App-server crashes on any tool call ("thread_tools" unknown feature key) — sudden onset, no local trigger #29140

What version of the IDE extension are you using?

26.616.32156

What subscription do you have?

I have ChatGPT Plus subscription

Which IDE are you using?

VS Code

What platform is your computer?

macOS (Intel, x86_64) — Intel Core i5-8279U

What issue are you seeing?

Summary
The Codex extension (openai.chatgpt) for VS Code crashes the app-server whenever a tool call is required (project analysis, file reading, etc.), while plain conversational prompts work fine. The bug appeared suddenly mid-session — everything was working normally, then a single prompt triggered the first crash with no local change on my end. The issue is fully reproducible and has persisted through a complete clean reinstall.

Environment

• macOS, Intel (x86_64), CPU: Intel Core i5-8279U
• IDE: VS Code
• Extension: openai.chatgpt (reproduced across multiple extension versions, including a fresh install of the latest version)

Important timeline detail
The first crash was not triggered by any local change, update, or configuration edit on my side — Codex was working normally, and then abruptly started crashing on a prompt mid-session. This strongly suggests a server-side/cloud-pushed change (e.g. a feature flag rollout) rather than anything client-side, since nothing in my environment changed between the last working prompt and the first crash.

Clear reproduction steps

1. Open Codex in VS Code with a project folder open
2. Send a plain conversational prompt, e.g. "Tell me a story" → works perfectly, no crash
3. Send any prompt that requires tool use, e.g. "Analyze my project" or "Analyze this project and its recent changes" → app-server crashes immediately

This confirms the crash is specific to the tool-calling pathway, not the model/chat pathway itself.

Troubleshooting already performed (ruled out as causes)

• Fixed a local SSL_CERT_FILE environment variable pointing to an invalid/oversized custom certificate bundle
• Fixed a local SSL_CERT_FILE pointing to a nonexistent Homebrew path
• Removed an unrelated stray certificate from the macOS login keychain
• Fully uninstalled the extension, renamed/cleared ~/.codex entirely, reinstalled the extension from scratch, re-authenticated
• Tested on both a downgraded and the latest extension version
• None of the above resolved the crash

Issue 1 — Crash on any tool call: "unknown feature key in config: thread_tools"
config.toml contains only model, model_reasoning_effort, and a [projects."..."] trust_level block — no reference to thread_tools anywhere on disk (confirmed via recursive grep across ~/.codex/ and VS Code's settings.json). Despite this, any tool-requiring prompt crashes the app-server. I've seen this surface in two slightly different forms:

Form A (with explicit signal):

Codex process errored: Codex app-server process exited unexpectedly (code=unknown, signal=SIGTRAP). Last CLI error: codex_features: unknown feature key in config: thread_tools
codex_features: unknown feature key in config: thread_tools

Form B (process unavailable, no signal shown), triggered by the prompt "Analyse ce projet et ces récentes modifications" ("Analyze this project and its recent changes"):

Codex process errored: Codex process is not available
codex_features: unknown feature key in config: thread_tools

I suspect thread_tools is a server-side/cloud-managed feature flag tied to enabling tool calls within a thread, which the currently shipped app-server build does not recognize, causing a crash instead of graceful degradation. The sudden, config-less onset strongly points to a rollout that started pushing this flag to my account/client.

Issue 2 — Related crash on tool use: rollout state db discrepancy
On a separate tool-requiring prompt, the same crash occurred with a different last log line:

Codex process errored: Codex app-server process exited unexpectedly (code=unknown, signal=SIGTRAP). Last CLI error: codex_rollout::list: state db discrepancy during find_thread_path_by_id_str_in_subdir: falling_back
codex_rollout::list: state db discrepancy during find_thread_path_by_id_str_in_subdir: falling_back

This looks like a logged fallback/recovery path that may precede the actual thread_tools crash when tool invocation kicks in — possibly the same underlying bug surfacing via different log lines depending on timing.

Issue 3 — Bundled plugin manifest violates your own validation rule (separate bug, found during troubleshooting)

Codex process errored: Codex process is not available
codex_core_plugins::manifest: ignoring interface.defaultPrompt[0]: prompt must be at most 128 characters path=/Users/[user]/.codex/.tmp/plugins/plugins/ngs-analysis/.codex-plugin/plugin.json

The bundled ngs-analysis plugin (author: OpenAI) ships with a defaultPrompt[0] string well over your documented 128-character limit. The log says it's being "ignored" but the process becomes unavailable instead of gracefully skipping the invalid field.

Request
Given the clean reproduction (any tool call crashes the app-server via thread_tools, appearing suddenly with no local trigger) could you confirm whether this is a feature flag rollout issue affecting tool-calling, and provide a fix or rollback? This currently makes the extension unusable for any real coding task, since project analysis, file edits, and command execution all require tool calls. Also please fix the ngs-analysis bundled plugin manifest so it complies with your own 128-character limit.

Happy to provide full logs (~/.codex/logs_2.sqlite) or a Feedback ID if needed.

What steps can reproduce the bug?

1. Open VS Code with a project folder loaded

2. Open the Codex panel (openai.chatgpt extension)

3. Send a plain conversational prompt, e.g.:
   "Tell me a story"
   → Works correctly, no crash

4. In the same or a new session, send a prompt that requires tool use, e.g.:
   "Analyze my project"
   or
   "Analyse ce projet et ces récentes modifications"
   → The app-server crashes immediately with:

   Codex process errored: Codex app-server process exited unexpectedly (code=unknown, signal=SIGTRAP).
   Last CLI error: codex_features: unknown feature key in config: thread_tools
   codex_features: unknown feature key in config: thread_tools

   (Sometimes surfaces instead as:
   Codex process errored: Codex process is not available
   codex_features: unknown feature key in config: thread_tools)

Note: config.toml contains no reference to "thread_tools" (verified via recursive
grep across ~/.codex/ and VS Code settings.json — only present in the internal
log database). The crash also persists after a full clean reinstall (uninstalled
extension, wiped ~/.codex, reinstalled, re-authenticated), and first appeared
suddenly mid-session with no local config change beforehand.

What is the expected behavior?

Codex should successfully process tool-requiring prompts (e.g. project analysis,
file reading/editing) the same way it handles plain conversational prompts.

If "thread_tools" is an unrecognized or unsupported feature key (e.g. pushed via
a server-side feature flag rollout), the app-server should log a warning and
gracefully ignore/skip it rather than crashing the entire process with SIGTRAP.
A single unknown config/feature key should never take down the whole session.

Additional information

No response