Using context-kit with GitHub Copilot

context-kit is a plugin marketplace that GitHub Copilot CLI installs from directly — the same plugins Claude Code and APM use, from one shared source. No manual copying of skill folders.

GitHub Copilot CLI installs via copilot plugin; Claude Code installs via /plugin and APM via apm install. The retrieval SKILL.md bodies, their references/, the agents, and the commands are identical across hosts.

Install for GitHub Copilot

Register this marketplace, then install the plugins you want:

copilot plugin marketplace add mbeacom/context-kit
copilot plugin install code-search@context-kit      # auto-installs retrieval-core
copilot plugin install indexkit@context-kit
copilot plugin install obsidian@context-kit
copilot plugin install adr-bridge@context-kit
copilot plugin install plan-execute@context-kit
copilot plugin install context-steering@context-kit
copilot plugin install verify@context-kit           # auto-installs retrieval-core
copilot plugin install runtime-evidence@context-kit # pulls verify, then retrieval-core
copilot plugin install context-handoff@context-kit  # pulls verify, then retrieval-core
copilot plugin install memory@context-kit           # pulls handoff, verify, retrieval-core
copilot plugin install corpus-review@context-kit  # pulls plan-execute, verify
copilot plugin install deep-review@context-kit    # pulls plan-execute, verify
copilot plugin install token-economics@context-kit # pulls verify, retrieval-core
copilot plugin install plugin-forge@context-kit

runtime-evidence and context-handoff depend on verify; verify transitively pulls retrieval-core. memory depends on context-handoff.

Manage them with copilot plugin list, copilot plugin update <name>, and copilot plugin uninstall <name>.

What you get

Repository asset	GitHub Copilot	Claude Code
plugins/*/skills/<name>/SKILL.md (+ references/)	Installed via copilot plugin install	Installed via /plugin install
plugins/retrieval-core/agents/retrieval-strategist.md	Installed with the plugin	Installed as a subagent
plugins/indexkit/bin/indexkit	Auto-bootstrapped by the plugin's SessionStart hook	Auto-bootstrapped by the same hook
plugins/memory/scripts/memory-provider.py	Explicit commands, plus opt-in hooks — Copilot loads hooks/hooks.json	Explicit commands plus the same opt-in hooks
.claude-plugin/* manifests	Used to resolve the marketplace	Marketplace packaging

Copilot's discovery-critical skill frontmatter is name and description, which this repo's skills keep compatible.

Using the skills

Once installed, ask Copilot naturally, for example:

• "Use the retrieval strategy to find where retry backoff is handled."
• "Use code-search to find structural React useEffect cleanup issues."
• "Use indexkit to query my notes for billing open questions."
• "Use the Obsidian RAG bridge to search notes linked to Project X."
• "Analyze the prospective impact of changing this schema."
• "Collect bounded runtime evidence for this unable-to-check claim."
• "Write a context handoff for the next session."
• "Recall why this project changed its retry policy, then verify current evidence."

Bootstrapping the indexkit runtime

indexkit's CLI runs on a uv-managed venv. Both Claude Code and GitHub Copilot CLI bootstrap it automatically from the plugin's SessionStart hook, so no manual step is normally needed.

Bootstrap it yourself only when the venv is missing or stale — on APM, which does not deploy hooks, or after an indexkit upgrade. Check first:

bash plugins/indexkit/scripts/bootstrap.sh --check   # exit 0 ready, 3 needs bootstrap

The fix depends on which one it reports. When the venv is missing, the quickest route needs no clone — the launcher falls back to an indexkit on PATH:

pip install indexkit          # or: uv tool install indexkit

When it reports stale, that shortcut does not apply: the launcher prefers an existing venv over PATH, so a packaged install would sit unused behind the stale copy. Rebuild the venv from a clone of this repo:

export CONTEXT_KIT_DATA="$HOME/.local/share/context-kit"
bash plugins/indexkit/scripts/bootstrap.sh
export PATH="$PWD/plugins/indexkit/bin:$PATH"
ollama pull nomic-embed-text
indexkit index /path/to/vault --name notes
indexkit query "open questions about billing" --name notes --k 8
indexkit query "open questions about billing" --name notes --k 8 --hybrid

Supported neutral environment variables:

