herdr resume — service handover completed, TCC cost measured, jdmbair13m5 disk-full root cause

Session: claude@rdmbair15m5 · 2026-09-06 14:00:24 → 14:20 EDT (~20 min) · resume of the 2026-09-05 rollout

Headline

The launchd handover predicted on 09-05 completed by itself, and it had a real cost that only became measurable once it did.

• herdr is now launchd-owned on rdmbair15m5: PID 1107, ppid=1, brew services = started. HERDR-20260905-07 → RESOLVED, no intervention.
• Fleet state: 6/6 herdr 0.8.2 · 5/5 remote services started · config sha 26c720d1d2a5 byte-identical on 6/6 · reload-config = applied on all.
• Cost of the handover: the TCC probe that read all five OK on 09-05 under Terminal.app reads all five MISS today under launchd — and in the Claude context, not just herdr's own.

That the Claude result changed proves claude is not its own responsible process here — so making herdr a service stripped Screen Recording / FDA / Accessibility from every agent running inside herdr. Granting the herdr binary should restore it for all panes (inference, untested).

Verified mitigation for the FDA half: ssh localhost '<cmd>' — sshd holds FDA fleet-wide. Direct read DENIED, via ssh localhost OK. Also found: this Mac cannot ssh to itself by hostname (times out) — corroborates HERDR-20260905-08.

Cannot be automated (circular): computer-use MCP needs Accessibility + Screen Recording to drive System Settings — the very grants that are missing. MDM PPPC also needs approval on a non-enrolled Mac. Needs Rich at a keyboard.

jdmbair13m5 — socket bind failure was a FULL DISK (HERDR-20260906-17)

fleet_herdr_attach.zsh failed there with remote port forwarding failed for listen path. Ruled out in order: not the stale-socket trap (a fresh path failed, and so did /tmp); not sshd config (byte-identical to a working host). Actual cause — ssh -R cannot bind a socket it cannot create:

/dev/disk3s5   926Gi   901Gi   100Mi   100%   /System/Volumes/Data
touch ~/.config/herdr/writetest -> No space left on device

No APFS snapshots (count 0). Real data: Pictures 332 GB, Documents 260 GB, Library 55 GB. Reclaimed to 1.4 GiB using vendor-managed maintenance only (brew cleanup --prune=all = 756 MB, plus tool caches). No user data touched. Bind then verified BIND-OK and the pane re-attached — causally clean in both directions (fails at 100 MiB, succeeds at 1.4 GiB). jdmbair13m5 has a 926 GiB disk while every other fleet Mac has 3.6–7.3 TiB; 0.15% headroom will refill. Needs Rich.

Other findings

• HERDR-20260906-16 — Claude is NOT LOGGED IN on rdmbair13m5 (Not logged in · Run /login). The pane is decorative until an interactive OAuth. ~/.claude/.credentials.json exists there, so the file alone is not proof of a usable session. Measurement trap recorded: security find-generic-password over ssh returns MISSING on all six hosts — an ssh session has no GUI context to unlock the login keychain, so that reading is an artifact, not evidence.
• HERDR-20260906-15 — backticks in a heredoc COMMENT executed hostnames — FIXED. The identity guard's comment contained bare `rdmbair13m5` inside an unquoted heredoc, so both hostnames ran as commands at generation time and their empty output replaced the text, corrupting the comment in every generated launcher. Same class as the rc-through-a-pipe trap.
• HERDR-20260905-06 corrected — the claude slow-start MIGRATES. rdmbair13m5 71–146s → 1s; jdmbair13m5 20s → 1s; rdmsm4x 1s → 45s. Transient host contention, not a property of the 13m5 Macs. The original title was too host-specific.
• Sidebar rebuilt. The 09-05 ssh panes died with the old server; --close removed exactly the five fleet workspaces by label inclusion and left Rich's four live local sessions untouched. Identity guard printed "verified by the far end" for rdmsm4x, rdmpw3265m, rdmpw3275m, jdmbair13m5.
• Stale note corrected: SESSION-STATE said config.toml was not pushed fleet-wide; a 15:34 push on 09-05 did push it, all six match, each host kept one config.toml.bak-fleetpush-*, and there is no theme drift (no explicit theme name in the deployed config).

Budget

usage_status returned advice=unknown — usage.json 18 min stale AND the Anthropic vendor read is unavailable ("Claude credential has no access token"), which on a host with a collector is itself a fault. Local measurement: Anthropic 7d 44.4%; OpenAI 100% until 2026-09-07. No subagents spawned; all work inline.

Records

fleet/herdr/ISSUES.md (19 headings, was 16) · fleet/herdr/SESSION-STATE.md (245 lines) · both pushed to canonical rdmsm4x. No secrets written.