Skip to content

docs(skill): add the first automation eval — ADR-0044 revise loop#2300

Merged
os-zhuang merged 1 commit into
mainfrom
docs/automation-eval-revise-loop
Jun 24, 2026
Merged

docs(skill): add the first automation eval — ADR-0044 revise loop#2300
os-zhuang merged 1 commit into
mainfrom
docs/automation-eval-revise-loop

Conversation

@os-zhuang

Copy link
Copy Markdown
Contributor

What & why

Part of #2274 (ADR-0044 AI-authoring awareness). Adds the first concrete eval under the skills' evals/ directories — until now every skills/*/evals/ held only a placeholder README documenting the intended format.

skills/objectstack-automation/evals/approvals/test-revise-loop.md follows that documented format:

  • Scenario — author an approval flow with send-back-for-revision (cap at 2 send-backs).
  • Expected Output — the full ADR-0044 shape: revise branch → signal wait node → resubmit edge typed type: 'back' + maxRevisions (mirrors showcase_budget_approval).
  • Common Mistakes — the three footguns, each mapped to what catches it (the lint-flow-patterns rules from feat(cli): lint ADR-0044 approval revise-loop footguns at compile time #2279, registerFlow, design review).
  • Validation CriteriaregisterFlow-registers + revise branch + back-edge + wait window + maxRevisions >= 1 + no lint findings.

Closes the authorable half of #2274's eval slice: it pins, as a checkable spec, the behaviour the skill (#2294) teaches and the lint (#2279) enforces.

Scope (important)

This is the eval spec (markdown), not an executable runner. The skills' eval harness does not exist yet — every evals/README.md says "Not yet implemented." Building a runner that feeds prompts through the agent, registers the output, and scores pass-rate across runs is a separate platform project (it benefits all skills and involves non-deterministic-LLM CI concerns). Tracked as the remaining (b) item on #2274.

Verification

  • check:doc-authoring ✓ (164 files clean) · check:skill-docs in-sync ✓ · markdown-only, no changeset (matches prior skill-doc PRs).

Refs #2274, #1770. 🤖 Generated with Claude Code

Adds skills/objectstack-automation/evals/approvals/test-revise-loop.md, the first
concrete eval under the skills' (previously placeholder) evals/ format — Scenario
/ Expected Output / Common Mistakes / Validation Criteria. It pins the expected
send-back-for-revision shape (revise branch + signal wait + type:'back' resubmit +
maxRevisions), the three footguns the compile-time lint flags, and registerFlow-
based pass criteria.

This is the authorable (a) half of #2274's eval slice. The executable eval RUNNER
(b) remains a separate platform project — the skills' evals/ harness is still
"not yet implemented" per each dir's README.

Refs #2274, #1770 (ADR-0044). Builds on the lint (#2279) and skill (#2294).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@vercel

vercel Bot commented Jun 24, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
spec Ready Ready Preview, Comment Jun 24, 2026 4:36pm

Request Review

@github-actions github-actions Bot added documentation Improvements or additions to documentation size/s labels Jun 24, 2026
@os-zhuang
os-zhuang merged commit a9363b4 into main Jun 24, 2026
14 of 15 checks passed
@os-zhuang
os-zhuang deleted the docs/automation-eval-revise-loop branch June 24, 2026 16:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation size/s

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant