Vibekit Doctor

vibekit-doctor

A read-mostly health check for the vibekit plugin and the repo it lives in. Answers one question: does vibekit actually work here, end to end?

The doctor does not install, upgrade, or migrate. It diagnoses. It writes only when the user passes --fix, and even then only to safe, deterministic targets (rebuild an index, recreate a docs/ subdir).

When to invoke

• The user reports vibekit "isn't working" / "skill not found" / "AGENTS.md is wrong".
• After pulling a new vibekit version, before the first run.
• After hand-editing skills/, .vibekit/, or a runtime registration file.
• Periodically as a hygiene check on long-lived projects.

Do NOT invoke for:

• Diagnosing a feature failure (failed verify-gate, broken plan). That is the relevant skill's failure path, not a doctor concern.
• Diagnosing a Claude Code / Codex / Gemini / opencode runtime bug. The doctor checks vibekit's contracts; runtime bugs are out of scope.

Operations

check

Run every check below in order. Output a single YAML report. Read-only.

verdict: ok | warn | critical
checks:
  - name: <check name>
    status: ok | info | warn | critical
    detail: <one-line, verbatim>
    evidence: <path / line / command output, optional>
summary:
  ok: <int>
  info: <int>
  warn: <int>
  critical: <int>

verdict is the worst status across all checks. Any critical → critical. Any warn and no critical → warn. Otherwise ok. info rows are recorded but never escalate the verdict — they exist for optional/gitignored files (CLAUDE.md, GEMINI.md) that may legitimately be absent.

check --fix

Run all checks, then auto-repair the items below. Other findings remain in the report; only the safe set is touched.

Anything else — registration mismatches, missing skills, broken [[links]], contract drift — is reported but never auto-fixed. Those changes touch user intent (which skills exist, what AGENTS.md should say) and require human judgment.

check --strict

Same as check, but treats warn as critical for the verdict. Useful in CI.

Checks

Each check produces one row in the report.

C1 — Skill file integrity

For each directory under skills/ (excluding _authoring/):

• A SKILL.md exists.
• SKILL.md has YAML frontmatter with name: and description: keys.
• name: matches the directory name.

critical if any skill directory lacks a parseable SKILL.md. warn if name mismatches dir.

C2 — Runtime registration parity

For every skill directory under skills/ (excluding _authoring/):

• AGENTS.md lists the skill in the skill table.
• GEMINI.md has an @./skills/<name>/SKILL.md line — only if GEMINI.md exists. GEMINI.md is gitignored as a per-user local context file (parity with CLAUDE.md); a missing GEMINI.md is not a defect.

Reverse: every skill referenced in AGENTS.md (and GEMINI.md when present) exists on disk.

Claude Code (.claude-plugin/plugin.json) and Codex (.codex-plugin/plugin.json) declare "skills": "./skills/" directory globs; they auto-discover by scan and need no per-skill row. opencode (.opencode/plugins/vibekit.js) adds the skills directory to config.skills.paths; same auto-discovery.

warn on any one-sided registration; critical if a required runtime entry doc is missing (see C3).

C3 — Plugin manifest presence

Existence check.

Required (committed to the repo):

• .claude-plugin/plugin.json
• .claude-plugin/marketplace.json
• .codex-plugin/plugin.json
• gemini-extension.json
• AGENTS.md
• .opencode/plugins/vibekit.js

Optional (gitignored, per-user local context):

• CLAUDE.md — Claude Code project-level context.
• GEMINI.md — Gemini CLI context file (declared as contextFileName in gemini-extension.json). Users running on Gemini author this locally; the C2 GEMINI.md parity check only fires when this file exists.

critical for any missing required file. info (not warn) when an optional file is absent. The doctor does not parse contents beyond what C2 needs; deeper schema validation is the runtime's job.

C4 — docs/ subdirectory presence

The pipeline writes to docs/specs/, docs/plans/, docs/reviews/, docs/verifications/. Each must exist as a directory.

warn for any missing dir (not critical — fresh repos won't have them yet, and --fix creates them).

C5 — .vibekit/ directory health (if present)

If .vibekit/ does not exist, status is ok (clean repo, no memory yet).

If .vibekit/memory/ exists:

• INDEX.md exists; every <key>.md is in INDEX; every INDEX row has a file. warn on drift; --fix rebuilds.
• Every <key>.md parses with required frontmatter (key, type, scope, created, updated, confidence). warn on malformed entries.

