Skip to content

2142-architect-progress-report-overnight-jul30

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

2142-architect-progress-report-overnight-jul30

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 

README.md

topic agents
type report
status research-complete
last-validated 2026-07-30
related-docs 2138, 2139, 2140, 2141, 2137, 2124
original-query Overnight autonomous upgrade run 2026-07-29/30 - Brandon's standing Architect's Progress Report
tier DEEP

2142 - Architect's Progress Report - overnight run 2026-07-29/30

The standing report Brandon's directive requires at the end of every cycle. Every "shipped" below has a merged PR number; every proof was executed, not claimed.

1. Implemented (all merged)

PR What
ZAOOS #2704 DreamNet Phase 3 cross-runtime conformance: 47/47 checks, both directions, LIVE SDK execution; fixed 4 real canonicalize divergences (integer-key ordering, toJSON, NaN/Infinity fail-open) in both copies; permanent DreamNet-generated fixtures + verifiable conformance receipt (doc 2138)
ZAOOS #2705 Heart fleet extraction: dependency-free packages/heart-fleet (fencing tokens, retry ceiling -> quarantine, execution receipts, metrics, guardIrreversible), Memory + Supabase stores, 24 property tests (doc 2139)
ZAOOS #2706 Killed the Telegram truncation CLASS: every unbounded send chunks (brief/recap fallbacks, zoldraft, draft actions, relay pushes, preflight); router + direct sends share one chunker
ZAOOS #2707 Board #51 decision-UX: colon-form contextual buttons, multi-choice toggle keyboard, typed-answer capture (option-shaped text only - chat never eaten)
ZAOOS #2708 Board #49: ZABAL August Finals doc 2137 (concept: August = the July submitters) + deployable page draft + FC/X/YT copy + rollout (Zaal posts)
ZAOOS #2709 Heart canary beat wired into the scheduler, flag-gated (ZOE_HEART_FLEET_CANARY, default OFF - inert until Zaal flips)
ZAOOS #2710 Board #48 drift verify-report (doc 2140): 333 active / 130 overdue measured, timestamp-clobber data-integrity finding, evidence-backed closes proposed
ZAOOS #2711 Board #45 Downloads triage (doc 2141): 44 files / 15.6 GB bucketed, zero deleted
zao-website #27 thezao.com SEO: canonical on 13 pages, metadataBase, box-derived title/description (the icm-grounding drift fixed in page metadata)

2. Proof produced

  • Conformance: scripts/spore-conformance.ts 47 PASS / 0 FAIL with SPORE_SDK_PATH set - the DreamNet runtime itself recomputed ZAO bytes, subject hashes, and digests (fixture values pinned at SDK commit 4072102). Attested by a committed receipt both runtimes can verify.
  • Heart: the 5 safety properties are 24 green tests (mutual exclusion, killed-worker recovery incl retry-ceiling quarantine, stale-fence rejection incl the pre-reclaim zombie window, crash-retry effect dedup, receipt/state reconciliation with recomputing digests).
  • Every PR: bot vitest 130/131 files green (the 1 failure is a pre-existing meetings date test, proven identical on untouched main), bot tsc diffed against the 46-error baseline (0 new), app typecheck 0, SEO changes verified in BUILT html.

3. Remaining failure modes (honest)

  1. The effect ledger is in-memory only. guardIrreversible dedups within a process lifetime; a cross-process replay needs the durable effect_intents table (migration = ask-first, not shipped).
  2. App-side Heart still runs its own lease impl. Migration deferred on purpose: the app vitest is broken (rolldown native binding), so its 14-scenario recovery suite cannot gate a refactor. Fixing the app test runner unblocks this.
  3. Canary is dormant until the env flips + bot restarts (gated operator actions). Until then the shared lease layer runs only under test.
  4. receipt.v1 interop digest is ZAO-proposed - the SDK defines the shape but no digest formula; Brandon has not ratified ours. If DreamNet ships a different formula, the interop layer (one function) changes, fixtures regenerate.
  5. Historical digests over payloads with integer-like keys or non-finite numbers no longer recompute after the canonicalize fix. No stored payload with either shape was found, but it is a theoretical break.
  6. Board timestamps are untrustworthy (bulk-rewritten) - any hygiene automation keyed on them silently no-ops (doc 2140 finding 1).

4. Operational overhead added

Near zero: one cron entry (no-ops while flag off), one shared chunker (replaced a duplicate), receipts collect in-memory unless a sink is injected. No new services, no migrations, no deps.

5. Highest-leverage next milestone

Make the Heart REAL in production: flip the canary, then lease-wrap the first real consumer (repo-improver tick) and ship the effect_intents migration + DurableEffectLedger. That closes the standing "nothing acquires leases" gap (doc 2124) AND gives every irreversible bot action crash-safe at-most-once semantics - the single biggest safety upgrade available to the fleet.

6. Three ranked alternatives

  1. Transactional outbox (send-side counterpart of the intents ledger): every outbound TG/FC/mail action written to an outbox row first, dispatched by a lease-guarded worker. Strongest consistency story; needs a migration + a dispatcher loop.
  2. Spore claims (Phase 5): claim.v1 referencing receipts, DreamNet federation deepens - high strategic value with Brandon, lower operational value this week.
  3. Fleet-wide Cortex: central advisory brain reading receipts/metrics across loops - premature until the Heart is actually in the execution path (advice without enforcement).

7. Iteration roadmap (1-3)

  1. Canary live -> watch a week of receipts/collision logs -> lease-wrap repo-improver.
  2. effect_intents migration + DurableEffectLedger + reconciler on the recovery cron (the doc 2139 protocol, built).
  3. App-side Heart delegates to the package once the app test runner is fixed; then transactional outbox.

8. Next focus (the standing question)

Transactional-outbox over Spore-claims over fleet-wide-Cortex - the outbox is the productionized version of the idempotent side-effect protocol and directly protects real actions; claims federate value we have already proven; Cortex without an enforced execution path is theater.

9. Self-critique + simpler alternatives

  • The canonicalize fix now exists in THREE copies (eyes, bot, package), each pinned to shared fixtures. Simpler: one file, but the bot's isolated build makes a literal shared import the exception not the rule today; the fixtures make drift loud. Revisit when bot/app builds converge.
  • The multi-select grill keyboard may be over-built - typed "1 and 3" already covered multi-choice; the toggle UI adds surface for a flow Zaal may rarely tap. It cost ~60 lines + tests; if unused in a month, delete it (code-restraint).
  • The board report proposes an amnesty rather than executing anything - deliberately underpowered, because "done" without proof is fabrication. The simpler alternative (mass-close by age) was WRONG twice over: ages are clobbered and the tasks are real.
  • One process slip caught mid-run: a secret-scan piped through head reported the pipe's exit, not the scan's (silent-failure-guard rule 1, my own). Re-ran the gate standalone on the merged commit (clean) - the lesson is already codified in the rule; I violated it once under speed.

Sources

  • PRs #2704-#2711 (ZAOOS), #27 (zao-website) - all merged, all listed gates run in-session
  • Docs 2137-2141 (this run's artifacts), doc 2124 (the federation spec)