Overnight Fleet Report

Sunday evening → Monday morning · Jun 21 18:00 → Jun 22 ~04:30 ET
5 commits shipped 3 self-healing infra fixes 2 audits run theme: RAM ceiling + cron-wake waste fleet: healthy

Infrastructure & Self-Healing — the night's main theme

live MCP RAM watchdog Forge · e4e83b39

New launchd service com.cortextos.mcp-rss-watchdog (5-min interval) hard-kills any chrome-devtools / playwright MCP node tree exceeding 500MB RSS, so the wrapper restarts it with a clean heap. Wired into mem-headroom-guard.sh to fire under memory pressure.

Why it matters: the always-on MCP servers were the structural RAM leak behind this week's HEADROOM WARNs. This is the real enforcement lever (the old --max-old-space-size flag was only a GC hint). RAM ceiling is now actively defended.

fixed GBP Chrome respawn loop ad751c1b

The orphan-reaper was killing the clinic's GBP Chrome (intentionally persistent, watchdog-managed) every 10–20 min, and the watchdog instantly relaunched it — a tight loop that showed up as "9 Chrome instances" during Saturday's HEADROOM WARN. Reaper now excludes it.

Why it matters: was silently burning CPU/RAM all day and masking the true memory picture. GBP Chrome is now stable.

fixed Overdrive false-wake suppression bf979848

A stopped agent always carries the daemon [STALE] marker, which false-fired the overdrive sweep every 30 min — waking a full Opus Sage session for nothing. Now only stale-while-running counts as a real wedge.

Why it matters: direct Opus-token savings, every half hour, around the clock.

Cron Efficiency — killing wasteful Opus wake-ups

done codex-usage-publish → shell cron Sage

Converted from an agent-prompt cron to a silent type:shell cron. Eliminates 48 Opus wake-ups/day (every 30 min) that previously spun up a model just to publish a usage number.

done postpartum-form-poller → shell + exit-0 wrapper Hippocrates · cba4c6b3

Converted set -euo pipefail to a soft error handler so transient GWS/JotForm hiccups never leave the agent stuck in an error state, and re-registered as a shell cron (no Claude PTY). Matches the proven pto-form-poller pattern.

Audits & Knowledge Hardening

done Efficiency Audit (05:00 UTC) Atlas/Analyst

Found and routed the cron-wake waste above (codex-usage-publish, postpartum poller) and put the MCP RAM watchdog at the top of the morning cascade.

done Internal-Workings Audit + 3 skill fixes Atlas · 1b591981

The Mon/Thu audit (the email you screenshotted) found 3 knowledge gaps living only in agent memory; all 3 are now permanent skills:

• ~/.secrets/cache/ credential-decode pattern → env-management skill
• cron "Before Flipping" 5-step safety checklist → cron-management skill
• Vertex AI x-goog-user-project header gotcha → env-management skill

Status of your question: yes — all of it is implemented (as direct commits, not GitHub PRs). The op-freeze stale-directive item is also cleared: Librarian updated; Argus updates the instant it wakes.

done Theta wave + Forge reliability cycle Atlas + Forge

Nightly theta wave scored 8/10 HOLD (6th consecutive — system stable). Forge registered its first build-reliability theta cycle: baseline 29.2% verified-live rate across 144 cortextOS PRs (Apr–Jun), closing the specialist-diversity gap that had pinned theta for six cycles.

RAM Relief

your call 3 agents parked overnight

Mercury, Argus, and Hippocrates were stopped overnight to hold the memory ceiling while the watchdog landed. Open question for you: wake all three back up now? (Y/N — surfaced in the morning brief, still pending.)

Content & Clinical (externally gated, progressed overnight)

live Scribe — VE guide embeds Scribe

embed-66 + 6 follow-ons done and live-verified: 69 VE guides now embedded. 3 embed-on-publish deps queued (ocd / depression / webster). Marietta TTC email held on your venue + early-bird facts; anatomy series held on Dr. Saylor's neck-video sign-off.

done Hippocrates — clinical pipeline Hippocrates

TWP pipeline built, tested, Apollo-approved, cron goes live Jun 27. VE conditions clinical audit: 86 pages, 0 critical flags, ingested to the knowledge base.

Still In Flight This Morning

deploying VE address → "4203 S Rochester Rd"

You flagged the address should read S Rochester everywhere (matches Google Maps). The fix was committed to the wrong branch (perfect-storm-staging); production builds from main. Forge cherry-picked it to main (a9dbaa5) and the Cloudflare Pages build is running now — I'll confirm the moment the live site flips.

staged VE DNS cutover — awaiting your "approve DNS"

Both gates green (Resend verified since Mar-19; canonical address locked). One word from you and Forge cuts vaneverychiropractic.com over to Cloudflare Pages. approval_1782047253_sn8l5