Couplet

An agentic coding assistant that runs entirely inside VS Code (and VS Code-based editors like Cursor), powered by SAP AI Core. No separate server process — the agent loop, tools, and streaming chat client all run in the extension host. The only network calls it makes are to your own SAP AI Core endpoint.

Install

You need one file: couplet-<version>.vsix. Pick whichever is easier.

From the editor (no terminal):

1. Open the Extensions view (Cmd/Ctrl+Shift+X).
2. Click the ... menu at the top of that panel → Install from VSIX…
3. Select the .vsix file.
4. Reload the window when prompted (Cmd/Ctrl+Shift+P → Developer: Reload Window).

From a terminal:

code --install-extension couplet-0.0.22.vsix     # VS Code
cursor --install-extension couplet-0.0.22.vsix   # Cursor

Then reload the window.

To upgrade later, install the new .vsix the same way — the version number must be higher, otherwise the editor keeps the cached copy.

Set up credentials

Everything is entered in the panel; you never have to edit settings.json.

1. Click the Couplet icon in the activity bar (left edge).

2. Click the gear icon in the panel header.

3. Fill in the fields from your SAP AI Core service key:

   Field	Where it comes from
   Client ID	clientid in the service key
   Client secret	clientsecret in the service key
   AI Core base URL	serviceurls.AI_API_URL
   Auth URL	url
   Resource group	usually default

   There is no deployment ID to find: Couplet lists the running deployments in your resource group and resolves the right one from the model you pick.

4. Click Test connection. It lists your deployments and then sends one trivial message, so you know credentials, endpoint, and streaming all work before you start a real task rather than finding out midway through one.

5. Close settings (✕), pick a model from the composer, and send a message.

The client secret is stored in your OS keychain via the editor's SecretStorage, not in settings.json. The other values are plain configuration, written to your user settings so they follow you into every workspace.

Using it

Type a request and press Enter. The agent reads files, searches, edits, and runs commands to complete the task.

• Approval mode — the Ask/Auto pill in the composer. Ask (default) prompts before every shell command. Auto runs a small allowlist of read-only and test commands unattended and still prompts for everything else.
• Always allow — the caret next to Approve offers a standing approval for that kind of command, e.g. npm run, either everywhere or in this workspace only. See below for what is and is not eligible.
• Actions fold away — while the agent works, each tool call is listed as it runs. Once the agent replies, the whole run collapses into one line (6 actions · read_file, bash) that expands on click, so a long turn does not bury the answer.
• Rewind (⟲) — hover any of your messages to undo that turn: the files it changed are restored and the chat is truncated back to that point.
• History — the clock icon lists past sessions in this workspace; each row has a delete button.
• Touched files — chips above the composer link to a diff, with a revert button per file.

Standing approvals

The Always allow option stores a pattern — the program plus a bare subcommand, so approving npm run build grants npm run, not npm. Four rules bound it:

• Anything carrying shell metacharacters (&&, |, ;, backticks, $(, >) is never eligible, so a grant for npm test can never cover npm test && curl … | sh.
• Anything the policy treats as destructive or noteworthy — rm, sudo, git push, curl, wget, redirects, absolute paths — is never eligible. The caret simply is not offered.
• Every rule is re-checked when a command is matched, not only when the pattern was stored, and matching is on the derived pattern rather than a string prefix. npmfoo run does not match npm run.
• Neither scope can be reached by a repository. Always allow writes to your user settings, where you can see and revoke it; Allow in this workspace is kept in the extension's own per-workspace storage, outside the repo. couplet.alwaysAllow is honoured only from your user settings — a workspace value is ignored, so a cloned repository cannot ship a .vscode/settings.json granting itself permission to run its own commands unattended.

Keeping the agent away from certain files

Add a .coupletignore to the workspace root, same syntax as .gitignore. Listed paths are excluded from read_file, search_replace, grep, and list_dir. .gitignore is honoured too. Note this does not restrict shell commands the agent runs.

What gets written to your workspace

A .couplet/ directory holding session transcripts (sessions/*.jsonl) and rewind checkpoints (checkpoints/). Checkpoints contain copies of file contents, so add .couplet/ to your .gitignore.

Configuration reference

Most people never need these — the settings panel covers the common ones.

The client secret is deliberately absent from this table — it lives in SecretStorage and has no settings.json key.

There is no deployment ID setting to fill in. Couplet lists the running deployments in your resource group and picks the one serving your chosen model, re-resolving automatically if a redeploy issues a new id. If your tenant somehow exposes a deployment that cannot be discovered, adding "couplet.deploymentId" to settings.json by hand still pins that id — it is read but intentionally unlisted, so it never shows up as a setup step.