[BUG] /auto-mode-setup always fails: propose side query returns HTTP 429 rate_limit_error at 2% utilization

### Preflight Checklist

- [X] I have searched [existing issues](https://github.com/anthropics/claude-code/issues?q=is%3Aissue%20state%3Aopen%20label%3Abug) and this hasn't been reported yet
- [X] This is a single bug report (please file separate reports for different bugs)
- [X] I am using the latest version of Claude Code

### What's Wrong?

`/auto-mode-setup` always fails. It prints its banner, then a few seconds later:

```
Auto-mode setup scan failed: The model call didn't complete. This is usually temporary - re-run to try again.
```

It is not temporary. It has failed on every attempt since at least 2026-08-15, and no `autoMode` block has ever been written on this account.

Under `--debug`, the cause is visible: the propose side query is refused with HTTP 429 `rate_limit_error`, three times on the primary model and three times on the fallback, each in roughly 300ms.

This is not ordinary rate limiting. At the moment of failure, reported utilization was **2% of the 5-hour window** (56% weekly), and the interactive session's own requests were succeeding in the same second the side query was refused. Only `source=side_query` requests were rejected.

I varied everything reachable from the client, and the result never changed:

| Variable | Values tried | Result |
| --- | --- | --- |
| Repository | 3 different repos | 429 in all |
| Classifier model | `claude-opus-5[1m]`, its auto-selected fallback, `claude-sonnet-5` | 429 in all |
| Recon scope | `scope=all depth=both` and `scope=project depth=here` | 429 in both |
| Entry point | interactive wizard and headless `--propose` | 429 in both |
| Date | 2026-08-15 and 2026-08-22 | 429 on both |

The `claude-sonnet-5` run carries no 1M-context beta and still fails, so the beta is not implicated. No proxy and no `ANTHROPIC_BASE_URL` are set on the process.

### What Should Happen?

`/auto-mode-setup` should return a drafted auto-mode proposal for review. Failing that, it should surface an actionable error. A permanent entitlement or quota refusal should not be reported to the user as a transient failure with an instruction to re-run, since no number of re-runs can clear it.

### Error Messages/Logs

```
18:37:25.461 [DEBUG] [API REQUEST] /v1/messages source=side_query
18:37:26.237 [DEBUG] [API REQUEST] /v1/messages source=side_query
18:37:27.632 [DEBUG] [API REQUEST] /v1/messages source=side_query
18:37:27.935 [ERROR] auto-mode-setup sideQuery failed: 429 {"type":"error","error":{"type":"rate_limit_error","message":"Error"},"request_id":"req_011CeJGjtLpKA52vNtVJpBFk"}
18:37:27.936 [WARN] auto-mode-setup propose: primary model failed; retrying on fallback
18:37:27.939 [DEBUG] [API REQUEST] /v1/messages source=side_query
18:37:28.768 [DEBUG] [API REQUEST] /v1/messages source=side_query
18:37:29.803 [DEBUG] [API REQUEST] /v1/messages source=side_query
18:37:30.061 [ERROR] auto-mode-setup sideQuery failed: 429 {"type":"error","error":{"type":"rate_limit_error","message":"Error"},"request_id":"req_011CeJGk3cfqMpmn7bvQNiJN"}

A separate run on a different model produced:
req_011CeJGxHFnvw18UFaXrwAsg
req_011CeJGxSipA9kQcs2X82Ktf
```

### Steps to Reproduce

The headless form reproduces it in about ten seconds and writes nothing to settings:

1. Open a terminal in any git repository.
2. Run:
```
claude --debug -p "/auto-mode-setup --wizard posture=mixed scope=project depth=here --propose"
```
3. Observe stdout:
```json
{
"ok": false,
"code": "api_failed",
"reason": "The model call didn't complete. This is usually temporary - re-run to try again."
}
```
4. Open the debug log named at session start and search for `sideQuery failed`. The 429 and its request ID are there.

The interactive path fails identically: run `/auto-mode-setup`, arrow down to Continue, press Enter, and the scan failure appears within about 40 seconds.

### Claude Model

Opus

### Is this a regression?

No, this never worked

### Last Working Version

N/A

### Claude Code Version

2.1.240 (Claude Code)

### Platform

Anthropic API

### Operating System

macOS

### Terminal/Shell

iTerm2

### Additional Information

**Account context, in case it is the cause.** This account has `hasExtraUsageEnabled: false` and `cachedExtraUsageDisabledReason: "org_level_disabled"` (Max subscription held through an organization). In the same debug log, Fast mode is refused with `Fast mode requires usage credits`. If the propose call is metered against the extra-usage path, an entitlement refusal surfaced as a bare 429 would explain an instant rejection at 2% utilization, on every model, indefinitely. I cannot confirm that from the client side, and the redacted `"message":"Error"` body gives the user nothing to act on either way.

**Why this took so long to diagnose.** The user-facing string is a catch-all around the API call, so a permanent refusal and a transient blip render identically, and the text advises the only action that cannot work. The real status is reachable only via `--debug`, which the message does not mention. Noting this as context on the same bug rather than as a separate report.

**Not a duplicate.** #86778 reports a hang with no error surfaced. #87078, #87608 and #88450 fail at `removeFromPermissionsAllow` validation, which runs only after a proposal has come back, so their propose call succeeded and mine never returns one. #65054 and closed #83037 are in the same family of confusingly surfaced entitlement states, but neither covers this command.

Start at the `/auto-mode-setup --propose` entry point and trace the `sideQuery failed` path shown in the debug logs. Check how HTTP 429 responses become the generic transient failure, then verify the behavior for entitlement or quota refusals versus temporary errors. Done means a successful proposal is returned when possible, while persistent refusal produces an actionable message instead of recommending retries.

Escrito por el modelo de indexación a partir del texto del issue.