hermes - ✅(Solved) Fix [Bug] _copy_reasoning_content_for_api: cross-provider reasoning promotion leaks stale content to DeepSeek/Kimi [3 pull requests, 1 comments, 2 participants]

hermes2026-04-25 18:05:55

ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

GitHub issue URL

Helpful · Quick feedback

GitHub stats

NousResearch/hermes-agent#15748•Fetched 2026-04-26 05:25:20

View on GitHub

Comments

Participants

Timeline

Reactions

Author

Zjianru

Participants

alt-glitch

Zjianru

Timeline (top)

labeled ×5cross-referenced ×4commented ×1

Root Cause

In _copy_reasoning_content_for_api (run_agent.py), the normalized_reasoning promotion path returns before the DeepSeek/Kimi empty-string guard can execute:

# Current (buggy) ordering in origin/main:
explicit_reasoning = source_msg.get("reasoning_content")
if isinstance(explicit_reasoning, str):        # False when switching from MiniMax
    api_msg["reasoning_content"] = explicit_reasoning
    return

normalized_reasoning = source_msg.get("reasoning")  # "MiniMax chain of thought..."
if isinstance(normalized_reasoning, str) and normalized_reasoning:
    api_msg["reasoning_content"] = normalized_reasoning  # ← LEAKED
    return                                            # ← DeepSeek/Kimi guard never reached

# This guard is unreachable for cross-provider histories
if source_msg.get("tool_calls") and (self._needs_kimi...() or self._needs_deepseek...()):
    api_msg["reasoning_content"] = ""

When the session history contains a tool-call message from MiniMax with reasoning="MiniMax thinking..." and no reasoning_content, switching to DeepSeek causes:

reasoning_content key absent → first check returns False
reasoning key present → promoted to reasoning_content
DeepSeek receives MiniMax's reasoning → HTTP 400

Fix Action

Fix

Reorder the logic so that:

reasoning_content already set → preserve verbatim (including DeepSeek's own "" placeholder written at creation time)
Tool-call turns with neither reasoning_content nor reasoning → inject "" for DeepSeek/Kimi (poisoned history from prior provider)
reasoning present → promote to reasoning_content (healthy same-provider path)

The key addition is not has_reasoning in the guard, which distinguishes "this provider has reasoning to promote" from "a prior provider left reasoning that should not be forwarded".

PR fix notes

PR #15760: Fix DeepSeek reasoning_content propagation

Repository: NousResearch/hermes-agent
Author: LMTRADES
State: open | merged: False
Link: https://github.com/NousResearch/hermes-agent/pull/15760

Description (problem / solution / changelog)

This PR fixes two issues with DeepSeek reasoning_content. 1. Extracts reasoning_content from model_extra for ZenMux responses. 2. Adds empty string fallback for DeepSeek to non-tool-call messages.

Changed files

run_agent.py (modified, +32/-10)

PR #15761: fix: ensure DeepSeek reasoning_content guard runs after tool_calls are populated

Repository: NousResearch/hermes-agent
Author: unknown-gezi
State: open | merged: False
Link: https://github.com/NousResearch/hermes-agent/pull/15761

Description (problem / solution / changelog)

Summary

Fixes recurring HTTP 400 error when using DeepSeek thinking mode:

The reasoning_content in the thinking mode must be passed back to the API.

Root Cause

The _build_assistant_message method in run_agent.py had a timing bug: the DeepSeek reasoning_content guard at L7642 checked msg.get("tool_calls") but tool_calls were only added to msg at L7672 — 30 lines later. This rendered the guard permanently dead.

Consequence: DeepSeek tool-call messages were persisted to the session store WITHOUT reasoning_content. On session reload + replay, _copy_reasoning_content_for_api injected an empty-string fallback (""), which DeepSeek"s API rejects with HTTP 400.

Changes

`run_agent.py` — 2 fixes

Move the DeepSeek guard to AFTER tool_calls are populated (after msg["tool_calls"] = tool_calls). This ensures the guard actually fires when needed, pinning reasoning_content: "" at creation time for tool-call messages that lack raw reasoning from the API response.
Remove empty-string injection from _copy_reasoning_content_for_api. The reply path no longer injects "" for poisoned history replays — the creation-time pin (fix 1) prevents new poison, and empty strings are rejected by DeepSeek anyway.

`tests/run_agent/test_deepseek_reasoning_content_echo.py` — 4 test updates

Updated tests to expect reasoning_content ABSENCE rather than empty-string injection:

test_deepseek_tool_call_poisoned_history_skipped (was _gets_empty_string)
test_kimi_path_left_alone (was test_kimi_path_still_works)
test_kimi_moonshot_base_url_left_alone (was test_kimi_moonshot_base_url)
test_deepseek_custom_base_url_left_alone (was test_deepseek_custom_base_url)

Test Plan

All 21 tests in test_deepseek_reasoning_content_echo.py pass
All 11 TestBuildAssistantMessage tests in test_run_agent.py pass

Notes

Existing sessions already poisoned by the old bug may need /new to reset — the persisted messages still lack reasoning_content. New messages will not be poisoned.
The same fix also covers Kimi/Moonshot thinking mode behaviour via the shared _copy_reasoning_content_for_api path.

