rdmsm4x-changelog-20260928-0001-coordinator-overnight-resume-all-work
rdmsm4x — coordinator run: "resume all work, fix all open items" (EPIC-20260927-04)
Summary: claude@rdmsm4x (session dev-b4) coordinated the fleet from 22:11 to 00:01 EDT, alongside peer lanes (tyrell-64, dev-28's overnight leads, dev-8c, dev-40, dev-f9, dev-4b). It ran 16 workers (Opus for hard fixes and releases, Sonnet otherwise), each claiming its ticket first so nothing was done twice. Shipped:
- dev_update 2.18
- ReplicantDB Build 54, which fixes the daemon memory leak (two causes)
- the unified usage view (Sunday brief #1), fleet-wide
It also contained the untrusted GLM-Hermes work Rich flagged and handled one credential exposure (tokens printed into a transcript, then scrubbed and rotated).
- Host: rdmsm4x, with changes on all six Macs
- Agent: claude@rdmsm4x (Claude Code, session 4fec0f54)
- Window: 2026-09-27 22:11 → 2026-09-28 00:01 EDT
- Lane map / record:
~/dev/_handoff/coordinator-20260927/LANES.mdplus a W1–W16 report each
Shipped / resolved
- ReplicantDB 1.22.4 (54) on 6/6 (W1 found it, W15
shipped it):
- two daemon leaks fixed: hash chunks retained per batch (341f8b4), and per-file walk objects outside any pool (f1051af);
- daemon on rdmpw3265m now 86–93 MB while crawling, versus 2.7 GB on Build 53;
- notarized via ASC key, DR unchanged, no prompts.
- Tickets: ISSUE-20260927-26 and INC-20260927-01 resolved.
- Usage view, brief #1, on 6/6 (W9, W16, L-libs,
lead): ECSQuotaKit v1.1.0 (e1cc315) is signed, notarized and installed.
- Grok and Hermes collectors added; Hermes measured live (USD 218.32); Grok honestly unmeasured (no cap set).
- The launchd agent's stacked deprioritisation (Nice 15 + nice 15 +
Background + LowPriorityIO) had starved it into an all-unmeasured hub
file. It is fixed on 6/6:
zsh -f, Nice 0, Standard, no LowPriorityIO. - Tickets: FEAT-20260927-20 and ISSUE-20260927-38 resolved.
- dev_update 2.18 on 6/6 (ae0e84c): beta simulator runtimes are retired only while in state Ready.
- Small fixes (W2, W6, W14):
- ticket-store
closed/handling: TASK-84 and its duplicate ISSUE-28 (24b92a3); - fd-limit test, ISSUE-22 (301c9b6, riding Tyrell);
- TASK-87 (FLEET-SPECS);
- fleet_audit agent-kind mismatch, ISSUE-29 (6f92b67);
- notes_changelog publish lock, ISSUE-30 (d60d4b7 / 3005856);
- ISSUE-16, and RFC-04/05/06 (df9f48e).
- ticket-store
- Reviews and verdicts:
- orchestrator10 retired: FEAT-27, tag
archive/orchestrator10-wip-20260927(W4); - mem0 "leak" not a leak: 147 processes = one per live session (W8, ISSUE-31);
- XEntropy "undeployed build" was stale
dist/; build 6 is on 6/6 (W13); - LogTTY PR #2 merged, 9b45c58, flag off, 290/0 (W11);
- xattr consolidation PRD, 9de37dd (W12), DEC-17 filed. W12 also caught a credential-reading regression on the unmerged xattr lane-A branch and warned its owner.
- orchestrator10 retired: FEAT-27, tag
- Ops batch (W10): STATE files refreshed in 6 projects; 4 idle tmux sessions captured and closed; TASK-79 and TASK-20260926-69 resolved.
Rich's directive: distrust the GLM Hermes sessions
- Sessions: 20260927_173836_586a2f, 20260927_191907_51e7a3, and an rdmbair15m5 GLM session.
- Audit (W7): no commits, no destructive host actions, no credential exposure, no installs.
- Quarantined (moved, not deleted) to
~/dev/_quarantine/glm-hermes-20260927/: their rollout SPEC.md and rollout.sh (wrong, and a destructiverm -rfrollback) and HANDBACK.md. - Build dirs: archived.
- Running app: the GLM-built unsigned Hermes.app that was running on rdmbair15m5 was quit cleanly.
- Status-bar patch: removed on rdmsm4x, with the diff saved; it stays on rdmbair15m5 for Rich's v9 trial.
- Trusted hermes (8d6bdc1): redirected to redo the rollout via brew.
Incidents
- SEC-20260927-01: W16's probe printed the Hermes
Nous access token, agent_key and refresh token into its transcript.
- Scrubbed: 14 occurrences, 0 remaining, JSON valid.
- Rotated with
hermes auth refresh nous; the refresh-token hash changed. - Residual: server-side revocation of the old token is unverified, and the transcript may have been copied before the scrub. Rich can re-login for certainty.
- Tyrell Build 48 regression (shipped by this session
earlier): hub chat POSTs returned 500 from 21:31 to 23:40,
because of a missing column in the old hub table.
- tyrell-64 repaired it live; code fix fcf2413 ships in Build 50.
- Spokes verified unaffected.
- Recorded on TASK-83 and in memory
(
tyrell-release-traps).
- Locked console: notarization looked "deleted"
because the data-protection keychain is unavailable while the screen is
locked. The ASC API-key path (v1.1.3,
ECS_NOTARY_AUTH=asc-key) works while locked, and codesign works while locked.
Handed off / in flight
- tyrell-64:
- Tyrell Build 50: chat fix, chat UI, usage wire-in, usage-agent template, and the ReplicantDB 54 helper.
- ReplicantDB Build 55: ISSUE-41 getStats, eedbd68.
- ISSUE-42: bus-index I/O.
- ISSUE-50: Apple Event denied at launch.
- dev-28 overnight leads: L-apps (TASK-103 done), L-libs (done), L-rtty (LogTTY 23), L-ops; routed TASK-104.
- Scheduled: 03:41 check of the 03:07 dev_update sweep (ISSUE-20260927-20 watch item).
Needs Rich
- DEC-20260927-16: Hermes default model, GLM → Sonnet (per-token spend).
- DEC-20260927-17: xattr owns discovery; XEntropy stays the password manager (recommended).
- DEC-20260927-08: tyrelld via SMAppService.
- Optional:
hermes auth logout nousthenhermes auth add nous(SEC-01). - Unlocking the console lets the ecs-notary profile path work again (the ASC-key path already works).
Undo
- ReplicantDB: each host keeps a Build 53 rollback.
- ecs-quota:
~/.tyrell/bin/ecs-quota.prev-8ae63a1and.prev-0221. - Usage plist and wrapper:
*.bak-20260927-*next to each file. - dev_update: rollback v2.17 is kept on every host.
- GLM items:
~/dev/_quarantine/glm-hermes-20260927/and~/.hermes/archive/unverified-glm-20260927/.