[Bug / Potential Fix] DeepSeek V4 Flash/Pro: reasoning_content must be passed back to the API (400) on assistant messages with tool_calls

# [Bug]: DeepSeek V4 — 400 error caused by `reasoning_content` missing on edge-case assistant messages

> **Author's note:** This report builds on the foundational DeepSeek V4 support added by **Jatmn** in commit `ff2a380` (`Add DeepSeek V4 flash/pro support and DeepSeek thinking compatibility #877`). That commit appears to have provided the core infrastructure: DeepSeek URL detection, `thinking: { type: 'enabled' }` support, and the `preserveReasoningContent` mechanism. We are very grateful for that work.
>
> **What we found:** After testing extensively with `deepseek-v4-pro`, we found what appear to be **four edge cases** not covered by `ff2a380` where `reasoning_content` was not being set, which in our DeepSeek reproductions led to `400` errors in multi-turn conversations with tool use. The original commit seems to handle the happy path well (assistant messages that **do** contain thinking blocks), but may miss assistant messages without thinking blocks, string-content messages, and synthetic interrupt messages.
>
> Our validation for this report was performed against **DeepSeek**. While the proposed explanation and impact assessment seem reasonable, we are **not fully certain** there are no secondary effects in other paths or providers, so review from maintainers with deeper context would be greatly appreciated.

---

## Summary

When using OpenClaude with **DeepSeek V4** (`deepseek-v4-flash`, `deepseek-v4-pro`) as an OpenAI-compatible provider, **conversations that involve tool calls can fail** with a `400 invalid_request_error`. The error message is:

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

This appears to happen because:
1. DeepSeek V4 enables **thinking mode by default**, which emits `reasoning_content` on assistant messages.
2. Commit `ff2a380` added the infrastructure to capture `thinking` blocks and re-attach them as `reasoning_content`.
3. **However, what appear to be four edge cases were not covered**, causing `reasoning_content` to be `undefined` on some assistant messages.
4. In the DeepSeek scenarios we validated, the request is rejected if an assistant message in thinking mode omits `reasoning_content`.

This issue appears to affect **tool-calling conversations with DeepSeek V4**, especially **subagents / Task tool invocations** where tools are used repeatedly across turns.

---

## Environment

| Component | Version / Setting |
|---|---|
| OpenClaude | `v0.6.0` |
| Provider | `openai-compatible` |
| Model | `deepseek-v4-flash`, `deepseek-v4-pro` |
| Base URL | `https://api.deepseek.com/v1` |
| `OPENAI_MAX_TOKENS` | `262144` |

---

## Error Message

```text
API Error: 400 {"error":{"message":"The `reasoning_content` in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}
```

---

## DeepSeek's Documented Rule

> *"When `thinking: { type: 'enabled' }` is active, **all** `role=assistant` messages in the conversation history must include `reasoning_content`."*
>
> — [api-docs.deepseek.com/guides/thinking_mode](https://api-docs.deepseek.com/guides/thinking_mode)

**Based on the docs and the API behavior we observed, the practical rule seems to be:**
- Assistant message **has** a thinking block → echo the exact `reasoning_content`.
- Assistant message **does NOT have** a thinking block → send `reasoning_content: ""` (empty string).
- Assistant message **omits the property entirely** (`undefined`) → **400** in the DeepSeek scenarios reproduced here.

---

## Validation: 5 CURLs Against Real DeepSeek API

### CURL 1 — First request (user asks for tool use)

```bash
curl https://api.deepseek.com/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $DEEPSEEK_API_KEY"   -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "user", "content": "What time is it? Use the get_current_time tool."}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_current_time",
          "description": "Get current time",
          "parameters": {"type": "object", "properties": {}}
        }
      }
    ]
  }'
```

