vv-validate-dev-loop Skill

vv-validate-dev-loop

When to use

Use this skill for the day-to-day coding loop: editing code, running validation, reading extracted errors, fixing, and revalidating. It covers vv validate, vv pre-commit, vv state, vv run, and vv sync-check.

vibe-validate provides confidence checkpoints, not mandatory gates. Adopters choose where the checkpoint goes — as a pre-commit hook, a pre-push hook, an ad-hoc command an agent runs after each change, or in CI. This skill describes the mechanics; the placement is a project choice.

The core loop

edit code
  ↓
run validation   (vv validate | vv pre-commit | vv run <cmd>)
  ↓
  ├─ pass  → reach a checkpoint (commit, push, hand back to user, continue)
  └─ fail
       ↓
       vv state           (read extracted errors — instant, no re-run)
       ↓
       fix the errors shown
       ↓
       re-run validation  (mostly cache hits; changed code re-executes)
       ↓
       repeat until pass

Two rules keep this loop efficient:

Query vv state instead of re-running to read errors. The last run's result and extracted errors are already cached.
Re-run validation after every fix before declaring the change complete. Caching makes this cheap; skipping it misses side effects.

`vv validate` — full pipeline

Runs every validation step defined in vibe-validate.config.yaml (typecheck, lint, tests, custom steps), honoring parallelism and dependencies. Result is cached by git tree hash — same content, same cached answer.

vv validate                   # run full pipeline (cache-aware)
vv validate --force           # re-run everything, ignore all cache
vv validate --retry-failed    # re-run only failed steps; keep passed steps cached
vv validate --check           # exit 0 if already cached passing, 1 otherwise; no execution

When to reach for each:

Plain vv validate — default. Fast on unchanged code; runs only what changed otherwise.
--retry-failed — a transient failure (network blip, flaky test, resource contention). Keeps passed steps cached; re-runs only the failures. Warns when a step passes on retry with no code change (flaky indicator).
--force — you suspect stale cache or want a clean-room run before pushing a release candidate.
--check — scripting: "is this commit already known-good?" without executing anything.

For why validation re-ran (or didn't), see vibe-validate:caching-and-locking.

`vv pre-commit` — branch-synced validation

A superset of vv validate that adds a branch-sync check (is the branch behind origin/main?) and, if configured, a staged-secret scan. It's one natural place to put the checkpoint — wired as a git pre-commit hook — but the same result is available from vv validate + vv sync-check invoked anywhere in the workflow.

vv pre-commit                 # branch sync + full validation + secret scan

If it fails with "branch behind origin":

git fetch origin
git merge origin/main         # or: git rebase origin/main
vv pre-commit                 # re-run (cached where possible)

Equally valid alternatives to vv pre-commit:

Run vv validate manually after each logical change — agent-driven loop.
Run vv validate on a pre-push hook instead of pre-commit — faster local commits, gate before sharing.
Run vv validate --check in CI as a belt-and-suspenders layer.

`vv state` — query cached state

Reads the cached result of the last validation run on the current tree. Instant; never executes validation steps. Output is YAML with extracted errors.

vv state                      # compact YAML: pass/fail, failed step, extracted errors
vv state --verbose            # include full failedStepOutput

Typical failing output:

passed: false
failedStep: TypeScript
rerunCommand: tsc --noEmit
extraction:
  totalErrors: 2
  errors:
    - file: src/feature.ts
      line: 42
      message: "Type 'string' is not assignable to type 'number'"

The loop: vv validate fails → vv state shows which step failed and the extracted errors → fix those files → vv validate again.

`vv run <command>` — wrap any command

Wraps any shell command (tests, lint, typecheck, build) to add two things the raw command doesn't provide:

Tree-hash caching. Re-running on the same tree is instant; new output only when code actually changes.
Error extraction. Verbose tool output collapses into a concise YAML block with file, line, message — typically 90–95% smaller than raw output, massively reducing context usage.

vv run npx vitest path/to/foo.test.ts
vv run pnpm --filter @pkg test
vv run tsc --noEmit
vv run eslint src/
vv run --check  npx vitest foo.test.ts      # cache lookup only; no execution
vv run --force  npx vitest foo.test.ts      # ignore cache; re-execute

Example output on failure:

exitCode: 1
extraction:
  totalErrors: 1
  errors:
    - file: foo.test.ts
      line: 42
      message: "Expected 5 to equal 6"
  summary: "1 test failure"
  metadata:
    detection:
      extractor: vitest

Use vv run for:

Ad-hoc test/lint/typecheck invocations during development.
Any command you'll likely repeat and whose output is bulky.
Commands in any language ecosystem — built-in extractors cover TypeScript, ESLint, Vitest/Jest, Prettier, Maven, OpenAPI, with a generic fallback.

Don't use vv run for:

Watch modes (vitest --watch, next dev, tsc --watch).
Already-orchestrated commands (vv validate — it already extracts).
Interactive commands (git log, npm init).
One-shot commands whose output you want verbatim.

If vv run returns totalErrors: 0 but exitCode: 1, the extractor fell back to generic — see vibe-validate:authoring-extractors.

`vv sync-check` — branch divergence

Reports whether the current branch is behind, ahead of, or in sync with its upstream (usually origin/main). vv pre-commit calls this internally; run it directly when you want the answer without triggering validation.

vv sync-check                 # YAML status: ahead/behind/in-sync/no-remote

Typical recovery when behind:

git fetch origin
git merge origin/main         # or rebase

Secret-scanning prevention

When configured, vv pre-commit runs a staged-only secret scan (Gitleaks by default) before validation, blocking secrets from entering git. Once a secret is committed and pushed, it's compromised regardless of how fast it's detected — prevention at the pre-commit checkpoint is the only real fix.

This skill describes the runtime behavior. For enabling, choosing a scanner, managing false positives (.gitleaksignore, gitleaks:allow inline comments, baselines for legacy repos), see vibe-validate:setting-up-projects.

Operational reminders for AI agents

After fixing errors, always re-run validation before considering the change complete. Caching makes the re-run instant when correct; it catches side effects when wrong. Don't skip it on the theory that "the fix was obvious" — that's how regressions sneak into commits.

Read state, don't re-run, to see errors. When validation just failed, vv state gives you the extracted errors for free. Running validation a second time to "see what failed" wastes time and context.

Wrap ad-hoc commands with vv run. Raw npm test output can be thousands of lines; the extracted form is dozens. The token savings compound across the iterate-fix-revalidate loop.

Invocations — use whichever matches the adopter's environment:

vv validate, vv pre-commit, vv state, vv run …, vv sync-check — installed CLI (short alias).
vibe-validate validate etc. — installed CLI (long name).
npx vibe-validate validate, pnpm dlx vibe-validate validate, bunx vibe-validate validate — one-off without a local install.

Many adopter repos wire project wrapper scripts (e.g., pnpm validate, npm run validate, just validate). If the repo's CLAUDE.md or README documents one, prefer it for consistency with the team's workflow — but the canonical vv invocations above always work.

Common failure patterns

"Branch behind origin/main" from vv pre-commit. Merge or rebase from origin/main, then re-run. Cached validation results for the pre-merge tree still apply to unchanged files.
Validation appears to re-run every time on unchanged code. Usually uncommitted build artifacts (dist/, coverage/, .tsbuildinfo) are changing the tree hash. Add them to .gitignore. For deeper diagnostics (tree-hash mismatch, ignored-file leakage, locking conflicts), see vibe-validate:caching-and-locking.
Validation locked by a concurrent run. By default each worktree locks independently. For one-off bypass: vv validate --no-lock or vv validate --no-wait. For persistent config, see vibe-validate:caching-and-locking.
A test passes on --retry-failed with no code change. That's a flaky test — vv validate emits a warning. Investigate the root cause rather than papering over it with retries.
Validation fails but vv state shows totalErrors: 0. The extractor didn't recognize the tool's output and fell back to generic. Add or tune a custom extractor — see vibe-validate:authoring-extractors.
"Which commit was this tree last validated at?" vv history list --limit 5 shows recent snapshots and their tree hashes. For recovery from those snapshots, see vibe-validate:recovering-work.

Agent Skills: vv-validate-dev-loop

Install this agent skill to your local

Skill Files