agent-skills

Role-diverse local council. Dispatches multiple expert roles (Devil's Advocate, Integration Checker, Domain Experts, Synthesis Lead) within one runtime/model family to review, audit, or brainstorm. Complements Model Council, Runtime Council, and Deep Council; use when role diversity and local adversarial debate are needed.

Shared contract for all bridge adapters — pre-flight SOP, standardized input/output schemas, artifact format, and status semantics. Read by any bridge or orchestrating skill. Not invocable standalone.

Reference adapter for OpenCode — model-agnostic multi-provider bridge. Read by any orchestrating skill via the Read tool. Covers pre-flight checks with interactive advisory, HTTP API server path (preferred), CLI run path as fallback, correct flags and model format embedded. Usable by deep-council, deep-review, deep-audit, or any future skill that needs multi-model review.

Shared vocabulary for Agent Council, Model Council, Runtime Council, and Deep Council. Use before designing or invoking council workflows so role, model, runtime/toolchain, and debate-layer diversity are not conflated.

Generic structured adversarial protocol for review, audit, research synthesis, and brainstorm/design councils. Supports finding validation plus proposal brainstorming lifecycles, packetized exchanges, and auditable manifests for nested councils.

Multi-agent standards and compliance auditing with pass/fail verdicts. Checks against security, accessibility, code standards, regulatory requirements, and performance benchmarks.

Council-of-councils orchestrator. Dispatches discovery-first packets to bridge adapters (Claude, Gemini, Codex, OpenCode), each of which runs a local Agent/Model/Runtime Council, then performs cross-council debate using debate-protocol. Supports review, audit, research, and brainstorm/design modes.

Git-based codebase exploration with delta analysis. Performs full exploration on first run, then incremental delta exploration tracking committed and uncommitted changes.

Generic multi-domain research framework with domain-aware scheduling. Model-agnostic and domain-agnostic - perform comprehensive research on any topic.

Multi-agent quality improvement review with constructive feedback. Provides suggestions for best practices, code quality, alternatives, and performance optimization.

Generic multi-agent verification framework with balanced expert analysis. Model-agnostic and domain-agnostic - verify any work through expert analysis.

Reference library of domain definitions used by deep-* skills to select appropriate expert agents. Not invocable standalone — read via the Read tool by context, deep-audit, deep-verify, deep-review, deep-research, deep-explorer, and deep-council.

Reference adapter for Claude (Anthropic) dispatch. Read by any orchestrating skill via the Read tool. Defines how to invoke Claude sub-agents or the Anthropic API, with availability checks. Usable by any AI orchestrator — Claude Code, OpenCode, Codex, Gemini, or custom agents.

Reference adapter for Codex multi-agent review. Read by any orchestrating skill via the Read tool. MCP server path (preferred) with auto-setup option, CLI path as fallback, interactive pre-flight advisory when not configured, correct flags embedded. Usable by deep-council, deep-review, deep-audit, or any future skill that needs Codex-based review.

Reference adapter for Gemini CLI. Read by any orchestrating skill via the Read tool. Defines how to invoke Gemini CLI in non-interactive mode, timeout estimation, and fallback behavior. Usable by deep-council, deep-review, deep-audit, or any future skill that needs Gemini-based review.

Analyze conversation context to classify what is being discussed, detect relevant domains from domain-registry, and determine optimal routing (parallel-workflow vs debate-protocol vs deep-council). Can be invoked directly or called by other skills as a pre-flight step.

Lightweight scope clarifier. Invoked by deep-* skills when conversation context is too sparse to determine what to analyze. Asks 1–3 targeted questions one at a time, then returns a structured scope_clarification block for the calling skill to use. Adapted from superpowers brainstorming skill (https://github.com/obra/superpowers) — same principle of asking before acting, scoped to review/analysis intent rather than feature design.

Use when executing multiple tasks with dependencies - automatically determines optimal parallel execution order while respecting task dependencies through DAG scheduling

Optional web search enrichment via Brave Search MCP. Provides web, news, and local search with an independent index (no Google dependency). Invokable standalone or called by deep-research, deep-explorer, or any skill needing web search. Non-blocking — falls back to other search tools if unavailable.

Optional AI-synthesized web search via Perplexity MCP. Returns AI-grounded answers with inline citations rather than raw search results — best for synthesis-heavy questions, current consensus queries, and cross-validation of research findings. Non-blocking — falls back to other search tools if unavailable.

Optional codebase intelligence skill — wraps Devin's DeepWiki MCP server to provide structured documentation lookup, topic discovery, and natural-language Q&A about any GitHub repository indexed by Devin. Supports public and private repositories (private requires Devin API key). Falls back gracefully to Glob/Grep/Read + QMD if not configured. Invokable standalone or called by other skills as a data source.