# Issue 25021 runtime transcript

Real T3 dev server, current main 313ca2715f, browser at localhost:6514, isolated worktree state. CLI provider boundary fixture replaces external Claude responses; no shared gateway settings or credentials changed. This is controlled protocol reproduction, not a claim that the live pool was exhausted.

Live baseline: Claude Opus 5.5 started native background Agent with sleep 15, root ended, task finished and wake answered WAKE OK. No cooldown occurred.

Controlled before: CLI fixture emits background task started, a root wrapped gateway 503 with reset after 10s, then task completion plus a wake wrapped 503 with reset after 1m. Real adapter buffers the wake result, real orchestration requests continuation, and the continuation replays the result.
- Wake fixture emitted: 2026-10-25T22:16:22.529Z
- Continuation started: 2026-10-25T22:16:22.603Z
- Persisted failure resetAt: 2026-10-25T22:17:22.987Z
- Reset anchored at replay, 458ms later than fixture emission +60s.

Controlled after, same CLI messages and running app:
- Wake fixture emitted: 2026-10-25T22:17:39.148Z
- Persisted failure resetAt: 2026-10-25T22:18:39.148Z
- Reset equals receipt +60s; turn completion still records replay completion.

The fixture emission timestamp is immediately before the stdout write, not an instrumented adapter receipt timestamp. The deterministic TestClock regression test proves the exact one-hour queue-delay case without network/scheduling ambiguity:
Expected 1970-01-01T02:54:21.000Z; before received 1970-01-01T03:54:21.000Z; after passes.

No render code changed. Proof matrix is empty for this server-only fix. No screenshots are required. The runtime JSON contains only synthetic data. The CLI fixture is evidence-only and is not committed.