Settled Magic Context child sessions are not deleted on OpenCode 2 (teardown delete is directory-scoped; orphan sweep reads the v1 session table) #558

Summary

On Windows + OpenCode 2 (OpenChamber), settled Magic Context internal child sessions (dreamer*, historian, dreamer-classifier, dreamer-memory-mapper) are not deleted after success. They linger in the session store for many hours (observed 0.1h–27h) and are shown as normal, clickable sessions in the host UI.

keep_subagents is at its default false, i.e. the intended behavior is "delete every settled Magic Context child session after success". The teardown path attempts that delete but it does not take effect, and the failure is swallowed by a .catch().

Environment

• OS: Windows 11
• Harness: OpenCode 2.x via OpenChamber 2.0.0
• Magic Context: @cortexkit/opencode-magic-context@0.43.2
• Stores: ~/.local/share/opencode/opencode.db, ~/.local/share/cortexkit/magic-context/context.db

Evidence

All lingering rows carry the internal marker and a terminal succeeded outcome:

metadata: {"magic_context":"hidden-run","role":"dreamer"|"dreamer-curate"|"historian"}
outcome:  succeeded (most), failed (some)
age:      0.1h … 27.4h
count:    22 lingering at one point in time (after a full day of scheduled tasks)

keep_subagents in config: unset → default false (description: "Debug: keep every settled Magic Context child session instead of deleting it after success").

Root cause (from source)

teardownChildSession (dist) does:

if (promptSettled && !shouldKeepSubagents()) {
  await client.session.delete(request).catch((error) => {
    log(`${context}: session cleanup failed: ${String(error)}`);
  });
  return;
}
if (!promptSettled) {
  // archive + "left to the age-gated sweep"
}

Two problems combine:

1. The delete is directory-scoped. client.session.delete({ path: { id }, query: { directory: sessionDirectory } }) only resolves the session when the caller's directory matches the session's directory. Reproduced directly: deleting a child session from the wrong cwd returns Session not found (exit 1); deleting the same id from the correct directory returns deleted (exit 0). If the spawn/teardown directory and the session directory ever diverge (worktrees, per-chat directories, transient cwd changes), the delete silently no-ops.

2. The fallback sweep cannot clean them either. sweepOrphanedRetrospectiveChildren reads from the v1 session table:

   SELECT id, time_created FROM session
   WHERE directory = ? AND (…age predicates…)

   On an OpenCode 2 store the live rows are in session_v2, so the sweep never sees them. The stale rows also only match by exact directory, which misses the divergent-directory case above.

Additionally, the host log shows the child sessions failing to drain at boot:

Failed to drain Session
cause="HiddenCompletionRefusal: hidden_prompt_unrecognized:
       Refusing an unregistered prompt on a Magic Context hidden-run session"

So on OpenCode 2 the hidden-run lifecycle is not completing end-to-end, and nothing reclaims the leftovers.

Impact

1. Session-list pollution. Every scheduled Dreamer task in every tracked project leaves a permanent, clickable session in the host UI. Over a day this is dozens; the list becomes unusable.
2. No user-facing filter. The host does not read the magic_context: "hidden-run" metadata marker, so there is no way to hide them.
3. Storage growth. The rows (and their messages) accumulate indefinitely.

Requests

1. Make the teardown delete robust on OpenCode 2 — resolve the child session's own directory (it is known at spawn time) instead of relying on the caller's cwd, and surface delete failures instead of swallowing them.
2. Point the orphan sweep at session_v2 on OpenCode 2 (or make it store-aware), and match children by their internal marker/title rather than only by exact directory.
3. Consider a user-level keep_subagents-adjacent visibility option so operators can suppress internal child sessions in hosts that do not read the metadata marker.

Notes

• The magic_context: "hidden-run" marker already exists and is reliable — a host that honors it could filter these out with no store changes.
• Related: internal child lifecycle / OpenCode 2 readiness (OpenCode V2 readiness: existing-state migration and remaining V2 capability gaps #492).