**Response (200 OK):**
```json
{
    "id": "eeae50f5-f00c-4556-a999-200239f6ad0d",
    "object": "chat.completion",
    "created": 1777007567,
    "model": "deepseek-v4-flash",
    "choices": [{
        "index": 0,
        "message": {
            "role": "assistant",
            "content": "",
            "reasoning_content": "The user wants to know the current time. Let me use the get_current_time tool.",
            "tool_calls": [{
                "index": 0,
                "id": "call_00_9nLfpIQ6VO1JCaEsJcU8WZnS",
                "type": "function",
                "function": {
                    "name": "get_current_time",
                    "arguments": "{}"
                }
            }]
        },
        "finish_reason": "tool_calls"
    }],
    "usage": {
        "prompt_tokens": 267,
        "completion_tokens": 47,
        "total_tokens": 314,
        "prompt_tokens_details": {"cached_tokens": 256},
        "completion_tokens_details": {"reasoning_tokens": 18},
        "prompt_cache_hit_tokens": 256,
        "prompt_cache_miss_tokens": 11
    }
}
```

DeepSeek returns `content: ""`, `reasoning_content`, and `tool_calls` — the standard agent format. `reasoning_tokens: 18` in usage suggests that thinking mode is active.

---

### CURL 2 — Second request WITH `reasoning_content` → 200 OK

```bash
curl https://api.deepseek.com/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $DEEPSEEK_API_KEY"   -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "user", "content": "What time is it? Use the get_current_time tool."},
      {
        "role": "assistant",
        "content": "Let me check the current time for you.",
        "reasoning_content": "The user wants to know the current time. Let me use the get_current_time tool.",
        "tool_calls": [
          {
            "id": "call_00_9nLfpIQ6VO1JCaEsJcU8WZnS",
            "type": "function",
            "function": {
              "name": "get_current_time",
              "arguments": "{}"
            }
          }
        ]
      },
      {"role": "tool", "tool_call_id": "call_00_9nLfpIQ6VO1JCaEsJcU8WZnS", "content": "10:30 AM UTC"},
      {"role": "user", "content": "Thanks! What about tomorrow?"}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_current_time",
          "description": "Get current time",
          "parameters": {"type": "object", "properties": {}}
        }
      }
    ]
  }'
```

**Response (200 OK):** Validated — conversation continues normally.

---

### CURL 3 — Second request WITHOUT `reasoning_content` → 400 ERROR

```bash
curl https://api.deepseek.com/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $DEEPSEEK_API_KEY"   -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "user", "content": "What time is it? Use the get_current_time tool."},
      {
        "role": "assistant",
        "content": "Let me check the current time for you.",
        "tool_calls": [
          {
            "id": "call_00_9nLfpIQ6VO1JCaEsJcU8WZnS",
            "type": "function",
            "function": {
              "name": "get_current_time",
              "arguments": "{}"
            }
          }
        ]
      },
      {"role": "tool", "tool_call_id": "call_00_9nLfpIQ6VO1JCaEsJcU8WZnS", "content": "10:30 AM UTC"},
      {"role": "user", "content": "Thanks! What about tomorrow?"}
    ],
    "tools": [
      {
        "type": "function",
        "function": {
          "name": "get_current_time",
          "description": "Get current time",
          "parameters": {"type": "object", "properties": {}}
        }
      }
    ]
  }'
```

**Response (400 ERROR):**
```json
{"error":{"message":"The `reasoning_content` in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}
```

In this reproduction, the relevant difference from CURL 2 is the **absence of `reasoning_content`** on the assistant message.

---

### CURL 4 — Agent format `content: ""` + `tool_calls` WITHOUT reasoning → 400

Exact format used by OpenClaude agent flows:

```bash
curl https://api.deepseek.com/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $DEEPSEEK_API_KEY"   -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Run ls"},
      {"role": "assistant", "content": "", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "Bash", "arguments": "{"command": "ls"}"}}]},
      {"role": "tool", "tool_call_id": "call_1", "content": "file.txt"}
    ],
    "stream": false
  }'
```

**Response (400 ERROR):**
```json
{"error":{"message":"The `reasoning_content` in the thinking mode must be passed back to the API.","type":"invalid_request_error","param":null,"code":"invalid_request_error"}}
```

---

### CURL 5 — Agent format `content: ""` + `tool_calls` WITH reasoning → 200 OK

```bash
curl https://api.deepseek.com/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $DEEPSEEK_API_KEY"   -d '{
    "model": "deepseek-v4-flash",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Run ls"},
      {"role": "assistant", "content": "", "reasoning_content": "The user wants me to run ls. I will call the Bash tool.", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "Bash", "arguments": "{"command": "ls"}"}}]},
      {"role": "tool", "tool_call_id": "call_1", "content": "file.txt"}
    ],
    "stream": false
  }'
```

