ADAW

Acceptance-Driven Agent Workflow.

ADAW gives coding agents a protocol surface centered on human acceptance criteria and evidence. The user sees whether the goal is satisfied, why, and what is still missing. The agent can still plan internally, but progress is determined by acceptance evidence.

Shape

.adaw/
  manifest.json
  protocol.md
  active/
    <goal>.acceptance.md
    <goal>.evidence.json
  completed/
  blocked/
  reports/
  brainstorms/

Quick Start

node ./bin/adaw.js install --root . --dry-run --json
node ./bin/adaw.js install --root . --skill --json
node ./bin/adaw.js doctor --root . --json
node ./bin/adaw.js skill export --pack --json
node ./bin/adaw.js brainstorm --idea "Explore a fuzzy goal before acceptance" --root . --json
node ./bin/adaw.js draft --from-brainstorm explore-a-fuzzy-goal-before-acceptance --candidate A --root . --json
node ./bin/adaw.js draft --goal "Ship a user-visible task" --root . --json
node ./bin/adaw.js approve --root . --summary "User approved acceptance criteria." --json
node ./bin/adaw.js init examples/adaw-self.json --root . --json
node ./bin/adaw.js resume --root . --json
node ./bin/adaw.js next --root . --json
node ./bin/adaw.js status --root . --json
node ./bin/adaw.js criterion update --root . --criterion AC-P-1 --user-story "作为用户，我能判断当前验收缺口。" --json
node ./bin/adaw.js evidence add --root . --criterion AC-P-1 --kind test-summary --summary "User can inspect the acceptance contract." --result passing --basis tool-observation --source '{"type":"command","label":"npm test","command":"npm test","outcome":"passed"}' --reviewability "User can rerun the command." --limitations "Does not prove visual UX." --json
node ./bin/adaw.js profile add --root . --type skill --name design-taste-frontend --strength must --purpose "Use this Skill for the design read and theme token pass." --install-policy existing_only --json
node ./bin/adaw.js profile add --root . --type stack --name radix-ui --strength prefer --purpose "Use accessible primitives for custom components." --install-policy ask_before_install --json
node ./bin/adaw.js profile show --root . --json
node ./bin/adaw.js profile evidence --root . --item skill-design-taste-frontend --result satisfied --summary "Agent used the required Skill before implementation." --json
node ./bin/adaw.js evaluate --root . --json
node ./bin/adaw.js report --root . --json

adaw install writes .adaw/manifest.json with the ADAW version, managed files, active goals, optional project Skill Pack state, and supported protocol capabilities. adaw doctor reports whether a project is ready, needs-action, or broken, with recovery actions for missing structure, stale manifest data, broken active goals, and stale Skills.

adaw install --dry-run returns an install_plan that shows each planned action, asset kind, managed status, write intent, destructive overwrite flag, and reason. Dry-run plans report will_write: 0 so users can review install impact before applying it.

Real adaw install --force requires --confirm. Preview destructive actions with adaw install --force --dry-run --json before applying a confirmed overwrite.

adaw uninstall --dry-run --json previews removals. Default uninstall removes entry assets while preserving active goals, evidence, reports, archives, and brainstorms; deleting .adaw state requires adaw uninstall --include-state --confirm --json.

Evidence is intentionally open-ended. Agents can use tests, diffs, screenshots, logs, artifacts, URLs, human confirmation, AW doctor, or any other useful signal. ADAW only requires the submitted evidence to be reviewable by the user: include a basis, one or more sources, reviewability, confidence, and limitations when the evidence affects completion.

adaw install --skill installs the ADAW Skill Pack under .agents/skills/: the root adaw router plus focused Skills for acceptance, evidence, capability profile, project health, and reporting. Users keep using natural language; the Skills route agent behavior to the right CLI loop.

Development

npm test
npm run check