Mirrored verbatim from utopia-models/knowledge wi/decisions/ADR-182-win-rebuild-architecture.md by the allowlisted publish pipeline. Edit it there, not here.

artifact_type: adr
adr_id: ADR-182
adr_status: accepted
status: active
decision_date: 2026-07-23
supersedes:
superseded_by: ADR-2026-07-31
issue: PLAN-182
references:
  - world-infrastructure/specs/3511-one-brain-shape/decisions/001-guards-are-advisory-not-required.md

🔴 RESTORED 2026-09-24 (PLAN-313 P4). Deleted from disk in a KB reorg (cdfea10e8) while still live-cited by adrs: [182] frontmatter across multiple components. Confirmed still relevant, not historical prose: C3 (transient systemd-run dispatch, no resident processes) is verbatim what dispatch.sh/task-runner.sh still do today and is the reasoning that made deleting the ttyd/ claude-agent@ units correct this same session; C2/C5/C8 describe the files+git memory, symlinked doctrine, and portfolio-first tree this repo still runs on. Content below is verbatim from the original 2026-07-23 record — not edited. One known drift: C8 names world-infrastructure/ as the portfolio root; that directory has since been shortened to wi/ (a rename, not a reversal of C8's decision).

DR 2026-07-23 — The WIN Rebuild Architecture (ACCEPTED)

Status: ACCEPTED (Tyler, 2026-07-23 — plan-mode approval, the explicit decision event). Immutable. A change of mind = a NEW superseding record naming the new information — never an in-flight edit. C7 Superseded-by: world-infrastructure/specs/3511-one-brain-shape/decisions/001-guards-are-advisory-not-required.md (2026-07-31, #3514) — C7's (guard.yml, **required**) is not achievable on this plan tier: both /branches/main/protection and /rulesets return 403. Guards are ADVISORY. C1-C6 and C8 stand unchanged. Full working plan (options, inventory, migration): a private working document (playful-purring-sunrise.md), no longer in the tree since the plan documents were deleted with the plan registry (ADR-10109). This record keeps the decisions; the rejected alternatives are not reproduced here.

Context

Masters completed ~20% of delegated work. Root causes (10-workflow inventory, 2026-07-23): no grounding (masters can't orient in a portfolio), thin tasks (titles, not briefs), and a board that lies (manual status). Deeper: the fleet's custom machinery (~1,000 scripts, 180+ hooks, 21 CI gates, 3 memory systems, 5 comms generations) rots silently, and its detectors hardcode the same stale assumptions as the docs — the fleet cannot see its own drift. Tyler's #1 time-sink was repairing the machinery, which falsifies the 1000x thesis.

Goal: Tyler + fleet execute at 1000x speed while Tyler is present (phone, 8-10h/day). Non-goals (the scope fence): walk-away autonomy · new revenue features inside the rebuild · multi-human scaling. Cost is a soft consideration, not a driver.

The Decisions (per component; alternatives + rejection reasons were recorded in the private plan file)

• Principle: no stored copies of state · no resident processes · docs are the config (read at runtime). Drift classes are removed, not monitored. (1-way door in spirit — reversing this recreates the disease.)
• C1 Work system: native GitHub Feature + ordered sub-issues; the full brief in committed spec.md + plan.md beside the work; plan.md checkboxes ARE the work state; merged PR (Closes #N) closes the issue and moves the board — status is derived, never typed.
• C2 Memory: files + git only. All three memory services (Cognee, mem0+Qdrant, SQLite second-brain) are deleted. Auto-memory points at a git-synced dir; one append-only decisions file. (Corpus measured 360KB — three orders below where vector/graph earns its cost; every credible fleet uses files+git+issue-tracker.)
• C3 Runtime: pure on-demand dispatch. No resident agent processes. Dispatch = a transient systemd-run unit running task-runner.sh (▶ comment → claude -p under the master's own Max OAuth → trap-EXIT ✔/✘ comment). ralph + watchdogs + heartbeats deleted. Routines revisited when out of research preview.
• C4 Comms: GitHub-native — issue comments + @mentions on the work item; receive = a mentions query injected at run start (durable read-state, at-least-once). NATS stack deleted. (The bus as wired was discarding mail inside a >/dev/null health probe.)
• C5 Doctrine: prose doctrine lives IN the knowledge repo; each master's ~/.claude/rules/ is a symlink into its clone (doc/config drift structurally impossible); rules slim 380K → ~50K behavioral core; the plugin shrinks to hooks-only (SHA-pinned code).
• C6 Surfaces ("one reading, no publishers"): every Tyler surface (OCC · board · dmvresearch.shop · raw.utopiamodels.ai · RAW phone · Drive) reads its source live at view time or IS the source. Upstash + all publisher crons deleted; liveness derived from observable work, never self-reported; two failure-only smoke Actions whose only output is a board issue; RAW phone = raw.utopiamodels.ai as PWA + Tailscale/Taildrop.
• C7 Gates: minimal — fail-closed only for stable invariants: gitleaks secret-scan + binary-guard (guard.yml, required) + CI-green where a real build exists. ONE PreToolUse survivor (hook-syntax-guard). One daily advisory lint upserting a single drift issue. ~19 of 24 workflows + all content gates deleted.
• C8 Knowledge repo: portfolio-first tree (world-infrastructure/ raw/ dmv/ govcon/), each rooted by a portfolio CLAUDE.md map; specs beside the work; aggressive prune (git history is the archive).

Consequences

• The fleet is rebuilt by strangler-fig (M0-M8 in the plan file): guards first, dispatch proven on one master before ralph dies anywhere, each OCC section dual-run before its publisher dies, the one 1-way door (Hetzner box deletion) gated on a snapshot + 7 quiet days, and a DATED old-system-deletion milestone.
• Success metrics: dispatched-work completion >80% over the first 10 dispatches · Tyler stops filing fleet-repair issues · zero silent-death incidents · prefix 380K→~50K · scripts ~1,000→<100 fleet-wide.
• The planning method itself (phases, options-before-decision, decisions-lock-via-event, ceremony-by- reversibility, never-big-bang) is now the standing process for org-wide changes.