**Response:** 200 OK

---

## Validation Summary

| # | Format | `content` | `reasoning_content` | Result |
|---|---|---|---|---|
| 1 | First request (user only) | N/A | N/A | 200 OK |
| 2 | Multi-turn WITH reasoning | `"Let me check..."` | ✅ Present | **200 OK** |
| 3 | Multi-turn WITHOUT reasoning | `"Let me check..."` | ❌ Absent | **400 ERROR** |
| 4 | Agent format WITHOUT reasoning | `""` | ❌ Absent | **400 ERROR** |
| 5 | Agent format WITH reasoning | `""` | ✅ Present | **200 OK** |

---

## What Appear to Be the Four Edge Cases Not Covered by `ff2a380`

> **Important:** These edge cases only appear to matter when **thinking mode is enabled** (`thinking: { type: 'enabled' }`). When thinking mode is OFF, DeepSeek does not seem to require `reasoning_content`, so these code paths may be irrelevant.

Commit `ff2a380` appears to correctly handle the **happy path**: assistant messages that contain a `thinking` block. However, in real-world multi-turn conversations with thinking mode ON, we found four situations where `reasoning_content` was not being set:

### Edge Case 1: Assistant with array `content` but NO thinking block

When the model calls tools without emitting a visible thinking block, the `hasThinkingBlock` guard prevented `reasoning_content` from being set at all.

```typescript
// BEFORE (ff2a380, suspected bug):
if (preserveReasoningContent) {
  const thinkingText = (thinkingBlock as { thinking?: string })?.thinking
  if (typeof thinkingText === 'string' && thinkingText.trim().length > 0) {
    assistantMsg.reasoning_content = thinkingText  // <-- only if thinking exists
  }
}

// AFTER (proposed fix):
const reasoningContent = preserveReasoningContent
  ? ((thinkingBlock)?.thinking ?? '')  // <-- empty string if no thinking block
  : undefined
const shouldAttachReasoning = preserveReasoningContent  // <-- always for DeepSeek
```

### Edge Case 2: Assistant with `content` as a string (not an array)

When `content` is a plain string (not an array of blocks), the `else` branch in `convertMessages()` created the assistant message without `reasoning_content`.

```typescript
// AFTER (proposed fix):
} else {
  const assistantMsg = {
    role: 'assistant',
    content: ...,
    ...(preserveReasoningContent && {
      reasoning_content: '',
    }),
  }
}
```

### Edge Case 3: Synthetic `"[Tool execution interrupted by user]"` message

The shim injects a synthetic assistant message when a `tool` message is followed by a `user` message (required for OpenAI/Mistral role alternation). This message was created as a bare object literal outside of `convertMessages()`, so it never received `reasoning_content`.

```typescript
// BEFORE (suspected bug):
coalesced.push({
  role: 'assistant',
  content: '[Tool execution interrupted by user]',
})

// AFTER (proposed fix):
coalesced.push({
  role: 'assistant',
  content: '[Tool execution interrupted by user]',
  ...(preserveReasoningContent && {
    reasoning_content: '',
  }),
})
```

### Edge Case 4: `conversationRecovery.ts` stripping thinking blocks

`stripThinkingBlocks()` removed all `thinking`/`redacted_thinking` blocks during conversation deserialization for 3P providers. This prevented the shim from ever seeing the thinking blocks to convert them to `reasoning_content`.

```diff
- function stripThinkingBlocks(messages) { ... }  // REMOVED
- const thinkingStripped = isThirdPartyProvider
-   ? stripThinkingBlocks(filteredThinking)
-   : filteredThinking
- const filteredMessages = filterWhitespaceOnlyAssistantMessages(thinkingStripped)
+ const filteredMessages = filterWhitespaceOnlyAssistantMessages(filteredThinking)
```

---

## Why We Expect Limited Impact on Other Providers

All four changes are gated behind `preserveReasoningContent`, which is **only** `true` for DeepSeek and Moonshot:

```typescript
// openaiShim.ts
preserveReasoningContent:
  isMoonshotBaseUrl(request.baseUrl) || isDeepSeekBaseUrl(request.baseUrl),
```

