Claude + Mac fleet discovery; CLAUDE.md stale dev-root path corrected on all six hosts

Read-only discovery of Claude Code settings across the six-Mac fleet, followed by one authorized config change: the retired ~/dev/LLM/Claude/Claude/Projects path was corrected in ~/.claude/CLAUDE.md and propagated to all six hosts.

• Session: 2026-08-23 03:02:13 -> 03:5x EDT, America/New_York
• Origin host: rdmbair15m5 (M5, 32 GB, macOS 27.0 / 26A5416b)
• Scope: rdmsm4x, rdmbair13m5, rdmbair15m5, rdmpw3265m, rdmpw3275m, jdmbair13m5

Files touched

~/.claude/CLAUDE.md on all six hosts.

Pre-edit backup written on every host before the push: ~/.claude/CLAUDE.md.bak.20260823-03*-pre-devpath-fix

The five edits (path corrections only; no policy/permission/secrets text changed)

Verified live that ~/dev/LLM/Claude/Claude/Projects/ exists on NO host, rdmsm4x included. Only rdmsm4x:~/dev/LLM/Claude/changelogs/ (119 files) survives from that tree.

1. "Starting new work" step 1 -> rdmsm4x:~/dev/PROJECTS.md.
2. "Memory: project folders" -> rdmsm4x:~/dev/<domain>/<project-name>/, domain in apps|agy|fleet|sites|design|ai|arista|concepts|net|macos-tools. SESSION-STATE.md and ISSUES.md live inside the project folder (verified against 21 real instances).
3. HOST CAVEAT rewritten: path gone everywhere; changelogs branch named as sole survivor; added "ls -d the target before you write".
4. "Maintain PROJECTS.md" -> rdmsm4x:~/dev/PROJECTS.md; sub-registries noted.
5. Fleet-update run reports -> rdmsm4x:~/dev/fleet/_review/mac-fleet-updates/ (verified present).

Commands run

shasum -a256 ~/.claude/CLAUDE.md        # pre/post on all six
scp ~/.claude/CLAUDE.md richh@<host>.ts.dataroo.net:~/.claude/CLAUDE.md
# jdmbair13m5 relayed through rdmsm4x (no direct SSH from rdmbair15m5)
~/.agent-coordination/agent_msg.zsh broadcast --topic claudemd-devpath-fix-propagated-20260823

Verification evidence

All six hosts byte-identical, shasum -a256 prefix f024908f1626:

rdmbair15m5 f024908f1626    rdmpw3265m  f024908f1626
rdmsm4x     f024908f1626    rdmpw3275m  f024908f1626
rdmbair13m5 f024908f1626    jdmbair13m5 f024908f1626

A peer edit to the dev_update.zsh section ("identical on all six hosts", 9aba7a07cc1a) landed mid-push and was merged forward, not clobbered. Diff against the pre-edit backup confirmed only the 5 intended changes plus that peer text.

How to undo

On any host: cp ~/.claude/CLAUDE.md.bak.20260823-03*-pre-devpath-fix ~/.claude/CLAUDE.md

Discovery findings (no action taken)

• All six hosts: Claude Code 2.1.241, cask claude-code@latest, model: opus[1m]. Uniform: verbose, remoteControlAtStartup, crossSessionInbound=accept, both notify flags, 60s timeouts, SessionStart hook, zero custom agents/commands.
• permissions.defaultMode split 2/4: auto on rdmsm4x + rdmbair15m5, bypassPermissions on the other four. Unchanged. Escalated to owner; a peer message from claude@rdmsm4x (20260823-021736) declines to change it on peer request.
• Duplicate Tailscale node registrations on every Mac (stale <host> offline ~1d beside live <host>-1). MagicDNS verified to resolve the bare name to the LIVE node, so nothing is broken. Cleanup is a fleet-identity change - owner decision.
• jdmbair13m5 is unmanaged relative to the fleet: 1 skill, 0 plugins, 0 MCP, Developer mode disabled, macOS 26.6.2.
• rdmbair15m5 load: 15-min average 79 on 10 cores. Cause is NOT Claude. See below.
• MCP servers exist only on rdmbair15m5 (rdworkbench, chrome-devtools).
• superpowers hooks/run-hook.cmd is mode 755 here; the exit-126 bug is not present.

Raycast runaway - diagnosed, NOT remediated (needs owner consent)

PID 36395 Raycast Backend, elapsed 1d 08h 33m, 80-285% CPU, 41 threads, 3,851 open fds of which 3,662 are on ~/Library/Application Support/com.raycast.macos/index/db.

It is Raycast's Node.js extension host running a full-text file index: 2.8 GB across 3,678 .seg/.post segment files. It is ACTIVELY writing (file count grew during inspection; journal.wal current), so it is not deadlocked - it is a 32-hour indexing run that is not converging. Raycast 2.0.5.0. Folder read permissions include cloudStorage, the likely amplifier given the FileProvider-backed cloud mounts on this host.

Suggested remedy, owner's call: narrow Raycast's file-search scope (disable cloud storage indexing) in Raycast Settings, then quit and relaunch Raycast to abort the current run. No process was killed and no setting was changed.

Outstanding owner actions

1. SSH key install BLOCKED. rdmbair15m5 cannot SSH to jdmbair13m5; that host's authorized_keys holds all five other Macs but not this one. Appending the key was denied by the Claude Code permission classifier. To grant it, run from rdmbair15m5: ssh-copy-id -i ~/.ssh/id_ed25519.pub [email protected] (or relay the append through rdmsm4x). All jdmbair13m5 access this session was relayed.
2. Decide permissions.defaultMode policy (see split above).
3. Decide whether to delete the six stale Tailscale node registrations.
4. Decide on the Raycast index scope.

No secrets were read, written, or transmitted. No credential values appear in this record.