| name | agent-e2e-adoption-smoke |
| description | Run an independent adoption smoke test for Agent E2E Harness by spawning or directing a fresh agent in a clean consumer app, giving it only the app idea plus the agent-e2e-harness repo/skill pointer, then observing whether it can install the skill, build a journey, run agent-e2e verify, prove Dev MCP with tool calls, and report adoption learnings. |
Agent E2E Adoption Smoke
Use this skill when the user wants to dogfood how easy Agent E2E Harness is to adopt from a fresh app, validate docs/skill quality, or learn where a real agent gets stuck.
The goal is not to help the worker succeed by leaking internal knowledge. The goal is to run a realistic adoption tracer bullet and collect evidence.
Core Rule
The worker should think it is doing ordinary implementation work. Do not tell it it is being evaluated. Give it a normal product brief, a technology stack, and only the public repo/skill pointer.
Workflow
-
Prepare a clean consumer repo
- Prefer
/tmp/<slug> for throwaway tests.
- Initialize git.
- Do not copy files from the official harness repo.
- Use a short AOE title; long titles can be truncated in tmux targets.
-
Spawn or direct a persistent worker
- Use the
aoe skill for AOE sessions.
- Prefer
codex when testing Codex adoption.
- Use
--yolo unless the user asks otherwise.
- Decide sandbox with the user when needed; for browser/npm-heavy local smoke tests, no sandbox is often cleaner.
- Complete AOE readiness before sending the brief.
-
Send a neutral adoption brief
- Use
references/brief-template.md.
- Include app idea, stack, repo pointer, skill install requirement, and outcome criteria.
- Do not include known workarounds unless the first attempt gets blocked.
-
Observe without contaminating
- Use
references/observer-protocol.md.
- Poll capture and filesystem state.
- Do not interrupt while it is making observable progress.
- If it stalls, record the stall before intervening.
-
Require real proof
agent-e2e dev startup is not enough.
- The worker must either use a configured MCP client or
mcporter to call tools.
- Use
references/mcp-proof-checklist.md as the minimum proof loop.
-
Collect learnings
- Ask for the callback in
references/learning-callback.md.
- Capture commands, artifacts, confusing docs, package/API gaps, and recommended fixes.
- Keep the report grounded in what the worker actually tried.
Success Criteria
The smoke is successful only when all are true:
- app runs locally;
agent-e2e.config.ts exists;
npm run e2e:verify or equivalent passes;
- Dev MCP tools are called successfully, not merely started;
- artifact paths are reported;
- cleanup restores seeded state or reports a concrete gap;
- the worker reports adoption blockers and documentation improvements.
Stop Conditions
Stop when the worker provides a complete learning report, asks for human input, hard-stalls without filesystem/capture progress, or leaves unsafe processes running. If servers are still listening, ask the worker to stop them or stop them yourself before closing.