Scheduled routine sessions stuck at zero progress (worker idle, no tool calls, no error) — even minimal single-line prompts

Summary

Scheduled routines (via the Routines/Triggers API) consistently fail to make any progress — the CCR session provisions, logs "Claude Code process started", and then sits at worker_status: idle indefinitely. The web UI reports these runs as "completed" but the transcript shows only "initialized session" — no tool calls, no model response, no error message anywhere.

Reproduction

1. Created a new routine ("Daily Briefing") on a freshly auto-created environment (this was the first-ever routine on this account, so the environment was provisioned moments before the first attempt).
2. Ran it via POST /v1/code/triggers/{id}/run three times over ~20 minutes, regenerating the event UUID between attempts in case of an idempotency issue. All three sessions failed identically: status: active, worker_status: idle, no events past "Claude Code process started".
3. To rule out prompt/tool/MCP complexity as the cause, created a second, minimal diagnostic routine with a trivial single-line prompt ("Reply with exactly: 'Diagnostic routine is working.' Do nothing else."), no MCP connectors, no tools beyond Read/Write, scheduled via run_once_at a few minutes out. This also failed identically — same idle-at-zero-progress state.
4. Checked https://status.claude.com — all systems operational at time of testing, no active incident (though incidents on Aug 14 and Aug 16 had explicitly listed "Routines" as affected, now resolved).

Expected behavior

The routine should process the initial user event and begin executing the prompt (tool calls, model response), or fail with a visible error if something prevented that.

Actual behavior

Session silently never starts real work. No error surfaces via get_run_log, list_runs, or the web UI (which just shows "completed" / "initialized session").

Environment

• Trigger type: cron-scheduled routine + one-time run_once_at diagnostic
• Environment: freshly auto-created "Default" anthropic_cloud environment, first-ever use on this account
• Model: claude-sonnet-5
• Sources: none (no repo attached)
• Observed via: RemoteTrigger/Code Triggers API (list_runs, get_run_log, get)

Suspected cause

Given this was the very first routine ever run on a brand-new auto-provisioned environment, and the failure is 100% reproducible (4/4 sessions, including a trivial no-MCP prompt) within a single account/environment, this looks like a first-use provisioning bug specific to new environments rather than general platform flakiness.

Related but not identical

#24048 describes interactive sessions going idle mid-task after initial progress. This report is different: zero progress ever occurs, and it's specific to routine-triggered sessions, not interactive ones.