Skip to content

PRIORITY.md

Latest commit

 

History

History
189 lines (147 loc) · 8.08 KB

PRIORITY.md

File metadata and controls

189 lines (147 loc) · 8.08 KB
Raw

PRIORITY.md — CTO Directive

Issued: 2026-05-22 03:27 UTC
Updated: 2026-05-22 10:05 UTC — P0 complete

Project State

  • 222K lines Go (151K prod + 71K tests), 5,368 functions
  • 199 packages, 276 directories, 172 commands
  • Build/Vet/Tests: Clean. Go 1.26.3.
  • Architecture docs: 12 in docs/architecture/
  • P0 wiring: COMPLETE

P0 Status

Task Status Commit
P0-1: Org Bootstrap → Persistent Org ✅ DONE 4e67df9
P0-2: Quality Gate Blocks Bad Code ✅ DONE 24f4615
P0-3: Dashboard Shows Real Data ✅ DONE c8ad6dc
P0-4: 60-Second Demo ✅ DONE this commit

Assessment

The codebase has passed through invention phase. 199 packages cover every gap in VISION.md through #181. The architecture is coherent. The strategy is clear. The moat is defensible.

The single critical gap: end-to-end wiring. Every subsystem works in isolation. None work as a system. forge org init creates an in-memory struct. qualitygate.Evaluate() returns a score that nothing blocks on. guard.CheckBudget() returns a boolean nobody reads. correlator.Correlate() produces signals that go nowhere.

The next phase is not more packages. It's connecting what exists into a working product.

P0 — The Working Pipeline (Next 48 Hours)

These four items, completed, produce a demonstrable product. Nothing else matters until these are done.

1. Org Bootstrap → Persistent Org

Why: forge org init must create a living org that survives restart. Wire: cmd/org.go → internal/org/org.go → SQLite persistence → internal/openclaw/session.go

  • forge org init creates 4 divisions + 4 head agents + 4 division channels
  • Each division gets a cost budget allocation
  • Org memory seeded with defaults (quality standards, coding conventions)
  • forge org status reads from SQLite, shows real data
  • Test: Run forge org init, kill process, run forge org status → data persists.

2. Quality Gate Blocks Bad Code

Why: "Code MUST pass review" is the #1 value prop. Currently it's a score nobody reads. Wire: internal/pipeline/executor.go → internal/qualitygate/ → internal/review/ → internal/guard/

  • Pipeline executor calls qualitygate.Evaluate() before marking step complete
  • If score < threshold → step fails, retry with higher-trust agent
  • If retry fails → escalate to division head
  • Every gate result recorded in genealogy + audit log
  • Trust score updated on pass/fail
  • Test: Agent produces garbage → pipeline step fails → trust score drops.

3. Dashboard Shows Real Data

Why: Users need to see their org working, not mock data. Wire: internal/dashboard/websocket.go ← internal/costlive/ + internal/org/ + internal/trust/

  • Agent status from org.ListAgents() (real, not mock)
  • Cost tracking from tokentracker (real spend, not random numbers)
  • Quality scores from qualitygate (real evaluations)
  • Division health from correlator (real signals)
  • WebSocket pushes updates on every state change
  • Test: Start org, create session, send task → dashboard updates in real-time.

4. 60-Second Demo

Why: Zero adoption without a demo. This is the growth blocker. Record: forge org init → agents working → quality gate catches bad code → dashboard shows activity

  • <90 seconds, zero errors
  • Published as docs/demo.md with asciinema recording

P1 — Production Enforcement (Next 72 Hours)

5. Cost Budget Actually Enforced

Wire: internal/guard/guard.go cost_cap → internal/cost/tracker.go → internal/openclaw/session.go

  • Guard checks budget before every session send
  • At 80% budget: model downgrade (GPT-4 → GPT-4-mini, etc.)
  • At 100% budget: session stop, division head notified
  • Every enforcement action → immutable ledger entry
  • Test: Agent hits cap → model downgrades → session stops at 100%.

6. Memory Compounds Across Sessions

Wire: Task completion → internal/memory/memory.go → internal/orglearn/ → internal/alignment/

  • Auto-store outcomes: what worked, what failed, what it cost
  • New agent onboarding reads org memory before first task
  • Nightly forge dream processes accumulated knowledge
  • Pattern extraction: "this type of task usually costs $X and takes Y minutes"
  • Test: Agent A completes task → Agent B (new) reads A's learnings → acts on them.

7. Compliance Blocks Violations

Wire: internal/compliance/compliance.go → internal/consent/consent.go → internal/auditlog/

  • Policy engine evaluates before external actions (email, API calls, deployments)
  • GDPR: PII detection blocks data export
  • Financial: amount limits block unauthorized payments
  • IP: license scanner blocks copypasta
  • Blocked actions → audit log + human notification
  • Test: Agent tries to export user emails → blocked → logged → human notified.

8. Feedback Signals Update Trust

Wire: internal/correlator/correlator.go → internal/trust/trust.go → internal/qualitygate/

  • Error rate spike → trust -5 per incident
  • Cost anomaly → trust -3
  • Quality drop → trust -3, tighten quality gate
  • Stuck 30min → stuck.Escalate() → division head notified
  • Positive outcomes → trust +3
  • Test: Agent's code causes error → trust drops → quality gate tightens for that agent.

P2 — Product Polish (This Week)

9. CLI Grammar Audit

  • Every command follows forge <noun> <verb> pattern
  • --output=json on every command
  • Remove duplicate commands
  • Consistent error messages

10. Payment E2E Wiring

Wire: internal/integration/payment.go → internal/banking/ → internal/cost/

  • Agent actions → cost records → division budgets → org billing
  • forge cost live shows real spend per agent/division/org

11. Documentation Website

  • Command reference (auto-generated from Cobra)
  • Quickstart guide
  • Architecture guide (from docs/architecture/)
  • Comparison pages (vs Cursor, Copilot, LangGraph, CrewAI)

P3 — Growth (Next 2 Weeks)

12. Plugin Marketplace MVP

  • Git-based registry
  • forge skill publish/install/version

13. Comparison Pages (SEO)

  • vs Cursor, vs Copilot, vs LangGraph, vs CrewAI, vs AutoGen

14. Conference Talk Submissions

  • GopherCon, AI Engineer Summit, KubeCon

Architecture Coherence Notes

What's Strong

  • Package surface: 199 packages, every VISION.md gap covered through #181
  • Test coverage: 71K test lines (47% of production). Above industry average.
  • Code quality: No stubs. Every package has real implementations with real logic.
  • Security: QA audits on schedule. Data races fixed proactively.
  • Bridge pattern: HTTP primary, local fallback, in-memory cache. Clean abstraction.
  • Architecture docs: 12 comprehensive docs covering every subsystem.

What's Not Working

  • End-to-end wiring: THE critical gap. Packages work alone, not as a system.
  • Dashboard: Mock data instead of real.
  • Quality enforcement: Scores computed but not used to block anything.
  • Cost enforcement: Budgets tracked but not enforced.
  • Feedback loops: Signals detected but not routed to trust/quality.
  • Org bootstrap: Creates structs in memory, doesn't persist.
  • Pipeline: Steps execute but skip the gate checks.

New Architecture Docs (This Sprint)

  • EXECUTION-PIPELINE.md — Full request-to-completion flow, wiring status matrix
  • DATA-ARCHITECTURE.md — Storage substrates, data ownership, consistency model
  • AGENT-LIFECYCLE.md — Hire to fire lifecycle, trust dynamics, cost budget lifecycle

Success Metrics

Metric Current Target (48h) Target (1 week)
E2E demo working No Yes Published
Org bootstrap persists No Yes Yes
Quality gate blocks bad code No Yes Yes
Dashboard shows real data No Yes Yes
Cost budget enforced No No Yes
Compliance gates active No No Yes
Memory compounds No No Yes
Feedback → trust updates No No Yes
Published docs 12 arch docs 12 + website + comparisons

Next CTO sync: After P0 items complete.