Skip to content

comses/odder

Repository files navigation

odder

Agent skills that generate complete ODD+2 protocol documents from agent-based model code and documentation. 4-phase, human-in-the-loop workflow designed to preserve author's intent.

Compatible with any agent that supports the Agent Skills standard: Claude Code, GitHub Copilot, Cursor, Gemini CLI, OpenCode, VS Code, OpenAI Codex, and many more.

Quick Start

# Install to your project:
npx skills add comses/odder

# Install to specific agents:
npx skills add comses/odder -a claude-code -a codex

# Install globally:
npx skills add comses/odder -g

Requires Node.js (npx runs the skills CLI automatically).

Run

/odd-interview .   # Phase 1: research + interview
/odd-plan          # Phase 2: generation plan
/odd-draft         # Phase 3: ODD document
/odd-check         # Phase 4: verification
/odd-feedback      # Optional: tell us how it went

Review artifacts in odder/ between each phase.

Skills

For model developers

Phase Skill What happens Output
Research /odd-interview Agent reviews files, interviews modeler odder/research/findings.md
Plan /odd-plan Creates standalone generation instructions odder/plan/odd-generation-plan.md
Draft /odd-draft Fresh agent executes the plan odder/draft/odd.md + odder/draft/traceability-matrix.md
Verify /odd-check Independent agent checks the ODD odder/checked/verification-report.md
Feedback /odd-feedback Interactive questionnaire (~2 min) odder/feedback/{datetime}.md

Each phase builds on the previous one's artifacts, and human review between phases catches problems early.

Design Decisions

Source-based confidence

Every claim in the generated ODD is annotated:

Category Meaning
CODE_VERIFIED Verified by reading code at specific file:line
DOC_STATED Explicitly stated in documentation
MODELER_CONFIRMED Confirmed during interview
INFERRED Reasonably inferred, inference chain documented
UNVERIFIABLE Cannot be verified, reason stated

Dual traceability

Inline citations in the ODD text ([source: model.py:42]) AND a companion traceability matrix mapping every claim to its source.

Reimplementability as quality bar

"Could someone rebuild this model from this ODD alone?" This catches vague prose, missing parameters, and implicit assumptions.

Adaptive interview

Input quality Strategy
Code + comprehensive docs Gap-driven — only ask about what's unclear
Code + minimal docs Conceptual-first — establish purpose, then verify against code
Docs only Deep interrogation — probe for precise process logic
Code only Reverse-engineer — present understanding for confirmation

Independent verification

The verification agent does NOT see the generation plan — only the ODD and original sources. This ensures verification against sources, not just plan compliance.


For SKILL developers

Skill What happens Output
/odd-integrate-feedback Aggregates feedback, interviews developer, generates change plan odder/suggested-changes/{datetime}.md

The improvement loop: scientists run /odd-feedback → developer runs /odd-integrate-feedback → implement changes → repeat.

You can also edit the SKILL.md files directly — they're plain markdown.

Creating Your Own SKILL Pack

These skills were built through an agent-driven interview process. You give the agent reference materials and a goal, then it interviews you until it has enough context to design the full workflow.

  1. Create a project and a .devcontext/ directory

  2. Place reference materials into .devcontext/ (domain docs, protocol specs, etc.)

  3. Research your agent's capabilities and save those docs to .devcontext/ too — this is the step most people skip

  4. Launch your agent:

    Read all files in .devcontext/
    
    I want to design [your process]. Conduct an interview with me
    until you fully understand how to achieve this task.
    

After 20-30 minutes of Q&A, the agent generates a full specification, multi-phase workflow, verification procedure, and file structure. The interview forces deep context before generation — without it, you get generic templates.

Tips

  • Research agent capabilities BEFORE the interview. We designed a solution then discovered the plugin system worked differently than assumed.
  • Commit your interview transcript. It documents why your process looks the way it does.
  • Start with a known model (Schelling, Wolf-Sheep) before tackling your research model.
  • Run /odd-feedback after every generation. It takes 2 minutes and drives improvements.

Development

This repo uses qlty for code quality. Formatting and linting run automatically on commit via git hooks.

# Set up git hooks (auto-formats on commit):
qlty githooks install

# Manual check:
qlty check --all

# Manual format:
qlty fmt --all

Plugins: prettier and markdownlint (formatting), trufflehog (secret scanning).

Agents Tested

Agent Notes
Claude Code (Opus) Strong code analysis, good at structured interviews
GitHub Copilot (VS Code) Picks up skills automatically
Gemini CLI Works with explicit prompting
OpenCode Full skill support
OpenAI Codex Works well
pi.dev Full skill support via .agents/skills/
More coming PRs welcome with experience reports

Related Links

About

Agent skills for generating ODD+2 protocol documents from agent-based model code through a multi-phase, human-in-the-loop workflow

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages