[liuhao-labs]

Summary

Long full-history prompt rendering blocks the agent adapter's aiohttp event loop. Move the existing _session_prompt_ids() call to one adapter-owned worker thread, retaining exact upstream rendering, saved sampled-token prefixes, and safe fallback for changed histories/tool schemas.

The worker bounds submitted running/queued work. Canceled requests retain their executor capacity until physical completion, discard queued jobs, return promptly during admission cancellation, and avoid starting SGLang or recording a trajectory after cancellation. App cleanup drains the worker. Session prefix lists/template options are snapshotted before offload; the live session is used for generation and records only flushed turns.

Bounded recent-window metrics at GET /debug/adapter_metrics include prompt length, total prompt-render latency, queue wait, worker utilization, event-loop lag, cancellations and discarded work, with P50/P95/P99 distributions.

Related to #2285.

Scope

This improves event-loop responsiveness and observability. It preserves upstream full-history rendering and canonical-boundary checks, including original sampled prompt/response token prefixes introduced by #2457. It does not remove the cumulative O(N²) tokenization cost or claim general incremental tokenization. Sampler score-centering/top-p/R3 metadata, output-token validation, and protocol behavior are unchanged.

Validation

• PYTHONPATH=. python -m pytest -q tests/test_agent tests/test_grouped_rewards.py: 169 passed, 2 skipped against main 58b08786.
• Direct CPU runners: worker tests 12 passed; prompt continuity 14 passed. Real in-memory BPE tests compare synchronous and worker IDs for tool continuation, changed tools and compaction, and assert both full-history and splice-boundary renders run off the event-loop thread.
• Cancellation handoff, repeated cancellation with delayed cleanup, stable session snapshots, capacity ownership, health/metrics responsiveness and executor shutdown regressions passed.
• The upstream sampler replay/token contracts and Responses/Anthropic/Codex harness tests passed.
• Local synthetic continuity workload: 16,250 prompt IDs, 12 renders per mode, identical IDs. Synchronous heartbeat p95/max lag was 47.17/48.64 ms; worker p95/max lag was 0.63/1.09 ms. Total render time was 0.197/0.206 s respectively. These single-host timings show event-loop isolation, not a tokenizer throughput improvement or production guarantee.
• Repository-pinned pre-commit, workflow regeneration, YAML and git diff --check passed. Sunabako 8-GPU/35-minute E2E registration, container init/cidfile cleanup, proxy/mount setup and artifacts remain unchanged. GPU E2E execution is not claimed locally.