VS Code Agent Extension

The L2M VS Code extension turns any L2M workflow into a coding agent inside the editor. Streaming chat, structured patch previews with workspace-path safety, command approval, and bounded context bundling — backed by a workflow you own and can edit visually.

What you get

• A sidebar L2M Agent view with persistent chat per workspace.
• Streaming responses via POST /api/webhooks/execute/stream (SSE).
• Bounded context bundling: active file, selection, open tabs, diagnostics, git state, pinned files. Configurable budgets.
• Structured action handling: code blocks render with copy/insert, suggested file patches preview before apply with workspace-path validation, suggested commands run only after explicit approval in a VS Code terminal.
• Local session memory with deterministic compaction — older turns get summarised, recent turns sent in full.

Install (development build)

Until the extension ships to the Marketplace, install the .vsix directly.

From CI

Every CI run on main produces a vscode-l2m-agent-vsix workflow artifact. Download the latest one from the Actions tab, unzip, and:

code --install-extension vscode-l2m-agent.vsix

From source

pnpm install
pnpm --filter ./apps/vscode-l2m-agent build
pnpm --filter ./apps/vscode-l2m-agent run package
code --install-extension apps/vscode-l2m-agent/vscode-l2m-agent.vsix

Connect it to your L2M instance

The extension talks to L2M over the standard webhook routes (no special server changes needed). You'll wire up three things:

1. Pick or import a workflow — the easiest start is the bundled VS Code Coding Agent template.
2. Mint an API key so the extension can authenticate.
3. Configure the extension to point at your API and the workflow.

1. Import the VS Code Coding Agent template

Open the L2M web UI → Templates → click Use Template on VS Code L2M Coding Agent Flow (in the Agents category). Save it. Note the resulting workflow ID (visible in the editor URL or the workflow header).

The template implements the full structured-response contract the extension expects: message, actions[] (patch + command), codes[], attachments[], context_update, follow_up_question. It uses the built-in echo provider by default — replace the LLM Call provider with OpenAI / Anthropic / Ollama once you've verified the wiring.

2. Mint an API key

In the L2M web UI: Settings → API Keys → Create. Copy the full key (ao_<prefix>.<secret>) — it's only shown once.

3. Configure the extension

Cmd/Ctrl + , to open VS Code Settings, search for L2M Agent. Set:

Setting	Value
l2mAgent.apiBaseUrl	http://localhost:4000 (or wherever your L2M API runs)
l2mAgent.workflowId	The workflow ID from step 1
l2mAgent.authToken	The API key from step 2
l2mAgent.streamResponses	true (recommended)

workflowId is optional — leaving it blank falls back to looking up a workflow by l2mAgent.webhookPath (default vscode-l2m-agent).

4. Open the chat

Click the L2M Agent icon in the Activity Bar (left rail). Type a message in the sidebar chat. You'll see streaming progress events, then the assistant's response. Code blocks come with copy/insert buttons; patch and command actions show a preview before anything touches your workspace.

Commands

Command	Description
L2M Agent: Open Chat	Reveal the sidebar chat
L2M Agent: New Session	Clear visible chat history and reset session memory
L2M Agent: Pin Active File	Attach the current file to subsequent context bundles
L2M Agent: Send Selection	Send the current editor selection as a chat message
L2M Agent: Reset Memory	Clear the local compacted-memory cache (keeps chat history)

Context budget tuning

The extension truncates predictably to fit a model's context. The defaults are conservative; raise these if you have a large-context model (Claude, GPT-4o):

• l2mAgent.maxContextChars — total approximate chars sent per request (default 60000)
• l2mAgent.maxFileChars — per active or pinned file (default 12000)
• l2mAgent.maxGitDiffChars — per git diff (default 20000)
• l2mAgent.maxDiagnostics — max VS Code diagnostics included (default 50)
• l2mAgent.recentTurnCount — recent turns sent in full before older turns are summarised (default 12)

Lower these for small-context local models (8B Llama, etc.).

Safety model

By design:

• No file edit happens without explicit user approval. Every patch action shows a preview with the diff and the target file path. Approval is a modal dialog, not a notification toast.
• Patches cannot write outside the workspace. Paths are validated against the workspace root; absolute paths and .. segments are rejected.
• Commands are visible and approved before execution. The agent's suggested command renders in the chat with a Run button; clicking it runs in a VS Code terminal you can watch. The extension never auto-runs.

Troubleshooting

Streaming hangs or times out — bump l2mAgent.requestTimeoutMs. The extension enforces this on the full stream, not just the initial response.

"Workflow not found" in the chat — the configured workflowId doesn't exist on the connected instance. Open the L2M web UI → Workflows and re-copy the ID.

401 / 403 on every send — the API key is missing, mistyped, or revoked. Re-create one in Settings → API Keys and re-paste.

Patch preview shows the wrong content — the workflow returned a malformed diff. The extension only applies single-file unified diffs; complex multi-file patches are out of scope. Tighten the system prompt in your workflow to constrain the output.

What it isn't

• It does not call OpenAI/Anthropic/etc. directly. The agent's reasoning lives in the L2M workflow you point it at — swap models server-side without touching the extension.
• It does not give the workflow unrestricted filesystem access. Only the bounded context bundle is sent; the extension is a client, not a runtime.
• It does not run shell commands automatically. Approval is required.

Related

• Why L2M? — the agent-composability story the extension plugs into
• Expose a Workflow as an MCP Tool — the inverse direction (other agents calling your L2M workflows)