19, 2026

Uh oh!

Vasanthdev2004 left a comment

Uh oh!

kevincodex1 left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

Meetpatel006 commented May 19, 2026

Summary

Changes

Test plan

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Vasanthdev2004 commented May 19, 2026

Blockers

Non-Blocking

Looks Good

Uh oh!

Vasanthdev2004 left a comment

Choose a reason for hiding this comment

Uh oh!

kevincodex1 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants