Skip to content

doodooms/agentic-core

v0.3.5

Canonical agents, skills, MCP integrations, and Expertise Pack authoring for Plugin Factory.

GitHub

Agentic Core

This Agent Plugin bundles the canonical agents, domain skills, and their MCP integrations.

Control is an optional capability detected from available Control MCP tools. Agentic Core has no manifest dependency on Control or the separate substrat-control plugin; workflows without a durable Control requirement continue with local tools, while workflows requiring durable Control semantics are blocked when those tools are unavailable.

The multi-harness skill coordinates bounded Copilot/Codex handoffs. Its two host workflows cross-read each other before communication; see skills/multi-harness/workflows/.

The Orchestrator can also use skills/orchestration/workflows/chatgpt-work-handoff.md to prepare a verified GitHub PR for a ChatGPT Work event-triggered status report. Codex/Copilot retain implementation and validation ownership; the Work task is read-only and returns a discussion summary.

Packaged lifecycle CLI

Build the installable package from the Plugin Factory repository root with uv run --no-project python scripts/materialize_agentic_core.py. The package is written to .build/agentic-core, which is the source used by the local marketplace. The build materializes canonical expertise/ support code into runtime/expertise/ for pluginctl; the source plugin itself contains no Factory implementation.

Run uv run --script /absolute/path/to/.build/agentic-core/runtime/pluginctl_cli.py from any working directory, passing explicit --store-root and (when needed) --workspace-root paths. The script declares Python 3.10+ and PyYAML 6.x; a direct python invocation works when PyYAML is already installed.

Local Expertise Pack authoring

Scaffold and smoke-test an Expertise Pack directly in the Plugin Factory repository root (the pack workspace root), without a Control Plane connection, manifest, or history paths:

uv run --no-project --with pyyaml python -m expertise scaffold local-workflow-check --name "Local Workflow Check" --description "A locally authored Expertise Pack." --capability local.check --skill-id local-check --skill-description "Check local authoring." --publisher "Local Workspace" --source "Workspace authoring" --target portable
uv run --no-project --with pyyaml python -m expertise test local-workflow-check

The scaffold is written under expertise/packs/local-workflow-check/; the test validates and smoke-compiles its declared targets in memory.

Rebuild agent projections

From the Plugin Factory repository root, regenerate and verify the Copilot, Codex, or Claude projection from the canonical Agentic Core source. These commands write only under the selected output directory; Codex profile installation is a separate explicit operation.

uv run --no-project --with pyyaml python scripts/project_plugin_agents.py copilot --source-root agentic-core --output-dir agentic-core/com.github.copilot/agents
uv run --no-project --with pyyaml python scripts/project_plugin_agents.py copilot --source-root agentic-core --output-dir agentic-core/com.github.copilot/agents --check
uv run --no-project --with pyyaml python scripts/project_plugin_agents.py codex --source-root agentic-core --output-dir agentic-core/projections/codex
uv run --no-project --with pyyaml python scripts/project_plugin_agents.py claude --source-root agentic-core --output-dir agentic-core/projections/claude
uv run --no-project --with pyyaml python scripts/project_plugin_agents.py claude --source-root agentic-core --output-dir agentic-core/projections/claude --check

GitHub App MCP

The github-mcp-server MCP is started by the plugin host as a dedicated Docker container using stdio. GitHub App authentication is supported by the official server in stdio mode; the HTTP mode does not support this authentication flow.

The canonical launcher asset is runtime/mcp/github-app-stdio.sh, referenced by mcp.json. Before starting Copilot, Codex, or Claude, make these bindings available in the host process environment:

  • GITHUB_APP_ID
  • GITHUB_APP_INSTALLATION_ID
  • GITHUB_APP_PRIVATE_KEY_PATH: absolute path to the App's PEM file on the host

The launcher requires Docker, passes the App ID and installation ID to the container, and mounts the PEM file read-only at /secrets/github-app.pem. It uses the pinned image ghcr.io/github/github-mcp-server:v1.12.2. No credential values or private key files belong in this package. Do not use GITHUB_APP_PRIVATE_KEY inline.

After setting the environment, start or restart the host/MCP session so it can launch the server. If the server or its tools are unavailable, inspect the host's MCP diagnostics and fix the host configuration; agents must report the GitHub operation as blocked rather than inspect credentials or switch to another GitHub identity.

Agent access

  • Orchestrator uses the GitHub MCP surface for remote GitHub API and lifecycle operations; local git remains for local repository and worktree operations.
  • Researcher has an explicit read-only GitHub MCP allowlist for repository evidence.
  • Other agents do not receive GitHub MCP access by default.