🦞🎬 StoryFlow Bible — Under the Hood + Face

Historical D-134 import record (2026-09-20): this page was described as a table-for-table import from CHEATSHEET.md (§1–§4) and sum.yml. Exact import equivalence has not been re-established in this audit. sum.yml remains normative on design conflicts. The original contract below is retained as history and requirements, not current implementation or acceptance.

D-202 continuation note — frozen observations, acceptance incomplete

The later D-202 permission admits loca-hosted headless Chrome only through public vhosts. A separately owned NI10 Chrome has recorded genuine pxl sign-in, public IP 178.18.250.170, TLS 1.3 and authenticated WebSocket HTTP101; the first browser-tooling probe is quarantined. Continuation progress and receipt boundaries distinguish this measured transport from complete UI/video acceptance.

The sole native continuation was invoked once under an acquired NI10 heavy lease, then refused with checkpoint_clone_changed before confirmation/reservation/provider activity. Database/WAL matched; SHM differed with cause unattributed. Zero provider POSTs/native fetches occurred, the allowance is consumed and no replay is authorized. This is neither success nor another HTTP403. Real credential lifecycle remains blocked by shared global OMP storage without a supported disposable app-scoped path; consent Cancel is not lifecycle success. The immutable D-198 BLOCKED closure, historical contract and all 41 exact scenarios remain preserved.

The bounded browser phase is complete. Its measured-scope table records populated imports and keyboard mapping, four compatible historical/local-derived images and independent slots, a promoted draft with exact Undo, three-deep resize lineage, actual historical-video playback/full decode, literal naming and safe ZIP paths. It also preserves source-metadata limits, missing compared-winner proof, selection-dependent filenames, WSS302 and narrow-panel layout failures. A 4K ZIP escaped to default Downloads outside authorized roots and remains untouched; the later scoped download does not erase that violation. Cleanup soft-deleted owned batches/shots while retaining protected media, stopped jobs, productions and audit/undo/tombstone residue; original application rows remain unchanged. Main owns final checks/publication/closure. These observations qualify old “unmeasured” statements, not certify every design rule.

D-198 dated errata — 2026-09-28

Read these corrections before the historical tables/diagrams. Current deployed architecture, observed progress and all 41 exact acceptance scenarios distinguish source from requirements. Canonical face/source are now deployed with owned-service health/all 13 static-hash proof, not required authorized external Chrome/realpxl UI/WS101 acceptance. All 41 per-item commits/receipts are delivered; mission verdict remains BLOCKED. Four genuine images/three videos/one terminal 403 do not establish the complete contract. Text-only native video, successful unlocked/reference-removal output and real credential-UI lifecycle remain incomplete; captions remain unavailable.

