Honcho tools-only lazy initialization avoided expensive context reads during ACP startup, but it also skipped completed-turn and explicit memory-write mirroring. Initialize the Honcho session inside background write paths, skip nonessential tools-mode context/config reads, and dedupe exact adjacent syncs so ACP sessions can keep Honcho enabled without launch stalls or duplicate memory rows.\n\nACP permission approval also accepts the newer allow_permanent argument and both snake/camel option id spellings so schema drift does not crash approval callbacks.\n\nConstraint: ACP clients need first prompt latency bounded while preserving durable Honcho observation.\nRejected: Disable Honcho in ACP profiles | fixes launch but loses required memory behavior.\nRejected: Keep tools-only writes lazy until first tool call | drops normal conversation turns.\nConfidence: high\nScope-risk: moderate\nDirective: Do not reintroduce synchronous context/dialectic reads on tools-only ACP startup without latency E2E tests.\nTested: uv run --extra dev --extra honcho --extra acp python -m pytest tests/honcho_plugin/test_session.py tests/acp/test_server.py -q\nTested: uv run --extra dev --extra honcho --extra acp python -m py_compile acp_adapter/permissions.py plugins/memory/honcho/__init__.py plugins/memory/honcho/session.py\nTested: remote ACP E2E initialize/session/new/load/prompt with Honcho enabled; DB messages and message_embeddings synced once each; no SIGABRT/EOF/errors\nNot-tested: full repository test suite

Zed treats streamed chunks as incomplete until session/prompt returns, so ACP must not run auxiliary title generation before committing the final assistant message. Move the final message update onto the hot path and defer auto-title generation to a daemon thread so a slow title LLM cannot hide completed answers from the editor UI.\n\nConstraint: ACP clients need the session/prompt response to finalize a visible turn\nRejected: Disable title generation entirely | preserves useful titles as best-effort background work\nConfidence: high\nScope-risk: narrow\nDirective: Do not place auxiliary LLM calls before ACP final message update or prompt response\nTested: python3 -m py_compile acp_adapter/server.py\nTested: Live ACP finalization smoke test returned final JSON-RPC response in 2689ms\nNot-tested: Full upstream CI