Skip to content

copilot-install.md

Latest commit

 

History

History
148 lines (106 loc) · 6.18 KB

copilot-install.md

File metadata and controls

148 lines (106 loc) · 6.18 KB
Raw

Installing LemonCrow into Copilot

Support level: MCP config + Copilot instructions + agent + tasks

Quick Install

make install

By default this installs VS Code user/global MCP, instructions, and task presets. For project-local Copilot artifacts:

bash scripts/install_copilot.sh --workspace /path/to/workspace

To configure project-specific role models after the global install, run:

uv run lemoncrow init --configure-models

Inside a git workspace, the init wizard writes <workspace>/.lemoncrow/settings.json and refreshes the workspace Copilot agent at <workspace>/.github/agents/lemoncrow.code.agent.md. Those workspace role agents are the surface Copilot uses for the current project, so it can carry explicit model pins even when the global install stays neutral.

What Gets Installed

Artifact Global install --workspace DIR install
MCP server config VS Code user mcp.json <workspace>/.vscode/mcp.json
Copilot instructions ~/.copilot/instructions/lemoncrow.instructions.md <workspace>/.github/copilot-instructions.md
Agent not installed globally <workspace>/.github/agents/lemoncrow.code.agent.md plus the other lemoncrow.<role>.agent.md files
Task presets VS Code user tasks.json (merged) <workspace>/.vscode/tasks.json (merged)

The MCP config registers LemonCrow as a stdio server:


## Project-local model config

The intended flow is:

1. Install LemonCrow globally so Copilot can see the MCP server and default instructions.
2. Run `uv run lemoncrow init --configure-models` inside a repository.
3. Let the wizard write `.lemoncrow/settings.json` plus a workspace-local
   `.github/agents/lemoncrow.code.agent.md` and the other `lemoncrow.<role>.agent.md` files.

`"auto"` means omit an explicit model pin for that host surface. Concrete values
write the selected model directly into the workspace agent frontmatter.json
&#123;
  "servers": &#123;
    "lc": {
      "type": "stdio",
      "command": "lc mcp",
      "args": ["--host", "copilot"],
      "env": {
        "LEMONCROW_WORKSPACE_ROOT": "<workspace>"
      }
    }
  &#125;
&#125;

Verify

make verify

First Task

Open Copilot Chat and either run LemonCrow: Copilot Preflight from the VS Code task picker or, if developer mode is enabled for the MCP server, ask:

Use LemonCrow context for this task and record a trace summary when it is done.

Additional workspace helpers:

  • LemonCrow: Session Summary prints the latest LemonCrow per-session cost and savings breakdown after a Copilot task or chat run.
  • LemonCrow: Worktree Bootstrap writes .env.worktree with stable per-worktree ports and container names.
  • LemonCrow: Runtime Evidence captures health, analytics/summary, and v1/traces into reports/runtime-evidence/latest.json.

Expected Behavior

  • Copilot Chat can invoke LemonCrow MCP tools through the local LemonCrow MCP wrapper
  • copilot-instructions.md provides LemonCrow context to every Copilot session
  • lc agent is available from the agent picker
  • The installed Copilot instructions and agent explicitly map native search/codebase, search, edit/editFiles, and execute/runInTerminal back to LemonCrow MCP equivalents
  • LemonCrow: Copilot Preflight runs the shell-based preflight before you continue in Copilot Chat
  • LemonCrow: Worktree Bootstrap makes local stacks easier to boot from multiple worktrees
  • LemonCrow: Runtime Evidence provides a repeatable validation artifact for service behavior
  • Active context/retrieval/verify tools require LEMONCROW_DEV_MODE=1; otherwise some tools may be visible but return passive noop
  • Projection links in the session UI open /projection?path=..., where Copilot users can inspect compact vs exact views, token savings, and mapping segments before retrying an edit

Why Tasks Still Use Shell

Copilot's MCP support applies to chat/tool calls inside the Copilot session. VS Code tasks.json entries are shell tasks, so the preflight task has to spawn the lc CLI rather than invoke MCP directly. The MCP server remains the primary integration surface for in-chat context, trace, memory, and related LemonCrow tools.

Projection Workflow

When Copilot reads code through LemonCrow, compact reads can carry projection_mapping metadata. If a projection edit is ambiguous, the failure now returns machine-readable retry_with guidance. Follow that reread instead of guessing a transformed span.

Reload Required

After install, reload the VS Code window:
Ctrl+Shift+P → Developer: Reload Window

Troubleshooting

Problem Fix
MCP tools not loading Reload VS Code window; check user mcp.json or workspace .vscode/mcp.json
code CLI not found Install VS Code CLI: in VS Code, run "Install 'code' command in PATH"

MCP Tools and Dev Mode

With developer mode enabled (LEMONCROW_DEV_MODE=1), the active LemonCrow MCP surface for Copilot includes all stable tools plus dev-only tools such as rescue and verify.

Without developer mode, the full stable surface is still active, including code, while dev-only tools remain passive/noop.

VS Code's native search, file reads, and editing remain the preferred raw-access tools when you do not need LemonCrow-specific context or trace behavior.

Disable LemonCrow cache: LEMONCROW_CACHE_DISABLED=1.

Uninstall

bash scripts/uninstall_copilot.sh
bash scripts/uninstall_copilot.sh --workspace /path/to/workspace