Mac single writer, staged: client layer now (#382), services after measurement; memory decided on merit - #383

What

Records the user's 2026-09-27 decision: native-agent-stack becomes the Mac's single writer, in two stages. It supersedes the 2026-09-24 agent-ecosystem#28 rule for mac-coordinator-64gb-20260925.

• Stage 1, now ([mac-coordinator] other: Stage 1 client layer on the 64 GB Mac (no service changes) #382):
  • Installs the Claude and Codex client layer: agents, guard hooks, MCP, settings, skills, plugins and workflows, plus the full token practice ([mac-coordinator] other: report and apply the token-efficiency coverage check (#264) on mac-coordinator-64gb-20260925 #276).
  • Changes no running service: the agent-ecosystem Qdrant and ai-memory launchd agents and Ollama are kept, and no second copy is started.
  • Memory hooks point at the running ai-memory or stay off.
  • ~/.claude and ~/.codex are backed up first.
• Stage 2, later: services move after the [mac-coordinator] model hosting: qualify Mac-native memory/RAG model hosting on the 64 GB Mac #379 capacity and parity numbers and the memory-layer head-to-head, by a cold-copy cutover with rollback.
• Memory on merit: the durable-memory layer is decided by a preregistered head-to-head over the user's starred repositories, the research-convergence layer, the memory-stack catalog and awesome-list discovery. ai-memory stays the reference arm, not a default.
  • Evidence so far, on the Mac only: agentmemory 0.821 recall_all@5, ai-memory production 0.570, BM25 0.747.
  • C4 and D2h have never run.
• Cross-host: GitHub host requests remain the durable record. Remote Control plus cross-session messaging is the live channel while the setup settles; messaging is not control.

docs/next-host-stages.md step 0 now points at the decision.

Not changed: adoption/host-roles.json (waits for #379's numbers), the macOS bootstrap, and any service.

Checks

• python3 scripts/validate.py: passed (7,145 hashed files, 159 receipts).
• python3 scripts/evidence_manifest.py --check: passed.
• python3 scripts/build_ecosystem.py --check: passed.
• python3 -m unittest tests.test_adoption_docs_consistency: OK (1 skipped).

SOTA sources

• Claude Code docs:
  • Remote Control: https://code.claude.com/docs/en/remote-control
  • Cross-session messaging: https://code.claude.com/docs/en/cross-session-messaging
  • Settings reference: https://code.claude.com/docs/en/settings-reference (crossSessionInbound, isolatePeerMachines, remoteControlAtStartup)
• Codex app-server daemon at rust-v0.157.1: https://github.com/openai/codex/blob/rust-v0.157.1/codex-rs/app-server-daemon/README.md
• Memory candidates: https://github.com/rohitg00/agentmemory (v0.9.29), https://github.com/vectorize-io/hindsight (v0.10.1)
• Repository records:
  • evidence/artifacts/host-upgrade-20260924/foundation-alignment.json
  • evidence/artifacts/memory-stack-20260925/experiment.json
  • docs/decisions/2026-09-25-host-request-lane.md

Merge timing: after the token lane's stacked queue, in a main-free window it offers. The coordinator merges it.

🤖 Generated with Claude Code

https://claude.ai/code/session_012qwXNtFyUhY5x2dtrm7tG5