Refs #15250, #15353

Changed files

run_agent.py (modified, +21/-15)
tests/run_agent/test_deepseek_reasoning_content_echo.py (modified, +17/-15)

PR #15830: fix(reasoning): promote reasoning field before empty-pad for DeepSeek/Kimi (#15812)

Repository: NousResearch/hermes-agent
Author: briandevans
State: open | merged: False
Link: https://github.com/NousResearch/hermes-agent/pull/15830

Description (problem / solution / changelog)

Summary

_copy_reasoning_content_for_api had a step-ordering bug introduced by PR #15749: the tool-call empty-string injection (old step 2) ran before the reasoning field promotion (old step 3)
A tool-call message with a valid reasoning field but no reasoning_content would hit the empty-string path, get reasoning_content="" injected, and return early — the reasoning value was silently discarded
Fix: move the reasoning → reasoning_content promotion to step 2, before the empty-string fallbacks

The bug

run_agent.py — _copy_reasoning_content_for_api (pre-fix):

# Step 1: explicit reasoning_content → preserve ✓
# Step 2: tool_calls + DeepSeek/Kimi → inject ""  ← fires here
#         RETURNS EARLY
# Step 3: reasoning field → promote               ← never reached

A message like {"role": "assistant", "reasoning": "thought trace", "tool_calls": [...]} on a DeepSeek session would return reasoning_content="" instead of reasoning_content="thought trace".

The fix

Reorder steps 2 and 3 so reasoning promotion runs first:

# Step 1: explicit reasoning_content → preserve
# Step 2: reasoning field present → promote to reasoning_content  ← moved up
# Step 3: tool_calls + DeepSeek/Kimi + no reasoning → inject ""
# Step 4: all other DeepSeek/Kimi + no reasoning → inject ""
# Step 5: non-string cleanup

Empty-string injection now only fires when neither reasoning_content nor reasoning is present — the poisoned-history case it was designed for.

Test plan

Before: test_deepseek_reasoning_field_promoted → AssertionError: assert '' == 'thought trace'
After: all 21 tests in test_deepseek_reasoning_content_echo.py pass
Regression guard: reverted run_agent.py → test_deepseek_reasoning_field_promoted failed with '' != 'thought trace'; restored → 21 passing
Broader suite: tests/run_agent/ — 1048 passed, 1 pre-existing baseline failure (test_inf_stays_string_for_integer_only reproduces on clean origin/main)

Fixes #15812
Regression introduced by #15749
Related: #15748 (same code path)

🤖 Generated with Claude Code

Changed files

run_agent.py (modified, +14/-12)

Code Example

# Current (buggy) ordering in origin/main:
explicit_reasoning = source_msg.get("reasoning_content")
if isinstance(explicit_reasoning, str):        # False when switching from MiniMax
    api_msg["reasoning_content"] = explicit_reasoning
    return

normalized_reasoning = source_msg.get("reasoning")  # "MiniMax chain of thought..."
if isinstance(normalized_reasoning, str) and normalized_reasoning:
    api_msg["reasoning_content"] = normalized_reasoning  # ← LEAKED
    return                                            # ← DeepSeek/Kimi guard never reached

# This guard is unreachable for cross-provider histories
if source_msg.get("tool_calls") and (self._needs_kimi...() or self._needs_deepseek...()):
    api_msg["reasoning_content"] = ""

RAW_BUFFERClick to expand / collapse

Bug: Cross-provider reasoning promotion leaks stale content to DeepSeek/Kimi thinking mode

Background

DeepSeek V4 and Kimi/Moonshot thinking modes require reasoning_content="" on every assistant tool-call message. If the field is missing on replay, the API returns HTTP 400:

The reasoning_content in the thinking mode must be passed back to the API.

When a session switches providers mid-conversation (e.g. MiniMax → DeepSeek), the internal reasoning field from the prior provider can leak into the reasoning_content field sent to the new provider, causing HTTP 400.

Root Cause

In _copy_reasoning_content_for_api (run_agent.py), the normalized_reasoning promotion path returns before the DeepSeek/Kimi empty-string guard can execute:

# Current (buggy) ordering in origin/main:
explicit_reasoning = source_msg.get("reasoning_content")
if isinstance(explicit_reasoning, str):        # False when switching from MiniMax
    api_msg["reasoning_content"] = explicit_reasoning
    return

normalized_reasoning = source_msg.get("reasoning")  # "MiniMax chain of thought..."
if isinstance(normalized_reasoning, str) and normalized_reasoning:
    api_msg["reasoning_content"] = normalized_reasoning  # ← LEAKED
    return                                            # ← DeepSeek/Kimi guard never reached

# This guard is unreachable for cross-provider histories
if source_msg.get("tool_calls") and (self._needs_kimi...() or self._needs_deepseek...()):
    api_msg["reasoning_content"] = ""

When the session history contains a tool-call message from MiniMax with reasoning="MiniMax thinking..." and no reasoning_content, switching to DeepSeek causes:

reasoning_content key absent → first check returns False
reasoning key present → promoted to reasoning_content
DeepSeek receives MiniMax's reasoning → HTTP 400

Fix

Reorder the logic so that:

reasoning_content already set → preserve verbatim (including DeepSeek's own "" placeholder written at creation time)
Tool-call turns with neither reasoning_content nor reasoning → inject "" for DeepSeek/Kimi (poisoned history from prior provider)
reasoning present → promote to reasoning_content (healthy same-provider path)

The key addition is not has_reasoning in the guard, which distinguishes "this provider has reasoning to promote" from "a prior provider left reasoning that should not be forwarded".

Affected Scenarios

Scenario	Before fix	After fix
DeepSeek tool_call with `reasoning_content=""` set	preserved	preserved
DeepSeek tool_call with `reasoning="DeepSeek reasoning"`	promoted	promoted
DeepSeek tool_call after MiniMax (MiniMax `reasoning` present)	MiniMax content sent → 400	`""` injected
DeepSeek tool_call after MiniMax (both fields absent)	`""` injected	`""` injected

Related Issues

#15250 — DeepSeek V4 Flash Discord session poisoned when tool-call assistant lacks reasoning_content
#15213 — HTTP 400 in cron/auxiliary path (same root cause, different trigger)
PR #15228 — earlier fix that introduced this ordering issue
PR #14973 — extract DeepSeek reasoning_content from model_extra

Test Coverage

All 21 existing tests in tests/run_agent/test_deepseek_reasoning_content_echo.py pass with this fix.

extent analysis

TL;DR

Reorder the logic in _copy_reasoning_content_for_api to prioritize checking for reasoning_content and injecting an empty string when necessary, before promoting reasoning to reasoning_content.

Guidance

Verify that the reasoning_content field is being set correctly for each provider, especially when switching between providers like MiniMax and DeepSeek.
Check the tool_calls history to ensure that the reasoning_content field is not being overwritten with stale content from a previous provider.
Update the _copy_reasoning_content_for_api function to follow the reordered logic:
1. Preserve reasoning_content if already set.
2. Inject "" for DeepSeek/Kimi if neither reasoning_content nor reasoning is present.
3. Promote reasoning to reasoning_content if reasoning is present.
Review the test coverage in tests/run_agent/test_deepseek_reasoning_content_echo.py to ensure that all scenarios are properly tested.

Example

def _copy_reasoning_content_for_api(source_msg, api_msg):
    # Check if reasoning_content is already set
    if "reasoning_content" in source_msg:
        api_msg["reasoning_content"] = source_msg["reasoning_content"]
        return

    # Check if tool-call turns have neither reasoning_content nor reasoning
    if source_msg.get("tool_calls") and (self._needs_kimi() or self._needs_deepseek()):
        api_msg["reasoning_content"] = ""
        return

    # Promote reasoning to reasoning_content if present
    normalized_reasoning = source_msg.get("reasoning")
    if isinstance(normalized_reasoning, str) and normalized_reasoning:
        api_msg["reasoning_content"] = normalized_reasoning

Notes

This fix assumes that the reasoning_content

Vote matrix · Quick signals

Works

Did the solution work? Tap to confirm.

Easy Fix

Was it a quick fix?

Time Saver

Did it save you time?

Blocking

Was it severely blocking?

Common Issue

Are others likely hitting this too?

Flaky / Intermittent

Is it intermittent?

Verified / Reproducible

Can you reproduce it reliably?

#api #output truncation #response parsing #generation error #database connection

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

hermes - ✅(Solved) Fix [Bug] _copy_reasoning_content_for_api: cross-provider reasoning promotion leaks stale content to DeepSeek/Kimi [3 pull requests, 1 comments, 2 participants]

Recommended Tools

GitHub issue graph ai analysis

Root Cause

Fix Action

Fix

PR fix notes

PR #15760: Fix DeepSeek reasoning_content propagation

Description (problem / solution / changelog)

Changed files

PR #15761: fix: ensure DeepSeek reasoning_content guard runs after tool_calls are populated

Description (problem / solution / changelog)

Summary

Root Cause

Changes

run_agent.py — 2 fixes

tests/run_agent/test_deepseek_reasoning_content_echo.py — 4 test updates

Test Plan

Notes

Changed files

PR #15830: fix(reasoning): promote reasoning field before empty-pad for DeepSeek/Kimi (#15812)

Description (problem / solution / changelog)

Summary

The bug

The fix

Test plan

Related

Changed files

Code Example

Bug: Cross-provider reasoning promotion leaks stale content to DeepSeek/Kimi thinking mode

Background

Root Cause

Fix

Affected Scenarios

Related Issues

Test Coverage

extent analysis

TL;DR

Guidance

Example

Notes

Still need to ship something?

RELATED_DISCOVERY

TRENDING

`run_agent.py` — 2 fixes

`tests/run_agent/test_deepseek_reasoning_content_echo.py` — 4 test updates