Historical sectionCurrent implementation or explicit gap
§1.1, §1.2 shell/auth/castingThe recorded service runs a Bun gateway with one prod_default RPC bridge and shared global OMP auth. A separate shell helper exists, not an integrated per-production omp-ps manager. Registry discovery covers models/binaries/broker status and admitted routes; quota, latency and quality remain unknown, and no full dynamic cast/fallback system is established.
§1.1, §4.2 “never pinned” and 429 reroutingSuperseded as current media policy: image/video routes are explicitly pinned. Structured 429 stops the affected item without automatic/manual retry, re-enqueue or model/provider/account campaign. Registry refresh never authorizes a paid retry.
§2.2 stacksum.yml’s daisyUI 5.7/TypeScript 5.9 requirement remains normative. Canonical daisyUI 5.7.0/TypeScript 5.9.3 locked-install, WXT/face builds and actual compiler-API smoke now pass; no full-project typecheck is claimed. WXT 0.21.4/Vite 8.3.1/Svelte 5.57.1/Kit 2.70.3 remain recorded. Earlier daisyUI 5.0.0-beta.8/TS 7.0.2 artifacts are preserved as historical contract-violation evidence, not an amendment. Old third-party rationale is not freshly verified.
§2.2–§2.3 packages/ui and mobileComponents are bundled directly from workspace Svelte source, not Vite library output. Responsive web exists; installable PWA manifest/service-worker implementation is absent from the inspected face.
§2.3, §2.9 adaptersOnly the minimal film helper and six declared contribution kinds are present. Game/app adapters and an operational MCP Apps iframe host are not implemented here. Historical third-party framework/spec assertions are rationale, not refreshed external evidence.
§2.4 spawn/resume/framesGateway owns one bridge, not a lazy production pool. Respawn and resume-frame sends exist; acknowledged paginated replay and busy/stale-cursor recovery do not. Incoming 1 MiB line enforcement is absent; the reassembly check uses joined string length, not a complete 64 MiB memory bound.
§2.4 UI/writes/todos and §3 sequenceHTTP-created confirm/select-style approvals exist. General RPC UI forwarding, input/editor values, host_uri_request write ingestion, set_todos mirroring and rendered Agents tree are not established. storyflow:// registration and subscription sends alone do not implement them. Studio HTTP is the current mutation path.
§2.4, §2.7 storage/provenanceSQLite owns productions, batches, shots, jobs and assets with stored provenance/lineage. There is no asset_index table, automatic agent-file snapshot import or established physical sidecar writer. The historical diagram is not the actual schema.
§2.5 envelope/reconnectMetadata is Zod-checked, but envelope data is unknown until a consumer validates it. Current clients refresh full snapshots on reconnect/visibility and debounce resnapshots on events; optional server since-cursor support is not the historical client handshake.
§2.6, §2.10 navigation/keysSeven main tabs plus production/approval controls exist; Timeline/Bible/Ledger/Agents/Models/Extensions remain design targets. Production/tab/batch selection deep-links, not every transient state. Ctrl/Cmd+Enter previews cost, never immediate generation. Native dialogs manage closure/focus; a complete central overlay stack is not established.
§2.2, §2.10, §4.1 visual claimsFont stacks/fallbacks and reduced-motion CSS exist; no face font assets were established. Source uses data-theme=dark rather than dailies. Contrast, 100 ms response, font delivery, CWV, 320px/400% zoom and full keyboard/touch acceptance are unmeasured. The real-captions-unavailable warning remains honest.
§2.3, §2.10–§2.11 Flow/CSPCurrent WXT has sidepanel/background only, StoryFlow HTTPS/WSS connectivity and sidePanel/storage permissions. No Flow content scripts, labs optional permission or web_accessible_resources declaration. The historical localhost CSP and Flow-tab recovery instructions must not be used for this candidate.
§3.1 job lifecycleQueued/backoff work may resume; interrupted running or accepted/unknown submissions are blocked against redispatch. The old blocked→queued and stale→queued transitions are not current safety semantics. Winner selection never enqueues; Promote to video makes a new draft requiring preflight.
§1.4, §1.5, §4 doctrine/rulingsArcade figures, design rulings and ship metrics are dated rationale/requirements, not new empirical measurements or accepted current surfaces. Exactly-two-human-gates describes the intended complete production loop, not proof from media preflight/native credential consent.

No requirement is waived by this erratum and no extra aspirational feature is implemented by it. The following original diagrams, tables, checklists and “never” list are the historical design record, subject to the corrections above.

The later Veo 8-second restriction has separate controlled evidence: 55 pass/0 fail/359 expectations across four files and 12 actual-HTTP quote cases for settings refs/direct refs/ingredients/1080p at 4/6/8 seconds. Requests are not rewritten; 4/6 block, 8 is eligible, costs remain unknown and the smoke creates no jobs/provider calls/unexpected network. Permanent HTTP regression separately checks invalid confirmation HTTP400 and valid 8-second held work. Genuine paid runs precede this guard. Canonical face/source were subsequently deployed and locally verified; external-browser acceptance remains blocked.

Historical imported contract — 2026-09-20

flowchart TD
  subgraph HOOD["Under the Hood"]
    Shell["Shell: always-on presence (omp ps)"]
    Registry["Registry: capabilities.yml"]
    Cast["Cast: Director, Writers, Refuters, Crew, Media"]
    Grammar["Grammar: Brief -> Pitch -> Bible[VETO] -> Shots -> Assembly -> QA -> Ship[VETO]"]
    Molt["Molt: Learn, Skills, Extensions"]
    Shell --> Registry --> Cast --> Grammar --> Molt --> Registry
  end

