tk-research

tk-harness

Claude Code plugin marketplace implementing a strict research-plan-stub-test-implement-eval (RPSTIE) workflow with Tier-2 nearly-pure functional discipline.

Status: M0 — skeleton. Plugin manifests, structure, conventions in place. No agents/skills/commands implemented yet.

Workflow Shape

Rough Idea
    |
    v
/start-design-plan          ---> Design Document (committed)
    | (clear context)
    v
/start-implementation-plan  ---> Phase files + worktree
    | (clear context)
    v
/execute-implementation-plan ---> per task: stub-author -> test-author -> body-implementor
                                  per phase: code-reviewer -> bug-fixer (3-strike cap)
                                  end: librarian -> code-reviewer -> test-analyst
    |
    v
/finish-branch               ---> merge / PR / discard

Verification: /verify runs the project's configured linters/typecheck/tests. Linter configs are project-owned (set up once per project, not bundled by the harness). No per-language verify skills.

Main thread runs at top-line model (Opus). Subagents default to Sonnet; reviewer escalates to Opus for subtle catches.

Plugins

Plugin	Purpose
tk-foundation	Generic subagents + base skills
tk-research	Codebase + internet research subagents
tk-house-style	Tier-2 nearly-pure FP discipline + per-language quick cards
tk-verify	Thin /verify runner + edit-validation hook
tk-rpie	RPSTIE workflow commands and subagents
tk-hooks	Security hardening, CLAUDE.md reminder, skill reinforcement

Subagent Roster

Stage	Agent	Model
Orchestrate	main thread	opus
Research	codebase-investigator, internet-researcher	sonnet
Clarify	clarifying-questioner	sonnet
Stubs	stub-author	sonnet
Tests	test-author	sonnet
Bodies	body-implementor	sonnet
Review (per phase)	code-reviewer	opus
Fix	bug-fixer	sonnet
Final tests	test-analyst	sonnet
Docs	librarian	sonnet

House Style: Tier-2 Nearly-Pure Functional

Default for projects using this harness. Project may override via .tk-harness/style.md.

• Result/Either for fallible operations; Option/Maybe for absence
• No exceptions in functional core
• Algebraic data types + exhaustive pattern matching
• Immutable data; persistent data structures for "updates"
• Smart constructors at boundaries (parse-don't-validate)
• Effects pushed to imperative shell

See plugins/tk-house-style/skills/nearly-pure-functional/SKILL.md (M3).

Installation (once published)

/plugin marketplace add https://github.com/TK-Evans01/tk-harness.git
/plugin install tk-rpie@tk-harness
/plugin install tk-house-style@tk-harness
# ... etc

Roadmap

• M0 Skeleton (done)
• M1 Foundation + research plugins (port from ed3d, adapt) — IN PROGRESS
• M2 tk-rpie core — agents + key skills + commands
• M3 tk-house-style — nearly-pure-functional skill + per-lang cards
• M4 tk-verify — thin /verify command + edit-validation hook
• M5 tk-hooks — port ed3d hook trio
• M6 Dogfood on a real project, tighten

Note: triage gate deferred. Every task currently routes through full RPSTIE.

RPSTIE vs tk-rpie. "RPSTIE" (Research-Plan-Stub-Test-Implement-Eval) is the workflow acronym, capturing all six stages including the explicit Stub and Test split. "tk-rpie" is the plugin name; we kept the shorter form for namespace ergonomics in skill / agent references (tk-rpie:writing-stubs-and-docs, etc.). The plugin name is not changing.

Attribution

Inspired by and partially derived from ed3dai/ed3d-plugins by Ed Ropple, which itself derives from obra/superpowers by Jesse Vincent.

Files derived from those projects retain their original MIT license (see per-plugin LICENSE files).

License

CC-BY-SA-4.0 for original content. MIT for files derived from obra/superpowers (marked per-file or per-plugin).