Explore Feature

Turns "here's my change, cover it" into one committed, locally-green Playwright spec.

It is a thin orchestrator: it owns two phases — resolve what to cover, and confirm the result — and delegates the actual authoring (analyze FE, discover live UI, write POM/spec, run green) to the writing-e2e-tests skill.

Announce at start: "I'm using the explore-feature skill to add an e2e test for X."

What this does — and doesn't

• Does: resolve the change surface → pick the one flow worth covering → hand it to writing-e2e-tests → confirm the spec is tagged, in the taxonomy, and green.
• Doesn't: deep bug-hunting or edge-case coverage (that's a separate QA activity); CI wiring.
• Cheap per-PR happy-path only. One flow, green locally in minutes.

The output is an ordinary spec: tests/<area>/<name>.spec.ts, a tier tag, an @area:, a @cap: per test — exactly what every other spec in the estate looks like, and picked up by the same suites. There is no separate lane and no version stamp.

The loop

digraph explore_feature {
    rankdir=TB;
    "1. Resolve scope (GATE)" [shape=box];
    "2. Local-run gate (GATE)" [shape=box];
    "3. Delegate authoring to writing-e2e-tests" [shape=box];
    "4. Confirm tagged + green" [shape=box];

    "1. Resolve scope (GATE)" -> "2. Local-run gate (GATE)";
    "2. Local-run gate (GATE)" -> "3. Delegate authoring to writing-e2e-tests";
    "3. Delegate authoring to writing-e2e-tests" -> "4. Confirm tagged + green";
}

Phase 1 — Resolve scope (gate)

Normalize whatever the dev pointed at into one ScopeSpec before authoring anything.

Input modes (auto-detect from the argument; ask if ambiguous):

Mode	Trigger	Resolve
Local diff	no arg / dirty tree / "my branch" / "my changes"	see Local-diff mode below — this is the default when no PR is named, incl. "I haven't opened a PR yet"
A PR	#<n> or a PR URL	PR diff + linked ticket via gh pr view <n> --json … (or the GitHub MCP)
Multi-PR / multi-ticket	a list	union of the diffs

Local-diff mode — a dev running this on their branch before (or without) a PR. Capture the full change surface, not just committed work — pre-PR work is often uncommitted:

base=$(git merge-base origin/main HEAD)
git diff --name-only "$base"...HEAD          # committed on the branch
git diff --name-only HEAD                     # unstaged working-tree changes
git diff --name-only --cached                 # staged but uncommitted
git ls-files --others --exclude-standard      # new untracked files

Union those for changedFiles. If all four are empty, there's nothing to cover — say so and stop.

Produce the ScopeSpec:

• tickets[] — for naming and the PR description, if there are any. A ticket key in the branch name (andreic/OPIK-1234-… → OPIK-1234) is the usual source. Never invent one; the spec name should describe the behaviour anyway, not the ticket.
• changedFiles[] — the FE/BE change surface.
• area — the taxonomy area the change belongs to, from tests_end_to_end/coverage/taxonomy.yaml. This decides the directory: tests/<area>/.
• targetPath = tests_end_to_end/e2e/tests/<area>/<name>.spec.ts — new, or an existing spec in that directory to extend. Prefer extending: a new test() in the area's existing spec beats a new file when the setup is the same.
• capabilities[] — the @cap: keys this will cover, grepped from the taxonomy. They usually already exist as covered: false. If nothing fits, add the entry — a kebab-case name for the user-facing capability.
• tier — @t1-smoke for fast deterministic core checks, @t2-cuj for multi-step journeys and anything destructive, @t3-nightly for slower/broader. Tier is chosen by how often it should run, not by importance; anything spending real LLM budget is not t1.
• happyPath — the one end-to-end flow to cover. Multi-PR → the combined assembled-feature flow, as one test.

Three things to resolve while shaping the happy path — each caught a false or unbuildable test in piloting:

• Fix PRs — cover the repro, not the easy path. For a fix:, the happy path must exercise the exact condition the bug needed. If the state can be reached two ways and only one triggered the bug (e.g. a trace shows the bug only when source=sdk via manual reference-linking, not via evaluate()), seeding the easy way makes the test pass against the pre-fix code too — a vacuous test. Identify the repro condition from the PR's root-cause description and seed that shape.
• N equivalent surfaces — cover the most representative one. If the change fixes the same behavior in several places (e.g. experiment "Go to logs", a shared sidebar, and a Playground cell link), don't try to cover them all — pick the single most representative entry point and tell the dev which ones you left. Keeps it cheap.
• Seeding is part of the deliverable, not a precondition. Work out early how the happy path's state gets created — an existing fixture, an SDK client, or the bridge (services/opik-sdk-driver). If the shape the repro needs isn't reachable through the current surface (e.g. the bridge only exposes evaluate() but the bug needs a manual client.trace(source=...) + ExperimentItemReferences shape), that seeding support is yours to add as part of authoring — extend the bridge route / add a fixture / use the SDK client directly — then write the test on top of it. Adding a fixture is normal, not scope creep. The only real stop is if the state cannot be produced through any public SDK / bridgeable path at all (rare) — then flag it, because it likely means the feature isn't end-to-end testable yet.