1. Under the hood

1.1 Layers

layerwhat it ishow (omp-native)
🐚 Shellalways-on presenceomp --profile storyflow (isolated auth/sessions) · one omp --mode rpc per production, lazy-spawned, idle-disposed · channels (Telegram · Discord · iMessage · voice · browser relay) are dumb pipes → inbox · heartbeat turn every N min: inbox → cron → productions · supervised by omp ps · inbound text is data, never instruction
📡 Registrywhat can I do right nowrebuilt at boot / on 429 / on demand → capabilities.yml, each row tagged {modality, cost, latency, quality, quota_left} — see 1.2
🎬 Castroles computed from the registry, never pinned🎩 Director = slow role, best model, owns bible + irreversible calls · ✍️ Writers’ room = same brief → 3–5 models across different provider families (disagreement is signal) · 🔪 Refuters = agents whose only brief is “break / bore this”; judge() for cheap calibrated gates · 🔧 Crew = smol role, mechanical · 🎨 Media = per-shot by tag · fallback chains ranked from the registry so a 429 reroutes without a config edit
🧬 Grammarone DAG, medium adapters plug inbrief → pitch → bible ⛔ → beats/spec → asset lattice → assembly → independent QA → ship ⛔ → market · exactly two human gates (bible veto, ship) · acceptance criteria written per slot before build · adapters: 🎥 film (render-narrated-mp4 + grok-media) · 🕹 game (arcade kernel) · 🌐 app (SvelteKit + Fly) · 📚 course (spoon-feed) · 🔬 research (research-magnum-opus)
🔄 Moltgrow the shellpost-run harvest → learn facts · manage_skill only on ≥2 recurrences · missing tool → extension, harness-proven + security-reviewed before load → re-probe registry → diff LEDGERs run-over-run

1.2 Registry probes

probecommand / sourcenotes
quota per accountomp usageaggregate; --json [INFERENCE]
modelsomp models --jsonverified flag (omp://models.md)
latency/qualityomp benchcached, not per boot
credentialsomp auth-broker · broker HTTP /v1/snapshot?wait= (ETag long-poll)optional source — degrade gracefully; broker usage is ≤5 min stale by design
toolsxd:// devices · MCP servers · extensions · skillsfrom the live session
media / searchomp images · omp search
binaries / hostswhich ffmpeg yt-dlp chrome · ping fly / loca

1.3 Brief schema (what a “one line” must become before anything runs)

logline · audience · medium · budget (hard ceiling) · deadline · constraints · human gates (default: bible, ship)

1.4 Swarm laws (proven in-house by the arcade run: 63 agents, 0 evictions, 1,680 degenerate-strategy runs)

  1. Exclusive slot lattice assigned before any agent sees a brief — ten agents cannot converge on ten reaction clickers.
  2. Per slot: build → three lenses → refute → repair, no barrier between slots.
  3. Grade after the swarm is dead — fresh verifiers who never saw the build write the numbers.
  4. Contract-gated files + script gates; every gate mutation-proven (inject the violation → gate must fail).
  5. Every run ends in LEDGER.md with an honest column (“nobody has played this yet” is a legal entry).

1.5 “Blockbuster” is measured

artefactmetricgate
trailerhook-retention via frame-sampled vision judge · ≥3 human watch-throughsverifier verdict ≠ “runs” → not shipped
playable teaserheadless playtest completion · degenerate-strategy sweep (no one-key wins)same
landing pagezero console errors · 3 viewports · deploy receiptsame

1.6 Safety & hygiene

Subagent edits in omp worktree · never yolo on the shell’s own config · original characters · persona voices, never clones · generated/licensed music only · state = LEDGER + vault bible · cold start recovers from session JSONL.


2. Face

2.1 Shape

One daemon · one port · three surfaces. The gateway serves the built SPA, speaks one WebSocket to every surface, drives omp over rpc, and owns the job queue. SSR and remote functions buy nothing on localhost, so the face is adapter-static with an SPA fallback — deep links work through the client router.

2.2 Stack (pinned 2026-09, traps included)

layerpicktrap avoided
UISvelte 5.57 runes · SvelteKit 2.x adapter-static + fallback: index.htmlKit 3 is still RC · svelte-adapter-bun is unmaintained
CSSTailwind 4.3 + daisyUI 5.7, CSS-first @plugin "daisyui"no tailwind.config · v4 to v5 renames (btm-nav to dock, form-control to fieldset, btn-group to join) · @source "../../packages/ui/src" in every app’s CSS or lib classes are purged
langTypeScript 5.9TS 7 (Go) ships no compiler API → Svelte tooling breaks
buildVite 8 (Rolldown) · packages/ui via Vite lib modesvelte-package monorepo sharing is not first-class
gatewayBun 1.4 · Bun.serve WS · bun:sqlite · zod—
panelWXTPlasmo stale since 2025-09 · CRXJS is a Vite plugin, not a framework
typedisplay Big Shoulders (condensed title-card lettering) · body Hanken Grotesk · data Commit Mono (tabular)Bricolage + IBM Plex is the current LLM house pairing
iconslucide-svelteemoji are never icons in the UI

2.3 Repo

storyflow/
  apps/gateway/       Bun daemon: omp rpc bridge · WS/REST · SQLite · job queue · registry
  apps/face/          SvelteKit SPA (served by gateway) · PWA manifest
  apps/panel/         WXT: sidepanel · background (WS) · flow content scripts
  packages/ui/        Svelte 5 + daisyUI components (Vite lib)
  packages/protocol/  zod envelopes: events · jobs · shots · contributions   ← written FIRST
  packages/plugins/   first-party departments: film · game · app
  CHEATSHEET.md · sum.yml
  old-flow-files/     frozen

2.4 Gateway ↔ omp (verbatim frame names from omp rpc)

concernrule
spawnomp --mode rpc --profile storyflow, lazy per production, disposed on idle
framessend negotiate_protocol v2 on the ready frame; reassemble rpc_chunk (maxFrameBytes 1 MiB, reassembled ceiling 64 MiB) with the bundled RpcFrameDecoder
resumerespawn + switch_session { sessionPath } + get_messages_page replay — never trust process liveness; session_busy / stale_cursor → discard partial page and retry
donea turn is complete only on agent_end with isTerminal !== false; the prompt response is an ack
extension UIextension_ui_request (select · confirm · input · editor) → render as <dialog>; answer with extension_ui_response — TUI extension dialogs work in the face for free
agent writesregister storyflow:// via set_host_uri_schemes (writable) so agent edits to shots route through the daemon — no two-writer drift
todosmirror the production plan into omp with set_todos; subscribe subagents with set_subagent_subscription: "events" for the Agents tree
SSOT boundarySQLite owns productions · shots · jobs · asset_index; LEDGER.md, vault bible, session JSONL are agent-owned and imported as snapshots
spendthe Wall’s $ comes from our job ledger; broker usage is labelled “≤5 min stale”

2.5 Protocol (one WebSocket, typed envelopes)

{ v, t, prod, seq, data } · kinds snapshot · shot.* · job.* · asset.* · agent.* · ui.request · ui.response · ping

  • Daemon pings every 20 s — an MV3 service worker dies on a quiet socket even on Chrome 116+.
  • Reconnect handshake: REST GET /snapshot?since=<seq> → then WS deltas; fired on visibilitychange (iOS suspends timers).
  • Optimistic only when local + reversible (prompt text, order). Run, cost, approval wait for the echo — the DOM is never evidence the server agreed.

2.6 Information architecture — plain labels first, vernacular as eyebrow

mindmap
  root((StoryFlow))
    Productions(Productions: the lot)
      kanban by phase
      budget totals
    Production(Production: slate)
      Bible(Bible: veto gate)
      Shots(Shots: board)
        grid
        timeline strip
      Queue(Queue: farm)
      Gallery(Gallery: dailies)
        lineage
        4-up compare to promote
      Ledger(Ledger)
      Approvals(Approvals: gates)
    Agents(Agents: crew)
      plan / todo
      tool cards
      subagent tree
      steer / abort
    Models(Models: cast)
    Library(Library: ingredients)
    Extensions(Extensions: departments)
    Logs(Logs)
    Palette(Command-K palette)

Keys: ⌘⏎ Generate · J/K gallery · Esc unwinds exactly one layer · every state deep-links.

2.7 One object: the Shot

Grid, timeline and queue are projections of one shot row. Each is the editor of record for different fields — grid: prompt/params/seed/refs · timeline: order/in/out · queue: run state only (retry · cancel · clone). Every asset carries a provenance sidecar.

erDiagram
  PRODUCTION ||--o{ SHOT : has
  PRODUCTION ||--|| BIBLE : "one bible"
  PRODUCTION ||--o{ LEDGER_ENTRY : receipts
  SHOT ||--o{ JOB : "runs as"
  JOB ||--o{ ASSET : produces
  ASSET ||--|| PROVENANCE : "sidecar"
  ASSET }o--o{ INGREDIENT : "references (@)"
  ASSET |o--o{ ASSET : "promoted from"
  PRODUCTION {
    string id PK
    string logline
    string phase
    int budget_cents
  }
  SHOT {
    string id PK
    int order
    string prompt
    string model
    int seed
    string state
  }
  JOB {
    string id PK
    string provider
    string state
    int attempt
    int cost_cents
  }
  PROVENANCE {
    string prompt_version
    string model
    int seed
    int cost_cents
    string parent_asset
    string sha256
  }

