| name | verify-on-behalf-of-user |
| description | Run human-style scenario verification after an agent claims work is done: git status, build into ~/.sandbox/bin, sandbox HOME=~/.sandbox/default-home, real integration/E2E scenarios (start servers, browser-agent for UI), always-labeled depth, markdown transcript written and inlined for review. Use for "verify on my behalf", before commit. Complements doctests; does not replace them. Load a sub-topic with: agent-pro skill --show verify-on-behalf-of-user/<topic> |
Verify on behalf of user
After an agent claims work is done and automated tests passed, run the same steps
a human would before committing — real user journeys, not only smoke CLI checks.
Write a markdown transcript and print its full contents in the reply so the
user can review without opening the file.
This is outside doctests — it mimics one-time manual verification.
When to use
- Agent finished a feature/fix and reported GREEN doctests
- User would otherwise start servers, exercise the product, and open the UI
- Before user checks in or commits
When NOT to use
- Deep adversarial code review → use project-specific review skills
- Automated regression → use
doctest test (headless tools like playwright-debug belong there)
- User said
no verify: or only wants a diff review
Topics
sandbox — ~/.sandbox paths, enter-sandbox.sh, forbidden writes, strict mode
workflow — phases: scope → git → build → runtime → scenarios → evidence → transcript
scenario — depth labels, surface gate, browser-agent UI rules, FAIL when UI path broken
transcript — format rules, template, write file and inline full content
Usage
agent-pro skill verify-on-behalf-of-user --show
agent-pro skill --show verify-on-behalf-of-user sandbox
agent-pro skill verify-on-behalf-of-user workflow --show
agent-pro skill --show verify-on-behalf-of-user/scenario
agent-pro skill --show verify-on-behalf-of-user/transcript
Utility files install with the skill (not in topic bodies):
scripts/enter-sandbox.sh, scripts/sandbox-verify.sb
templates/transcript.md
Core rules (non-negotiable)
- Always label depth — every transcript and final reply must include
Depth: smoke | scenario | full plus a one-line reason. Never leave depth unlabeled.
- Scenario-first — default is real integration (bring-up + user journey). Smoke is a
labeled downgrade when justified (e.g. pure CLI with no runtime surface).
- UI uses browser-agent only — for any UI/frontend surface, drive a live browser via
browser-agent. Do not use playwright-debug or other headless runners here;
those are for doctests/CI.
- Missing UI path = FAIL — if the surface requires UI and browser-agent/Chrome/session
is unusable, verdict is FAIL (not skip, not BLOCKED).
- Transcript = file + inline — write
~/.sandbox/transcripts/<timestamp>-<slug>.md,
then put the full file body in the agent reply for direct review.
Relationship to other skills
| Skill | Role |
|---|
| doctest-tdd | Automated sealed tests (may use headless browsers) |
| verify-with-prototype | POC before risky host integration |
| verify-on-behalf-of-user | Human-shaped sandbox scenarios + transcript (file + inline) |
Suggested order: doctests GREEN → this skill → user commits.
Project recipes
Consumer repos add docs/verify-recipes/<feature>.md with surface, depth, bring-up,
scenarios, teardown, and expected observables. See references/README.md.