GenAI Development Stack

Archon orchestration + SPARC methodology + Multimodal RAG — backed by GitHub & Linear

Project Backbone source of truth
📋
Linear
Issue tracking, cycles, roadmap
Issues = atomic work units for agents Cycles = sprint boundaries Labels: agent/auto human/review parallel-ok Status automations synced with GitHub PRs
Linear API Webhooks GitHub sync
🐙
GitHub
Code, PRs, Actions, review gates
Repos = all project codebases PRs = agent output review surface Actions = CI/CD + agent trigger pipelines Branch protection: agents can't merge to main
gh CLI Actions Webhooks
Bidirectional sync: Linear issue created → triggers GitHub Action → agent picks up work → opens PR → PR merge auto-closes Linear issue → cycle progresses. Magic words in commits (Fixes LIN-123) link everything.
Linear issue assigned to agent → webhook firesArchon picks up
Orchestration Layer Archon OS brain
🏛️
Archon
Agent orchestrator + knowledge backbone
Planner: Reasoning LLM (o3-mini / Claude) decomposes Linear issue into SPARC phases Router: Assigns phases to specialized agents, detects parallelizable work Knowledge: Surfaces docs, codebase context, and RAG results to each agent Feedback loop: Reviews agent output, triggers refinement if quality gate fails
LangGraph Pydantic AI MCP Multi-provider LLM
🧠
Multimodal RAG
Cognee + Honcho + Hermes
Feeds codebase knowledge to agents FTS5 / LanceDB / FalkorDB triple backend Honcho personalizes per-agent context → See RAG deployment diagram
Cognee SDK D1 + R2 FalkorDB
Archon decomposes → assigns SPARC phases to coding agents
Execution Methodology SPARC by ruvnet
S Specification
Define objectives, requirements, constraints, edge cases from the Linear issue
→ Researcher agent
P Pseudocode
High-level plan, algorithm design, API contracts before writing real code
→ Planner agent
A Architecture
System design, data flows, file structure, dependency mapping from RAG context
→ Architect agent
R Refinement
Iterative implementation, TDD, linting, type-checking, review loops
→ Coder agent(s)
C Completion
Tests pass, docs written, PR opened, Linear issue updated, deploy readiness
→ Deployer agent
Boomerang pattern: Refined code loops back to Architecture validation. If structural drift detected, Archon restarts from phase A. Each phase gate is a quality checkpoint — code only advances if tests, lint, and type checks pass.
SPARC phases execute via coding agentsparallel when independent
Coding Agents execution parallelizable
Claude Code
Primary coding agent
CLI agentic coding with MCP tools Deep codebase understanding SPARC-compatible via system prompts
CLI MCP Opus / Sonnet
🦘
Roo Code
SPARC-native coding agent
Native SPARC mode integration Boomerang task orchestration VS Code embedded execution
VS Code SPARC modes
🔮
Cursor / Cline
Secondary coding agents
IDE-integrated AI coding Good for UI/frontend work SPARC prompts via .cursorrules
IDE Multi-model
Parallel Execution — Independent Work Streams
Lane A — Claude Code
Bug fix: auth token refresh race condition (LIN-142)
Branch: fix/auth-token-142
running
Lane B — Claude Code
Bug fix: rate limiter not respecting burst config (LIN-143)
Branch: fix/rate-limiter-143
running
Lane C — Roo Code
Feature: add FTS5 search to settings page (LIN-144)
Branch: feat/fts5-settings-144
running
Lane D — Queued
Refactor: extract shared utils (LIN-145) — depends on A + B merging
Blocked: waits for fix/ branches
queued
Parallelism rule: Archon checks file-level overlap between issues. If no shared files → parallel lanes. If overlap → sequential with dependency. Each lane = separate git branch, separate agent instance.
agents push branches → open PRsGitHub Actions CI → human review
Quality Gates automated + human
GitHub Actions CI
Automated quality checks
Lint + type-check + unit tests Integration tests (non-overlapping = parallel) Security scan (SAST/dependency audit) SPARC Completion gate: all phases validated
Actions pytest / vitest ruff / eslint
🔄
Archon Feedback Loop
Auto-refinement on CI failure
CI fails → Archon reads error logs Routes back to SPARC Refinement phase Agent fixes → re-pushes → CI re-runs Max 3 retries before escalating to human
Webhook Auto-retry
👁️
Human Review
Final merge authority
PR review with AI-generated summary Approve → merge → Linear auto-closes Request changes → back to Archon Agents never merge to main directly
PR review CODEOWNERS
End-to-End: Issue → Merged Code
1
Issue Created
Linear issue with description, acceptance criteria, labels
Linear
2
Archon Picks Up
Webhook → decompose into SPARC phases → check for parallel opportunities
Archon
3
RAG Context
Cognee retrieves relevant code, docs, graph context for the task
RAG stack
4
S → P → A
Spec → Pseudocode → Architecture phases produce implementation plan
SPARC agents
5
Code (parallel)
Claude Code / Roo Code implement on separate branches. Non-overlapping = parallel
Coding agents
6
CI + Refinement
Actions run tests. Fail → auto-retry via Archon (max 3x). Pass → PR ready
GitHub Actions
7
Review + Merge
Human reviews PR, approves → merge → Linear auto-closes → cycle updates
GitHub + Linear
Integration Direction Protocol Trigger Purpose
Linear → Archon webhook HTTPS Issue created / updated Agent picks up new work, detects priority changes
Archon → GitHub API gh CLI / REST SPARC phase starts Create branch, push commits, open PR
GitHub → Archon webhook HTTPS CI pass/fail, PR review Trigger refinement loop or mark complete
GitHub ↔ Linear bidirectional Native sync PR merge / commit keywords Auto-close issues, update status, link PRs
Archon → RAG in-process Python SDK Before each SPARC phase Feed codebase context, docs, graph relationships to agents
Archon → Coding Agents spawn CLI / API SPARC Refinement phase Launch Claude Code / Roo Code with task prompt + context
Agents → GitHub push git + gh CLI Code written Commit to branch, open/update PR with SPARC summary
Where Things Run infrastructure
Fly.io / VPS
Archon — orchestrator process Hermes + Cognee — RAG engine (co-located) Honcho — sidecar memory service Local SQLite FTS5 — conversation cache
Cloudflare
D1 — Cognee KB (FTS5 relational) R2 — LanceDB vector files Workers — webhook receivers, D1 API adapter
Separate Host
FalkorDB — graph DB (FalkorDB Cloud / Docker on Fly) Postgres — Honcho user store (Neon / Supabase free)
Local Dev Machine
Claude Code CLI — interactive coding Roo Code / Cursor — IDE agents falkordblite — embedded graph for dev