CEO Review: SkyTwin Full Build Plan

Generated by /plan-ceo-review on 2026-04-01 | Mode: SCOPE EXPANSION | Branch: jayzalowitz/add-gstack

---

Vision

10x Check

The 10x SkyTwin is the OS layer between you and everything digital. Every app is a data source for the twin. Every API is an action surface. The twin doesn't just answer "what would I want done?" — it does it before you ask. You wake up and the day is already organized. The 10x version is invisible. You only notice it when it's wrong — and it's almost never wrong.

Platonic Ideal

A personal nervous system. The twin model is a continuously updated world model of one person's judgment. Temporal reasoning, social graph awareness, financial patterns, cross-domain synthesis. The UX is almost nothing: morning briefing, evening summary, occasional approval requests that get rarer over time. The feeling: someone competent is handling my life.

---

Scope Decisions

#	Proposal	Effort	Decision	Reasoning
1	Twin Query API + whatWouldIDo()	M	✅ ACCEPTED	Core driving function — this IS the product
2	OpenClaw Adapter + ExecutionRouter	L	✅ ACCEPTED	Execution breadth is the thesis. Create IronClaw issues when only OpenClaw has a skill.
3	Proactive Mode + Morning Briefing	L	✅ ACCEPTED	Transforms from reactive to proactive
4	Preference Archaeology	M	✅ ACCEPTED	Accelerates twin learning by surfacing implicit patterns
5	Cross-Domain Synthesis (foundation)	XL	✅ ACCEPTED	Foundation only — cross-domain event correlation engine
6	Twin Export + Portability	S	✅ ACCEPTED	Reinforces user sovereignty, cheap to build
7	Undo-with-Learning	S	✅ ACCEPTED	Fast feedback with reasoning accelerates twin quality

---

Architecture Decisions (13 issues resolved)

#	Section	Issue	Decision
1	Architecture	ExecutionRouter location	New @skytwin/execution-router package — clean separation from adapter packages
2	Architecture	Proactive scanner location	Part of @skytwin/decision-engine — reuses scoring/ranking infrastructure
3	Architecture	whatWouldIDo() pipeline depth	Full pipeline, no execution — prediction matches what system would actually do
4	Architecture	Cross-domain event storage	New signals table in @skytwin/db — CRDB will have memory AI later
5	Data Model	Preference proposals storage	New preference_proposals table — clean separation from confirmed preferences
6	Data Model	Undo reasoning capture	Extend feedback_events with undo columns — unified feedback stream
7	Safety	Explanation logging for queries	Log as predictions with type='prediction' — valuable for twin quality eval
8	Safety	OpenClaw risk adjustment	Adapter risk modifier — adapters declare trust characteristics, router bumps risk
9	API Surface	/ask endpoint protection	Token-scoped rate limiting, scales with trust tier (60/hr → 600/hr)
10	Testing	E2E test strategy	Golden path test covering full pipeline: signal → outcome → feedback → twin update
11	Sequencing	Build phase sequencing	5 phases — approved
12	Product/UX	Briefing delivery format	Dashboard-first + email digest opt-in
13	Risks	Risk mitigations	Time-box cross-domain (1 week), mock OpenClaw first, HIGH confidence for proactive

---

Build Phases

Phase 1: Foundation Extensions (DB migrations — unblocks everything)

• New signals table (id, userId, source, type, domain, data JSONB, timestamp, TTL)
• New preference_proposals table (id, userId, domain, key, value, confidence, evidence JSONB, status, timestamps)
• New proactive_scans, briefings, twin_exports, skill_gap_log tables
• Extend feedback_events: add undo feedbackType, undo_reasoning JSONB, undo_step_id
• Add source column to decisions (reactive | proactive | query)
• Add adapter_used column to execution_results (ironclaw | openclaw | direct)
• Add type column to explanation_records (action | prediction)

Phase 2: Core New Capabilities (parallel tracks)

Track A: Execution Router

• New @skytwin/execution-router package
• ExecutionRouter with adapter selection (risk level, skill availability, credentials)
• OpenClawAdapter implementing IronClawAdapter interface (mock-first)
• Adapter trust characteristics + risk modifier system
• Fallback chains (IronClaw → OpenClaw → Direct for low-risk)
• Skill gap detection and logging

Track B: Twin Query API

• whatWouldIDo(userId, situation) in @skytwin/decision-engine
• Creates synthetic DecisionObject, runs full pipeline, no execution
• Logs ExplanationRecord with type='prediction'
• POST /api/v1/twin/:userId/ask endpoint
• Token-scoped rate limiting (60/hr observer → 600/hr high_autonomy)

