GPT-5.6 Codex Desktop new tasks can start without workspace tools #32436

What version of the Codex App are you using?

ChatGPT Desktop / Codex Task: 26.709.11516

What subscription do you have?

Unknown from collected evidence.

Which model were you using?

GPT-5.6, observed in new Codex Workspace Tasks. Publicly related issues specifically involve gpt-5.6-sol.

What platform is your computer?

Windows 11.

What issue are you seeing?

Every newly created GPT-5.6 Codex Desktop task lacks workspace capabilities. The assistant reports that the session has no filesystem access, no terminal execution, and no file editing tools.

Typical assistant response:

This session is not attached to file access, terminal execution, or file editing tools...

The issue reproduces across newly created GPT-5.6 Codex tasks and is not limited to one repository, one workspace, or one AGENTS.md.

The same machine can still run a current Codex task with workspace tools, and the CLI path can run a shell command with gpt-5.6-sol, so this does not appear to be a simple local config or OS-level filesystem permission failure.

Steps to reproduce

1. On Windows 11, open ChatGPT Desktop / Codex Task app version 26.709.11516.
2. Create a new Codex Workspace Task using GPT-5.6.
3. Ask the assistant to inspect the workspace or run a simple read-only command, for example:

Please run Get-Location and list the current workspace files.

4. Observe the assistant response.

Expected behavior

Every Codex Workspace Task should receive workspace tools appropriate to its permission profile, including:

• filesystem access for the selected workspace;
• terminal execution;
• file editing capability.

The assistant should be able to read the current workspace, execute targeted commands, and edit files when allowed by the task permission profile.

Actual behavior

New GPT-5.6 Codex Desktop tasks report no attached file access, terminal execution, or file editing tools. The assistant refuses normal workspace operations before attempting a tool call.

Local control evidence

Current task is not affected:

• C:\Users\black\.codex\config.toml contains sandbox_mode = "danger-full-access" and approval_policy = "never".
• This current task successfully read files, executed shell commands, and created the investigation report.
• codex --version reports codex-cli 0.144.1.

CLI control probes on the same machine:

codex exec --skip-git-repo-check --ephemeral --json 'Get-Location'
codex exec -m gpt-5.6-sol --skip-git-repo-check --ephemeral --json 'Get-Location'

Both returned C:\Users\black\.codex via command execution.

This narrows the suspected failure away from machine-wide filesystem/terminal denial and toward Desktop new-task provisioning, Desktop/runtime state, or GPT-5.6-specific capability negotiation.

Related public issues

• openai/codex#31894: gpt-5.6 Responses Lite turns do not expose exec/code-mode tools in codex exec

  • This is the closest public match. The report states that gpt-5.6-sol can respond that no shell execution tool is available while gpt-5.5 works.
  • It also identifies the Responses Lite request shape: GPT-5.6 has no top-level tools, while additional_tools are placed in the input.
  • URL: gpt-5.6 Responses Lite turns do not expose exec/code-mode tools in codex exec #31894

• openai/codex#31870: Codex with GPT-5.6-Sol through Azure fails every turn with X-OpenAI-Internal-Codex-Responses-Lite

  • Related to GPT-5.6 Sol and the internal Responses Lite route.
  • URL: Codex with GPT-5.6-Sol through Azure fails every turn with X-OpenAI-Internal-Codex-Responses-Lite #31870

• openai/codex#32049: Windows app 26.707.31428 regression: active GPT-5.6 sessions fall back to "Custom", new sessions lose GPT-5.6 in the picker, and account/workspace state becomes inconsistent

  • Related Windows Desktop regression involving GPT-5.6 model state and new-task behavior.
  • URL: Windows app 26.707.31428 regression: active GPT-5.6 sessions fall back to "Custom", new sessions lose GPT-5.6 in the picker, and account/workspace state becomes inconsistent #32049

• openai/codex#32291: Tool-backed Codex Desktop ignores prompt model steering and cannot select named custom agents

  • Shows Desktop tool-backed paths can diverge from native CLI behavior for model/capability propagation.
  • URL: Tool-backed Codex Desktop ignores prompt model steering and cannot select named custom agents #32291

Suspected regression surface

Open-source Codex code at commit 5c19155cbd93bfa099016e7487259f61669823ff shows:

• codex-rs/models-manager/models.json: gpt-5.6-sol sets use_responses_lite: true, tool_mode: "code_mode_only", shell_type: "shell_command", apply_patch_tool_type: "freeform", and minimal_client_version: "0.144.0".
• codex-rs/core/src/client.rs: when model_info.use_responses_lite is true, build_responses_request puts tool schemas into a developer ResponseItem::AdditionalTools prefix and sends top-level tools as None.
• codex-rs/core/src/client.rs: the client also sets the X-OpenAI-Internal-Codex-Responses-Lite header for Lite models.

If the Desktop/backend/model route does not promote or preserve additional_tools as callable tools, GPT-5.6 sees no callable workspace tools and reports that no filesystem/terminal/editing tools are attached.

Requested engineering investigation

Please inspect the request payload and resolved capability state for affected Desktop new tasks:

1. Confirm whether thread/start receives the intended cwd, runtimeWorkspaceRoots, permission profile, and selected capability roots.
2. Confirm whether the turn prompt for GPT-5.6 includes workspace tools in the Lite additional_tools item.
3. Confirm whether the Responses Lite backend/model runtime exposes those additional_tools as callable tools.
4. Compare a working GPT-5.5 task and failing GPT-5.6 task from the same Desktop app/workspace/account.
5. Compare Desktop-created new tasks against native codex exec -m gpt-5.6-sol.

Expected fix direction

If the failure is in Responses Lite tool promotion, ensure GPT-5.6 code-mode tools are exposed as callable tools. Candidate approaches:

• preserve additional_tools for the Lite input contract and also send compatible client-executed tools at top-level tools, if accepted by the API;
• fix backend Responses Lite promotion so additional_tools is reliably converted into callable tools for GPT-5.6;
• gate GPT-5.6 Desktop availability behind a capability handshake that verifies workspace tools are visible before the first turn;
• fall back to the non-Lite tool path or an unaffected model route when code-mode tool negotiation fails.

Test plan

Add regression tests covering:

• GPT-5.6 Sol new Desktop task gets non-empty workspace/code-mode tools.
• GPT-5.6 Sol codex exec can execute a read-only command.
• GPT-5.6 Sol Desktop thread/start + turn/start preserves cwd, runtimeWorkspaceRoots, and permission profile.
• Responses Lite requests either include top-level callable tools or have a server-side assertion that additional_tools was promoted before inference.
• A model must not produce the "no shell execution tool is available" response for a simple command when the resolved permission profile allows shell execution.

Workaround

No reliable user-side workaround is proven by this investigation. Do not recommend cache clearing or reinstalling unless a specific affected case is shown to recover after that action.