Real-app replay comparison (UTC; Linux, Chromium 151, 1440x900, light theme, Claude Haiku 5.5)

Fixture: identical 14,000 untracked text files, 568,890,000 bytes (542.54 MiB).
Dataset SHA256 (file names and bytes): 95de043d47b6451e8678843d746aee7beb8b7b1f2f1aaf492a41f0aedc133717
The smaller ~236 MiB fixture reproduced the base but sometimes finished its head checkpoint before the wakes. The final pair uses this larger identical fixture to ensure overlap on both commits.

BASE 9ab1d22d8fc0bbf491d92995f92fb6eb94efe8b0:
Initial checkpoint: 30,372.7 ms (timeout).
ALPHA result: 22:05:51.518. BRAVO result: 22:05:54.142.
CHARLIE root init: 22:05:55.462.
Continuation adapter attached: 22:05:56.626.
Provider turn settled: 22:05:56.632, using older BRAVO result.
CHARLIE delegation request: 22:06:03.263; returned parent_not_active.
CHARLIE native result: 22:06:07.672; replayed in a later continuation.
Earlier ALPHA usage lost: continuation outputTokens=9 (BRAVO only), despite two earlier 9-token results.

HEAD e7b32a77e310ddcd666594736106ef647fdc2e33:
Initial checkpoint: 30,726.0 ms (timeout).
ALPHA result: 22:03:15.887. BRAVO result: 22:03:19.052.
CHARLIE root init: 22:03:19.971.
Continuation adapter attached: 22:03:23.222.
CHARLIE delegation request: 22:03:27.842; child completed with CHILD READY.
Provider turn stayed active until CHARLIE native result: 22:04:06.187.
Earlier usage retained: continuation outputTokens=803 = ALPHA 9 + BRAVO 9 + CHARLIE 785.
Persisted cachedInputTokens=190598 = 13388 + 34041 + 143169.
Persisted cacheCreationTokens=25536 = 20653 + 403 + 4480.

The following sections are extracted from real native provider logs and read-only server projections, not tests. Native redacted logs are provided separately.

BASE
All timestamps UTC. Synthetic project and real Claude Haiku 5.5.

2026-11-14T22:05:08.704Z App run 1 startedAt; final status=completed
2026-11-14T22:05:13.809Z Provider turn 1 adapter attached
2026-11-14T22:05:14.519Z Native root init
2026-11-14T22:05:20.824Z Native result success num_turns=4 usage={"input_tokens": 4, "cache_creation_input_tokens": 20474, "cache_read_input_tokens": 46083, "output_tokens": 635, "output_tokens_details": {"thinking_tokens": 233}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 20474, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=STARTED
2026-11-14T22:05:20.825Z Provider turn 1 complete; billed token usage={"usageScope": "main_agent", "cachedInputTokens": 46083, "cacheCreationTokens": 20474, "reasoningTokens": 233, "hasSubagents": false, "models": ["claude-haiku-5-5"], "otherUsage": {"inputTokens": 0, "outputTokens": 0}, "usageStatus": "complete", "inputTokens": 66561, "outputTokens": 635}
2026-11-14T22:05:20.857Z App run 1 completedAt; final status=completed
2026-11-14T22:05:49.508Z Native root init
2026-11-14T22:05:51.479Z App run 2 startedAt; final status=completed
2026-11-14T22:05:51.518Z Native result success num_turns=1 usage={"input_tokens": 4, "cache_creation_input_tokens": 20635, "cache_read_input_tokens": 13388, "output_tokens": 9, "output_tokens_details": {"thinking_tokens": 0}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 20635, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=ALPHA handled
2026-11-14T22:05:52.423Z Native root init
2026-11-14T22:05:54.142Z Native result success num_turns=1 usage={"input_tokens": 4, "cache_creation_input_tokens": 399, "cache_read_input_tokens": 34023, "output_tokens": 9, "output_tokens_details": {"thinking_tokens": 0}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 399, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=BRAVO handled
2026-11-14T22:05:55.462Z Native root init
2026-11-14T22:05:56.626Z Provider turn 2 adapter attached
2026-11-14T22:05:56.632Z Provider turn 2 complete; billed token usage={"usageScope": "main_agent", "cachedInputTokens": 34023, "cacheCreationTokens": 399, "reasoningTokens": 0, "hasSubagents": false, "models": ["claude-haiku-5-5"], "otherUsage": {"inputTokens": 0, "outputTokens": 0}, "usageStatus": "complete", "inputTokens": 34426, "outputTokens": 9}
2026-11-14T22:05:56.772Z App run 2 completedAt; final status=completed
2026-11-14T22:06:03.263Z Native delegate_task request: {"task": "Reply only CHILD READY", "mode": "wait", "target": {"providerInstanceId": "claude", "driverKind": "claudeAgent", "model": "claude-haiku-5-5"}}
2026-11-14T22:06:07.672Z Native result success num_turns=4 usage={"input_tokens": 10, "cache_creation_input_tokens": 4731, "cache_read_input_tokens": 144038, "output_tokens": 1004, "output_tokens_details": {"thinking_tokens": 349}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 4731, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=CHARLIE arrived, so I called `orchestrator_capabilities` and then `delegate_task` for the Claude Haiku child. The delegation failed, so the child never ran.

