Steward handoff cannot execute in a cloud run: create_session fails with "permission mode is not yet available"

[peterdrier]

What happens

memory/process/steward-thin-session.md requires the session that built a PR to hand it to a thin steward session at ready-for-review and stop. In an unattended cloud run that handoff is mcp__Claude_Code_Remote__create_session, and it fails every time with:

the parent session's permission mode is not yet available (it is recorded
shortly after the parent session starts); retry, or run the parent in auto mode.

Four attempts across the 2026-09-22 section-doctor run on Email (#1790), spread over roughly eleven hours — the first three around PR creation at 11:06Z, the fourth at 22:20Z. Identical error each time, with and without an explicit permission_mode.

The error text says the condition is transient ("recorded shortly after the parent session starts"). It is not: the fourth attempt was eleven hours in.

Why it matters

The handoff is the whole point of the thin-steward rule — the building session carries a large context, and the steward is supposed to be a cheap sonnet session that does nothing but watch. When create_session fails there are only two outcomes, and both are bad:

1. The building session stays subscribed and runs the wake protocol itself, burning its own context on CI and review events. That is what happened on doctor(Email): guard the drain's failure path, cut the dead backfill arm, retire the prose about a section that moved #1790, and it is the skill's documented fallback — but it is the fallback, not the design.
2. Nobody watches the PR at all.

It also means memory/process/steward-thin-session.md describes a workflow that cannot currently run unattended in the cloud, which is where section-doctor runs.

Reproduce

From an unattended cloud run (CLAUDE_CODE_REMOTE=true), after opening a PR, call mcp__Claude_Code_Remote__create_session with any prompt. It fails regardless of elapsed time since session start.

Possible directions

Not investigated — recording the defect, not the fix.

• If "run the parent in auto mode" is the real remedy, the routine that fires section-doctor needs its permission mode set accordingly, and steward-thin-session.md should say so.
• If the permission mode genuinely is never recorded for trigger-fired sessions, that is a platform bug rather than a repo one.
• Either way the skill should state what to do when the handoff is unavailable, rather than leaving it to the fallback paragraph.

Context

Found during the 2026-09-22 section-doctor run on Email; the deviation is recorded in docs/health/runs/2026-09-22-Email.md and was surfaced to Peter twice during the run.