Web search returns "Found no tool response" for valid keyword queries while direct URL open works

What version of Codex CLI is running?

0.153.2

What subscription do you have?

ChatGPT Plus

Which model were you using?

gpt-5.6-sol

What platform is your computer?

Windows 11, 64-bit (10.0.26100)

What terminal emulator and version are you using (if applicable)?

Windows Terminal with PowerShell. Terminal version was not captured by codex doctor.

Codex doctor report

Relevant sanitized summary from codex doctor --json:

{
  "overallStatus": "warning",
  "codexVersion": "0.153.2",
  "config.load": {
    "status": "ok",
    "model": "gpt-5.6-sol",
    "model provider": "openai"
  },
  "network.provider_reachability": {
    "status": "ok",
    "summary": "active provider endpoints are reachable over HTTP"
  },
  "network.websocket_reachability": {
    "status": "ok",
    "summary": "Responses WebSocket handshake succeeded",
    "provider name": "OpenAI",
    "wire API": "responses"
  },
  "runtime.provenance": {
    "status": "ok",
    "platform": "windows-x86_64",
    "version": "0.153.2"
  },
  "system.environment": {
    "status": "ok",
    "os": "Windows 10.0.26100 (Windows 11 CoreCountrySpecific) [64-bit]"
  },
  "terminal.env": {
    "status": "ok",
    "terminal": "Windows Terminal"
  }
}

The overall warning is due to unrelated local-state / endpoint-protection warnings. Network reachability and the active OpenAI provider checks are OK.

What issue are you seeing?

Keyword web search consistently fails, while directly opening a URL through the same web tool works.

For a normal search request, Codex generates a valid tool call such as:

const r = await tools.web__run({
  search_query: [{ q: "OpenAI official website" }],
  response_length: "short"
});

The search extension then runs for about 6-7 seconds and completes with an empty result set:

{
  "type": "Extension",
  "kind": "web.search",
  "query": "OpenAI official website",
  "action": {
    "type": "search",
    "query": "OpenAI official website",
    "queries": null
  },
  "results": []
}

The caller then receives only:

Found no tool response. This likely means the arguments you provided were not valid.

The same behavior occurs with an unrelated query such as:

What is the capital of France?

However, directly opening a URL through the same web tool succeeds. For example, opening https://www.openai.com returns the page normally (OpenAI | Research & Deployment).

This makes the generic "arguments ... were not valid" message look misleading: the search action is actually created and runs before returning no results / no tool response.

As a control, intentionally passing the wrong argument type (a JSON string instead of an object) fails immediately with the specific validation error:

tool `webrun` expects a JSON object for arguments

So the failing keyword-search path appears different from ordinary argument validation.

What steps can reproduce the bug?

1. Run Codex CLI 0.153.2 using ChatGPT auth and the OpenAI provider.

2. Start a fresh session.

3. Ask Codex to search the web for OpenAI official website.

4. Observe that Codex invokes web__run with a normal search_query object.

5. The web.search extension runs for about 6-7 seconds and completes with results: [].

6. Codex receives:

   Found no tool response. This likely means the arguments you provided were not valid.
   

7. Repeat with another unrelated keyword query (for example, What is the capital of France?) and observe the same result.

8. In the same session, directly open https://www.openai.com through the web tool and observe that the page loads successfully.

What is the expected behavior?

A valid keyword web search should return search results. If the search backend fails, Codex should surface the actual search/backend error rather than suggesting that valid tool arguments were invalid.

Direct URL opening and keyword search should both work when the web tool is available.

Additional information

• Reproduced multiple times in a fresh Codex CLI session.
• This is using the official OpenAI provider, not a custom provider.
• Direct URL access works in the same session.
• Provider HTTP reachability and Responses WebSocket handshake both pass in codex doctor.
• I searched the openai/codex issue tracker for the exact error text before filing and did not find an existing report.
• Full rollout logs are available if maintainers need additional sanitized event details.

Start with the web__run entry point and reproduce the search_query request from the issue, comparing its results: [] response with a successful direct URL open. Trace how the web.search extension result becomes “Found no tool response”; done means valid keyword searches return results or expose the actual backend error instead of an argument-validation message.