If .vibekit/wiki/ exists:

• index.md and log.md exist. warn if missing; --fix rebuilds.
• Every page has frontmatter (title, slug, category, tags, created, updated, status). warn on malformed.
• Every [[slug]] cross-link resolves. warn per broken link, listing source page + target.

If .vibekit/notepad.md exists:

• Has the three sections (## Priority Context, ## Working Memory, ## Manual). warn on missing section.
• Priority Context ≤ 500 chars. warn on overflow.

C6 — Authoring contract alignment

skills/_authoring/peg-cheatsheet.md and skills/_authoring/karpathy-principles.md exist. critical if either is missing — they are the source of truth for skill authoring.

The Karpathy principles file declares an injection map (which principles are enforced by which skills). For each principle row, every named enforcing skill must exist on disk. warn on mismatch.

C7 — external/ integrity

external/ is the read-only references dir.

• It exists. warn if missing; the references inform skill authoring.
• It is gitignored. warn if not (would commit a large external tree by accident).
• The expected sub-references exist if referenced in CLAUDE.md (caveman, superpowers, oh-my-claudecode, oh-my-codex, Prompt-Engineering-Guide, andrej-karpathy-skills). warn per missing sub-tree.

C8 — Skill count consistency

AGENTS.md opens with a skill count ("vibekit is an N-skill plugin"). The number must match the count of skill directories under skills/ (excluding _authoring/).

warn on mismatch. The doctor does NOT auto-edit AGENTS.md prose; the user edits it.

C9 — Hooks/sessions hygiene

If hooks/ directory contains any files, they exist as committed code (not stray artifacts). The vibekit hooks/ is reserved but currently empty; any unexpected file is warn.

C10 — Git state

• The repo is a git repo. critical if not (vibekit assumes git for isolate, finish-branch, commit-based plan resolution).
• The current branch and worktree state are reported as informational, not graded.

Discipline rules

• Read-mostly. check never writes. check --fix writes only to the auto-fix table above.
• Evidence first. Every non-ok row carries evidence — the path, the line, the command output. No "trust me."
• Verbatim output. Check names and statuses are byte-identical across runs. Tools and CI parse this report.
• Never auto-edit user intent. AGENTS.md skill counts, skill registrations, ship-order phases — these are decisions, not state. Report them; do not rewrite them.
• No network. No npm view, no remote version checks. The repo is the source of truth.

Karpathy alignment

• Surgical Changes. --fix touches only the safe set; it never reorganizes. Any non-safe finding requires user action.
• Simplicity First. Plain-text checks against plain-text artifacts. No daemons, no caches, no introspection beyond ls / cat / git.
• Goal-Driven. Every check has a checkable success criterion (file exists; frontmatter parses; row matches).
• Think Before Coding. check defaults to read-only; the user must opt into --fix. The doctor never assumes its judgment beats the user's.

Compression policy

The full report is YAML for parseability. Every finding is one row; no nested prose.

Pipeline integration

vibekit-doctor is utility, not pipeline. vibe does NOT call it; verify-gate does NOT call it. It runs only when the user invokes it.

The doctor MAY be invoked by:

• A pre-flight check before a long vibe run.
• CI as a --strict health gate.
• The user, ad hoc, when something feels off.

Cross-runtime portability

All operations are file I/O and git shell-outs. Verified portable to:

• Claude Code — auto-discovers via .claude-plugin/plugin.json.
• Codex — auto-discovers via .codex-plugin/plugin.json.
• Gemini CLI — referenced from GEMINI.md via @./skills/vibekit-doctor/SKILL.md.
• opencode — auto-discovers via .opencode/plugins/vibekit.js (skills directory scan).

No runtime-specific logic. The doctor does not introspect the runtime it is running under.

Anti-patterns

• Auto-fixing a missing skill registration. The user wrote the registrations; doctor reporting drift is the contract.
• Running check --fix without first running check. The user should see what would change before opting in.
• Treating warn as critical outside --strict. Inflated severity erodes trust in the report.
• Network calls for "latest version" comparisons. The repo is the source of truth; this is not an installer.
• Silently rewriting AGENTS.md prose to match disk. That is user intent; the user fixes it.
• Caching results. Each check is independent; the cost is small and freshness matters.

Output of this skill

For check and check --strict: the YAML report exactly as specified above, then exit. For check --fix: the YAML report, then a fixes_applied: block listing each safe-set repair, then exit.

No conversational narration around the report. The skill is a diagnostic, not a chat partner.