Track C: Twin Export

• Export function in @skytwin/twin-model
• Serializes: profile, preferences, inferences, patterns, traits, temporal profile
• JSON + Markdown formats
• GET /api/v1/twin/:userId/export endpoint
• twin_exports audit table

Phase 3: Higher-Order Features (depend on Phase 2)

Track A: Proactive Mode + Morning Briefing

• ProactiveEvaluator in @skytwin/decision-engine
• Loads full user context from all connected accounts
• Ranks "what needs doing most" using existing scoring
• Morning briefing generation (ranked actions, confidence, reasoning)
• Dashboard-first delivery with inline approve/reject/edit
• Email digest opt-in with deep links to dashboard
• Worker cron integration (configurable schedule)
• Requires HIGH confidence for any proactive auto-execution

Track B: Preference Archaeology

• Pattern analyzer in @skytwin/twin-model
• Scans evidence + feedback history for implicit patterns
• Generates PreferenceProposal records
• GET /api/v1/preferences/:userId/proposals endpoint
• POST /api/v1/preferences/:userId/proposals/:id (accept/reject)
• On accept: creates real Preference with source='inferred'
• On reject: records as negative signal for twin learning

Track C: Undo-with-Learning

• Extended feedback recording in @skytwin/twin-model
• Captures: which step went wrong, user's reasoning, preferred alternative
• Triggers rollback via ExecutionRouter if action is reversible
• Feeds reasoning back to twin model as high-weight correction signal (2x weight)

Phase 4: Integration Layer

• Cross-domain correlation engine (time-boxed: 1 week)
  • Signal persistence in signals table with retention policy (30 days)
  • Cross-domain event matching (email about meeting + calendar conflict)
  • Correlation results feed into DecisionContext enrichment
  • Foundation only — not a general-purpose CEP
• Skill gap logger → IronClaw issue creation
  • When ExecutionRouter finds OpenClaw has a skill but IronClaw doesn't
  • Creates structured issue with skill description, frequency, user demand
• Golden path e2e integration test
  • Mock signal → interpret → twin query → candidates → risk + adapter modifier → policy → router → mock execution → explanation logged → feedback recorded → twin updated
  • Validates every package boundary in one test

Phase 5: Polish

• Rate limiting implementation on /ask endpoint
• Briefing schedule configuration UI
• Dashboard updates for proposals, briefings, skill gaps, undo history
• Export format refinements

---

New API Endpoints

Method	Endpoint	Purpose
POST	/api/v1/twin/:userId/ask	whatWouldIDo() prediction (rate-limited)
GET	/api/v1/twin/:userId/export	Twin profile export (JSON/Markdown)
GET	/api/v1/briefings/:userId	Latest morning briefing
PUT	/api/v1/briefings/:userId/preferences	Configure briefing schedule/format
GET	/api/v1/preferences/:userId/proposals	Proposed preferences from archaeology
POST	/api/v1/preferences/:userId/proposals/:id	Accept/reject proposed preference
GET	/api/v1/skill-gaps	View skill gap log

New Packages

Package	Purpose
@skytwin/execution-router	Adapter selection, routing, skill gap detection, risk modifiers

Schema Changes

Change	Type
signals	New table
preference_proposals	New table
twin_exports	New table
skill_gap_log	New table
proactive_scans	New table
briefings	New table
feedback_events.undo_reasoning	New JSONB column
feedback_events.undo_step_id	New column
decisions.source	New column
execution_results.adapter_used	New column
explanation_records.type	New column

---

Risk Mitigations

Risk	Mitigation
Cross-domain scope creep	Time-boxed to 1 week. Ship whatever's done. Foundation only.
OpenClaw API uncertainty	Mock-first adapter. Real integration deferred until API stabilizes.
Proactive false positives	Require HIGH confidence for any proactive auto-execution. Conservative start.

---

Planning Artifacts

Detailed issue specs written to planning/ directory:

• issue-011-execution-router.md
• issue-012-twin-query-api.md
• issue-013-twin-export.md
• issue-014-proactive-mode.md
• issue-015-preference-archaeology.md
• issue-016-undo-with-learning.md
• issue-017-cross-domain-correlation.md
• issue-018-phase1-migrations.md
• milestone-1.5-scope-expansion.md

---

🤖 Generated with Claude Code via /plan-ceo-review