›_ Claude Code Issues
Claude Code Issues / #65863

Claude Code v2.1.167: Agent() spawn fails with "400 thinking options type cannot be disabled when reasoning_effort is set" on DeepSeek Anthropic-compatible endpoint

Open 💬 14 comments Opened Jun 6, 2026 by lzllget5154321-gif

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

Environment

  • Claude Code: v2.1.167 (npm global install)
  • OS: Windows 11 x64
  • API Endpoint: ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
  • Main model: ANTHROPIC_MODEL=deepseek-v4-pro (works fine for main conversation)
  • Sub-agent models: ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-flash, ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash (set at Windows User env level)

Steps to Reproduce

  1. Configure Claude Code to use DeepSeek's Anthropic-compatible endpoint
  2. ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
  3. ANTHROPIC_MODEL=deepseek-v4-pro
  4. Run any Agent spawn: Agent({ model: "sonnet", subagent_type: "general-purpose", ... })
  5. Observe error: API Error: 400 thinking options type cannot be disabled when reasoning_effort is set

Same error occurs with:

  • WebSearch tool
  • WebFetch tool
  • All three Agent model tiers (sonnet/haiku/opus)

What Works

  • Main conversation on deepseek-v4-pro → ✅ perfect
  • All non-subprocess tools (Read, Write, Edit, Bash, Grep, Glob, etc.) → ✅

What We Tested (Direct API Verification)

We ran 11 direct curl tests to DeepSeek's /anthropic/v1/messages endpoint with thinking: {type: "disabled"} combined with various parameters. All returned 200:

| # | model | extra params | Result |
|---|-------|-------------|--------|
| 1 | v4-pro | — | ✅ |
| 2 | v4-pro | tools:[...] | ✅ |
| 3 | v4-pro | tools + system | ✅ |
| 4 | v4-pro | tools + system + tool_choice:{type:"auto"} | ✅ |
| 5 | v4-pro | reasoning_effort:"high" | ✅ |
| 6 | v4-pro | budget_tokens:0 | ✅ |
| 7 | v4-flash | — | ✅ |
| 8 | v4-flash | tools + system | ✅ |
| 9 | claude-sonnet-4-20250514 | Anthropic model ID (mapped to v4-flash by DeepSeek) | ✅ |

This demonstrates that DeepSeek's Anthropic-compatible endpoint fully supports thinking: disabled — the 400 error is specific to Claude Code's Agent subprocess API request path.

Environment Variable Investigation

We tried setting ANTHROPIC_DEFAULT_SONNET/HAIKU_MODEL=deepseek-v4-flash via three different paths:

  1. settings.json env block
  2. Windows User-level registry (permanent)
  3. Post-restart process inheritance

All three correctly set the env vars (confirmed via echo in subprocesses), but Agent spawn still fails with the same 400 — suggesting the Agent subprocess mechanism bypasses these env vars entirely.

Hypothesis

Claude Code's Agent subprocess API request construction differs from the main conversation path. The subprocess may:

  • Send a parameter or header combination not covered by our direct API tests
  • Use a different HTTP endpoint path
  • Use a different SDK code path that triggers the error

Workaround

Direct execution by Team Lead (no Agent spawn). Quality is unaffected — verified with Critic review scoring 9.0/10 on directly-executed page reviews.

Additional Context

This bug was discovered during systematic /lint --system auditing of a production Obsidian knowledge base (~41K files). The Agent system (Writer/Critic/Researcher with state files, memory, and inter-agent cache channels) was fully deployed but had zero usage over 5 days because every spawn attempt failed silently. The knowledge base has comprehensive self-monitoring protocols that detected the anomaly.

What Should Happen?

Environment

  • Claude Code: v2.1.167 (npm global install)
  • OS: Windows 11 x64
  • API Endpoint: ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
  • Main model: ANTHROPIC_MODEL=deepseek-v4-pro (works fine for main conversation)
  • Sub-agent models: ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-flash, ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash (set at Windows User env level)

Steps to Reproduce

  1. Configure Claude Code to use DeepSeek's Anthropic-compatible endpoint
  2. ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
  3. ANTHROPIC_MODEL=deepseek-v4-pro
  4. Run any Agent spawn: Agent({ model: "sonnet", subagent_type: "general-purpose", ... })
  5. Observe error: API Error: 400 thinking options type cannot be disabled when reasoning_effort is set

Same error occurs with:

  • WebSearch tool
  • WebFetch tool
  • All three Agent model tiers (sonnet/haiku/opus)

What Works

  • Main conversation on deepseek-v4-pro → ✅ perfect
  • All non-subprocess tools (Read, Write, Edit, Bash, Grep, Glob, etc.) → ✅

What We Tested (Direct API Verification)

We ran 11 direct curl tests to DeepSeek's /anthropic/v1/messages endpoint with thinking: {type: "disabled"} combined with various parameters. All returned 200:

| # | model | extra params | Result |
|---|-------|-------------|--------|
| 1 | v4-pro | — | ✅ |
| 2 | v4-pro | tools:[...] | ✅ |
| 3 | v4-pro | tools + system | ✅ |
| 4 | v4-pro | tools + system + tool_choice:{type:"auto"} | ✅ |
| 5 | v4-pro | reasoning_effort:"high" | ✅ |
| 6 | v4-pro | budget_tokens:0 | ✅ |
| 7 | v4-flash | — | ✅ |
| 8 | v4-flash | tools + system | ✅ |
| 9 | claude-sonnet-4-20250514 | Anthropic model ID (mapped to v4-flash by DeepSeek) | ✅ |

This demonstrates that DeepSeek's Anthropic-compatible endpoint fully supports thinking: disabled — the 400 error is specific to Claude Code's Agent subprocess API request path.

