e2e-validate
Validate that a change runs and fulfills its success criteria —
not just that it compiles. Discover the real harness; never invent
commands. When no harness fits the change kind, that is a reportable
finding, not a silent pass.
Interface
Caller passes (via the invoking prompt):
mode — chunk (validate a specific change) or full-repo.
changed_files — paths in scope (when mode=chunk).
success_criteria — what "done" means for this change.
ui_work — true if a UI surface must be visually verified.
stack — TypeScript / React / Vite / Next / Electron / backend API /
static HTML, if known.
Returns:
status — one of the enum values in Status below.
report — commands run, output excerpts, screenshot paths, and the
specific reason for the status.
Process
- Discover the harness — do not invent commands. Read
package.json scripts, Cargo.toml, Makefile, CI config
(.github/workflows), and CLAUDE.md. Also read any project
harness-quirks memory (e.g.
~/.claude/projects/<repo-key>/memory/*harness*.md) — if a failure you
hit is documented there with a fix, apply it or cite it in one line
rather than re-deriving it. Use what exists.
- Pick the smoke recipe by stack (see Recipes).
- Run it. Capture output.
- Run the project's format/lint check on the changed files. Discover
it the same way as the harness — the lint/format script in
package.json (biome, eslint, prettier), or ruff/gofmt/etc.
A non-zero result is Code Errors, not a pass, even if tests are
green. On Windows, distinguish a real content failure from a
line-ending-only blob difference (git ls-files --eol, or check the
staged/index content) so you don't chase CRLF noise. If the repo has
no lint/format tool, skip — don't invent one.
- If
ui_work, capture a screenshot and verify each
success-criterion feature is actually visible (e.g. "dark-mode
toggle present in settings").
- Judge against
success_criteria, not just exit codes. Code that
runs but doesn't meet the criteria is Requirements Unmet.
- Route stubborn failures to
/diagnosing-bugs if it's installed and the
failure's root cause isn't obvious.
- Write the report with the status and evidence.
Status (the routing signal)
| Status |
Meaning |
E2E Validated and Passing |
Harness ran clean AND success criteria met. |
E2E Validation Failed: Code Errors |
Tests fail, build broke, or runtime errored. |
E2E Validated but Requirements Unmet |
Runs fine, but doesn't fulfill the success criteria. |
Unable to Validate: No Harness |
No test runner and no smoke recipe fits — escalate; re-implementing won't add a harness. |
The status is a routing signal; the detailed reasons live in the
report. Always write an informative report regardless of status.
Recipes (smoke fallbacks when there's no first-class test suite)
See per-stack detail in recipes/.
- Static HTML/CSS/JS —
node --check on JS; parse HTML and
regex-assert critical IDs / aria attributes / script tags.
(recipes/static-html.md)
- Node / TS, no tests —
node --check, npx tsc --noEmit,
smoke-require/import the entry point.
(recipes/ts-node.md)
- Vite / Next / SPA, no tests —
npm run build; boot the dev
server and curl -fsS / for a non-5xx; web-visual-loop screenshot
if installed. (recipes/vite-spa.md)
- Electron, no tests — build;
electron-visual-loop screenshot +
element-existence check. (recipes/electron.md)
- Backend API, no tests —
curl -fsS representative inputs; DB
migrations dry-run. (recipes/backend-api.md)
- None fit — the
Unable to Validate: No Harness finding, with a
concrete remediation suggestion (e.g. "add tests/toggle.test.ts
covering [behavior]"), is the minimum delivery.
Discipline
- Missing harness is a finding, never a silent skip. Always emit a
concrete remediation suggestion with the no-harness status.
- Documented quirks aren't fresh findings. Before labeling a failure
"environmental / out of scope" or carrying it forward to the human,
check the project's harness-quirks memory. If it's documented with a
fix, apply it (one-time setup is in scope) or cite the memory in one
line — don't re-derive it as a new finding or escalate it.
- Lint/format failures in the changed diff are
Code Errors. A green
test run over a lint-failing diff is not validated — surface it so the
caller fixes it before shipping.
- Don't invent commands. If
npm test isn't a real script, don't
run it. Discover first.
- UI claims need pixels. For
ui_work, a screenshot showing the
feature is the evidence — not "the code looks right."