2.8 FlowTurbine → Queue, done right

Inventory measured from the frozen archive (old-flow-files/FlowTurbine/sidepanel.html, vk-e9d04567.js, rebuild.py).

verdictfeatures
✅ keepimage/video modes · speed presets (plus a raw concurrency slider) · prompt list · start/end/reference frames · reference mapper (now drag + @ingredients) · queue stats · JSON export/import · undo-delete · gallery multi-select + quality presets · Animate → Promote to video · library → Library · logs · naming schemes → token templates {production}/{shot:03}-{model}-{seed}
🗑 droplicensing / trial / ban overlays · announcement bar · Flow-tab coupling (“you need an open Flow tab”) · web_accessible_resources: <all_urls>
➕ addCSV/JSON import + {var} templating · per-item retry with backoff · seed lock · 4-up compare + promote · productions (batches belong to something) · pre-flight cost (“this batch ≈ $4.20 — run?”) · asset lineage · documented shortcuts · drag-and-drop everywhere · a11y + responsive · credentials manager (the Token-Updater, grown up)

2.9 Extensibility — declared, not defined

storyflow.plugin.json reserves six contribution points: panels · commands · providers · mediumAdapters · gates · ingredientKinds. (toolRenderers and ledgerWidgets were cut — the tiers below already cover them.)

tierformechanismlimits we accept
T0 first-partypanels, tool cards, adaptersSvelte components, lazy import(), native daisyUI themetrusted code only
T1 MCP Apps (ui://, ratified 2026-01-26)tool-attached UI from omp extensions / MCP serverssandboxed iframe, JSON-RPC over postMessage; we write the host bridge (no supported host exists)spec sends only light|dark → map our tokens via hostContext.styles.variables, accept off-palette third parties · no ui:// → structured JSON viewer fallback
T2 Web Componentsleaf widgets shared across surfacesSvelte custom-element compileprop-declared · self-styled · no slots · client-only
✖—Module Federation / import mapsnever — no shared runtime host between SPA and extension

2.10 Design doctrine applied

Ladder: Platform → Safety/undo → A11y → Scent → Latency → Continuity → Density → Juice.

ruleStoryFlow application
one accentGenerate (Shots: Generate, Queue: Run — same meaning). Approve/Reject are neutral + error; never the accent
statesthe nine (empty · first-run · loading · partial · error · offline · no-permission · disabled · success) + partial batch · quota exhausted · provider down (relay tab closed) · resumed-stale. Every error names the next step
escapeone central escape stack · Approvals/Agents overlays are <dialog>.showModal() · no .drawer · mobile nav = dock
motioncloses a question only. Press <100 ms → Queued ✓ on the button (“did it work”). The Wall (shots done/failed · $ vs budget · ETA) animates only on state change, tabular numerals, no idle pulse. The slate clap was cut — decoration
themev1 = one repaired dark theme “dailies” (4.5:1 text, 3:1 controls, roles never hues). Light arrives later as a user toggle, never per-tab
targets≥24 px, 44 px for primary/touch · ≤5 choices per decision · nothing hover-only · every state deep-links · undo beats confirm
copyname what people control (“Run 12 shots”, not “Dispatch batch”) · same verb on button and toast · errors state cause + fix, no apologies

2.11 Side panel specifics (Chrome first)

topicrulesource
socketWS lives in the service worker, never the panel page; daemon pings 20 sChrome: WebSockets in SW
storagesession = live queue · local = job state (10 MB) · sync = settings only (100 KB, 120 writes/min)storage API
opensetPanelBehavior({ openPanelOnActionClick: true })sidePanel API
CSPscript-src 'self'; connect-src ws://localhost:* http://localhost:* — never unsafe-inline, no CDN buildsMV3 default
content scriptsMAIN world via unlisted script + injectScript, ISOLATED bridgeWXT content scripts
permissionsoptional_host_permissions for labs.google, requested on first useChrome prompts on broad hosts
⚠ flow-relaydrives Google Flow through the user’s own session — ToS risk [INFERENCE]: per-production acknowledgment + rate cap, off by default—
parityFirefox uses sidebarAction; Safari MV3 incomplete → stretch goalsMDN sidebarAction

