Skynet — Delivery Plan (sequence, gates, cards — no dates)

For: Agon, Sead, Sven (CTO), Vince. Companion to: the functional design (what), the workflow source map (how work flows), the architecture and deployment design (how it is built and run). The executable text of every S1 card is in skynet-task-specs.md; this plan is the order, the gates and the load.

1. Rules of this plan

1. Delivered = 100 % complete. The project is done when every workflow in the source map runs natively in Skynet, the standalone apps are retired, ClickUp is replaced board by board, and the app is SaaS-ready. Nothing before that is "the delivery".
2. No calendar dates. Work is a sequence of slices (S1–S5) and, inside a slice, waves. A wave starts only when the previous wave's gate is true. A gate that is not met is a stop, not a slip.
3. Best implementation, no quick fixes. A card that would need a shortcut to close is split or re-planned, never closed short. Discovered work becomes a card before its code lands.
4. Load: one card, one owner; sizes S ≤ ½ day, M ≈ 1 day, L ≈ 2 days; no person carries more than 2 days per wave (a wave ≈ two working days of effort per person, not a date). Loads below are computed from the card JSON.
5. Three lanes: A = Agon (platform, deploy, boxes, later Intelligence), V = Vince (shell, Production, bindings, later Studio UI), S = Sead (core auth, contracts, creative code, read layer, readback, bridge, later Workspace).

2. Slice sequence and entry gates

2b. Bridge track B — standalone work for Mathew while S1 runs

Agon's decision: the standalone scraper and RescaleOS may be changed now for Mathew's needs, in parallel with S1; no freeze between the projects. Track B is temporary: it lives in the standalone repos, it does not redefine Skynet's ownership or end state, and every bridge path is retired by A19 / A20 / B0 once the native flows exist. Cards are in the task specs, section "Bridge track B"; ordered by what unblocks Mathew first.

Impact on Skynet cards (features, not dates): A18b becomes a port of the proven PubMed adapter, not a build; A18c consumes the proven evidence-packet contract; A12a takes the Grok extraction schema as its prompt asset seed; A13a takes the lander/destination filter as lens semantics; V14 takes the extraction schema, the "write one for me" contract and rules.md as layer inputs; V17 takes the reaction-banking rule as instinct-loop source; V15's evidence/teardown views follow the package shape the bridge proved. A19 and A20 gain explicit sunset duties; new card B0 · Bridge sunset checklist (A, S, after A19 + A20).

3. S1 exit gate

1. app.rescale.media serves the unified app from server.ecomprofits.io; one login for Skynet itself.
2. Production fully native at parity.
3. Creative-code loop proven with one real ad.
4. Bindings + spend guard + minimal admin.
5. Export package attached to the existing ClickUp Launch card via the workos bridge.
6. Intelligence and Studio entries link to the standalones.
7. Staging + CI/CD + backups + rollback drill + monitoring.

4. Card delta from the previous card set (48 → S1 set)

S1 card set = 43 cards (plus A4 in S2 and S0b/S6/S8b/V9/S11 in S4 kept in the same JSON). Specs for rewritten/added cards are updated in skynet-task-specs.md in the same review pass as this plan (marked "S1-rev" there).

5. S1 waves, gates and load

Load (S=½, M=1, L=2): W0 A 0.5 · S 0.5 — W1 A 2 · V 2 · S 2 — W2 A 2 · V 2 · S 2 — W3 A 2 · V 2 · S 2 — W4 A 2 · V 2 · S 2 — W5 A 2 · V 2 · S 2 — W6 A 2 · V/S QA. No cell above 2.

Hard dependency chains: S0a → A1a → A1b → A2 → A3 → A3b → A6 → X7 (deploy spine, Agon) · S1a → contract day → S3/S4 → V7 → S5a → S8a → S7 → S10 (loop spine) · V3 → M1 → V10 → S12 (Production data). Everything else floats inside its wave.

Within-wave order where a same-wave dependency exists:

• W4: S7 lands first; S10 runs on the readback output from S7 plus the stamped export from V7; V10 does parity/regression work in parallel, but its winners/code/lineage acceptance closes only after S7 is green.
• W5: S12 closes before the production migration checklist is signed.

6. S1 cards — one line each (full spec in skynet-task-specs.md)

Agon: A0 box prep · A1a repo + imports + CODEOWNERS · A1b CI + governance · A2 compose (prod / staging / dev) · A2c limits + Postgres tuning + concurrency caps · A3 Caddyfile (SPA + /api/*, request ids, limits, rate limit) · A3b deploy pipeline (build → staging → smoke → prod, digests, .env.release, Telegram) · A6 staging stack · X5 rollback drill + cutover matrix · X7 production stack prepared · X3 synthetic + backup freshness · X3b Supabase-protection smoke · X4 restore drill · X4b backup path validation · X1 auth policy tests · M3 email · M4 legacy-host monitoring · M6 comms.

Vince: V1 SSO in Production · V2a shell skeleton · V3 Production on unified PG + ops endpoints · V4 bindings/usage/spend guard · V6 shell with links · M1 settings migration · V7 creative code on export + lineage · V8 gateway client · V10 parity + winners with code/lineage · V5 minimal admin · V13 contract tests + docs · X2 proxy I/O tests · M2 R2 prefix ADR.