Two gates here, before expensive authoring:

1. Scope gate — state the resolved happy path + repro seed shape + target path + tier + the @cap: keys back to the dev and get a yes. If seeding the repro needs new bridge/fixture support, say so here so the dev knows this work also touches services/opik-sdk-driver or the fixtures. Multi-PR especially: "One test in <area>/<name>.spec.ts covering X→Y→Z, @t2-cuj. OK?"
2. Skip check — if the change is pure refactor / infra / docs with no user-facing behavior, say so and stop. Note two cases that are user-facing even though they look like config: a capability-map / constants change that adds a user-visible option (e.g. a new model in a dropdown → happy path: "open the page, the option is selectable"), and a backend-dominant change whose only visible effect is subtle (e.g. a trace that should not appear in a default list) — find the user-observable effect and cover that, don't skip. The opposite case also happens: a perf / internals change with no behavior delta by design (e.g. swapping a slow probe for a fast one, same rendered result). There's no PR-specific happy path — so either cover a generic regression on the affected page and say so, or skip-with-a-note if the suite already covers that page. Don't dress a generic regression up as PR-specific coverage. Two things to get right here: (a) skip vs cover turns on coverage of the specific state/decision the change governs, not the page as a whole — grep the existing suite (tests_end_to_end/e2e/tests) for that exact state. A page whose populated path is covered but whose empty/onboarding path (the branch a probe like this actually drives) is not is not "already covered" — cover the uncovered half. Skip-with-a-note only when the specific state is genuinely already asserted somewhere. (b) This is a fix: PR, so the "repro condition" mandate seems to apply — but a no-behavior-delta fix has no repro that renders differently pre/post. The perf-fix escape hatch overrides the repro mandate: say "N/A — no behavior delta; generic regression" and label it generic. A generic test that passes on both the pre- and post-fix build is correct, not a bug — say so when you report back.

Phase 2 — Local-run gate

Before authoring can be verified, confirm the dev has a local stack with their changes:

1. Probe for a running stack — the frontend (http://localhost:5173, or :5174 for FE-from-source) and the backend, the way the suite reaches it: GET <baseUrl>/api/is-alive/ver (e.g. http://localhost:5173/api/is-alive/ver). A standard opik.sh compose stack does not expose the backend on a bare :8080 — the FE proxies /api to it, and that proxied path returning a {"version": …} JSON is the real "backend is up" signal. A FE that answers on / but 000s on /api/is-alive/ver is a half-up stack: every seeded test fails on the first API call for env reasons, not the feature. Require the /api health check to pass, not just "/ answers on 5173." Note the returned version — it tells you which build is running (see step 2).

2. Gate the dev: confirm the running stack actually contains their changes. A stale prebuilt opik.sh stack won't show new data-testids — if the feature adds testids, the dev must be on FE-from-source :5174 (dev-runner --restart). Verify the change is actually in the served build, not just that a FE answers: for a FE-only PR merged to main, curl http://localhost:5174/src/<changed-file> and grep for a symbol the PR added (Vite serves source), and/or check git merge-base --is-ancestor <merge-sha> HEAD. "The dev server is up" is not "the fix is present."

3. If nothing is running / it's the wrong stack: offer to spin it (local-dev / dev-runner) or ask the dev to bring it up with their changes, then proceed. Never silently run against a stack lacking the feature — that produces false-green or false-missing-testid results.

   Worktree gotcha (FE-from-source against a prebuilt backend). dev-runner.sh is worktree-aware: in a worktree it offsets every port from a per-worktree hash and starts its own JAR-mode backend against a fresh, empty DB — so --restart there does not reuse the healthy opik.sh docker DB, and FE-from-source may not land on :5174. When you need FE-with-the-fix on top of an existing seeded docker backend, the reliable path is to run the Vite dev server directly with pinned ports and point its /api proxy at the running backend:

   • The opik.sh backend container usually publishes only its internal port to a random host port (docker port opik-<proj>-backend-1), not :8080. Vite's /api proxy strips /api and needs a bare backend, so bridge the container's app port to host :8080 on the compose network, e.g. docker run -d --name opik-be-8080 --network <compose_net> -p 8080:8080 alpine/socat tcp-listen:8080,fork,reuseaddr tcp-connect:opik-<proj>-backend-1:8080.
   • Then cd apps/opik-frontend && npm ci && VITE_DEV_PORT=5174 VITE_BACKEND_PORT=8080 npm run start.
   • Point the suite at it: OPIK_BASE_URL=http://localhost:5174 OPIK_DEPLOYMENT=oss. OSS needs no auth. Tear down the socat container when done.

Phase 3 — Delegate authoring to writing-e2e-tests

Invoke the writing-e2e-tests skill to do the analyze → discover-live-UI → write → run-green loop. Hand it the ScopeSpec: the target path, the tier + @area: + @cap: tags, the happy path and its seed shape, and "verify green against the dev's local stack."

That skill owns the conventions — test.step() wrapping, UI-first assertions, selector preference, SDK-only seeding, fixture-owned teardown, and the taxonomy update. Don't restate them here; read .agents/skills/writing-e2e-tests/conventions.md if you need them.

Phase 4 — Confirm

• The spec exists at targetPath, tagged with one tier + @area:<area>, with a @cap: per test.
• Those @area:/@cap: values resolve in tests_end_to_end/coverage/taxonomy.yaml, and the taxonomy was updated in the same change (spec added to specs:, covered capabilities flipped to covered: true with the tier).
• tag_lint.py reports 0 problem(s) — this is the CI tag-lint job, so a miss here is a red build:

  python3 tests_end_to_end/coverage/tag_lint.py --taxonomy tests_end_to_end/coverage/taxonomy.yaml --estate tests_end_to_end
  

• It runs green locally, and so does the rest of its feature directory — a shared POM or fixture is used by sibling specs:

  cd tests_end_to_end/e2e && npx playwright test tests/<area>/ --reporter=list
  npx tsc --noEmit
  

• Report the committed spec path back to the dev, plus anything you deliberately left uncovered (the other N surfaces, edge cases) so they know the boundary.

Ownership

QA owns this skill. When a generated test misses something, the fix lands in this skill's files — this is the feedback loop. Edit in .agents/skills/explore-feature/, then make claude to mirror for local testing.