fix(agent): persist streamed reasoning_content on assistant turns (#16844)#16892
Merged
Conversation
…6844) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in #16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs #16844, #16884, #15250, #15353, #15748.
This was referenced Apr 28, 2026
cluricaun28
referenced
this pull request
in cluricaun28/Logos
Apr 28, 2026
…6844) (#16892) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in #16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs #16844, #16884, #15250, #15353, #15748.
ulasbilgen
pushed a commit
to ulasbilgen/hermes-adhd-agent
that referenced
this pull request
May 1, 2026
…usResearch#16844) (NousResearch#16892) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (NousResearch#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (NousResearch#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (NousResearch#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in NousResearch#16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs NousResearch#16844, NousResearch#16884, NousResearch#15250, NousResearch#15353, NousResearch#15748.
donald131
pushed a commit
to donald131/hermes-agent
that referenced
this pull request
May 2, 2026
…usResearch#16844) (NousResearch#16892) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (NousResearch#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (NousResearch#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (NousResearch#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in NousResearch#16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs NousResearch#16844, NousResearch#16884, NousResearch#15250, NousResearch#15353, NousResearch#15748.
02356abc
pushed a commit
to 02356abc/hermes-agent
that referenced
this pull request
May 14, 2026
…usResearch#16844) (NousResearch#16892) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (NousResearch#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (NousResearch#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (NousResearch#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in NousResearch#16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs NousResearch#16844, NousResearch#16884, NousResearch#15250, NousResearch#15353, NousResearch#15748.
dannyJ848
pushed a commit
to dannyJ848/hermes-agent
that referenced
this pull request
May 17, 2026
…usResearch#16844) (NousResearch#16892) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (NousResearch#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (NousResearch#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (NousResearch#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in NousResearch#16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs NousResearch#16844, NousResearch#16884, NousResearch#15250, NousResearch#15353, NousResearch#15748.
gweeteve
pushed a commit
to gweeteve/hermes-agent
that referenced
this pull request
Jun 2, 2026
…usResearch#16844) (NousResearch#16892) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (NousResearch#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (NousResearch#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (NousResearch#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in NousResearch#16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs NousResearch#16844, NousResearch#16884, NousResearch#15250, NousResearch#15353, NousResearch#15748.
Egavasyug
pushed a commit
to Egavasyug/hermes-agent
that referenced
this pull request
Jun 10, 2026
…usResearch#16844) (NousResearch#16892) Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) emit reasoning through delta.reasoning_content chunks that get accumulated into the local reasoning_text string — but never land on the assistant message object as a top-level attribute. The prior guard at _build_assistant_message only wrote reasoning_content when the SDK exposed hasattr(msg, 'reasoning_content'), so these providers persisted the chain-of-thought under the internal 'reasoning' key and omitted the protocol-standard field. The poison was silent until the user later switched to a DeepSeek-v4 or Kimi thinking model, at which point replay failed with HTTP 400: 'The reasoning_content in the thinking mode must be passed back to the API.' One reported session store accumulated 4,031 poisoned messages across 1,101 files (NousResearch#16844). Fix: add an additive fallback that promotes the already-sanitized reasoning_text to reasoning_content when no earlier branch wrote it AND reasoning text was actually captured. Layered on top of the existing SDK-attr branch and DeepSeek ''-pad (NousResearch#15250) rather than replacing them, so every existing behavior is preserved: - SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still wins. - DeepSeek tool-call ''-pad still fires when the SDK exposes the attr but the value is None. - Non-thinking turns with no reasoning leave the field absent, so _copy_reasoning_content_for_api's cross-provider leak guard (NousResearch#15748), promote-from-'reasoning' tier, and thinking-pad tier remain live at replay time. - No empty '' gets eagerly written on every assistant turn (which would have bypassed the read-side ladder and triggered empty thinking-block insertion in the Anthropic adapter). Tests: three new TestBuildAssistantMessage cases covering the streaming promotion path, SDK precedence, and field-absent-when-no-reasoning invariant. Credit @Sanjays2402 for the original diagnosis and patch in NousResearch#16884; this is a scoped rework that preserves the existing read-side compensation code as defense in depth. Refs NousResearch#16844, NousResearch#16884, NousResearch#15250, NousResearch#15353, NousResearch#15748.
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.
Supersedes #16884 with a scoped rework that preserves existing read-side compensation.
Summary
Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via openai-compat shims) accumulate reasoning through
delta.reasoning_contentchunks but never expose it as a top-level attribute on the finalized SDK message. The existinghasattr-guarded block at_build_assistant_messagetherefore never wrotereasoning_contentfor those providers, so the chain-of-thought was persisted only under the internalreasoningkey.The poison is silent until the user later switches to a DeepSeek-v4 or Kimi thinking model, at which point replay 400s with "The reasoning_content in the thinking mode must be passed back to the API." Issue #16844 reports 4,031 poisoned messages across 1,101 session files on one install.
Changes
run_agent.py_build_assistant_message: additive fallback promotes the already-sanitizedreasoning_texttoreasoning_contentwhen no earlier branch wrote it and reasoning text was actually captured. Existing SDK-attr branch and DeepSeek""-pad ([Bug]: DeepSeek V4 Flash Discord session poisoned when tool-call assistant lacks reasoning_content #15250) are untouched.tests/run_agent/test_run_agent.py: 3 regression tests — streaming promotion path, SDK precedence, field-absent-when-no-reasoning invariant.Why not #16884 as-written
#16884 replaced the conditional with
msg["reasoning_content"] = ""as a universal fallback, which would have:isinstance(reasoning_content, str)branch to prepend an empty{"type":"thinking","thinking":""}block on every replayed assistant turn.reasoning_content: ""to every strict OpenAI-compatible provider (Mistral, Fireworks, stock OpenAI, GitHub Models)._copy_reasoning_content_for_api's step-1 string check on every turn, making tiers 2–4 dead code — including [Bug] _copy_reasoning_content_for_api: cross-provider reasoning promotion leaks stale content to DeepSeek/Kimi #15748's cross-provider reasoning-leak guard for DeepSeek/Kimi.The layered approach here solves the write-side bug without changing the field's presence semantics for non-thinking sessions. Existing read-side ladder (cross-provider leak guard #15748, promote-from-
reasoning, DeepSeek/Kimi thinking-pad) stays live as defense in depth.Validation
reasoning_content→ 400 on DeepSeek/Kimi replayreasoning_content""-pad (#15250)_copy_reasoning_content_for_apitiers 2–4TestBuildAssistantMessage+test_deepseek_reasoning_content_echoCredit @Sanjays2402 for the original diagnosis in #16884.
Refs #16844, #16884, #15250, #15353, #15748.