Skip to content
/
Phase D Batch 2

Phase D Batch 02 — Notes

10 frameworks in this batch

Phase D Batch 02 — Notes

Theme: Spec-driven (session-driven + multi-runtime SDD) — session continuity, multi-runtime adaptation, voice/NL input Frameworks analyzed: 10 Files produced: 110 (11 per framework) + this notes file


Framework Roster

# Slug GitHub Stars Language Archetype
1 trw-mcp wallter/trw-mcp 0 Python MCP-anchored toolserver (3)
2 sdd-session-ankushdixit ankushdixit/sdd 61 Python Skills-only behavioral (1)
3 gsd-for-antigravity toonight/get-shit-done-for-antigravity 862 PowerShell Mirror commands+skills (2)
4 vibe-architect mohdhd/vibe-architect 34 TypeScript Closed IDE-like + proprietary (5)
5 claude-bootstrap alinaqi/claude-bootstrap 670 Python Mirror commands+skills (2) — extreme
6 nl-spec-driven-elliotone ElliotOne/nl-spec-driven-ai-assisted-coding-agent-control 2 none Markdown scaffold zero primitives (4)
7 spec-gen clay-good/spec-gen 137 TypeScript MCP-anchored toolserver (3)
8 cogeet-ai-dev-specs cogeet-io/ai-development-specifications 32 none Markdown scaffold zero primitives (4)
9 speccrew charlesmu99/speccrew 3 JavaScript Skills-only behavioral (1) — multi-agent
10 vibedoc JasonRobertDestiny/VibeDoc 373 Python Markdown scaffold (4) + web UI

Cross-Cutting Observations

1. The Spec Generation / Spec Execution Divide

This batch splits cleanly into two camps:

Spec generators (produce documents, no execution):

  • vibe-architect (Next.js web app, 4-phase state machine)
  • cogeet-ai-dev-specs (6 .txt template files)
  • nl-spec-driven-elliotone (4 markdown files)
  • vibedoc (Gradio web app, NL → development plan)

Spec executors (take specs as input, produce running code):

  • trw-mcp (MCP server + CLI, 12 agents execute against AARE-F PRD)
  • sdd-session-ankushdixit / Solokit (session-driven, git history as spec)
  • gsd-for-antigravity (27 commands execute against SPEC.md)
  • claude-bootstrap / Maggy (67 skills execute against multi-tier context)
  • spec-gen / OpenLore (MCP tools extract specs from existing code + generate)
  • speccrew (8 agents execute PRD through testing pipeline)

This divide matters: spec generators solve the "I don't know how to write a good spec" problem; spec executors solve the "I have a spec, now automate the work" problem. Teams need both — but very few frameworks bridge them.

2. The Web UI Emergence

Two frameworks in this batch (vibe-architect, vibedoc) are browser-based web applications — anomalous for a field dominated by CLI tools and IDE integrations. Both target the "spec generation" phase. The pattern suggests a market belief that spec authoring benefits from richer UX (live previews, voice input, multi-format export) than terminal text entry allows.

3. ISA-95 as AI Methodology

SpecCrew's use of ISA-95 (enterprise integration standard) as its intellectual framework is unique in the entire corpus. No other seed or batch framework borrows from industrial automation standards. This represents a plausible but unproven hypothesis: that manufacturing-derived process standards have natural analogs in software delivery pipeline design.

4. Session Continuity Strategies

The batch theme explicitly includes "session continuity." Three distinct strategies emerged:

Strategy Framework Mechanism
Git-as-memory sdd-session-ankushdixit (Solokit) sk start reads last N commits as briefing
Iteration directories speccrew iterations/iXXX/ preserves full artifact lineage
Knowledge tiers speccrew L1/L2/L3 tiers accumulate cross-session context
Context-as-files cogeet-ai-dev-specs Specification files manually maintained
None (export-only) vibedoc User exports plan to file; no session continuity

The git-as-memory approach (Solokit) is the most elegant: the version control system becomes the session memory, requiring zero additional infrastructure.

5. Multi-Runtime Adaptation

The batch theme includes "multi-runtime adaptation." Results:

Framework Cross-tool support Mechanism
gsd-for-antigravity Yes adapters/CLAUDE.md, GEMINI.md, GPT_OSS.md
cogeet-ai-dev-specs Yes (high) Plain .txt files work with any LLM
nl-spec-driven-elliotone Yes (high) Plain markdown, no tool dependencies
trw-mcp Low MCP + Python CLI, Claude Code primary
claude-bootstrap Low Claude Code-specific hooks/skills
speccrew Low Claude Code agents/skills format
vibedoc High Generates docs; any tool can consume them

gsd-for-antigravity is the standout: explicit adapter files per LLM provider are a unique engineering choice that no seed or other batch framework implements.

6. Voice/NL Input (Batch Theme)

The batch theme included "voice/NL input." Only vibe-architect implements this: OpenAI Whisper voice input converts speech to text for spec description. All other frameworks accept text-only input. vibedoc accepts NL description but text-only. The voice input dimension is underrepresented — a genuine gap in the spec-driven development space.


