Shiplight

Shiplight

The single entry point for Shiplight QA work. This skill takes a subcommand and routes it to the right workflow. Everything Shiplight-branded comes through here; the description above is deliberately gated so this skill fires only when the user names "shiplight" or types /shiplight — never on a generic "write a test".

Routing contract

1. Identify the subcommand. Match the argument against the dispatch table — do not rely on the first token alone, because synonyms are often multi-word (yaml test, set up tests for my app) and intents may lead with a non-token word (create yaml test, show failing tests). In order: (a) an exact canonical token as the leading word; (b) the longest canonical-token or synonym phrase the argument contains; (c) overall intent against the synonym/intent column. Pick the single best-matching subcommand.
2. Pass-through context. Forward whatever the selector didn't consume to the subcommand as context. Natural phrasing is expected — users won't type the exact hyphenated token; match the intent and treat the rest as the target. Examples:

• /shiplight cover checkout flow → cover + context checkout flow
• /shiplight create a yaml test for login → create-yaml-tests (matched on "yaml test") + context for login
• /shiplight create agent verification for the signup flow → create-agent-verification (matched on "agent verification") + context for the signup flow

1. Dispatch. Read the matching references/<subcommand>.md (or references/<subcommand>/index.md for nested subcommands) and follow it, carrying the context forward.
2. Clarify, don't guess. If the selector is empty (/shiplight alone) or ambiguous (see Ambiguity notes), show the menu and ask one clarifying question — the user wants to act but didn't say how. This differs from help, which is informational: help lists/explains subcommands and never executes (see references/help.md).
3. Confirm destructive actions. Never auto-run init against a non-empty project — confirm first.

Shared layer

• On every subcommand invocation (skip for help and support), identify the test project root, then run references/_shared/update-check.md once (daily skill refresh + shiplightai CLI version gate). It can halt the run: a CLI behind the latest published version stops the subcommand before it starts, and offers an upgrade.
• Each subcommand names the references/_shared/ modules it needs (auth, mcp, evidence-and-report, project-layout, ground-truth, knowledge, secrets, vocabularies, test-spec-template). Read those before acting — they are the single source of truth, not restated per subcommand.

Subcommands (menu)

Show this grouped menu when invoked bare or when clarifying.

Setup

• init — scaffold a Shiplight test project + write specs/context.md
• setup-test-auth — configure login for the application under test and save reusable browser/Playwright state
• update — refresh installed Shiplight skills + the shiplightai CLI

Author

• create-yaml-tests — implement deterministic YAML E2E tests from a spec
• create-agent-verification — create a reusable agent-run verification script
• cover — decide test format + effort, plan, drive the producers, report

Maintain

• fix — reproduce and repair failing or drifted tests

Check

• verify — verify UI changes in the browser during local development

Review

• review — app-quality review (security, privacy, design, performance, …)

Ship

• ci — wire CI workflows + failure-triage pipeline
• cloud — read Shiplight Cloud test results (runs, failing/flaky tests, artifacts) and analytics (health summary, pass-rate/run trends, slowest/flaky rankings, failure attribution)

Help

• help — list subcommands, or help <subcommand> for details (does not execute)
• support — get human help: file a support ticket with session diagnostics, check replies

End the menu with one footer line: Stuck? /shiplight support reaches a human — the ticket drafts itself from this session.

Dispatch table

Ambiguity notes

• "test" / "write a test" → could be create-yaml-tests (deterministic, UI focused) or create-agent-verification (spans UI and backend state). Default to create-yaml-tests unless the user signals a core journey, backend/ cross-layer proof, release smoke, or live-env judgment — but if unclear, ask.
• "smoke test" → ambiguous alone. A quick UI pass over key screens is create-yaml-tests; a pre-release check that a core journey works and left the right backend state is create-agent-verification. Ask which, unless the user names a backend expectation or a release gate.
• "verify" / "verification" → the verb (check a change now) is verify; creating a reusable verification script is create-agent-verification. The create- framing is the tell. Ask if the user's phrasing doesn't disambiguate.
• "triage" → in Shiplight this means fix (repair failing tests). Do not confuse with review's internal triage/plan step.
• "failing tests" / "flaky tests" → reading them from CI ("in the cloud", "from the last run", plural reporting) is cloud (hosted results); repairing a broken test ("my test is failing", "fix this") is fix. Ask if the phrasing doesn't say which.
• "report a bug" / "X is broken" → depends on what is broken. The user's app misbehaving is ground truth to report (_shared/ground-truth.md), not a subcommand; Shiplight itself misbehaving (skill, CLI, cloud API) is support. Ask if unclear which one the user means.
• "login" / "auth" → ask what is being authenticated when the target is