1---2name: e2e-validate3description: Stack-aware end-to-end validation. Discovers the repo's test/build harness, runs the closest smoke recipe, and verifies a change against declared success criteria — including UI feature visibility via screenshots. Use when verifying that code changes actually run and fulfill their requirements, or when the user asks to validate/smoke-test a change. Usable standalone or as a reviewer leaf inside an orchestrated loop.4added_in: 0.14.05metadata:6 source: https://skillmd.com/skills/superliaye/e2e-validate7---89# e2e-validate1011Validate that a change **runs** and **fulfills its success criteria** —12not just that it compiles. Discover the real harness; never invent13commands. When no harness fits the change kind, that is a reportable14finding, not a silent pass.1516## Interface1718Caller passes (via the invoking prompt):1920- `mode` — `chunk` (validate a specific change) or `full-repo`.21- `changed_files` — paths in scope (when `mode=chunk`).22- `success_criteria` — what "done" means for this change.23- `ui_work` — `true` if a UI surface must be visually verified.24- `stack` — TypeScript / React / Vite / Next / Electron / backend API /25 static HTML, if known.2627Returns:2829- `status` — one of the enum values in **Status** below.30- `report` — commands run, output excerpts, screenshot paths, and the31 specific reason for the status.3233## Process34351. **Discover the harness — do not invent commands.** Read36 `package.json` scripts, `Cargo.toml`, `Makefile`, CI config37 (`.github/workflows`), and `CLAUDE.md`. **Also read any project38 harness-quirks memory** (e.g.39 `~/.claude/projects/<repo-key>/memory/*harness*.md`) — if a failure you40 hit is documented there with a fix, apply it or cite it in one line41 rather than re-deriving it. Use what exists.422. **Pick the smoke recipe** by stack (see Recipes).433. **Run it.** Capture output.444. **Run the project's format/lint check on the changed files.** Discover45 it the same way as the harness — the lint/format script in46 `package.json` (`biome`, `eslint`, `prettier`), or `ruff`/`gofmt`/etc.47 A non-zero result is `Code Errors`, not a pass, even if tests are48 green. On Windows, distinguish a real content failure from a49 line-ending-only blob difference (`git ls-files --eol`, or check the50 staged/index content) so you don't chase CRLF noise. If the repo has51 no lint/format tool, skip — don't invent one.525. **If `ui_work`,** capture a screenshot and verify each53 success-criterion feature is actually visible (e.g. "dark-mode54 toggle present in settings").556. **Judge against `success_criteria`,** not just exit codes. Code that56 runs but doesn't meet the criteria is `Requirements Unmet`.577. **Route stubborn failures to `/diagnosing-bugs`** if it's installed and the58 failure's root cause isn't obvious.598. **Write the report** with the status and evidence.6061## Status (the routing signal)6263| Status | Meaning |64|---|---|65| `E2E Validated and Passing` | Harness ran clean AND success criteria met. |66| `E2E Validation Failed: Code Errors` | Tests fail, build broke, or runtime errored. |67| `E2E Validated but Requirements Unmet` | Runs fine, but doesn't fulfill the success criteria. |68| `Unable to Validate: No Harness` | No test runner and no smoke recipe fits — escalate; re-implementing won't add a harness. |6970The status is a routing signal; the detailed reasons live in the71report. Always write an informative report regardless of status.7273## Recipes (smoke fallbacks when there's no first-class test suite)7475See per-stack detail in [recipes/](recipes/).7677- **Static HTML/CSS/JS** — `node --check` on JS; parse HTML and78 regex-assert critical IDs / aria attributes / script tags.79 ([recipes/static-html.md](recipes/static-html.md))80- **Node / TS, no tests** — `node --check`, `npx tsc --noEmit`,81 smoke-`require`/`import` the entry point.82 ([recipes/ts-node.md](recipes/ts-node.md))83- **Vite / Next / SPA, no tests** — `npm run build`; boot the dev84 server and `curl -fsS /` for a non-5xx; `web-visual-loop` screenshot85 if installed. ([recipes/vite-spa.md](recipes/vite-spa.md))86- **Electron, no tests** — build; `electron-visual-loop` screenshot +87 element-existence check. ([recipes/electron.md](recipes/electron.md))88- **Backend API, no tests** — `curl -fsS` representative inputs; DB89 migrations dry-run. ([recipes/backend-api.md](recipes/backend-api.md))90- **None fit** — the `Unable to Validate: No Harness` finding, with a91 concrete remediation suggestion (e.g. "add `tests/toggle.test.ts`92 covering [behavior]"), is the minimum delivery.9394## Discipline9596- **Missing harness is a finding, never a silent skip.** Always emit a97 concrete remediation suggestion with the no-harness status.98- **Documented quirks aren't fresh findings.** Before labeling a failure99 "environmental / out of scope" or carrying it forward to the human,100 check the project's harness-quirks memory. If it's documented with a101 fix, apply it (one-time setup is in scope) or cite the memory in one102 line — don't re-derive it as a new finding or escalate it.103- **Lint/format failures in the changed diff are `Code Errors`.** A green104 test run over a lint-failing diff is not validated — surface it so the105 caller fixes it before shipping.106- **Don't invent commands.** If `npm test` isn't a real script, don't107 run it. Discover first.108- **UI claims need pixels.** For `ui_work`, a screenshot showing the109 feature is the evidence — not "the code looks right."