Codex Desktop intermittently corrupts structured tool-call arguments, with related Computer Use app-signature failure

Open 💬 2 comments Opened Jun 29, 2026 by jakyeamos

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

26.623.61825

What subscription do you have?

Pro

What platform is your computer?

macOS arm64

What issue are you seeing?

Summary:
Codex Desktop has repeatedly emitted malformed structured tool-call argument objects. Instead of normal JSON like { "query": "...", "limit": 5 } or { "cmd": "..." }, the argument object sometimes contains long corrupted/random multilingual property names. This causes validation failures such as ObjectParam/property_name_above_max_length or missing required fields like query / cmd. In at least one case, the turn then completed with last_agent_message: null, silently dropping the task.

Affected sessions / occurrences:

Session ID: 019f1186-b4f8-75a3-a913-2888f848ee8c

Time: 2026-06-29T04:04:04Z
Project cwd: /Users/jakyeamos/projects/BBDSE
Failed tool: tool_search
Session log symptom:
failed to parse tool_search arguments: missing field query
UI/API validator symptom:
[ObjectParam] [property_name_above_max_length] Invalid property name in input[3].arguments; expected max length 256, got length 1252.
Result:
task_complete with last_agent_message: null

Session ID: 019f0a08-19e0-7eb0-a6c4-85c5dd4f7e2c

Symptom:
Same ObjectParam/property_name_above_max_length class.
UI/API validator symptom:
Invalid property name in input[5].arguments; expected max length 256, got length 3405.
The error text was captured in a Codex thread title/preview.

Session ID: 019f0b64-1d40-7c71-a0d1-b4669f648fb4

Time: 2026-06-27T23:31:07Z
Failed tool: tool_search
Session log symptom:
failed to parse tool_search arguments: missing field query
Same failure class:
Structured tool-call arguments were invalid before tool execution. I only found the missing-query parser symptom locally for this one, not the full ObjectParam text.

Additional related local app state / Computer Use failure:

Session ID: 019f118b-c232-7361-aba1-2f8f887558cf

Time: 2026-06-29 around 04:03Z
User saw macOS/Gatekeeper asking to delete or move "Codex Computer Use" to Trash.
Local check found the containing app bundle failed codesign verification:

/Applications/Codex.app: invalid signature (code or signature have been modified)
In architecture: arm64

Embedded helper path involved:
/Applications/Codex.app/Contents/Resources/plugins/openai-bundled/plugins/computer-use/Codex Computer Use.app

Recommended local fix:
Quit Codex, move /Applications/Codex.app to Trash, reinstall a fresh official Codex build. Do not just remove quarantine attributes.

Relevant local log/session paths:
~/.codex/sessions/2026/06/28/rollout-2026-06-28T23-57-48-019f1186-b4f8-75a3-a913-2888f848ee8c.jsonl
~/.codex/sessions/2026/06/27/rollout-2026-06-27T19-22-18-019f0b64-1d40-7c71-a0d1-b4669f648fb4.jsonl
~/Library/Logs/com.openai.codex/2026/06/29/codex-desktop-00cd09e2-a3f8-4243-b0e2-2badc911af64-32507-t0-i1-035548-0.log

Additional nearby desktop log symptoms:

Received turn/started for unknown conversation
Item not found in turn state
Failed to generate thread title: Timed out waiting for structured result
Browser route warnings around IAB_LIFECYCLE

Expected behavior:
Tool-call arguments should either be valid JSON matching the target tool schema, or the turn should recover with a visible assistant/tool error. It should not emit corrupted argument keys, fail schema validation, and then silently complete with no final assistant message.

Impact:
Tasks fail before meaningful work starts. The user sees opaque ObjectParam/property_name_above_max_length errors or missing-field parse errors, and may lose the agent turn without a useful final message.

What steps can reproduce the bug?

I do not have a deterministic minimal code snippet, but I have multiple local Codex Desktop sessions showing the same reproducible failure pattern when Codex attempts structured tool calls.
Observed reproduction pattern:

Start a Codex Desktop thread with a normal coding request in a local project.
Let the assistant begin its usual workflow. The assistant often decides to use a structured tool call, especially tool_search.
Instead of emitting valid tool arguments such as:

{ "query": "read_thread list_threads", "limit": 5 }
Codex emits an arguments object with a long corrupted/random multilingual property name.

The app/API validator rejects the tool call with ObjectParam/property_name_above_max_length, or the tool parser reports a missing required field like query or cmd.
In at least one case, the turn then completed with last_agent_message: null, so the task silently failed before useful work began.

What is the expected behavior?

Codex should only send valid structured tool-call arguments matching the target tool schema.

For example, tool_search should receive a small valid JSON object such as:
{ "query": "read_thread list_threads", "limit": 5 }

exec_command should receive:
{ "cmd": "rg -n 'pattern' path", "workdir": "/path" }

If structured argument generation fails, Codex should recover with a visible assistant/tool error and keep the turn alive or clearly mark it failed. It should not emit corrupted long property names, fail schema validation, and then complete the task with no final assistant message.

Computer Use should also not be blocked by a signature-invalid embedded helper inside /Applications/Codex.app. If the app bundle is damaged or stale, Codex should surface a clear reinstall/update prompt rather than macOS asking to delete the helper with no explanation.

Additional information