[BUG] Routine-fired runs that create a fresh cloud session crash at Claude Code startup (turns=0); interactive sessions and into-existing-session deliveries work

In claude.ai/code, every routine-fired run that must create a fresh cloud session crashes at Claude Code startup with error_during_execution, turns=0 — while interactive sessions and scheduled deliveries into an existing session work fine on the same account, same repos, same hour.

Crash signature (identical in all four runs, from GET /v1/code/sessions/{id}/events): provisioning is fully green (Allocating sandbox → Cloning repository → Finished processing sources → Claude Code executor created → Claude Code process started), then 3–6 seconds later:

assistant: An error occurred while executing Claude Code. You can try again by sending a new message or starting a new session.
result: error_during_execution is_error=true turns=0 duration=0s
env[error]: Claude Code execution failed

Reproduction: create any run_once_at routine via POST /v1/code/triggers with a trivial prompt; let it fire (or fire it via /run). 4/4 crashes on this account, 2026-08-17 22:46–23:27 UTC:

| Session | Fired | Repo | Variation |
|---|---|---|---|
| cse_011CQM1p2vzoRfosQvZSE5Ri | 22:46:00Z | private repo A | minimal allowed_tools (Bash/Read/Glob/Grep) |
| cse_01Gykfbz65kFT6FWfWxFGXHY | 22:49:30Z | private repo A | same config, manual /run retry |
| cse_018Byi54XRuZebVG3XfUkL1o | 22:52:41Z | private repo A | preset:default + full tools list |
| cse_01NNo33JqcRoAsHNJk5QG7Sh | 23:27:22Z | private repo A | after interactive success (below) |

Plus one on a different, public repo (cse_01A6acJgXFdcKaYoDrbX5tVR, 22:59:35Z) — same crash. Ruled out: repo choice, presence/absence of a committed .claude/settings.json (public repo has none), tools-list shape, scheduled-fire vs. manual run.

What works on the same account in the same hour (~23:10–23:35 UTC):

• A user-initiated interactive claude.ai/code session on repo A: started normally, ran Bash, and called the auto-attached Claude Code Remote MCP tools successfully (list_triggers and send_later — notably with none of the -32003 "MCP tool call requires approval" errors from #61044, which we'd reported on that issue earlier).
• That session's send_later trigger fired at 23:31:00Z and delivered into the existing session on time — so the scheduler and the deliver-into-existing-session path are healthy. Only the create-a-fresh-session path dies.

Possibly relevant observations:

• Every crashed run's event log contains one pre-crash system/hook_started event (no hooks configured in either repo or the environment; no setup script).
• Routines auto-attach the Claude_Code_Remote connector (connector_uuid bf7c680d-5fdc-5ef4-b4a0-abadb619bf0a, https://api.anthropic.com/v1/code/mcp/meta) with permitted_tools: [].

Possibly related to the #61044 / #61015 routine-degradation cluster (same surface, different symptom — that one is mid-session tool denial, this is startup death). Happy to provide more event logs or run further controlled probes on request.