P0 regression: prompt delivery broken after async delivery pipeline (#3253)

[OneStepAt4time]

Regression

After the async prompt delivery pipeline merged (2026-05-12), CC sessions start but never receive prompts. Dogfooding is broken — Hep cannot forge through Aegis.

Suspect PRs

• feat(sessions): async prompt delivery for ACP sessions (#3243) #3253 — feat(sessions): async prompt delivery for ACP sessions (merged 2026-05-12 15:51 UTC) — PRIMARY SUSPECT
• fix(acp): increase default prompt timeout 60s to 120s for BYO-LLM proxies (#3243) #3244 — fix(acp): increase default prompt timeout 60s to 120s (merged 2026-05-12 12:12 UTC) — superseded fix(acp): increase JSON-RPC timeout to 60s and make configurable (#3223) #3225
• fix(dashboard): promptDelivery schema missing status field (#3256) #3258 — fix(dashboard): promptDelivery schema missing status field (merged 2026-05-12 19:21 UTC) — dashboard fix only, unlikely

Diagnosis notes

• Regression timeline suggests async delivery (feat(sessions): async prompt delivery for ACP sessions (#3243) #3253), not timeout change
• If revert needed, revert feat(sessions): async prompt delivery for ACP sessions (#3243) #3253 (not fix(acp): increase default prompt timeout 60s to 120s for BYO-LLM proxies (#3243) #3244)
• Boss directive: diagnose before cutting anything

Impact

• All Aegis sessions affected (cannot receive prompts)
• Blocks Hep from forging through Aegis
• Blocks dogfooding entirely

Priority: P0 — core functionality broken

Refs: #3223, #3225, #3243, #3244, #3253, #3256, #3258

[OneStepAt4time]

Root cause diagnosis (Hephaestus)

NOT a timeout regression from #3225. Race condition in async prompt delivery (#3243/#3253).

Chain:

1. acpBackend.createSession() spawns CC process + completes ACP handshake (initialize + session/new)
2. Async deliverAsync() calls sendPrompt() → sessionService.getSession() to read acpAgentSessionId
3. acpAgentSessionId is null — ACP session record not persisted between handshake and async delivery

Error: prompt goes pending → failed with attempts: 1, error: 'no_agent_session'

Scope: All new sessions with prompts on running server (PID 107761, up since May 12). 11+ orphaned CC processes observed.

Fix path: Either persist acpAgentSessionId synchronously before dispatching async delivery, or add retry/wait logic in deliverAsync(). NOT a revert candidate — surgical fix in the async pipeline.

Server restart alone won't fix if the code has the race condition. Code fix + restart needed.