OpenCode 2: every failed historian run leaves a new hidden session behind (114 in 2.5 h with a quota-exhausted provider) #481

What happened

On OpenCode 2 (2.0.7, magic-context 0.42.6) the historian's models were all unavailable for a while: the primary (opencode-go/gpt-5.6-luna) returned "Go usage limit exceeded" (weekly quota), and the fallback (opencode/muse-spark-1.3-contributor-free) returned "OpenCode's free tier can only be used from within OpenCode". Every historian run failed, and every failed run left a new Magic Context historian session behind: 114 sessions in about 2.5 hours in one project, all with 3–9 messages, all root sessions, so they fill the session list.

The project's hidden-children meta showed active: {} and 114 retired_children, every one with reason hidden-run-failed.

Why

In packages/plugin/src/v2/hidden-completion.ts:

• close() retires the child on any failed run (hidden-run-failed).
• open() also retires an active child whose newest assistant row is not a successful completion (newest-assistant-not-reusable), and then creates a new session.

A provider error (quota, rate limit, refused request) is persisted as a settled assistant row with error set. So a quota-exhausted provider guarantees a new child session per historian trigger until the quota resets, with nothing to clean the sessions up.

Why reusing the child is safe in this case

The hidden-child hook replaces the child's whole draft on every hidden prompt (draft.system, draft.messages = [calibrated request], draft.tools = {}). The error row is never sent to a model again, so a child whose only problem is a settled provider error is as good as a fresh one.

Proposed fix (PR to follow)

• A run that failed only on settled provider-error rows keeps its child: no retire in close(), and open() accepts a child whose newest assistant is a settled provider error.
• Every other failure keeps today's behaviour and retires the child: dispatch failures (a thrown prompt), refusals, timeouts and aborts.

A backoff for repeated provider errors would also stop the historian from re-trying a quota-exhausted pool on every trigger. That is a separate, larger change, so it is not part of this PR.