codex in vscode doesn't work

Resolved 💬 2 comments Opened Oct 29, 2025 by lshgdut Closed Oct 29, 2025
💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What version of Codex is running?

codex-cli 0.50.0

What subscription do you have?

free

Which model were you using?

qwen3-coder

What platform is your computer?

Darwin 24.6.0 arm64 arm

What issue are you seeing?

I use codex vscode extension to test with the prompt write 1 to a.txt, the codex panel shows it had finished working, but actually nothing changed in my workspace.

here is the conversation output:

<|endoftext|><|im_start|>
<|im_start|>
<|endoftext|><|im_start|>
<|im_start|>
<|endoftext|><|repo_name|>
<function=shell>
<parameter=command>
["echo", "1", ">", "a.txt"]
</parameter>
<parameter=justification>
Writing the number 1 to a file named a.txt as requested
</parameter>
<parameter=with_escalated_permissions>
True
</parameter>
</function>
</tool_call>

here is the codex logs in vscode extension output panel:

2025-10-29 11:24:11.413 [error] Unknown MCP notification: sessionConfigured {"sessionId":"019a2de4-c3a0-7662-be67-05474349ccfc","model":"qwen3-coder","reasoningEffort":"high","historyLogId":116814424,"historyEntryCount":38,"initialMessages":[{"type":"user_message","message":"# Context from my IDE setup:\n\n## Active file: /Users/xxxx/.codex/config.toml\n\n## Active selection of the file:\ners.context7]\n# command = \"npx\"\n# args = [\"-y\"\n## Open tabs:\n- article.json: data/article.json\n- prompt.txt: prompt.txt\n- fileUtils.ts: src/utils/fileUtils.ts\n- config.toml: /Users/wenhua/.codex/config.toml\n\n## My request for Codex:\nwrite `1` to a.txt\n","images":[]},{"type":"agent_message","message":"I'll write '1' to `a.txt` using the shell."},{"type":"token_count","info":null,"rate_limits":null}],"rolloutPath":"/Users/xxxx/.codex/sessions/2025/10/29/rollout-2025-10-29T10-55-58-019a2de4-c3a0-7662-be67-05474349ccfc.jsonl"}
2025-10-29 11:24:19.649 [error] Unknown MCP notification: codex/event/user_message {"id":"0","msg":{"type":"user_message","message":"# Context from my IDE setup:\n\n## Active file: codex-rs/core/src/mcp_connection_manager.rs\n\n## Open tabs:\n- codex.js: codex-cli/bin/codex.js\n- codex_message_processor.rs: codex-rs/app-server/src/codex_message_processor.rs\n- mcp_process.rs: codex-rs/app-server/tests/common/mcp_process.rs\n- list_resume.rs: codex-rs/app-server/tests/suite/list_resume.rs\n- archive_conversation.rs: codex-rs/app-server/tests/suite/archive_conversation.rs\n- codex.rs: codex-rs/core/src/codex.rs\n- mcp_connection_manager.rs: codex-rs/core/src/mcp_connection_manager.rs\n\n## My request for Codex:\nwrite `1` to a.txt\n","images":[]},"conversationId":"019a2de4-c3a0-7662-be67-05474349ccfc"}

What steps can reproduce the bug?

Uploaded thread: 019a2e03-4eaf-71e2-bdcd-8feeee317ff9

What is the expected behavior?

the prompt can create the file a.txt with content 1

Additional information

I am using litellm-proxy as my provider with model qwen3-coder

my codex config content

# Codex CLI Configuration
# See: https://developers.openai.com/codex/local-config#ide

model_reasoning_effort = "high"
model_reasoning_summary = "detailed"
show_raw_agent_reasoning = false
model_verbosity = "high"
approval_policy = "untrusted"
sandbox_mode = "workspace-write"
hide_agent_reasoning = false
model = "qwen3-coder"
model_provider = "litellm"

[model_providers.litellm]
name = "litellm"
base_url = "http://localhost:4000"

codex extension version
Version
0.4.31
Last Updated
2025-10-29, 08:44:01

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 10 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

  • #5488
  • #5846
  • #5845

Powered by Codex Action

gpeal contributor · 10 months ago

This seems like an issue with your litellm which may not know how to use the exec and apply_patch tools that Codex uses. You're welcome to play around with your model and model config but we won't be able to help you with your specific model behavior issues that aren't from OpenAI models.