FlashFusion Agent Stack v1.9.9 — Build Pipeline Skill

Purpose

Provides the complete 5-phase build pipeline for FlashFusion Agent Stack v1.9.9. Each phase is a bounded, self-contained prompt that feeds the next via a BATON block. Phase 0 is mandatory first. Phases 1–3 are parallel-safe after Phase 0 clears. Phase 4 is the release gate.

---

Quick Navigation

Phase	Focus	Deliverables	Ref
0	Gaps + Assumptions Audit	GAPS_AUDIT.md	phases/PHASE_0.md
1	Foundation	ARCHITECTURE.md, SECURITY_MODEL.md, SQL migration	phases/PHASE_1.md
2	Orchestration	AGENT_CATALOG.md, GRAPH_SPEC.md, MEMORY_SPEC.md, TOOL_CONTRACTS.md	phases/PHASE_2.md
3	Operations	BILLING_GUARDRAILS.md, OBSERVABILITY_RUNBOOK.md, ROLLBACK_RECOVERY.md	phases/PHASE_3.md
4	Verification	API_CONTRACTS.md, TypeScript types, smoke tests, acceptance criteria	phases/PHASE_4.md

WSJF sprint backlog: WSJF_BACKLOG.md

---

How to Run This Pipeline

In Claude.ai Projects (recommended for doc generation)

1. Upload this zip to your Project via Settings → Skills → Upload Skill
2. Start a new conversation in that Project
3. Type: run phase 0 — the skill activates automatically
4. Paste the phase prompt from phases/PHASE_0_GAPS_AUDIT.md into the chat
5. Read the BATON block at the end before starting Phase 1
6. Repeat for each phase in order

In Claude Code CLI / Antigravity (recommended for implementation)

# Copy scaffolding stubs into your repo first
cp scaffolding/supabase/001_v199_foundation.sql ./supabase/migrations/
cp scaffolding/typescript/v199.ts ./src/types/
# Then paste phase prompts directly into Claude Code chat

Baton Pass Protocol

Every phase ends with a BATON block like this:

PHASE_N_STATUS: [COMPLETE / INCOMPLETE]
LOCKED_OUTPUTS: [what the next phase must reference]
OPEN_ITEMS: [gaps not yet resolved]
CONFIDENCE: [0-100%] — if below 80%, resolve before proceeding

Do not start the next phase if CONFIDENCE < 80% or STATUS = INCOMPLETE. Paste the full BATON block at the top of the next phase prompt.

---

Marking Conventions (used in all phases)

Marker	Meaning
[UNVERIFIED]	Claim made, source not confirmed
[SHALLOW]	Correct direction, needs expansion
[ASSUMED]	Constraint assumed — call it out
⚠	Adversarial finding — must resolve
!	Internal contradiction flagged

Rule: Shallow + honest beats deep + hallucinated. Use these markers.

---

Stack Reference (locked for v1.9.9)

Frontend:       Next.js (App Router)
Backend:        Node.js / TypeScript
Orchestration:  LangGraph (primary spine)
                CrewAI (content lanes)
                AutoGen (premium reasoning, optional)
Auth + DB:      Supabase (Auth, Postgres, RLS, Storage)
Hot State:      Redis (run locks, queues, session state)
Vector Memory:  pgvector via Supabase
Tools:          MCP-compatible adapters
Model Routing:  OpenAI / Anthropic / Gemini (abstracted)
Background:     Worker pool
Operator UI:    Agent Studio (read-only trace viewer scope)
Automation:     Make / n8n (external ops only)

---

Agent Cast (mandatory first wave)

Agent	Role
Router	Classify, gate, route requests
Planner	Decompose goal into bounded execution plan
Tool Executor	MCP calls, permission-scoped, idempotent
Memory	Decide what to retain / expire / redact
Validator	Schema, policy, plan, format checks
Publisher	External platform publish + post-actions
Recovery	Failure classification, rollback, resume

Second-wave agents (feature-flagged only, not in v1.9.9 golden path): Research, UI Critic, Brand Consistency, Commerce Optimizer

---

Non-Negotiable Build Rules

Security

• All tenant data RLS-protected — no exceptions
• Service-role key never in frontend or client bundle
• Tool inputs/outputs redacted before log/trace storage
• Secrets via env vars only — never hardcoded

Reliability

• Every side-effecting action requires idempotency key
• Every run produces a replayable trace
• All long-running jobs support checkpoint + resume
• Compensating action defined for every irreversible side effect

Cost Control

• Per-run token budget enforced in graph
• Max hops, retries, wall-clock duration set per route
• Premium model path blocked unless plan gate passes

Versioning

• Graph version + prompt version + schema version + model route stored per run
• Prompts treated as code: versioned, rollback-safe

---

Scaffolding Stubs

Pre-built stubs ready to drop into your repo:

• scaffolding/monorepo/FOLDER_TREE.md — full annotated monorepo structure
• scaffolding/supabase/001_v199_foundation.sql — migration stub with RLS
• scaffolding/typescript/v199.ts — typed interfaces stub

These are stubs, not finished code. Phase 1 and Phase 4 generate the real implementations.

---

References

• references/STACK_DECISIONS.md — why LangGraph over CrewAI / AutoGen / Make
• references/AGENT_ROLES.md — agent boundaries and interaction rules
• references/BATON_TEMPLATE.md — copy-paste baton block format
• references/MARKING_CONVENTIONS.md — full marker guide

---

Version Notes

v1.9.9 is a stability + hardening release, not a feature release.

Ships when:

• All 8 smoke tests pass
• All tenant data is RLS-protected
• All agent runs produce replayable traces
• All premium paths are plan-gated
• All long jobs are resumable or safely rollback-able
• No service-role key in any client bundle
• No sensitive field in any log or trace

Do not add experimental agents, new product features, or cosmetic improvements unless they directly serve stability, isolation, idempotency, or observability.