Skip to content
/
Phase D Batch 20

Batch 20 — Plan/Context Dev-Experience Frameworks

Batch 20 — Plan/Context Dev-Experience Frameworks

Roster (10)

slug stars distribution cli_binary local_ui orchestration multi_model tier
aegis 60 standalone-repo no no sequential no A
aide 106 standalone-repo no no none no A
aegis-folder-trees unknown methodology-doc no no unknown unknown C
clean-coder-ai 576 standalone-repo (Python) no (python scripts) no hierarchical yes A
polka-codes 16 npm-package yes (polka) no sequential yes A
tsumiki 965 claude-plugin no no parallel-fan-out yes (haiku sub-agents) A
deliberate-agentic-dev 26 standalone-repo no no sequential no A
flow-framework-khgs unknown unknown unknown no unknown unknown C
superclaude 23009 cli-tool (Python) yes (superclaude) no hierarchical no A
trellis-mindfold 8467 npm-package yes (trellis) no sequential yes A

Intra-batch patterns

All 8 analyzable frameworks in this batch share a common commitment to structured human-AI collaboration through markdown-file scaffolding — even the more sophisticated frameworks (Clean-Coder-AI, Polka Codes, Trellis) ground their state in git-tracked markdown files rather than databases. The batch splits cleanly into two tiers: simple .context/ drop-in conventions (Aegis, AiDE) that are Archetype 4 clones, and full-featured frameworks (Tsumiki, Trellis, SuperClaude, Clean-Coder-AI) with genuine multi-agent orchestration, hook injection, and external state. External PM tool integration is a theme across multiple frameworks: Clean-Coder-AI uses Todoist, Deliberate Agentic Dev uses Linear — a pattern absent from the 11 seeds entirely. Sub-agent model cost optimization appears in two independent frameworks (Tsumiki uses haiku for Explore sub-agents; Trellis routes all sub-agents to minimax-m2.7), suggesting this is an emerging best practice.

Most interesting find

Trellis — the PreToolUse hook architecture that intercepts Task/Agent calls to inject curated JSONL spec files into sub-agent prompts is the most architecturally novel mechanism in this batch. No seed framework uses hook-based automatic spec injection at sub-agent dispatch boundaries; this is a genuinely new pattern for "specs injected, not remembered." At 8,467 stars with 14-platform support and active development, it is also the most production-ready multi-platform framework in the corpus.

Clean-Coder-AI earns a second mention: the combination of Todoist as external state store (human-visible PM UI), LangGraph hierarchical multi-agent, multi-tier LLM routing (high/medium/mini per agent role), and frontend screenshot vision analysis is an unusually complete autonomous development system not represented in the seeds.

Items written as Tier C

  1. aegis-folder-trees — Reddit community reference with no public GitHub repository; insufficient public material
  2. flow-framework-khgshttps://github.com/khgs2811/flow is not publicly accessible; repository not found

Cross-references discovered

  • Tsumiki explicitly borrows EARS notation from Kiro (seed) — the first confirmed cross-seed borrowing in this batch
  • Tsumiki is also architecturally inspired by BMAD-METHOD (seed) full-lifecycle design
  • SuperClaude ships SuperGemini_Framework and SuperQwen_Framework forks (same pattern applied to Gemini and Qwen models)
  • Aegis and AiDE are structurally near-identical (both .context/ drop-ins) — AiDE predates or is independent of Aegis but both implement the same Archetype 4 pattern
  • Trellis CLAUDE.md verbatim matches patterns from the CLAUDE.md in this research project (same behavioral guidelines) — suggesting the CLAUDE.md content may have been borrowed from a shared source