moai-workflow-research

name	moai-workflow-research
description	Self-research workflow for optimizing moai-adk components through binary eval experimentation loops. Adapted from autoresearch pattern with 5-layer safety architecture.
user-invocable	false
allowed-tools	Read, Write, Edit, Grep, Glob, Bash
metadata	{"version":"1.0.0","category":"workflow","status":"experimental","updated":"2026-04-09","tags":"research, eval, experiment, self-improvement, autoresearch"}
progressive_disclosure	{"enabled":true,"level1_tokens":100,"level2_tokens":5000}
triggers	{"keywords":["research","eval","experiment","optimize skill","improve agent"],"agents":["researcher"],"phases":["research"]}

Research Workflow

Purpose

Optimize moai-adk components (skills, agents, rules, config) through iterative binary eval experimentation.

Data Locations

Data	Location
Eval suites	.moai/research/evals/{type}/{name}.eval.yaml
Baselines	.moai/research/baselines/{name}.baseline.json
Experiments	.moai/research/experiments/{name}/exp-NNN.json
Changelogs	.moai/research/experiments/{name}/changelog.md
Observations	.moai/research/observations/
Dashboard	moai research status (CLI)

Eval Suite Schema

target:
  path: .claude/skills/moai-lang-go/SKILL.md
  type: skill
test_inputs:
  - name: scenario-name
    prompt: "Test prompt for evaluation"
evals:
  - name: criterion-name
    question: "Binary yes/no question?"
    pass: "What yes looks like"
    fail: "What triggers no"
    weight: must_pass  # or nice_to_have
settings:
  runs_per_experiment: 3
  max_experiments: 20
  pass_threshold: 0.80
  target_score: 0.95

Safety Layers

1. FrozenGuard: Constitution files cannot be modified
2. Worktree Sandbox: All experiments in isolated worktrees
3. Canary Regression: Proposed changes tested against baselines
4. Rate Limiter: Max experiments per session/week
5. Human Approval: Required before merging to main

Common Rationalizations

Rationalization	Reality
"The experiment results are obvious, I do not need a baseline"	Without a baseline, you cannot distinguish improvement from noise. Always measure before and after.
"This change is clearly better, I can skip the canary check"	Canary checks catch regressions the author cannot predict. Clear-to-author is not clear-to-system.
"I will merge the research branch now and validate later"	Unvalidated merges bypass every safety layer. Validate in the worktree before proposing a merge.
"The rate limiter is too restrictive for rapid experimentation"	Rate limits prevent runaway self-modification. If you need more experiments, batch them.
"I can modify the constitution to improve it"	Constitution files are FROZEN. Propose amendments through the human approval channel.

Red Flags

• Experiment branch merged without canary regression results
• Baseline measurement missing from the experiment report
• Multiple experiments run in the same session exceeding the rate limit
• Worktree not used for experimental changes (changes made in main workspace)
• Constitution or frozen-zone file modified during a research session

Verification

• Experiment conducted in an isolated worktree (show worktree path)
• Baseline measurement recorded before any changes
• Canary regression check passed (show before/after comparison)
• Rate limit not exceeded (check experiment count in session)
• Human approval obtained before merge proposal
• No frozen-zone files were modified (verify with git diff --name-only)