preflight-check

Pre-flight Check

Universal pre-flight validation across commit, PR, release, and deploy boundaries. Orchestrates the validator suite that lives under .claude-plugin/skills/validation/ and dispatches the right check set for the current context.

When to Use

• User asks "is this ready to commit / PR / release / deploy?"
• User says "pre-flight", "validate", "check readiness", or runs /craft:check
• Before staging a commit, opening a PR, or kicking off the release pipeline
• After a session of edits, to confirm nothing regressed
• When the user wants to scaffold a new custom validator (gen-validator flow)

Scope Boundary (Read This First)

This skill is broader than its siblings — understand the seams:

Skill	Concern	Does NOT cover
preflight-check (this)	Cross-cutting validators: lint, tests, types, version sync, stale refs, link integrity, hook conflicts, CLAUDE.md health, docs staleness, badge URLs, formula desc	Detection logic, release orchestration, guard rule editing
project-detector (skills/ci/)	Detect project type / build tool / test framework to recommend a CI template	Validation orchestration. It answers "what kind of project is this?" not "is it ready?"
release (skills/release/)	The full release pipeline (version bump → PR → merge → GitHub release → Homebrew → verify). Calls this skill internally for its Step 2 pre-flight gate	Standalone commit/PR validation
guard-audit (skills/guard-audit/)	Tune branch-guard.sh rules via .claude/branch-guard.json to reduce false positives	Running checks. It edits guard config, doesn't validate the project

Rule of thumb: if the user wants to run validators, use this skill. If they want to change which rules block them, route to guard-audit. If they want a release, route to release (which calls back into this skill).

Backing Commands

• /craft:check — primary entry point; mode-aware, context-aware orchestrator
• /craft:check:gen-validator — scaffold a new validator into .claude-plugin/skills/validation/

The skill should prefer invoking the commands rather than re-implementing their logic. The commands handle dry-run, orchestration, and hot-reload validator discovery.

Context Flags (--for ...)

Pick the check set by what the user is preparing for. This is the primary axis of behavior:

--for	Run when user is about to...	Adds beyond default
commit	Stage and commit on a feature branch	Fast lint on changed files, fail-fast tests, basic version-sync, no-secrets scan
pr	Open a PR to dev/main	Full lint, full test suite, coverage ≥ 80%, internal links, merge-conflict detect, stale-ref scan, hook-conflict audit, CLAUDE.md health
release	Cut a release (called by skills/release/)	Full audit: strict lint, all tests + coverage ≥ 90%, all links, full version audit (fatal on drift), security audit, badge URL both-branch check, formula desc count check
deploy	Deploy docs site or downstream artifact	Same as release minus formula-desc; plus tag-exists check

If no --for is specified, run the "general" check set (lint + tests + git status + docs if changed).

Mode-Aware Behavior

Modes are orthogonal to --for — they tune verbosity, fail-fast, and threshold strictness:

Mode	Budget	Behavior
default	< 30s	Fast path, fail-fast, summary output
debug	< 120s	Verbose output, no fail-fast, surface all warnings
thorough	< 180s	Full lint + types + security + docs sweep
release	< 300s	Strictest thresholds (coverage ≥ 90%, version drift = fatal)

Mode applies to the hot-reload validators too via $CRAFT_MODE — see the validator template generator output for the per-mode threshold table.

Recommended Flow

1. Confirm intent. Ask which context (commit/pr/release/deploy) if the user didn't say. Default to the most conservative reasonable guess based on the current branch (feature/* → commit or pr; dev → release).
2. Dry-run first when stakes are high. For --for release or --for deploy, suggest /craft:check --for release --dry-run to preview the validator plan and time estimate before executing.
3. Run the appropriate command. Invoke /craft:check --for <context> [mode]. Let the command's Step 0 plan + AskUserQuestion gate handle confirmation.
4. On failure, route specifically. Don't just dump errors — point at the right follow-up:
   • Lint failures → suggest /craft:code:ci-fix or /craft:code:lint
   • Version drift → suggest /craft:check --version release for the focused validator
   • Stale refs from a rename → list the files that need updating
   • CLAUDE.md staleness → suggest /craft:docs:claude-md:sync
   • Hook-conflict false positives → route to guard-audit skill
5. On all-green, summarize and stop. Don't recommend "next steps" past the immediate gate the user asked about (per feature-branch-workflow rules: no PR-merge nudging).

Validator Generation (Separate Operation)

Generating a new custom validator is a distinct workflow from running checks. Trigger conditions:

• User says "create a validator", "add a custom validator", "generate a validator template"
• User wants to extend /craft:check with project-specific validation logic

In that flow:

1. Use /craft:check:gen-validator <name> [--languages ...] [--interactive]
2. The command scaffolds a file under .claude-plugin/skills/validation/<name>.md with hot_reload: true so /craft:check picks it up on next run with no restart.
3. After generation, suggest a smoke test: CRAFT_MODE=default bash .claude-plugin/skills/validation/<name>.md then /craft:check to confirm auto-discovery.

Do not auto-generate a validator while the user is running checks — these are independent concerns. If both are in flight, finish the check pass first.

Context Header (--context)

The check command supports a --context flag that skips validation and emits only a one-screen session header (project, branch, worktree, base, guard status, phase, test count, docs status, insights). When the user asks "what's the state of this branch?" or "what session am I in?", prefer /craft:check --context over running validators. It's read-only and < 1s.

Output Conventions

The backing command uses box-drawing output (see commands/check.md). When this skill summarizes results in chat, mirror that structure compactly:

Project: <name> (<type>)        Context: <for-value>      Mode: <mode>
Lint:    PASS / N issues
Tests:   N/M passed
Version: in sync / DRIFT (file:line)
Docs:    GREEN / YELLOW / RED
Status:  READY / BLOCKED — <reason>
Next:    <single actionable suggestion or none>

Keep summaries terse. The command's own output is the system of record.

Anti-Patterns

• Don't re-implement project detection here — defer to project-detector via the command.
• Don't edit branch-guard rules — that's guard-audit.
• Don't chain into the release pipeline — let the user invoke release explicitly. This skill stops at "READY FOR RELEASE."
• Don't silently expand scope. If the user asked --for commit, don't also run the release-tier validators just because "they're cheap." Mode budgets exist for a reason.
• Don't fix issues the validators report unless the user asks. Surface, then wait.

See Also

• commands/check.md — full check command spec (validators, modes, dry-run output)
• commands/check/gen-validator.md — validator scaffolding
• .claude-plugin/skills/validation/ — hot-reload validator directory
• skills/release/SKILL.md — orchestrates this skill at Step 2 of releases
• skills/ci/SKILL.md — project detection (used internally to pick check set)
• skills/guard-audit/SKILL.md — tune what blocks operations (not what validates them)