rdmsm4x-changelog-20260826-1857-unblock-fleet-mcp-backups

rdmsm4x changelog — 2026-08-26 18:26 → 18:57 EDT

Host: rdmsm4x · Session: claude@rdmsm4x · Elapsed: 31 minutes Ask: review open efforts and unblock them (rdWORKBENCH cited as stuck), review fleet state, confirm tyrell.app and replicantDB.app work as MCP on all agents/hosts. Mid-session Rich added: check whether global rules need a reset / whether to run /doctor; the dev.dataroo.net activity log is alarming; intervene on all blockers and unlock everything; keep delegating per-project work to codex because Claude usage is capped until Friday.

Headline

Almost nothing was actually stuck on the thing it appeared to be stuck on. Four separate "failures" all traced to the same root: the 2026-08 ~/dev consolidation moved files and nothing swept the references to them.

The most serious finding was not rdWORKBENCH. It was that the nightly GitHub backup had been covering 1 repo out of 87, and then not running at all — through exactly the window an unexplained process was destroying repos on jdmbair13m5.

Fixed and verified

Thing	Was	Now
rdWORKBENCH tests	exit 65, "Unable to resolve module dependency"	4/4 pass, `933c8fd`
dev.dataroo.net feed	200 cron heartbeat rows drowning real events	changed-only logging, 12 shown, 300s→1800s
`devbackup` / `claudestate`	exit 127, nightly, silent	repointed, status 0
`app-icon-gallery`	exit 78 (`WorkingDirectory` stale)	HTTP 200, status 0
backup coverage	1 repo of 87	v1.4: 77 to back up, 9 excluded by policy
Tyrell MCP	arm64-only on Intel host; absent on another	universal binary, handshakes on 5/5
replicantDB MCP	never loadable by any client	conformant `281c6e3`, deployed 5/5
Claude plugin registry	9 entries at `/Users/rich` (nonexistent)	repointed, 0 missing
LaunchAgents	3 failing	33 audited, 0 missing paths, 0 failing

Root causes worth remembering

rdWORKBENCH: build was green, test was red, on one error naming a module that plainly existed. .build/DerivedData still held RDWorkbench.swiftmodule, an entry created under the pre-rename product name. APFS is case-insensitive so the compiler wrote through it and the casing never changed; Swift's explicit-module scanner matches case-sensitively. Renaming a product does not rename its build artifacts. Applies to Tyrell (ex-LLMRoo) and replicantDB (ex-rooDB) too — filed to both.
replicantDB MCP: codex fixed the missing initialize and its own verification passed — but each probe fed a single message. Two pipelined messages produced zero output: the read loop used availableData (bytes available, not one line), so real clients had both messages concatenated and silently dropped. Caught by re-verifying with a realistic sequence.
launchd: correcting ProgramArguments was not enough for the gallery — WorkingDirectory is a path too, and launchd returns bare EX_CONFIG (78) without naming the key.
backup: the depth-1 glob ($DEV/*) was written for the pre-consolidation flat layout.

Fleet halt

Released 18:36 EDT on Rich's instruction — broadcast 20260826-183333-17877A8C, ACTIVE-DIRECTIVES.md updated. Three carve-outs deliberately kept gated: rdRECEIPT push (real financial corpus), the two customer-repo deletions (irreversible, needs an interactive scope grant), and delete nothing on jdmbair13m5.

Handoff filing

11 off-thread contexts written as ~/dev/todo/items/*.md and relayed by the hourly scanner into the owning project's ISSUES.md, each tagged [todo:<id>]: apps/Tyrell ×2, apps/replicantDB, sites/dev.dataroo.net, fleet/maintenance ×4, ai/claude-ops, fleet/portfolio-consolidation, concepts/icloud-cleanup (whose ISSUES.md was created so it could receive relays). All 11 confirmed present in their targets. The ~/dev/ISSUES.md claim that this scanner "not yet built" was stale and is corrected — it is live, 33 relays recorded.

Still needs Rich

First full backup run publishes ~77 private GitHub repos. Hold it until the concepts/icloud-cleanup Dropbox credential is rotated or that repo is excluded — push --all sends full history and the script's secret filter only guards the snapshot index, not history.
Two customer GitHub repo deletions (interactive gh auth refresh -s delete_repo).
Credential rotations still open: UDM root, Cloudflare Global API Key, X-Goog-Api-Key, Dropbox.
rdmpw3265m is SSH-unreachable and got none of today's repairs. It is Intel.

Notes on the two questions asked mid-session

Global rules do not need a reset. CLAUDE.md v4.0 is 12.6 KB; permissions.defaultMode reads bypassPermissions on rdmsm4x and all four reachable peers — no drift. Do not run /doctor to chase harness problems: its check 8 hardcodes defaultMode="auto" as "(recommended)", colliding with Rich's standing instruction; three sessions have already declined to act on it. The real harness rot was the /Users/rich plugin paths, found by direct audit instead.

Spend: Anthropic weekly bucket at 91.9% (usage_status advice = hold), OpenAI at 15%. Per-project implementation went to codex exec (1 job, replicantDB MCP), no Claude subagents spawned.

Apple Notes half not written — Notes is unresponsive fleet-wide per the standing ~/dev/ISSUES.md item; this file copy is the durable record.