Scheduled Tasks fail on every run with "issue with the selected model (OpenRouter/qwen/qwen3.8-max[1m])"

Scheduled Tasks fail on every run with "issue with the selected model (OpenRouter/qwen/qwen3.8-max[1m])"

Summary

Every run of every Scheduled Task on my machine fails immediately with a model-selection error referencing an OpenRouter model I never selected for these tasks. The task never does any work — it errors out on the very first turn. This has been happening consistently for 6+ days across two different scheduled tasks and 8+ runs, and survived both deleting/recreating the tasks and removing the local Third-Party Inference config.

Environment

• macOS (Darwin 24.6.0)
• Claude Code / Claude Desktop app, Scheduled Tasks feature (tasks stored under ~/.claude/scheduled-tasks/<taskId>/SKILL.md)
• Interactive sessions in the same app use claude-sonnet-5 and work normally — only unattended Scheduled Task runs are affected

Error

Every scheduled run's first (and only) assistant turn is exactly:

There's an issue with the selected model (OpenRouter/qwen/qwen3.8-max[1m]). It may not exist or you may not have access to it.

The run then ends (error (success), 1 turns) without executing any of the task's instructions — no tool calls, no file writes, nothing.

Steps to reproduce

1. Create a scheduled task via the Scheduled Tasks feature (any recurring cronExpression task; no model field is exposed when creating one).
2. At some point, open Developer → Configure Third-Party Inference and configure a provider (in my case, OpenRouter with a qwen/qwen3.8-max[1m]-style model).
3. Later, deactivate Third-Party Inference from the same screen.
4. Let the scheduled task fire (or trigger "Run now").
5. Observe the task fails instantly with the error above.

What I already tried (did NOT fix it)

1. Deleted and fully recreated both affected scheduled tasks (same prompts, fresh task IDs registered via create_scheduled_task/delete_scheduled_task). The identical error occurred on the very next run after recreation. This rules out anything stored on the task definition itself — the task-creation API doesn't even expose a model parameter.
2. Removed the local Third-Party Inference config directory (~/Library/Application Support/Claude-3p/) entirely and restarted Claude Desktop. Before removal, configLibrary/_meta.json showed an appliedId still pointing at an entry named "Default" (with empty {} content) even though Third-Party Inference had already been deactivated in the UI — suggesting the "applied" state wasn't cleared by the deactivate toggle. After removing the folder and restarting the app, the exact same error still occurred on the next scheduled run.

Since the error survives both a full task recreation and a full removal of the local third-party-inference config, the bad model reference ("OpenRouter/qwen/qwen3.8-max[1m]") appears to be cached or stored somewhere else — possibly account/server-side, or in another local cache not covered by the Claude-3p directory.

Expected behavior

Scheduled/unattended task runs should use the account's normal default model (or whatever is configured), the same as interactive sessions — not a stale reference to a previously-configured (and since-deactivated) third-party inference model.

Actual behavior

Every scheduled task run fails immediately and does zero work, referencing a specific OpenRouter model that was never selected for these tasks and that the account may not even have access to.

Additional context

• Confirmed reproducible across two independent scheduled tasks (a 30-minute recurring one and a once-daily one), 8+ runs total over 6 days (2026-08-23 through 2026-08-29).
• Confirmed by reading each failed run's own transcript — the error is character-for-character identical every time.
• Happy to provide the exact task SKILL.md files or further logs if useful.