openai / openai/codex

Windows Desktop 0.148-alpha.9: parent /root session has no exec, only wait + collaboration tools

Open
#38,960 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app app-server bug custom-model tool-calls windows-os
Dominant language
Rust
Stars
125k
Forks
19.5k
PR merge metrics
PR metrics pending

Description

What version of the Codex App are you using (From “About Codex” dialog)?

Codex Desktop, bundled CLI 0.148.0-alpha.9 (originator: Codex Desktop, source: vscode).

Desktop app build previously reported in local runtime env as 26.810.52044.

Desktop runtime still present at:

%LOCALAPPDATA%\OpenAI\Codex\bin\e305f1c75d8da435\codex.exe

Rolling the npm CLI back to @openai/codex@0.146.0 (codex-cli 0.146.0) fixes CLI sessions. The Desktop app does not follow that rollback and stays on 0.148.0-alpha.9.

What subscription do you have?

Custom Responses-compatible provider (model_providers.ccswitch, wire_api = "responses", local proxy). Not a ChatGPT-only first-party session. Model: gpt-5.6-terra.

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64 (Windows 10 Home China)

What issue are you seeing?

Every new Codex Desktop coding session is assembled as a broken /root orchestrator:

  1. Developer prompt includes You are /root, the primary agent in a team of agents... and collaboration tools (spawn_agent, send_message, followup_task, wait_agent, interrupt_agent, list_agents).
  2. A later instruction says do not spawn sub-agents unless explicitly asked.
  3. The parent agent is not given local file/command tools. In particular, exec is missing.

The model's actual tool surface on Desktop is only:

  • functions.wait
  • functions.request_user_input (Plan mode only)
  • collaboration.followup_task
  • collaboration.interrupt_agent
  • collaboration.list_agents
  • collaboration.send_message
  • collaboration.spawn_agent

There is no exec, no filesystem tools, no search/edit tools. Workspace permissions are fine (the project directory is writable). The agent then loops on functions.wait with invalid float args (1.0 / 10000.0) and eventually says it cannot read files.

This is not a repo-permission problem. It is a Desktop session tool-assembly bug.

Same machine, same ~/.codex/config.toml, CLI 0.146.0 works: it still receives the /root text, but exec is present and can read/write files.

config.toml already disables:

[features]
multi_agent = false
collaboration_modes = false

Desktop ignores this and still injects the /root collaboration surface without exec.

What steps can reproduce the bug?
  1. Use Codex Desktop on Windows with a custom Responses provider (wire_api = "responses").
  2. Open a normal coding task (not an explicit multi-agent / team entry).
  3. Ask: 列出你当前这一轮可用的所有工具
  4. Observed: only wait, Plan-only request_user_input, and collaboration tools. No exec.
  5. Ask to read any local file or 测试本地文件读写能力.
  6. Observed: agent tries functions.wait repeatedly and cannot read/write files.

Recorded Desktop sessions:

  • 01a0094e-5caf-7ed0-ab00-2798a4747c4e — user asked to read work/selections.json; agent only called functions.wait.
  • 01a00943-e967-7ff3-8c2d-e5cf711f693b — same missing-tool diagnosis on a new Desktop thread.
  • 01a00e65-fa9c-7c40-a973-07f1e22bbcb1 (2026-08-17T06:25:59Z) — originator: Codex Desktop, cli_version: 0.148.0-alpha.9. Agent listed only wait/collaboration tools, then failed a local IO test with functions.wait.

Working CLI control on the same machine, minutes earlier:

  • 01a00e63-362a-7bf2-9b31-b679e35e3949 (2026-08-17T06:22:58Z) — originator: codex-tui, cli_version: 0.146.0. Same prompt family (You are /root), but exec was available and used (custom_tool_call name=exec).

Desktop does not pick up the rolled-back npm CLI. It keeps using its private runtime under %LOCALAPPDATA%\OpenAI\Codex.

What is the expected behavior?

A normal Desktop coding session should expose local tools to the parent agent, at least exec (and the usual file/search/edit path), even if collaboration tools are also present.

If Desktop is going to force /root + collaboration tools:

  • the parent must still have exec, or
  • it must be allowed/required to spawn a worker that has exec.

The current combination is a deadlock: parent has no local tools, and it is told not to spawn sub-agents.

Rolling Desktop/app-server back to a 0.146-era tool assembly, or honoring multi_agent = false / collaboration_modes = false, should restore a single-agent Desktop session with exec.

Additional information

This looks related to, but stricter than, https://github.com/openai/codex/issues/37718 , https://github.com/openai/codex/issues/37075 , and https://github.com/openai/codex/issues/37376 .

In #37718 the captured request still included exec in the tools array, and the reporter said last-good Desktop was 0.147.0-alpha.1.2. On this machine, Desktop 0.148.0-alpha.9 does not expose exec to the parent at all. CLI 0.146.0 is a clean workaround; Desktop remains broken after CLI rollback.

Redacted config shape:

model_provider = "ccswitch"
model = "gpt-5.6-terra"

[model_providers.ccswitch]
name = "OpenAI"
base_url = "http://<redacted-local-proxy>:8080/v1"
wire_api = "responses"
requires_openai_auth = false

[features]
goals = true
multi_agent = false
collaboration_modes = false

Please treat this as a Desktop/app-server tool-assembly regression on Windows, not a workspace ACL issue.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Reproduce the Windows Desktop 0.148.0-alpha.9 session with the reported config.toml settings and inspect the assembled tool list, then compare it with the working CLI 0.146.0 session. Trace the Desktop/app-server tool assembly and feature-flag handling; done means a normal parent session has exec and local file tools, or can use a worker that does.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
desktop, devtools
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Mostly clear
Newbie friendliness
48/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.