Instructions file CodexOpenCode
Instructions for dentiny/kon, covering kon contributor instructions for ai agents, hard rules (apply to every agent, every command), core principles (design → impl → review), tone and agent & narrator emoji — quick-ref (always in context).
Plugin Claude Code
K-On! multi-agent dev workflow — explore, plan, implement, cleanup, review.
Command
Fetch review comments on the current branch's PR, triage with the user, route each item to /kon:quick or /kon:team, implement one work item at a time. Blocks if no PR.
Command
Read-only Q&A about the codebase. Azusa investigates and answers — zero repo writes (no code, no .kon/ artifacts); session JSON is tracked under /.kon/projects/.
Command
Start an interactive kon session. Plain chat is routed by intent — no /kon: prefix needed until /kon:finish.
Command
Bug investigation pipeline — reproduce with runtime evidence before fixing. Azusa investigates, Mugi proposes multiple fixes, user approves, then Yui implements. Mio reviews.
Command
Summarize a GitHub issue and all discussion comments. Jun only, read-only. Writes sessions/ /issue-summary.md.
Command
Design-only pipeline — Azusa explores, Mugi plans, then multi-agent debate (Azusa challenges, Mugi revises) before user confirms. No implementation.
Command
Mark the current session as completed. Called explicitly by the user when they are satisfied with the work done. Equivalent to clicking ✓ in the dashboard.
Command
Garbage-collect the codebase — Sawako identifies dead code, redundant comments, and bloated docs, presents an inventory for user confirmation, then cleans up. Mio reviews.
Command
Read-only bug hunt — Azusa analyzes bugs from source code and suggests best-effort repro SQL/tests. No fixes, no repo writes.
Command
Lightweight task entry point. Orchestrator calls Yui directly for small changes, skips Azusa / Mugi, Yui finishes and runs lightweight Mio (3-item checklist subset).
Command
External lookup — Jun searches docs and the web, writes .kon/research.md. Read-only for source code; session tracked.
Command
Session retro — propose preferences and conventions from the current session for public or repo memory. Human confirms each save. Orchestrator only.
Command
Holistic GitHub PR review — code diff, PR description, existing review comments, and linked issues. Mio only, read-only.
Command
Code review only — Mio runs strict-review on uncommitted or staged diff. No implementation.
Command
Write a clean session summary. Called automatically at the end of every run, or on-demand to summarize a past session.
Command
Full kon run — Azusa explores, Mugi writes a plan, Yui implements, Sawako cleans, Mio reviews. No automated testing.
Command
Project todo list — add tasks to .kon/todos.json; view and manage in the dashboard.
Command
Learn a codebase — Azusa maps it, Jun writes an interactive HTML guide (concepts + architecture) and study pack (flashcards + quiz).
Hook
Runs when a session starts, executing ensure_project_dir.py and start_dashboard.py via python3 (2 commands). From dentiny/kon.
Hook
Runs when you submit a prompt, before the agent sees it, executing init_kon_session.py, record_orchestrator_model.py and log_begin_prompt.py via python3 (3 commands). From dentiny/kon.
Hook
Runs before the agent uses a tool for Bash tool calls, executing no_git_write.py via python3. From dentiny/kon.
Hook
Runs when the agent finishes a response, executing log_begin_response.py via python3. From dentiny/kon.