Skip to content
/
Phase B Batch 4

Batch 04 — Spec-Driven (Aider/Cursor/Codex + research-grade SDD variants)

Batch 04 — Spec-Driven (Aider/Cursor/Codex + research-grade SDD variants)

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
codex-spec 45 cli-tool (npm) codex-spec none sequential no A
metaspec-acnet 47 cli-tool (pip/uv) metaspec none sequential no A
lean-spec 252 npm-package leanspec web-dashboard (port 3000) sequential no A
fpf 372 methodology-doc none none none no A
adversarial-spec 546 claude-plugin none none consensus yes A
quint-code 1333 cli-tool (Go) haft terminal-tui (alpha) hierarchical yes A
colign 4 standalone-repo none web-dashboard (port 3000) sequential no A
fspec 69 npm-package fspec terminal-tui sequential no A
mcp-server-spec-driven-development 430 mcp-server none none sequential no A
tessl-sdd-tile 38 template-bundle none none sequential no A

Intra-batch patterns

This batch divides cleanly into three sub-groups. The first (codex-spec, mcp-server-spec-driven-development) are minimal pipeline wrappers: one calls the OpenAI API sequentially, the other wraps three sentences as MCP prompts — both achieve SDD scaffolding with near-zero code. The second (lean-spec, fspec, colign) are full-stack infrastructure plays: lean-spec adds a provider abstraction and React web UI; fspec adds Gherkin, Example Mapping, TUI Kanban, and Git checkpoints; colign adds a PostgreSQL-backed team platform with real-time collaborative editing. The third (adversarial-spec, quint-code, FPF) are the "research-grade formal" frameworks: adversarial-spec implements multi-LLM consensus loops with litellm; quint-code (Haft) implements FPF's formal constructs (R_eff, parity enforcement, decision contracts) in Go with SQLite; and FPF itself is the epistemological layer underneath quint-code. MetaSpec is orthogonal to all — it generates SDD toolkits rather than applying SDD, and the Tessl SDD tile is the most evals-rigorous minimal methodology tile in the corpus.

Most interesting find

quint-code (Haft) is the most architecturally sophisticated framework in this batch and arguably in the entire corpus: it implements FPF's formal reasoning constructs (Characteristic Spaces, R_eff evidence decay, parity-enforced comparisons, WorkCommission lifecycle) in Go with a SQLite knowledge graph — essentially a programmable engineering governance system that treats "what to build and why" as a machine-queryable contract, not a chat history. The explicit Choose → Execute autonomy boundary in AGENT_CONTRACT.md is the most principled agent governance statement observed across all batches.

adversarial-spec is surprising for its 546 stars given that it is a single Python script and one skill file — the demand for multi-LLM spec review is clearly real. Its "Claude as participant not orchestrator" framing and quorum consensus mechanism are unique in the entire corpus.

Items written as Tier C

None. All 10 frameworks had sufficient public material for full 11-file reports.

Cross-references discovered

  • FPF → quint-code: FPF (ailev) is the explicit theoretical foundation for quint-code/Haft. The internal/fpf/ Go package implements FPF's formal constructs. FPF defines WorkCommission conceptually; Haft implements it in SQLite. This is the only example in the corpus of a framework explicitly implementing another framework's theory.
  • spec-kit → openspec → lean-spec: The codervisor organization built spec-kit first, then openspec, then lean-spec as successive generations. lean-spec explicitly names spec-kit and openspec as predecessors ("Phase 1: Universal SDD Tools") in the README's "Why MetaSpec?" section — wait, that's MetaSpec's README. Both MetaSpec (ACNet-AI) and lean-spec (codervisor) explicitly acknowledge spec-kit as an ancestor.
  • MetaSpec → spec-kit + openspec: MetaSpec explicitly credits building spec-kit and OpenSpec first ("Phase 1: Universal SDD Tools") then discovering the meta-framework need.
  • mcp-server-spec-driven-development + kiro: Both use EARS format for requirements, making them the only two frameworks in the corpus that use EARS notation as spec format. The Kiro influence on EARS adoption is clear.
  • colign + kiro: colign's two-layer spec architecture (Project Memory + Structured Proposal) mirrors Kiro's steering/ + per-feature spec pattern, though no explicit acknowledgment is made.