Kimchi

A Claude Code plugin that transforms vague ideas into validated, standalone task specifications (beads) for multi-agent execution.

Kimchi is prompt-native — all logic lives in markdown files. No TypeScript, no build step. Skills, agents, and personas work together through a 10-stage planning pipeline that produces self-contained task specs any agent can pick up and execute independently.

Install

claude plugin add https://github.com/Tromml/kimchi

Quick Start

/kimchi:plan "Add user authentication with magic links"

This runs the full pipeline through refinement: clarify, requirements, research, generate, review, refine. The output is a draft plan in .kimchi/PLAN-DRAFT.md.

For fully automated end-to-end (plan through bead generation):

/kimchi:plan "Add user authentication with magic links" --full-auto

How It Works

Kimchi breaks planning into discrete stages, each producing a markdown artifact:

 idea
  |
  v
clarify        Extract understanding through structured questions    -> CONTEXT.md
  |
requirements   Categorize into v1 (must), v2 (next), out of scope   -> REQUIREMENTS.md
  |
research       Investigate codebase patterns and frameworks          -> RESEARCH.md
  |
generate       Create plan with tasks, deps, complexity estimates    -> PLAN.md
  |
review         5 personas critique for scope creep, complexity, etc  -> PLAN-REVIEWED.md
  |
refine         Iterate until quality threshold or diminishing returns -> PLAN-DRAFT.md
  |
plan-revise    Fresh-eyes cross-model review (run per model)         -> PLAN-REVISED-{model}.md
  |
plan-synthesize Merge cross-model revisions into hybrid plan         -> PLAN-SYNTHESIZED.md
  |
beads          Convert to standalone YAML task specs                  -> .beads/
  |
validate       4 validators check standalone executability            -> enriched beads

Each stage can be run independently. The pipeline writes everything to .kimchi/ in your project root.

Halmoni (Self-Improvement)

"Taste, adjust, remember. Each batch teaches the next."

Halmoni is Kimchi's self-improvement system. Like a Korean grandmother who tastes her kimchi and adjusts the recipe each time, Halmoni observes what happened during execution, proposes improvements to skills and validators, validates them against history, and applies them with versioning.

/kimchi:halmoni --from-bead 008
/kimchi:halmoni --from-feedback "the review stage missed the N+1 query"
/kimchi:halmoni --taste-only

Pipeline Skills

Skill	Description
kimchi:plan	Full pipeline orchestrator (clarify through refine)
kimchi:clarify	Extract understanding through structured questions
kimchi:requirements	Categorize requirements into v1, v2, out of scope
kimchi:research	Investigate codebase patterns and frameworks
kimchi:generate	Generate plan with tasks, deps, complexity
kimchi:review	Multi-persona plan critique
kimchi:refine	Iterative improvement until quality threshold
kimchi:plan-revise	Cross-model fresh-eyes review
kimchi:plan-synthesize	Synthesize cross-model revisions
kimchi:beads	Convert plan to standalone bead YAML specs
kimchi:validate	Validate beads for standalone executability

Execution Discipline Skills

Skill	Description
kimchi:tdd	RED-GREEN-REFACTOR cycle enforcement
kimchi:systematic-debugging	4-phase root cause analysis before proposing fixes
kimchi:simplicity-enforcement	YAGNI, minimal code, duplication over wrong abstraction
kimchi:verification-before-completion	Evidence before assertions, always
kimchi:bead-protocol	Multi-agent bead execution coordination

Agents

Pipeline

Agent	Description
kimchi:pipeline:clarifier	Generates structured clarification questions
kimchi:pipeline:researcher	Discovers codebase patterns and conventions
kimchi:pipeline:bead-converter	Converts plan tasks into standalone bead YAML

Review Personas

Agent	Description
kimchi:review:complexity-detector	Identifies unnecessary complexity and over-engineering
kimchi:review:feature-trimmer	Identifies non-essential features to cut or defer
kimchi:review:scope-guardian	Catches scope creep and vague acceptance criteria
kimchi:review:test-coverage-advocate	Identifies missing or insufficient test specs
kimchi:review:premature-optimization-detector	Identifies optimizations not yet needed

Utility Skills

Skill	Description
kimchi:halmoni	Self-improvement system
kimchi:status	Check pipeline state and artifact status
kimchi:reset	Clear .kimchi/ and start fresh
kimchi:convert	Convert plugin to OpenCode/Codex formats

Format Conversion

Kimchi can convert itself into formats compatible with other AI coding tools:

/kimchi:convert --to opencode
/kimchi:convert --to codex
/kimchi:convert --to opencode --also codex

Requires bun runtime. Output goes to .converted/ by default.

Project Structure

.claude-plugin/marketplace.json
plugins/kimchi/
  .claude-plugin/plugin.json
  skills/           21 skills (pipeline + execution discipline + utility)
  agents/
    pipeline/       3 pipeline agents
    review/         5 review persona agents
  personas/         5 persona definitions
  validators/       4 bead validators
  lib/              4 reference libraries
  templates/        6 output templates
  hooks/            ACFS integration hooks

License

MIT