For providers where `preserveReasoningContent` remains `false` (OpenAI, Azure, Ollama, LM Studio, OpenRouter, Together, Groq, Fireworks, Mistral, Vertex, Bedrock, etc.), these changes are expected to be no-ops:

- `reasoningContent` evaluates to `undefined`
- `shouldAttachReasoning` is `false`
- All `...(preserveReasoningContent && ...)` spreads remain inactive

### Honest uncertainty

Our validation was performed against **DeepSeek**. We have **not** exhaustively validated these paths end-to-end against other providers, so the non-DeepSeek impact assessment should be treated as a reasoned expectation rather than a fully verified claim.

A few specific uncertainties:
- **Moonshot / Kimi**: Since `preserveReasoningContent` is also `true` there, these changes may affect it as well. We have **not** validated that behavior directly.
- **Future providers**: Any provider later added to the `preserveReasoningContent` list would inherit these rules.

---

## Files Modified

| File | Change |
|------|--------|
| `src/services/api/openaiShim.ts` | Edge cases 1, 2, 3: `reasoning_content` gating |
| `src/utils/conversationRecovery.ts` | Edge case 4: removed `stripThinkingBlocks()` |
| `src/utils/conversationRecovery.hooks.test.ts` | Updated test to reflect preserved thinking blocks |

---

## Why Subagents Are Most Affected

Agent flows in OpenClaude often use a pattern where **many turns involve tool calls**:

```text
Turn 1: user -> assistant (+ reasoning_1 + tool_call_1) -> tool result
Turn 2: user -> assistant (+ reasoning_2 + tool_call_2) -> tool result
Turn 3: user -> assistant (+ reasoning_3 + tool_call_3) -> tool result
...
```

Each assistant message with `tool_calls` can carry `reasoning_content` that must be preserved. In a long agent session, the history may contain **many assistant messages** that need their `reasoning_content` echoed back. The four edge cases described above appear more likely to show up as the conversation grows.

Simple chat without tools works fine because no `tool_calls` means no `reasoning_content` requirement.

---

## Request for Review

We are not core contributors and our understanding of the codebase is limited. We would greatly appreciate review from:

- **Jatmn** — author of the original DeepSeek support (`ff2a380`) @jatmn
- Anyone familiar with the OpenAI shim architecture
- Anyone with access to test Moonshot / Kimi (the other provider currently using `preserveReasoningContent`)

Please help verify:
1. Whether the gating logic is sufficient to protect other providers.
2. Whether sending `reasoning_content: ""` (empty string) is the right fallback for assistant messages that never had thinking content.
3. Whether removing `stripThinkingBlocks()` could have unintended effects on Moonshot or other 3P provider recovery paths.
4. Whether there are cleaner places in the codebase to address these edge cases.

---

## Related Issues

- #859 — bug: `reasoning_content missing problem` (Moonshot — fixed by commit `67de6bd2`)
- #876 — bug: DeepSeek v4 API returning error (same 400 error, this report appears related)

---

## References

- Commit [ `ff2a3807230101b6e170b5ecea8f6aef7a90ece9`](https://github.com/Gitlawb/openclaude/commit/ff2a3807230101b6e170b5ecea8f6aef7a90ece) — `Add DeepSeek V4 flash/pro support and DeepSeek thinking compatibility (#877)` by Jatmn
- Commit [`67de6bd2cffc3381f0f28fd3ffce043970611667`](https://github.com/Gitlawb/openclaude/commit/67de6bd2cffc3381f0f28fd3ffce043970611667) — `fix(openai-shim): echo reasoning_content on assistant tool-call messages for Moonshot (#828)`
- [DeepSeek Thinking Mode Documentation](https://api-docs.deepseek.com/guides/thinking_mode)
- [DeepSeek Models & Pricing](https://api-docs.deepseek.com/quick_start/pricing)


File	Change
`src/services/api/openaiShim.ts`	Edge cases 1, 2, 3: `reasoning_content` gating
`src/utils/conversationRecovery.ts`	Edge case 4: removed `stripThinkingBlocks()`
`src/utils/conversationRecovery.hooks.test.ts`	Updated test to reflect preserved thinking blocks

