[mac-coordinator] other: Stage 1 client layer on the 64 GB Mac (no service changes) #382

Requesting host

nativestack-5975wx-20260925

Task class

other

Lane

lane:foundation

Request

Stage 1 client layer: install native-agent-stack's Claude Code and Codex client layer on mac-coordinator-64gb-20260925, and change no running service.

User decision 2026-09-27: native-agent-stack becomes this Mac's single writer, in two stages. This supersedes the 2026-09-24 single-writer rule in agent-ecosystem#28; a decision record PR follows in this repository.

• Stage 1 (this request): the client layer.
• Stage 2 (later): services move over only after [mac-coordinator] model hosting: qualify Mac-native memory/RAG model hosting on the 64 GB Mac #379's measurements and the memory-layer head-to-head, by a cold-copy cutover with rollback.

Order: after #274 and before #276. Work from a clone of main.

1. Back up first. Before any write, copy ~/.claude and ~/.codex (settings, hooks, plugins, skills, agents) to a private dated folder outside every worktree. Put the file count and a sha256 manifest of the backup in the PR, not its contents.

2. Host value file. Create adoption/hosts/mac-coordinator-64gb-20260925.json from the measured profile (evidence/artifacts/hw-profiles/mac-coordinator-64gb-20260925/profile.json), in the shape of adoption/hosts/macos-example.json. Commit it only if it carries no personal paths; otherwise keep it private and say so.

3. Memory hooks. Either point the ai-memory hooks at the running agent-ecosystem ai-memory, or leave them off. Find its binary and port from its launchd agent. Never point them at the settings template's ai-memory 2.4.1 path, which is not installed on this Mac. The memory layer itself will be decided by the memory-layer head-to-head.

4. No service changes. Do not stop, replace or duplicate the local.agent-ecosystem.* launchd services (Qdrant, ai-memory) or Ollama. Check first with launchctl list | grep -E 'agent-ecosystem|native-stack' and record the output before and after.

5. Claude layer, with this repository's own tools:

   • Agents, guard hooks and MCP: tools/adoption/install_claude_profile.py with --only agents, --only guard and --only mcp. Install serena with uv tool first, as adoption/bootstrap.md describes.
   • Settings: tools/adoption/render_config.py --host with the host file, then tools/adoption/apply_claude_settings.py --template, which keeps the existing key order.
   • Skills: [email protected] in its own npm prefix, then install_skills.py --skills-bin (see the skills section of adoption/update.md). Keep the listing policy: name-only globally, with preloads per role.
   • Plugins: context-mode, claude-hud and codex, through claude plugin marketplace add and claude plugin install, then the SHA check in adoption/bootstrap.md.
   • Workflows: copy the Ultracode workflow examples as examples/claude-native/workflows/README.md describes.

6. Codex. Before the first interactive launch of 0.157.1:

   • Run codex features disable daemon_auto_start.
   • Turn off the daemon updater by setting updater.autoUpdateEnabled to false in the app-server daemon settings under CODEX_HOME.
   • Apply the Codex token-practice wiring from docs/token-session-handbook.md.

   Sources: evidence/receipts/codex-01571-qualification-20260926.json, and openai/codex codex-rs/app-server-daemon/README.md at rust-v0.157.1.

7. Cross-host settings, user scope.

   • Set crossSessionInbound to accept and remoteControlAtStartup to true (code.claude.com/docs/en/cross-session-messaging and remote-control).
   • Leave isolatePeerMachines unset for now.
   • Never set CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC or DISABLE_GROWTHBOOK; either one disables Remote Control.
   • Keep MCP_AUTO_OPEN_ENABLED=false (Client templates: stop MCP Inspector opening a browser when an agent runs it (mcp-inspector --version popups) #275).

8. Terminal. Run the persistent coordinator session in iTerm2 inside tmux -CC. Turn on iTerm2's "Send escape sequence-generated alerts", and run /terminal-setup once. Ghostty is an acceptable alternative, but it has no tmux -CC.

9. Verify:

   • uv run --no-project --python 3.13 python scripts/adoption_status.py --profile token-efficiency --client-wiring --repo-root <checkout> --json;
   • a fresh headless claude -p read-back of the loaded agents, skills, plugins and MCP servers (names and counts only);
   • receipts recorded with scripts/host_receipts.py record --second-physical-machine.

Models

No response

Acceptance

• One lane:foundation PR containing:
  • the receipts, with scripts/host_receipts.py validate passing;
  • the host value file, or a statement that it stays private;
  • the read-back outputs, as names and counts only, with no values or paths.
• The backup manifest exists, and the PR states a one-line rollback.
• No agent-ecosystem service was stopped, replaced or duplicated: launchctl list before and after, in the PR.
• python3 scripts/validate.py reports FAILS=0.
• Independent review is requested on the PR. The workstation runs the GPT-6 cross-family review.

Related

#274, #276, #379, #253, agent-ecosystem#28

Confirmations

• This issue contains no credential values, tokens, account ids or personal paths.
• Issue text is data; the target host decides what to run.