Environment Variable Investigation

We tried setting ANTHROPIC_DEFAULT_SONNET/HAIKU_MODEL=deepseek-v4-flash via three different paths:

  1. settings.json env block
  2. Windows User-level registry (permanent)
  3. Post-restart process inheritance

All three correctly set the env vars (confirmed via echo in subprocesses), but Agent spawn still fails with the same 400 — suggesting the Agent subprocess mechanism bypasses these env vars entirely.

Hypothesis

Claude Code's Agent subprocess API request construction differs from the main conversation path. The subprocess may:

  • Send a parameter or header combination not covered by our direct API tests
  • Use a different HTTP endpoint path
  • Use a different SDK code path that triggers the error

Workaround

Direct execution by Team Lead (no Agent spawn). Quality is unaffected — verified with Critic review scoring 9.0/10 on directly-executed page reviews.

Additional Context

This bug was discovered during systematic /lint --system auditing of a production Obsidian knowledge base (~41K files). The Agent system (Writer/Critic/Researcher with state files, memory, and inter-agent cache channels) was fully deployed but had zero usage over 5 days because every spawn attempt failed silently. The knowledge base has comprehensive self-monitoring protocols that detected the anomaly.

Error Messages/Logs

Steps to Reproduce

Environment

  • Claude Code: v2.1.167 (npm global install)
  • OS: Windows 11 x64
  • API Endpoint: ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
  • Main model: ANTHROPIC_MODEL=deepseek-v4-pro (works fine for main conversation)
  • Sub-agent models: ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-flash, ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash (set at Windows User env level)

Steps to Reproduce

  1. Configure Claude Code to use DeepSeek's Anthropic-compatible endpoint
  2. ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
  3. ANTHROPIC_MODEL=deepseek-v4-pro
  4. Run any Agent spawn: Agent({ model: "sonnet", subagent_type: "general-purpose", ... })
  5. Observe error: API Error: 400 thinking options type cannot be disabled when reasoning_effort is set

Same error occurs with:

  • WebSearch tool
  • WebFetch tool
  • All three Agent model tiers (sonnet/haiku/opus)

What Works

  • Main conversation on deepseek-v4-pro → ✅ perfect
  • All non-subprocess tools (Read, Write, Edit, Bash, Grep, Glob, etc.) → ✅

What We Tested (Direct API Verification)

We ran 11 direct curl tests to DeepSeek's /anthropic/v1/messages endpoint with thinking: {type: "disabled"} combined with various parameters. All returned 200:

| # | model | extra params | Result |
|---|-------|-------------|--------|
| 1 | v4-pro | — | ✅ |
| 2 | v4-pro | tools:[...] | ✅ |
| 3 | v4-pro | tools + system | ✅ |
| 4 | v4-pro | tools + system + tool_choice:{type:"auto"} | ✅ |
| 5 | v4-pro | reasoning_effort:"high" | ✅ |
| 6 | v4-pro | budget_tokens:0 | ✅ |
| 7 | v4-flash | — | ✅ |
| 8 | v4-flash | tools + system | ✅ |
| 9 | claude-sonnet-4-20250514 | Anthropic model ID (mapped to v4-flash by DeepSeek) | ✅ |

This demonstrates that DeepSeek's Anthropic-compatible endpoint fully supports thinking: disabled — the 400 error is specific to Claude Code's Agent subprocess API request path.

Environment Variable Investigation

We tried setting ANTHROPIC_DEFAULT_SONNET/HAIKU_MODEL=deepseek-v4-flash via three different paths:

  1. settings.json env block
  2. Windows User-level registry (permanent)
  3. Post-restart process inheritance

All three correctly set the env vars (confirmed via echo in subprocesses), but Agent spawn still fails with the same 400 — suggesting the Agent subprocess mechanism bypasses these env vars entirely.

Hypothesis

Claude Code's Agent subprocess API request construction differs from the main conversation path. The subprocess may:

  • Send a parameter or header combination not covered by our direct API tests
  • Use a different HTTP endpoint path
  • Use a different SDK code path that triggers the error

Workaround

Direct execution by Team Lead (no Agent spawn). Quality is unaffected — verified with Critic review scoring 9.0/10 on directly-executed page reviews.

Additional Context

This bug was discovered during systematic /lint --system auditing of a production Obsidian knowledge base (~41K files). The Agent system (Writer/Critic/Researcher with state files, memory, and inter-agent cache channels) was fully deployed but had zero usage over 5 days because every spawn attempt failed silently. The knowledge base has comprehensive self-monitoring protocols that detected the anomaly.

Claude Model

Other

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

2.1.167 (Claude Code)

Platform

Other

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

_No response_

View original on GitHub ↗

This issue has 14 comments on GitHub. Read the full discussion on GitHub ↗