- **Capabilities:** Claude provider `claude`, model `claude-haiku-5-5`, `canRunChildTask: true`.
- **Delegation:** `delegate_task` returned `OrchestratorMcpFailure`, code `parent_not_active`: "The calling provider no longer owns an active thread run."
- **Child reply:** none. "CHILD READY" was never produced.

This is a run-ownership error on the orchestrator side, not a missing provider or model. I did not retry or fall back to the native `Agent` tool, since you asked for no native delegation. A new turn should be able to re-run the delegation.

Earlier results: ALPHA handled, BRAVO handled.
2026-11-14T22:06:23.048Z App run 3 startedAt; final status=completed
2026-11-14T22:06:23.313Z Provider turn 3 adapter attached
2026-11-14T22:06:23.374Z App tool returned parent_not_active: The calling provider no longer owns an active thread run.
2026-11-14T22:06:23.374Z Provider turn 3 complete; billed token usage={"usageScope": "main_agent", "cachedInputTokens": 144038, "cacheCreationTokens": 4731, "reasoningTokens": 349, "hasSubagents": false, "models": ["claude-haiku-5-5"], "otherUsage": {"inputTokens": 0, "outputTokens": 0}, "usageStatus": "complete", "inputTokens": 148779, "outputTokens": 1004}
2026-11-14T22:06:23.973Z App run 3 completedAt; final status=completed


HEAD
All timestamps UTC. Synthetic project and real Claude Haiku 5.5.

2026-11-14T22:02:33.111Z App run 1 startedAt; final status=completed
2026-11-14T22:02:38.248Z Provider turn 1 adapter attached
2026-11-14T22:02:39.062Z Native root init
2026-11-14T22:02:46.535Z Native result success num_turns=4 usage={"input_tokens": 4, "cache_creation_input_tokens": 20533, "cache_read_input_tokens": 46081, "output_tokens": 680, "output_tokens_details": {"thinking_tokens": 278}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 20533, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=STARTED
2026-11-14T22:02:46.535Z Provider turn 1 complete; billed token usage={"usageScope": "main_agent", "cachedInputTokens": 46081, "cacheCreationTokens": 20533, "reasoningTokens": 278, "hasSubagents": false, "models": ["claude-haiku-5-5"], "otherUsage": {"inputTokens": 0, "outputTokens": 0}, "usageStatus": "complete", "inputTokens": 66618, "outputTokens": 680}
2026-11-14T22:02:47.068Z App run 1 completedAt; final status=completed
2026-11-14T22:03:14.033Z Native root init
2026-11-14T22:03:15.887Z Native result success num_turns=1 usage={"input_tokens": 4, "cache_creation_input_tokens": 20653, "cache_read_input_tokens": 13388, "output_tokens": 9, "output_tokens_details": {"thinking_tokens": 0}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 20653, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=ALPHA handled
2026-11-14T22:03:17.009Z Native root init
2026-11-14T22:03:18.101Z App run 2 startedAt; final status=completed
2026-11-14T22:03:19.052Z Native result success num_turns=1 usage={"input_tokens": 4, "cache_creation_input_tokens": 403, "cache_read_input_tokens": 34041, "output_tokens": 9, "output_tokens_details": {"thinking_tokens": 0}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 403, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=BRAVO handled
2026-11-14T22:03:19.971Z Native root init
2026-11-14T22:03:23.222Z Provider turn 2 adapter attached
2026-11-14T22:03:27.842Z Native delegate_task request: {"task": "Reply only CHILD READY", "mode": "wait", "target": {"providerInstanceId": "claudeAgent", "driverKind": "claudeAgent", "model": "claude-haiku-5-5"}}
2026-11-14T22:04:06.187Z Native result success num_turns=4 usage={"input_tokens": 10, "cache_creation_input_tokens": 4480, "cache_read_input_tokens": 143169, "output_tokens": 785, "output_tokens_details": {"thinking_tokens": 153}, "server_tool_use": {"web_search_requests": 0, "web_fetch_requests": 0}, "service_tier": "standard", "cache_creation": {"ephemeral_1h_input_tokens": 4480, "ephemeral_5m_input_tokens": 0}, "inference_geo": "global", "iterations": [], "speed": "standard", "fallback_credit": null} result=CHARLIE handled. The Haiku child delegation completed: it returned `CHILD READY` on `claude-haiku-5-5` (provider `claudeAgent`). Its thread is [child thread].

I used `orchestrator_capabilities` first, then `delegate_task` in wait mode, not the native Agent tool.
2026-11-14T22:04:06.187Z Provider turn 2 complete; billed token usage={"usageScope": "main_agent", "cachedInputTokens": 190598, "cacheCreationTokens": 25536, "reasoningTokens": 153, "hasSubagents": false, "models": ["claude-haiku-5-5"], "otherUsage": {"inputTokens": 0, "outputTokens": 0}, "usageStatus": "complete", "inputTokens": 216152, "outputTokens": 803}
2026-11-14T22:04:06.287Z App run 2 completedAt; final status=completed