Variable	Purpose	Claude fallback
CONTEXT_KIT_DATA	venv and index storage for indexkit	CLAUDE_PLUGIN_DATA
CONTEXT_KIT_EMBED_MODEL	ollama embedding model	CLAUDE_PLUGIN_OPTION_EMBED_MODEL
CONTEXT_KIT_OLLAMA_HOST	ollama base URL	CLAUDE_PLUGIN_OPTION_OLLAMA_HOST
CONTEXT_KIT_OBSIDIAN_VAULT	vault path used by examples/fallbacks	CLAUDE_PLUGIN_OPTION_VAULT_PATH
CONTEXT_KIT_RUNTIME_EVIDENCE_CONFIG	user-owned exact-ID JSON command allowlist	—
CONTEXT_KIT_RUNTIME_EVIDENCE_ROOT	installed runtime-evidence root	CLAUDE_PLUGIN_ROOT
CONTEXT_KIT_HANDOFF_PATH	handoff artifact override	—
CONTEXT_KIT_MEMORY_PROVIDER	none or optional mempalace provider	CLAUDE_PLUGIN_OPTION_PROVIDER
CONTEXT_KIT_MEMORY_HOME	reviewed records and project-isolated provider data	CLAUDE_PLUGIN_OPTION_MEMORY_HOME
CONTEXT_KIT_MEMORY_PROJECT	explicit durable-memory project scope (overrides Copilot auto-detection)	CLAUDE_PLUGIN_OPTION_PROJECT
CONTEXT_KIT_MEMORY_AUTO_CAPTURE	opt-in Claude lifecycle forwarding	CLAUDE_PLUGIN_OPTION_AUTO_CAPTURE
CONTEXT_KIT_MEMORY_ROOT	installed memory plugin root	CLAUDE_PLUGIN_ROOT
CONTEXT_KIT_TOKEN_ECONOMICS_ROOT	installed token-economics root	CLAUDE_PLUGIN_ROOT
CONTEXT_KIT_CLAUDE_PROJECTS	Claude Code transcript dir to read	—
CONTEXT_KIT_COPILOT_DB	Copilot CLI session store to read	—

Copilot plugin installs discover the memory plugin's stdio MCP server automatically. GitHub Copilot Desktop has no separate per-project MCP environment surface, so the memory hook creates a narrowly scoped session binding from trusted host data: SessionStart payload sessionId must exactly match COPILOT_AGENT_SESSION_ID, and payload cwd must resolve to a Git top-level with a canonical github.com/owner/repository origin. On POSIX, the binding contains only the session ID and normalized project, is mode-0600 below a mode-0700 directory, and is removed by a matching SessionEnd.

Explicit configuration still wins, with resolution order --project → CONTEXT_KIT_MEMORY_PROJECT → deprecated PRODUCTIVITY_SKILLS_MEMORY_PROJECT → CLAUDE_PLUGIN_OPTION_PROJECT → trusted Copilot binding. MCP cwd, PWD, prompt content, and tool arguments never select scope. Hosts without Copilot's matching hook/session-ID support still refuse until a project is configured. Windows, other Git hosts, and deeper namespaces also require explicit configuration because automatic routing cannot verify the owner-only ACL or preserve a unique identity there.

The Claude-specific variables remain supported so existing plugin installs keep working. The CONTEXT_KIT_* names are preferred for portable docs and shell profiles shared across agents. The former PRODUCTIVITY_SKILLS_* names (from before the marketplace was renamed to context-kit) are still honored as a deprecated alias — resolution order is CONTEXT_KIT_* → PRODUCTIVITY_SKILLS_* → the Claude fallback — but will be dropped in a future release; migrate when convenient.

Tooling expectations

Neither host installs the underlying CLI tools for you. Install the same tools the skills expect:

• Required for code-search: rg (ripgrep).
• Optional but useful: fd, ast-grep/sg, semgrep, jq, yq, gron, duckdb, sqlite-utils, rga, pandoc, pdftotext, difftastic, tokei, scc, and rtk.
• Required for indexkit: uv, ollama, and an embedding model such as nomic-embed-text. SQLite FTS5 is required only for opt-in --hybrid; the default semantic mode remains available without it.
• Required for runtime-evidence and context-handoff: Python 3. Their runner and validator use only the standard library. The runtime runner requires POSIX and refuses Windows before execution; the handoff validator is cross-platform. Runtime evidence's optional-tool path ships no runner, so it needs neither Python nor POSIX — only an approved, host-exposed observation tool. Copilot does not provide universal host-level command enforcement; runtime collection on the runner path remains bound to the plugin's reviewed exact-ID allowlist, while the optional-tool path is bound by operator approval plus host policy, with nothing enforced by the plugin.
• Optional for obsidian-rag-bridge: the official obsidian CLI with Obsidian running; otherwise use the rg/fd fallback over vault files.
• Required for local memory records: Python 3.10+ and Git on PATH. Optional provider recall uses a separately installed mempalace CLI. Copilot CLI does load the plugin's hooks/hooks.json, so opt-in session-start recall and boundary capture work here; both stay off until explicitly enabled.

Run plugins/code-search/scripts/check-tools.sh from a clone of this repository to see what is already installed and what brew install ... command would fill the gaps.

Authoring portable skills

When updating this repo, keep the reusable retrieval instructions agent-neutral:

1. Put cross-agent workflow knowledge in SKILL.md and references/.
2. Keep Claude marketplace mechanics in .claude-plugin/, hooks, and Claude-only docs.
3. Prefer CONTEXT_KIT_* in portable examples, with CLAUDE_PLUGIN_* documented as the Claude fallback.
4. State host behavior precisely. Claude Code and GitHub Copilot CLI both load a plugin's hooks/hooks.json; APM does not deploy hooks, so note when a step is automatic on the hook-running hosts but manual on APM.
5. Keep command output compact when possible (rtk is optional and safe to omit).