docs(state): session snapshots and status briefs

2026-08-29 05:22:43 -05:00
parent 79d93b6c49
commit bd0cea1757
8 changed files with 427 additions and 0 deletions
@@ -0,0 +1,75 @@
+# STATE SNAPSHOT #6 — 2026-08-29 ~05:30 CST (pre-wrap, crush update pending)
+
+READ THIS FIRST after relaunch. Then: `~/.coordinate/log.md` tail,
+`~/.coordinate/logs/queue.status` tail, `screen -ls`.
+
+## Who/where
+PMO agent reachableceo-PMO, host ultix-streaming, screen `RCEO-PMO`
+(relaunch PMO under EXACTLY this name — heartbeat stuffs into it).
+Charles = reachableceo (never "founder"; never track his time/availability).
+PMO = traffic cop ONLY: meaty TASK prompts -> headless workers, verify
+REPORTs, archive to ~/.coordinate/archive/, mirror Redmine. One item per
+turn, HUD table at end, medium verbosity.
+
+## Mission status (spec of record: harness/docs/SPEC-20260829-charles-brief.md)
+- MVP bar MET (8/28 19:29) and **V1 FEATURE-COMPLETE**: harness loop
+  (self-hosting PROVEN live, issue 487), events (:4100 webhooks), serve
+  (:8090 OWUI front door), keyproxy v0 (mpk_ refs, file/env backends,
+  BW/Vault stubs), mopac-bitwarden-go v0 (fake-server tested).
+- Timeline: **beta 8/31 0600 CST, production + first 3-COS briefing
+  9/1 0630 CST**.
+- Redmine backlog: tickets 490-499 (490 quota, 491 resource, 492 phase
+  routing, 493 queue selection, 494 multi-account deploy, 495 briefing,
+  496 review chain, 497 keyproxy BW live, 498 MCP fleet, 499 PDF).
+
+## QUOTA CRISIS — careful-weekend mode (Charles at 93% weekly)
+- Worker turns default **glm-4.7-flash** (dispatch-turn.sh, TURN_MODEL
+  override for escalation to glm-5.2). AFTER CRUSH UPDATE: check catalog
+  for glm-5.3-flash; if present switch the default in dispatch-turn.sh.
+- queue-after.sh has QUOTA WALL GUARD: any q-*.log with "Usage limit
+  reached" halts the gated chain + pings RCEO-PMO.
+- Do NOT freeze the weekend — careful sustained building (Charles's order).
+- 1900 tonight Charles connects: 1. Bitwarden creds (ticket 497), 2.
+  Signal/Discord, 3. Google, 4. **OpenRouter key** (separate quota pool).
+
+## Chain running (survives crush update; gates use fresh crush runs)
+- work-q9 RUNNING (quota/back-pressure + resource gate, tickets 490+491,
+  on glm-5.2 — sunk cost, let finish) -> work-q11 deploy (494, flash) ->
+  work-q12 briefing (495, flash) -> work-q10 dispatcher (492+493,
+  escalate to 5.2 if flash output garbage).
+- TASK files: ~/.coordinate/inbox-work/TASK-20260829-{0500-quota,
+  0700-dispatcher, 0900-deploy, 1100-briefing}.md
+- pmo-heartbeat screen: 1h interval, wakes PMO with chain-check prompt.
+- If q9 was mid-flight during the binary update: running proc unaffected
+  (old inode); gates launch NEW crush runs post-update.
+
+## Ops facts (hard-won)
+- LiteLLM http://192.168.3.78:4001, 13 z.ai models; /config/update
+  hot-reload LIES — docker restart litellm (~60s).
+- Harness vkey: ~/.coordinate/secrets/mopac-harness-vkey.env (0600,
+  alias mopac-harness-demo-v1). Redmine: source ~/.creds/redmine.env with
+  `set -a`, var REDMINE_API_KEY.
+- dispatch rc: 3=deferred(cap), 99=gate unavailable, 143=self-pkill
+  (happened once — TASKs now say never broad pkill), z.ai flake "Message
+  content shorter than read bytes" = redispatch.
+- dispatch-turn doorbell targets ${USER}-PMO (dead, cosmetic); real wake
+  = pmo-heartbeat.
+- PMO cannot ssh/sudo: ticket 494 deploy runs via Charles window (~10 min).
+- keyproxy repo: default branch fixed to main (was master).
+- Absolute paths for dispatch TASK args (relative path = empty prompt bug).
+- Gitea API pattern: tea token from ~/.config/tea/config.yml, never echo.
+
+## First moves for fresh PMO
+1. Confirm chain state (queue.status, screens). Verify q9 REPORT when it
+   lands -> archive -> Redmine 490+491 Done.
+2. Post-update: check `crush.json`/catalog for glm-5.3-flash; switch
+   dispatch default if present.
+3. Continue verify->archive->mirror->redispatch loop per heartbeat.
+4. At 1900: walk Charles's connection list (BW > Signal/Discord > Google >
+   OpenRouter); dispatch ticket 497 turn after BW creds land.
+
+## AMENDMENT ~06:30
+Charles upgraded to MAX PLAN — all quota buckets ZERO. Crisis over;
+quality-first restored (dispatch default glm-5.2; flash via TURN_MODEL for
+grind). q9 must verify its polling against MAX-plan bucket shapes. PDF
+pipeline (499) dispatched parallel (work-q13, new repo, no collision).