菜单
Manual end-to-end walkthrough — drive the live app first-hand (CLI, API, web UI) and query the datastore directly. Trigger on "e2e", "smoke test", "manually test".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Build, launch, and drive the mindwalk web UI end-to-end for verification.
Set up or troubleshoot Yassimba's curated agent skills, Pi packages, Herdr, and Herdr plugins through the ai-setup CLI. Use when the user asks to install this collection, configure Herdr, add one of its capabilities, update the setup, or diagnose installation problems.
Backlog management: use when the user mentions a backlog, asks what's next, wants work recorded before implementation, or wants queued ideas or specs prioritized, transitioned, completed, or removed; also use when another skill needs to record lifecycle changes.
Use when the user wants to brainstorm or explore an idea — a feature, product direction, or "what if" — before deciding whether it deserves a plan. Ideation only — ends in an idea brief, not a design.
Review the changes since a fixed point (commit, branch, tag, or merge-base) along two axes — Standards (this repo's documented coding standards) and Spec (what the originating issue/PRD asked for). Use when the user wants to review a branch, a PR, work-in-progress changes, or asks to "review since X".
Use when creating a git commit — the user asks to commit, or a unit of work is complete and ready to commit.
| name | e2e-test |
| description | Manual end-to-end walkthrough — drive the live app first-hand (CLI, API, web UI) and query the datastore directly. Trigger on "e2e", "smoke test", "manually test". |
Every "it works" claim in this walkthrough rests on first-hand evidence: an artifact on disk — ai-docs/e2e-output/**.txt, ai-docs/e2e-screenshots/**.png, a datastore query result — produced by you driving the live app, and read back by you afterward. A passing test suite is hearsay: a synthetic harness (TestClient, fixtures, fakes, mocked browsers) vouching for the app from inside. Hearsay is inadmissible — it misses what breaks in production: startup ordering, real network paths, real datastore constraints, real auth, real rendering. A step whose only support is a green test run is an untested step; go drive it.
The deliverables are the evidence trail and the report — plus, when you catch a bug, one regression unit test in the existing suite (see "When you find a real bug").
First-hand evidence per surface:
Bash runs the binary; read stdout, stderr, exit codecurl / httpx against the running server; read status + bodyagent-browser clicks / fills / screenshots; Read each screenshotDiscover what this project actually exposes:
[project.scripts], package.json bin, Cargo.toml [[bin]], go.mod + cmd/, user-facing Makefile / justfile targets, scripts in bin/net/http / gRPC …)vite.config.*, next.config.*, index.html, a package.json with dev/start) or static HTML mounted by a backend.env.example (never .env), config files, docker-compose.yml; note the client you'll query each with (psql, mysql, sqlite3, mongosh, redis-cli, aws s3 ls, ls, kafkacat, …)Write the inventory to ai-docs/e2e-output/00-inventory.md — every later phase is scoped to it. If nothing user-facing exists, stop: "This project has no user-facing interfaces — manual E2E needs something to drive." If a web UI exists, check agent-browser availability per references/browser-recipes.md.
Launch THREE sub-agents in parallel via the Agent tool in ONE message — app structure & journeys, datastore schema & data flows, bug hunt. Prompts in references/research-prompts.md; fill their placeholders from the pre-flight inventory. Wait for all three — their output drives every later phase.
Install deps and start each service with the project's own tooling (commands came from research), then capture the first evidence:
<runner> <start-command> &
until curl -sf http://localhost:<port>/<healthz-or-root>; do sleep 0.5; done
<cli-invocation> --help > ai-docs/e2e-output/00-help.txt 2>&1 # if a CLI exists
agent-browser open http://localhost:<port>
agent-browser screenshot ai-docs/e2e-screenshots/00-initial-load.png # if a web UI exists
TaskCreate one task per journey from sub-agent 1, only for surfaces that exist: [CLI] Test <journey>, [API] Test <journey>, [Web] Test <journey>, [Cross] <journey> (data flowing between two surfaces).
Then add the edge-case sweeps — full journeys in their own right, and where most real bugs live. Each recipe file defines its sweep; the task just names it:
[CLI] Error-handling sweep — references/cli-recipes.md[API] Auth & validation sweep — references/api-recipes.md[Web] Responsive sweep — references/browser-recipes.md[Cross] Data consistency across interfaces — if more than one surfaceThen add the probes — one [Probe] <suspected bug> task per high-priority finding from sub-agent 3's hunt. A finding from static analysis is hearsay until you drive the suspect path in the live app and it reproduces (or fails to).
The walkthrough is complete when every task — journeys, sweeps, AND probes — is completed.
Per task: TaskUpdate → in_progress, then drive it yourself. Recipes for the surfaces this project has: references/cli-recipes.md, references/api-recipes.md, references/browser-recipes.md, references/db-validation.md.
Every step has the same shape:
ai-docs/e2e-output/<journey>/<step>-*.txt, ai-docs/e2e-screenshots/<journey>/<NN>-*.png.Read the screenshot.references/db-validation.md).Tag everything you create with an e2e- marker (in names, emails, titles, keys, filenames) — Phase 5 finds and removes test data by that marker.
A journey closes only on admissible evidence: every step has an artifact on disk that you read back, and the datastore confirms every expected side effect. Then TaskUpdate → completed.
Real = a defect in the app, not a flaky harness or a missing dep.
tdd skill handles the watched-RED.diagnosing-bugs → GREEN.Kill the processes you started and agent-browser close. Then remove the test data: query the datastore for the e2e- marker, delete what it finds, and re-query — cleanup is done when the marker query returns nothing. Keep ai-docs/e2e-output/ and ai-docs/e2e-screenshots/ — that's the evidence trail.
Always emit a text summary in the final message:
## Manual E2E Walkthrough Complete
**Interfaces driven:** <e.g. "CLI, API (http://localhost:8000), Web UI (http://localhost:3000)">
**Journeys walked:** <count> (<N> CLI, <N> API, <N> Web, <N> cross — including <N> edge-case sweeps and <N> probes)
**Datastore validation queries run:** <count>
**Evidence files captured:** <count> outputs, <count> screenshots
**Issues found:** <count> (<fixed> fixed, <remaining> remaining)
### Issues fixed during the walkthrough
- [Interface] <description> — `<file:line>` — regression test at `<path>`
### Remaining issues
- [Interface] <description> — <severity> — `<file:line>`
### Probe results
- <sub-agent 3 finding> — confirmed (see issues above) / not reproduced — `<file:line>`
### Artifacts
- `ai-docs/e2e-output/`, `ai-docs/e2e-screenshots/`
Then ask the user (via AskUserQuestion) whether to write a full ai-docs/e2e-test-report.md with per-journey breakdowns, screenshots, datastore checks, and findings. If yes, write it.