Navigation Menu

[Bug / Potential Fix] DeepSeek V4 Flash/Pro: reasoning_content must be passed back to the API (400) on assistant messages with tool_calls #878

[Bug]: DeepSeek V4 — 400 error caused by `reasoning_content` missing on edge-case assistant messages

Summary

Environment

Error Message

DeepSeek's Documented Rule

Validation: 5 CURLs Against Real DeepSeek API

CURL 1 — First request (user asks for tool use)

CURL 2 — Second request WITH `reasoning_content` → 200 OK

CURL 3 — Second request WITHOUT `reasoning_content` → 400 ERROR

CURL 4 — Agent format `content: ""` + `tool_calls` WITHOUT reasoning → 400

CURL 5 — Agent format `content: ""` + `tool_calls` WITH reasoning → 200 OK

Validation Summary

What Appear to Be the Four Edge Cases Not Covered by `ff2a380`

Edge Case 1: Assistant with array `content` but NO thinking block

Edge Case 2: Assistant with `content` as a string (not an array)

Edge Case 3: Synthetic `"[Tool execution interrupted by user]"` message

Edge Case 4: `conversationRecovery.ts` stripping thinking blocks

Why We Expect Limited Impact on Other Providers

Honest uncertainty

Files Modified

Why Subagents Are Most Affected

Request for Review

Related Issues

References

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Component	Version / Setting
OpenClaude	`v0.6.0`
Provider	`openai-compatible`
Model	`deepseek-v4-flash`, `deepseek-v4-pro`
Base URL	`https://api.deepseek.com/v1`
`OPENAI_MAX_TOKENS`	`262144`

#	Format	`content`	`reasoning_content`	Result
1	First request (user only)	N/A	N/A	200 OK
2	Multi-turn WITH reasoning	`"Let me check..."`	✅ Present	200 OK
3	Multi-turn WITHOUT reasoning	`"Let me check..."`	❌ Absent	400 ERROR
4	Agent format WITHOUT reasoning	`""`	❌ Absent	400 ERROR
5	Agent format WITH reasoning	`""`	✅ Present	200 OK

[Bug / Potential Fix] DeepSeek V4 Flash/Pro: reasoning_content must be passed back to the API (400) on assistant messages with tool_calls #878

Description

[Bug]: DeepSeek V4 — 400 error caused by reasoning_content missing on edge-case assistant messages

Summary

Environment

Error Message

DeepSeek's Documented Rule

Validation: 5 CURLs Against Real DeepSeek API

CURL 1 — First request (user asks for tool use)

CURL 2 — Second request WITH reasoning_content → 200 OK

CURL 3 — Second request WITHOUT reasoning_content → 400 ERROR

CURL 4 — Agent format content: "" + tool_calls WITHOUT reasoning → 400

CURL 5 — Agent format content: "" + tool_calls WITH reasoning → 200 OK

Validation Summary

What Appear to Be the Four Edge Cases Not Covered by ff2a380

Edge Case 1: Assistant with array content but NO thinking block

Edge Case 2: Assistant with content as a string (not an array)

Edge Case 3: Synthetic "[Tool execution interrupted by user]" message

Edge Case 4: conversationRecovery.ts stripping thinking blocks

Why We Expect Limited Impact on Other Providers

Honest uncertainty

Files Modified

Why Subagents Are Most Affected

Request for Review

Related Issues

References

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

[Bug]: DeepSeek V4 — 400 error caused by `reasoning_content` missing on edge-case assistant messages

CURL 2 — Second request WITH `reasoning_content` → 200 OK

CURL 3 — Second request WITHOUT `reasoning_content` → 400 ERROR

CURL 4 — Agent format `content: ""` + `tool_calls` WITHOUT reasoning → 400

CURL 5 — Agent format `content: ""` + `tool_calls` WITH reasoning → 200 OK

What Appear to Be the Four Edge Cases Not Covered by `ff2a380`

Edge Case 1: Assistant with array `content` but NO thinking block

Edge Case 2: Assistant with `content` as a string (not an array)

Edge Case 3: Synthetic `"[Tool execution interrupted by user]"` message

Edge Case 4: `conversationRecovery.ts` stripping thinking blocks