Claude Code: AI Agent & Automation Workflows

Ranks candidate skills/agents by task fit using Sonnet LLM-as-judge AND classifies task complexity (model + effort) in same call. Input is union of cheatsheet + FTS5 candidates with EQUAL weight per D3+D9. Output is task-driven ranking + complexity classification + INSTALLBEST | INSTALLMULTIPLE | SYNTHESIZE recommendation. v0.5.0 replaces v0.4.0 BM25+Jaccard ranking with Sonnet-as-judge (Jaccard retained only for INSTALLBEST→MULTIPLE downgrade cross-check).

Audits the user's local skill collection (personal + plugin + project) for redundancy, broken descriptions, missing paths filters, skill bloat, and outdated components. Reports issues color-coded (red = must fix, yellow = recommendation, green = OK). Periodic use (monthly) or right after adding new plugins.

Compares the locally installed version of a skill or subagent against its upstream source (sha or semver) and reports whether an update is available. Cache-aware (1-hour TTL). Used by orchestrate before dispatching a subagent or invoking a skill, to guarantee the components are fresh.

Sweeps ~/.claude/skills//.maxv-source.json for ephemeral=true entries at SessionEnd. Deletes those below promotethreshold (default 5 uses); promotes those at/above threshold by rewriting ephemeral=false. Invoked automatically by the SessionEnd hook unless reason=resume. Backward compat — v1.0.0 manifests with no ephemeral field are treated as permanent.

Detects whether a user task contains multiple independent subtasks and splits it into a JSON array. Internal helper for orchestrate Phase 0.5; uses session credentials (no API key). Always returns STRICT JSON {ismulti, tasks[]} with cap N=5.

Looks up subagents by keyword/domain across all installed plugins (voltagent, pr-review-toolkit, feature-dev, code-review, superpowers, etc.) plus user-level agents. Twin of discover-skill but searches kind='agent' in the FTS5 index. Cross-references with CLAUDE-agents.md cheatsheet for trusted-mappings.

Looks up a skill by name, keyword, or domain in the curated catalog (skill-sources.json). Returns the source repo, tier of trust, install methods supported, and any guardrails (overlap warnings, deprecations). Use when the orchestrator detects a missing skill and needs to know where to fetch it from.

Dispatches a long-running task to a maxv- subagent in background via Claude Code's Agent View (claude --bg). Routes the task through route-task to pick the right agent, then forwards to bin/maxv-dispatch run. Use when the task is long-running and you want to keep the main session unblocked. The session appears as a row in claude agents.

Background-dispatch variant of execute-phase. Spawns one claude --bg --agent maxv-executor session per plan (with auto-worktree) instead of running waves of in-process Agent calls. Frees the main session immediately. Use when the phase has multiple independent plans, the work is long-running, and you want to monitor via claude agents instead of staying attached.

Bootstraps or refreshes the local FTS5 catalog at ~/.claude/cache/maxv-orchestration/index.db by downloading the latest release artifact (default) or rebuilding it locally with --rebuild. Use after first install OR when version-check shows the index is stale OR when you want to force a refresh.

Installs a Claude Code subagent on-demand from the FTS5 catalog (~/.claude/cache/maxv-orchestration/index.db) into ~/.claude/agents/ .md (user scope) or .claude/agents/ .md (project scope). Twin of install-skill for agents. Writes .maxv-source- .json sidecar for version tracking. Used by orchestrate Phase 7.3 when an agent is needed but not present locally, enabling lazy-load model that bypasses the cumulative-description token budget.

Coordinates a repo-wide mechanical migration (rename X → Y across N files, bump dependency version across a monorepo, port a deprecated API call site by site) by dispatching the change to Claude Code's bundled /batch skill, which fans out into 5-30 worktree-isolated subagents that each open an independent pull request. Use only for migrations that are (a) mechanical, (b) describable in one instruction, and (c) safely partitionable into independent file sets.

Internal helper. Runs a BM25 FTS5 query against ~/.claude/cache/maxv-orchestration/index.db and returns top-N candidates ranked by relevance. Called by route-task when the cheatsheet has no match. Not user-invocable directly.

Re-fetches metadata for catalog sources (skill-sources.json) and updates fields like lastupdated, star counts, latest commit sha, and license. Use periodically (monthly) to keep the catalog accurate. Does not change source IDs or tier classifications.

Decides which subagent and which skills should handle a task by reading CLAUDE-agents.md and CLAUDE-skills.md, applying the routing-rules.json heuristics, and returning a structured recommendation. Called by orchestrate after the local inventory phase.

A valid sample skill for unit tests covering common frontmatter patterns and SKILL.md body shape.

Creates a custom maxv- skill in ~/.claude/skills/ by combining capabilities from multiple already-installed source skills. Invokes anthropic-skills:skill-creator with a structured multi-source prompt, validates the generated frontmatter against claude-code-json-schema, and writes a multi-source .maxv-source.json plus NOTICE.md attributing each upstream. Always asks for confirmation before any disk write.

Removes a locally-installed Claude Code subagent from ~/.claude/agents/ (user) or .claude/agents/ (project). Deletes the .md file and the .maxv-source- .json sidecar. Leaves the FTS5 catalog intact (the agent stays discoverable for future install). Twin of install-agent.

Removes a locally-installed skill from ~/.claude/skills/ (user scope) or .claude/skills/ (project scope). Deletes the skill folder and its .maxv-source.json sidecar. Refuses orphans (no sidecar) and maxv- synthesized skills without explicit overrides. Called by cleanup-ephemerals at SessionEnd via --from-cleanup flag.

Updates a stale skill or subagent to its latest upstream version. Always asks for confirmation before any write. Updates .maxv-source.json (for manually-cloned skills) or runs /plugin update (for plugin-managed components). Used by orchestrate after check-version reports stale components.

Use when starting any conversation - establishes how to find and use skills, requiring Skill tool invocation before ANY response including clarifying…

Self-test the maxvision-orchestration plugin after install. Runs five read-only checks (index, cheatsheets, gh auth, jq+sqlite3 in PATH, sample BM25 query) and reports pass/info/fail. Use immediately after /plugin install, /plugin update, or whenever the orchestrator behaves unexpectedly.

Goal-backward phase verification + UAT conversational verification, consolidated. Use when checking phase achieves promised goal before milestone.