wcode Agent Plugin

This is a portable Agent Plugins 1.0 skill package. It intentionally does not contain mcp.json.

wcode must be started with the actual repository as an explicit Workspace. Agent Plugins stdio commands default to the plugin root and the portable specification has no host-workspace variable, so silently bundling wcode mcp-stdio would scope wcode to the wrong directory.

Install the Skill using one of the concrete project-local options below, then configure the wcode MCP server separately for the repository you are working in.

Install the Skill

Shared Agent Skills path — Codex, Gemini CLI, Grok Build, and compatible agents

mkdir -p .agents/skills
cp -R wcode-agent-plugin/skills/wcode-software-intelligence .agents/skills/

Claude Code

mkdir -p .claude/skills
cp -R wcode-agent-plugin/skills/wcode-software-intelligence .claude/skills/

For local plugin development without copying the Skill:

claude --plugin-dir ./wcode-agent-plugin

Gemini CLI

gemini skills link ./wcode-agent-plugin/skills/wcode-software-intelligence --scope workspace
gemini skills list

Grok Build

Grok reads .agents/skills/ and Claude-compatible skills directly. To test the whole plugin package without installing it:

grok --plugin-dir ./wcode-agent-plugin
grok inspect

ZCode

Add this directory as a local marketplace, then install and enable wcode. The ZCode-specific manifest starts wcode --workspace ${CLAUDE_PROJECT_DIR} mcp-stdio, so the open ZCode project remains the explicit Workspace and HTTP OAuth is not used for this local child process.

GitHub Copilot CLI

copilot plugins install --skill --scope project \
  ./wcode-agent-plugin/skills/wcode-software-intelligence/SKILL.md

The generated package contains instructions only: no hooks, executable scripts, credentials, or bundled MCP configuration. Keep the agent's own sandbox, trust, and approval controls enabled.

Configure wcode MCP separately

Recommended local MCP command shape:

wcode --workspace <ABSOLUTE_REPOSITORY_PATH> mcp-stdio

The stdio transport uses the same Workspace policy, Harness, Software Intelligence, Tasks, Prompts, Resources, and tools as the HTTP server. It does not use HTTP OAuth because the local child-process boundary is the transport trust boundary.

Repository-aware execution is still explicit. With process-wide --allow-risky-exec off, an exact risky operation can stop with a local authorization request; the operator approves it in the wcode TUI and retries. The Skill must never auto-approve that request. Use the flag only when the operator intentionally pre-authorizes repository-aware execution for the whole process.

The shared MCP runtime also exposes wcode Product Scopes. Start with workspace_info / scope_status / project_context; use scope_status.unmapped_files to surface architecture gaps, then pass relevant scopes to software_context when a task belongs to a bounded capability area. Product Scopes narrow context; they never widen Workspace or execution permissions. Clients that read MCP Resources can inspect wcode://runtime/product-scopes.

For source changes, language_quality_status reports language support as a capability matrix rather than one boolean: syntax, semantics, repository-declared/native formatter/linter/type/static/test/security providers, plus Property/Mutation/Fuzz/Runtime stages. language_quality_run executes only a declared, available, check-only provider through the normal authorization boundary and records current-revision Evidence; it never uses formatter fix/write mode.

The Skill treats always-on instructions as a short map and loads detailed context progressively. Host subagents/worktrees can be used for isolated independent research or review, but shared/dependent writes still belong behind wcode's Scheduler and SHA guards, and model consensus is never deterministic proof.

The Skill also carries wcode's maintainability gate: review_changes may emit deterministic maintainability-* findings, and medium-and-higher risk Verification Plans require an independent maintainability_review job. A correctness Pass does not replace that structural review.