Sead: S0a workos scrub · S1a core auth + contract · S1b invites + user import · S3 creative-code service · S4 lineage + outbox · S4b bindings contract · S5a read layer · S8a narrow bridge endpoints (POST /naming/ad-name, POST /launches/:id/package) in the old workos repo · S7 readback · S10 loop E2E · S12 migration dry run · S13 isolation tests · M5 retention plan.

7. Slices S2–S5 — full card sets (scoped to the same standard as S1)

Agon's decisions that shape these slices: S2 — scraper engine + transcriber run on server.ecomprofits.io as Skynet services with their own limits; real port of the meta_ads schema + RPC functions into Skynet's Postgres and the engine rewired to direct pg (no PostgREST shim); Intelligence UI rebuilt natively; scraping behaviour per the functional design (competitor entities ↔ products, lenses, schedules, alerts, auto-recreate rules, the scrape → classify → family → teardown → script chain). S3 — everything on the stack, skills/prompts as versioned in-app assets, subscription runner as an internal binding; needs Bian's agreement. S4 — only the 13 creative-workflow ClickUp lists are replaced; company-ops lists stay in ClickUp. S5 — multi-tenant, invite-only, manual invoicing from metered usage; Stripe self-serve out of scope. Executable specs for every card below are in skynet-task-specs.md (sections S2–S5); the JSON carries them.

S2 · Intelligence merge

S2 exit gate: Appendix A.1 at 100 % parity inside Skynet (settings, tools, research, teardowns, market intel, admin surfaces); standalone scraper off; static-ads' internal scraper retired.

Load (S=½, M=1, L=2): W0 A 1.5 · S 0.5 — W1 A 2 — W2 A 2 — W3 A 2 · V 2 — W4 A 2 — W5 A 2 — W6 A 2 — W7 A 2 · V 2 — W8 A 1.5. No cell above 2.

Risks → cards: RPC/browse edge cases hidden by Supabase → A4; scraper load on the shared box → A4, A2d; promotion semantics drift → D21, D22, A15; wrong VOC groups/export → D23, A18; wrong evidence classes / unsafe claims → D23b, A18b, A18c; UI parity gaps → V22, V15, A19; alert/auto-recreate rules misfire → D21, A13b.

S3 · Studio merge + strategy layer

S3 exit gate: Appendix A.3 at 100 % parity inside Skynet; Bian's loops run on the stack; KPIs come from readback; the script entity bridges to ClickUp until S4; every notification goes through the platform with quick actions; the team can tag and ask the agent in Telegram/Discord; the standalone console is off.

Load (S=½, M=1, L=2): W0 A 0.5 · S 0.5 · V 0.5 — W1 A 2 — W2 A 2 — W3 A 2 · S 2 · V 2 — W4 S 1 · V 2 — W5 S 1 · V 2 — W6 S 2 · V 2 — W7 S 2 · V 2 — W8 A 2 · V 1 — W9 A 1. No cell above 2.

Risks → cards: Bian rejects the model mid-build → D31; Postgres migration changes loop semantics → A10, A11; decorative coverage map → D32, V18; script bridge double truth → D33, S16; taste captured as shared truth → V17, V19, S9.

S4 · Workspace cutover

S4 exit gate: the 13 creative-workflow lists replaced by native boards; ClickUp read-only for them; workos in the stack; Railway off; company-ops lists untouched in ClickUp.

Load (S=½, M=1, L=2): W0 S 2 — W1 S 0.5 — W2 S 2 · V 2 — W3 S 0.5 — W4 S 1 · V 1 — W5 S 2 · V 2 — W6 S 2 — W7 S 2 · V 2. No cell above 2.

A.4 folds: naming (M3) stays in S1/S3 (S3, S8a, S16); classify knobs (M2) in S3 (S5b, S9), only hosted by S17; the presentation deck is archived inside S17.

Risks → cards: launch board built on guessed buyer behaviour → S0b, D41, S11; wrong board order → D42, S6; mirror drift on cascade paths → S8b, S26, S18, S19; hidden Railway-only secrets/crons → S17; company-ops lists touched by accident → S19, V26.

S5 · SaaS hardening

S5 exit gate (= 100 %): multi-tenant invite-only operation live; RLS enforced; per-workspace credentials/keys/quotas enforced; audit log + usage metering live; manual invoicing from metered usage; retention/deletion complete; admin/users/roles parity across Production, Intelligence, Studio, Workspace. Stripe self-serve out of scope.

Load (S=½, M=1, L=2): W0 A 0.5 · S 0.5 · V 0.5 — W1 S 2 — W2 A 1 · S 2 — W3 S 1 · V 2 — W4 A 2 · S 2 — W5 S 1 — W6 S 1.5. No cell above 2.

Risks → cards: late RLS breaks real workflows → S21; services still read global credentials → S22, A22; admin parity gaps → D52, V21; billing drifts into Stripe → D51, S24; deletion misses R2/lineage/sidecar data → S25b, A17.

8. Discovery and interviews (run during S1, feed S2–S4)

9. Confirmations still open (do not gate S1)

Bian — Studio stays standalone through S3; coexistence for A11. Leadership — D1–D4 (gate S4). Sven — bridge scope (S8a endpoints in the old repo, deployed on Railway). Cyrus — launch-pack minimum. R2 bucket versioning/lifecycle — infra input for the architecture doc §10.