Notable Technical Patterns

Pattern: JOURNAL.md as Append-Only Audit Log (gsd-for-antigravity)

The JOURNAL.md file — appended to with every agent action — creates a simple, version-control-friendly audit trail. This pattern (append-only markdown log) is more maintainable than structured logs and easier to read than JSON. No other framework in this batch uses it.

Pattern: Epistemic Lease / Decay Model (spec-gen / OpenLore)

OpenLore's "epistemic lease" model — where cross-module trajectory knowledge decays over time and must be renewed — is the most mathematically sophisticated context management approach in the batch. The decay model treats code knowledge like a lease that expires, with 4 escalation levels from advisory to full re-analysis. This is conceptually close to how human experts treat stale knowledge but operationalized for automated systems.

Pattern: 13-Tier Cost-Based Model Routing (claude-bootstrap)

Claude-bootstrap's model routing (T0 local free → T4 DeepSeek 80% of work → T12 Opus $25/M) is the most elaborate multi-model configuration in the entire corpus (both seeds and all batches). The "Telos IFS = F1×F2×F3" quality measurement approach and the iCPG (Intent-Augmented Code Property Graph) are novel engineering constructs with no analog in the seeds.

Pattern: "Each Requirement = Independent Project" (speccrew)

SpecCrew's iterations/iXXX-[name]/ pattern treats each requirement as a fully isolated project scope with its own artifact tree. This is isomorphic to git worktrees for requirements — preventing cross-requirement context contamination while maintaining shared technology/business knowledge in L1/L2 tiers.


Seed Comparison Summary

Seed Most similar batch framework(s) Key difference
spec-kit spec-gen/OpenLore, cogeet-ai-dev-specs spec-kit manages spec files; batch frameworks generate or execute them
openspec spec-gen/OpenLore openspec is API-focused; OpenLore reverse-engineers all spec types from code
spec-driver speccrew, trw-mcp spec-driver enforces via tests; batch frameworks execute via agent pipelines
bmad-method speccrew, trw-mcp bmad is Markdown; both batch frameworks add CLI + structured knowledge systems
agent-os cogeet-ai-dev-specs, gsd-for-antigravity agent-os ships commands to write the markdown; batch frameworks ship the markdown
claude-flow claude-bootstrap claude-flow is multi-agent orchestration; bootstrap adds multi-model routing + memory
taskmaster-ai speccrew, sdd-session-ankushdixit taskmaster breaks tasks; speccrew isolates per-requirement; Solokit uses git history
kiro vibe-architect both generate specs in IDE/UI; kiro integrates deeper; vibe-architect is standalone
ccmemory sdd-session-ankushdixit (Solokit) ccmemory is vector/KV memory; Solokit uses git history as memory
claude-conductor nl-spec-driven-elliotone, cogeet-ai-dev-specs all are pure markdown scaffolds with no tooling
superpowers claude-bootstrap both are large skills collections; bootstrap adds model routing + memory systems

Anomalies and Edge Cases

  1. spec-gen renamed to OpenLore — the npm package openlore is the same codebase as spec-gen, which was renamed. The GitHub repo is still clay-good/spec-gen. The openlore CLI name is what appears in npm and in the binary.

  2. ankushdixit/sdd product is named "Solokit" — the repo slug is sdd and the GitHub description mentions "spec-driven development," but the PyPI package, CLI binary, and product branding are all "Solokit." Documented under slug sdd-session-ankushdixit per assignment.

  3. vibedoc AGENTS.md is a developer guide, not an agent definition — unlike every other framework in the batch, VibeDoc's AGENTS.md defines module organization and coding style for contributors to VibeDoc itself, not agent behavior specifications.

  4. trw-mcp has 0 GitHub stars — this is the lowest star count in the batch and appears to be a private or unreleased project. The BSL 1.1 license (non-OSS) confirms it's not a community-first release. Analyzed based on available public content.

  5. enhanced_mcp_client.py in vibedoc — the presence of an MCP client module in an otherwise simple web app is an architectural signal: the author intends to evolve VibeDoc from a standalone spec generator into an MCP participant. Worth watching.


Files Produced

All 110 framework files confirmed written:

  • /spec-system/trw-mcp/ — 11 files ✓
  • /spec-system/sdd-session-ankushdixit/ — 11 files ✓
  • /spec-system/gsd-for-antigravity/ — 11 files ✓
  • /spec-system/vibe-architect/ — 11 files ✓
  • /spec-system/claude-bootstrap/ — 11 files ✓
  • /spec-system/nl-spec-driven-elliotone/ — 11 files ✓
  • /spec-system/spec-gen/ — 11 files ✓
  • /spec-system/cogeet-ai-dev-specs/ — 11 files ✓
  • /spec-system/speccrew/ — 11 files ✓
  • /spec-system/vibedoc/ — 11 files ✓