Skip to content
/
Phase B Batch 19

Batch 19 Notes — Spec-Driven: Ralph-Loop + Flywheel + Meta-Prompt SDD

Batch 19 Notes — Spec-Driven: Ralph-Loop + Flywheel + Meta-Prompt SDD

Theme: Frameworks that lean on iterative Ralph loops, flywheel patterns, meta-prompting, or WYSIWYG spec authoring.

Roster Table

slug stars distribution cli_binary local_ui orchestration multi_model tier
smart-ralph 339 claude-plugin no no hierarchical no A
get-shit-done 943 npm-package yes (gsd-tools) no parallel-fan-out no A
forge-lucasduys 29 claude-plugin no no sequential no A
vibe-kanban 26,518 desktop-app no (npx) yes (Electron+Svelte) none (UI only) yes (user-selects) A
panaversity-sdd-exercises 1 methodology-doc no no n/a no A
nimbalyst 604 desktop-app no yes (Electron+SvelteKit) none (UI only) yes (user-selects) A
specpact 13 cli-tool yes (specpact) no none no A
flywheel-sdd 1,488 bash-script-bundle no no n/a (infra) no A
ag-coding-flywheel 2 claude-plugin no no (Cytoscape.js browser viewer) swarm yes (opus+sonnet+codex) A
reap 41 npm-package yes (reap) no sequential-with-optional-subagent no A

Intra-Batch Patterns

This batch divides cleanly into five distinct archetypes:

UI Layers (no prompt primitives): vibe-kanban and nimbalyst are both Electron desktop apps that wrap multiple AI coding agents (10+ in vibe-kanban's case) in a visual task board. Neither ships commands, skills, or hooks — they are pure UI orchestration surfaces. Notably, vibe-kanban is actively sunsetting itself in favor of nimbalyst (formerly Crystal), making this an unusual case of a living framework with a planned successor within the same batch.

Prompt/Skill Frameworks (pure methodology): smart-ralph, forge-lucasduys, and specpact ship primarily as structured CLAUDE.md + slash-command + hook bundles. All three use different strategies to solve the same problem: smart-ralph uses a 9-persona agent roster + FSM state; forge-lucasduys uses git worktree isolation per task; specpact uses a 3-tier priority-ordered rule system.

Heavy Swarm Orchestrators: get-shit-done and ag-coding-flywheel are the most operationally complex frameworks in the batch. GSD ships 66+ commands, 33 agents, and a dedicated npm CLI with context budget engineering. ag-coding-flywheel ships the only formal adversarial duel review in the batch (2 agents cross-score each other) plus Zod-validated typed completion attestations.

Infrastructure (not methodology): flywheel-sdd (ACFS) is a Bash installer that provisions Ubuntu servers with all the tools needed to run agentic workflows — it is not a workflow framework itself. Its most confrontational AGENTS.md ("YOU HAVE A HORRIBLE TRACK RECORD OF NOT FOLLOWING INSTRUCTIONS") is the bluntest agent prompt in the batch.

Generation-Based Lifecycle Systems: reap stands alone in introducing a biology-metaphor architecture (Genome, Generation, Lineage, Fitness) where the knowledge base self-evolves between generations. Its nonce-chain stage enforcement at the code level — not just prompt instructions — is architecturally distinct from all other frameworks in the batch.

Two frameworks share a methodological ancestor: flywheel-sdd and ag-coding-flywheel both trace to Jeffrey Emanuel / Dicklesworthstone's Agentic Coding Flywheel methodology. They take it in entirely different directions — flywheel-sdd is infrastructure tooling; ag-coding-flywheel is a Claude Code plugin with a MCP server, adversarial review, and 36-code error contracts.

Most Interesting Find

Forge's automatic backpropagation of runtime failures into specs (forge-lucasduys) is the most architecturally novel mechanism in the batch. A PostToolUse hook monitors tool execution for runtime errors. When an error occurs, the hook automatically adds a new R-numbered acceptance criterion to the originating spec and creates a regression test. This closes the spec-code feedback loop in a way absent from all 11 seed frameworks — the spec is not authored once and then implemented; it grows with the implementation failures. The mechanism is the closest thing in the batch to a "living spec."

Runner-up: ag-coding-flywheel's adversarial duel review (2 agents independently review the same implementation, then score each other's review) is the only formal adversarial review protocol in the entire batch-19 corpus. The Zod-validated completion attestations in .pi-flywheel/completion/<beadId>.json — machine-readable typed certificates rather than conversational "done" signals — represent a level of production engineering rigor absent from all seeds.

Tier C Stubs

None. All 4 Reddit-only references were located via gh search repos:

  1. specpact → found as ivansglazunov/specpact (13 stars)
  2. reap → found as c-d-cc/reap (41 stars, "Recursive Evolutionary Autonomous Pipeline")
  3. flywheel-sdd → found as dicklesworthstone/agentic-coding-flywheel-setup (1,488 stars)
  4. ag-coding-flywheel → found as burningportra/agent-flywheel-plugin (2 stars, but v3.18.0)

Cross-References Discovered

  • vibe-kanban → nimbalyst: vibe-kanban's repo README explicitly states it is sunsetting and recommends nimbalyst (formerly Crystal/stravu) as the successor. Both are in this batch.
  • flywheel-sdd → ag-coding-flywheel: Both are implementations of the Jeffrey Emanuel (Dicklesworthstone) Agentic Coding Flywheel methodology. flywheel-sdd provisions the infrastructure; ag-coding-flywheel provides the Claude Code workflow layer.
  • ag-coding-flywheel → pi-agent: The burningportra repo forks to burningportra/pi-agent-flywheel (pi-agent variant) and vnnkl/agentflywheel (skills-only extraction). The vnnkl fork's competing-plans skill is the source of the multi-model-triangulation mechanic documented in the ag-coding-flywheel prompts file.
  • get-shit-done repo warning: The original glittercowboy/get-shit-done repository redirects to open-gsd/get-shit-done-redux with a note about a meme-coin promotion incident. The open-gsd fork (943 stars, next branch) is the canonical active version analyzed here.
  • smart-ralph → openspec ecosystem: smart-ralph's two-pass semantic skill discovery auto-invokes matching installed skills from other plugins (openspec, etc.), making it an integration layer for the broader ecosystem.