Opus-orchestrated, Codex-executed on its own quota. Four tracks. Everything here is a draft or a staged decision — nothing sent, no prod deploys, no memory writes, no merge, no Sorli deploy. The 10:30 Firestore deploy is untouched and still armed.
4tracks
6decisions
~37%weekly · gate 45
0writes / sends
Morning checklist · lightest first
Decisions needed · 6
💧 H2OS — Sorli scanner 2
The fix is built + adversarially verified but NEEDS-WORK. Core blocker is a design fork: is the real drop a clean USB re-enumeration (safely recoverable) or enumerated-but-silent (not, without a vendor liveness command)? Descope round 3 to disconnect-only?
docs/sorli-adversarial-verify.md
Deciding evidence is genie-122's dmesg-usb.log (the live observer). Want me to SSH in and read it to settle A vs B before any round 3?
48 open items adjudicated (10 topic proposals, ~19 re-routes, ~19 fold failures). Approve the ✅ set as-is and I execute via cosmo-mem? Two items are your call (sh-log tracker topic-vs-task; the B4 re-homes).
docs/dream-adjudication.md
📈 Finance — v3 1
"finance-v3" was an undefined placeholder. Three real directions; 2 and 3 both depend on 1 (fix council non-determinism + grade the 8 forward calls). Pick one and I spec it.
docs/finance-v3-direction.md
🤖 Cosmo / dev — merge 2
feature-solve-first-pipeline is 56 commits ahead of main (morning-brief + solve-first + enable-banking/Sabadell + more). Merge to main — go/no-go?
git log main..feature-solve-first-pipeline
Want me to scan the full 56-commit diff for anything risky before you merge, or is it clean by your read?
56 commits · no npm-test; per-service suites
The 4 tracks
💧 H2OSCodex fix + verify
Sorli self-heal — honest NEEDS-WORK
NEEDS-WORKfix built · adversarially verified · design fork surfaced
Findings
Codex built a full round-2 fix (device.py +333/−118): opt-in gate, tri-state presence, disconnect latching, udev rule, lifecycle locks, rotating+redacted logs, 11 tests. I ran the tests myself — 11/11 pass.
But a fresh adversarial Codex verify found the passing tests are largely tautological (none drive the real scanner loop through a silent recovery) and the fix still has real breaks: the one-shot latch can still wedge forever after one failed reset; the gate is captured only at startup; lifecycle-lock gaps remain.
The core blocker is a design limit, not a bug: with no verified SM-2D liveness command (Codex checked Newland's docs — none exists), a 5-min silence timeout can't tell a wedged scanner from an idle customer, so it would reset scanners mid-transaction.
The instrumentation brief shows the real failure is one of two modes: (A) USB re-enumeration ~135×/7d — safely recoverable; or (B) enumerated-but-silent — not, by timeout. The current fix tries to do both; the (B) path is what makes it unsafe.
My recommendation: don't polish the risky timeout path. Descope round 3 to the safe (A) disconnect→reappear recovery only, and read genie-122's dmesg to confirm (A) is the actual mode. If (B) is real, it needs a hardware/vendor fix, not software.
~19 routing questions: dev-lessons → AutoMem/claude-code-surface; ops-reliability facts → cosmo-starbase-observability; a few real facts re-homed; 2 stale/contradictory dropped (e.g. a "raise DreamPass gate" note for a gate that was removed 13 Jul).
~19 "failed 3 runs" fold problems: almost all granular HK/Momence login-bug debugging. Keep 4 load-bearing facts → humankind-digital, drop the ~14 blow-by-blow lines as session narrative.
The 7:30 brief will lead with "⚠ 23 fold problems" — that flag is correct; this sheet is the answer to it.
Approve the ✅ set and I execute via cosmo-mem (topics + tasks + queue clear). Nothing writes until you say go.
"finance-v3" existed only as a one-line placeholder — no captured vision. Rather than invent your strategy overnight, here's where v2 actually stands + three honest directions.
Option 1 (recommend): trustworthy-first — fix council non-determinism, let the 8 forward calls mature, grade them. Lowest budget, unblocks everything, matches the anti-p-hacking rails.
Option 2: finish the radar (Phases 2-5). Option 3: execution + book integration (real-money sizing). Both structurally depend on Option 1 first.
Pick 1/2/3 (or a blend) and I'll write the actual spec.
The branch is 56 commits ahead of main — far more than the recent morning-brief/solve-first work; it's become the de-facto trunk.
Recent work verified green by the weekend (morning-brief 30/30; solve-first test isolation fixed in 0ee4edf). No npm-test script — suites run per-service.
56 commits is a big state change to merge unattended, so I did NOT auto-merge. Staged as your call.
Say go and I merge — or I scan the full diff first for anything risky.