claude-plugins-validation

Fan out cache-optimizer-agent across every plugin in a marketplace, a list of plugins, or a single plugin. Accepts local paths and GitHub URLs. One cache-optimizer-agent per plugin running in batch_audit mode (read-only — Phase 1 of the cache pipeline). Detects the six prompt-cache-invalidation patterns (CA-01..CA-06) per plugin without applying fixes. Parallel main-session dispatch (default 8, cap 16).

Fan out cache-optimizer-agent across every plugin in a marketplace, a list of plugins, or a single plugin and APPLY the cache-invalidation fixes (CA-01..CA-07). Accepts local paths and GitHub URLs. One cache-optimizer-agent per plugin in batch_fix mode (Phase 1 audit + Phase 2 fix + Phase 3 re-validate; Phase 4 broader refactor SKIPPED in batch mode to avoid per-plugin user prompts). Parallel main-session dispatch (default 8, cap 16).

Parallel fix for one OR many plugins. Single-plugin inputs run the existing per-shard protocol (one plugin-fixer per shard of findings). Marketplace / list / URL inputs run a per-plugin fan-out (one plugin-fixer per plugin; each fixer may internally shard its own findings). Accepts local paths, GitHub URLs, marketplace URLs/folders, comma-separated lists, and @listfile shapes. Default 8 parallel agents per main-session message, cap 16. Each agent starts with a fresh context window (size determined by `plugin-fixer.model`).

Maximum-coverage same-turn scan + fix. One plugin-fixer per plugin reads each source file ONCE and runs validate + security + caching audit + caching optimize + verify-FPs + fix — all inline. Each per-file scan triggers every applicable checker; confirmed-real findings are fixed; confirmed-FPs are skipped. Cuts per-plugin token cost ~5× vs running the four separate batch skills sequentially. Default 8 parallel agents per main-session message, cap 16.

Same-turn diagnose + fix across a fleet of project folders. One cpv-doctor-agent per project scans the requested scope + verifies findings + applies obvious fixes inline (NIT and CRITICAL auto; MAJOR / MINOR auto when the recipe is safe). LOCAL paths only — URL inputs are CRITICAL errors. Default 8 parallel agents per main-session message, cap 16.

Self-sufficient cache-optimization WORK agent dispatched by the cpv-main-menu (menu-tree §3.3 cache rows) or by the /cpv-batch-caching-audit and /cpv-batch-caching-optimize commands. Accepts EITHER a pre-existing cache-audit report path OR a plugin/project path via the dispatcher's `<context>` block. Runs the full validate → fix → re-validate loop on its own. Fixes the seven documented prompt-cache invalidation patterns (CA-01..CA-07) AND, when the user asks (mode `audit_then_fix_broader`), performs Phase 4 broader cache-aware improvements to the plugin's skills/agents/commands/CLAUDE.md/rules. Loads cache-validation-skill and the fix-validation skill (cache-fixes references).

CPV doctor WORK agent invoked by the /cpv-main-menu main-session orchestrator (Diagnose category). The orchestrator renders the "Diagnose what?" first-contact menu and collects per-action follow-up; this agent receives a structured `<context>` block with the resolved `mode` and `target_path` and runs the matching diagnostic recipe. Runs BOTH the schema-correctness validator (validate_plugin.py et al.) AND nine deep design-correctness recipes (D1..D9): shape detection, command coverage audit, skill invocability audit, design-conflict scan, manifest/marketplace consistency, canonical-pipeline presence, README/CONTRIBUTING coverage, cross-reference integrity, the-skills-menu adoption. Findings land in a single report under $MAIN_ROOT/reports/plugin-diagnoser/. Free-form "Ask the doctor" mode routes the user's description to a diagnostic dialog.

Single-entry menu for every CPV command/skill/agent. Routes every menu (Validate, Fix, Optimize for Cache, Diagnose, Update, Create, Publish & Migrate, Manage, Help) through the claude-menu-system Stop hook, with `0 — Cancel / Exit` at every level and `B — Back` at every sub-menu. Loads cpv-main-menu-skill for the menu tree and per-leaf execution recipes. NEVER uses AskUserQuestion. NEVER prints menu tables inline — the Stop hook emits the menu post-turn via `systemMessage` (zero context cost). This agent only queues menu specs and parses integer/letter choices — fast, no analysis. Heavy lifting is dispatched to the specialised work agents (plugin-creator, plugin-fixer, plugin-diagnoser, marketplace-fixer, semantic-validator, cache-optimizer-agent) when a leaf is picked.

CPV's bundled lightweight implementation agent — focused fixes, quick tweaks, file creations, code changes inside a CPV workflow. Use for short-context dispatches from CPV menus and orchestrators where the task is bounded (one file or one symbol) and TDD-grade rigor would be overhead. Launch many in parallel for fan-out work. For larger, multi-file implementations escalate to specialised work agents (plugin-fixer, cache-optimizer-agent, etc.). Bundled into the plugin as part of v2.103.2 so CPV's dispatch chain no longer depends on a user-scope agent — every install gets cpv-spark out of the box with the SERENA MCP / TLDR tool surface its workflow requires. No `model:` pin (per CPV's CA-04 cache-warmth invariant — every CPV agent inherits the session model; the dispatch site decides the override with `model: opus` at Agent({...}) call time when bursts of Opus power are needed).

General-purpose CPV worker. Receives a free-form plugin-quality request and autonomously routes it to the right claude-plugins-validation skill, agent, or script, then executes the whole job in an isolated context. Dispatch it for "use CPV to <anything>" or "read the CPV skills menu and pick what is needed" — validate, security-scan, fix (delegates to plugin-fixer, never hand-edits), optimize prompt cache, create, publish to GitHub + add to a marketplace, migrate a marketplace, manage installed plugins, or AI-grade a skill. Reads the-skills-menu for the intent→action map and chains steps (validate → fix) as needed. For one bounded edit use cpv-spark instead; for an interactive numbered menu a human uses /cpv-main-menu.

Add a new component (skill / agent / command / hook / mcp) to an existing plugin. Use when scaffolding a single component into a plugin without re-running the generator. Used dynamically via the-skills-menu (TRDD-478d9687).

Add plugin dependencies to a target plugin (explicit --add specs or --from copy from another plugin's plugin.json). Use when adding/copying plugin.json::dependencies entries with atomic rollback on regression. Used dynamically via the-skills-menu (TRDD-478d9687).

Add a new hook entry to hooks/hooks.json in an existing plugin (idempotent — skips duplicate entries; cross-platform-aware). Use when adding a new event-handler that must run identically on Linux/macOS/Windows. Used dynamically via the-skills-menu (TRDD-478d9687).

Schema reference for the /cpv-batch-fix parallel-shard fix protocol — manifest format, status format, planner/aggregator contracts. Use when implementing a new consumer of the batch protocol or extending the planner/aggregator. Used dynamically via the-skills-menu (TRDD-478d9687) — any CPV agent can invoke.

Bump plugin version and run the full publish pipeline (plugin.json, pyproject.toml, README badge, CHANGELOG, push, release). Use when bumping version + publishing the current plugin via publish.py. Used dynamically via the-skills-menu (TRDD-478d9687).