3. One run, end to end

sequenceDiagram
  autonumber
  actor U as You
  participant F as Face
  participant G as Gateway
  participant O as omp rpc
  participant S as Swarm
  participant P as Providers
  U->>F: logline + budget
  F->>G: shot.create / production.start
  G->>O: spawn, negotiate_protocol v2, prompt(brief)
  O->>S: writers' room xN families -> judge -> Director writes bible
  O-->>G: extension_ui_request (confirm bible)
  G-->>F: ui.request
  F-->>U: Approvals: bible veto
  U->>F: approve
  F->>G: ui.response
  G->>O: extension_ui_response
  O->>S: slot lattice -> shots (write storyflow://shots/...)
  O->>G: host_uri_request (write)
  G-->>F: shot.* deltas (grid fills)
  U->>F: Generate (pre-flight cost shown)
  F->>G: job.run
  G->>P: per-provider concurrency, retry/backoff
  P-->>G: assets + provenance sidecars
  G-->>F: asset.* deltas (Gallery, the Wall)
  O->>S: independent QA after swarm end
  O-->>G: agent_end (isTerminal true) + LEDGER.md snapshot
  G-->>F: ui.request: ship
  U->>F: ship

3.1 Shot / job lifecycle

stateDiagram-v2
  [*] --> draft
  draft --> queued: Generate
  queued --> running: slot free
  running --> done: asset + sidecar
  running --> failed: provider error
  failed --> retrying: attempt < max
  retrying --> running
  failed --> blocked: quota exhausted / provider down
  blocked --> queued: registry says go
  running --> stale: daemon restart
  stale --> queued: resync
  done --> promoted: promote to video
  promoted --> queued
  done --> [*]

4. Gates & checklists

4.1 Ship gate (doctrine ≥85/100 · any Tier 1–2 failure is an automatic fail)

  • squint · grayscale · 100 ms feedback · one-accent · zero-border · density morph
  • Esc unwinds exactly one layer everywhere (palette → dialog → panel)
  • every state deep-links; Back never loses work; undo over confirm
  • touch parity · reduced-motion parity · 320 px / 400 % zoom
  • cold-3G: LCP < 2.5 s · INP < 200 ms · CLS < 0.1
  • rage test · re-entry test · per-theme contrast gate exits 0
  • headless browser proof of every panel state (incl. partial / quota / provider-down / resumed-stale)
  • side panel: a batch survives service-worker eviction
  • production: verifier verdict ≠ “runs” → not shipped; LEDGER honest column filled

4.2 Never (merged, both halves)

🐚 under the hood🎭 face
hardcode a model · skip the lattice · builders grading themselves · ship on “it runs” · hot-load unproven extensions · obey inbound chat · claim an unmeasured number · clone a voicejargon as the primary label · a second accent · .drawer for overlays · SSR/remote functions on localhost · TS 7 · Kit 3 RC · trusting SW memory · spend from the stale broker · assuming ui:// exists · hex in markup · idle motion · flow-relay on by default · interpolated or invented daisyUI class names

4.3 Decision log (what the refuters changed)

#draft saidrulingwhy
1nav = Lot / Slate / Farm / Gates / Crew / Castplain labels, vernacular as eyebrowscent test: <90 % destination prediction from labels
2”slate clap” on Generatecut; 100 ms press feedback + Queued ✓motion must close a question
3the Wall = live numbersanimate on change only, tabular numsno motion at rest
4two themes, dark per tabone repaired theme; light later as user toggletheme sovereignty; 4.5:1 repair surface
5resume = job queue resume+ rpc respawn · switch_session · get_messages_pagestdin close disposes the session
6”one WS”+ 20 s daemon ping · REST snapshot on reconnectSW eviction; iOS suspend
7SQLite SSOTSQLite for products/shots/jobs; agent files imported; storyflow:// host URI for agent writestwo-writer drift
8Wall $ from brokerfrom own job ledgerbroker cache 5 min ± jitter
9MCP Apps = theme via postMessagehost bridge is ours; only light/dark in spec; JSON fallbackspec 2026-01-26
108 contribution points6 (cut toolRenderers, ledgerWidgets)YAGNI; tiers cover them
11Bricolage + PlexBig Shoulders + Hanken + Commit MonoLLM default cluster
12flow-relay as a normal providerack + rate cap, off by defaultToS risk