CLIProxyApi Inline Completion

Inline completions (ghost text) and reviewable next edit suggestions in VS Code powered by CLIProxyApi.

Requirements

VS Code 1.90 or newer and a running CLIProxyApi server. By default the extension targets http://127.0.0.1:8317. The selected model must support chat completions; Next Edits also requires reliable JSON output.

Install from source

npm install
npm run package
code --install-extension cliproxy-inline-completion-0.1.3.vsix

Usage

Type in an eligible file and suggestions appear as ghost text. Press Tab to accept. Completions use the code before and after the cursor, with compatible IntelliSense ranges and cancellation of obsolete requests.

Run CLIProxy Inline: Set CLIProxyApi URL to change the server. Run Select Inline Completion Model and Select Inline Completion Reasoning Effort to configure completions, and Select Next Edits Model and Select Next Edits Reasoning Effort to configure Next Edits separately; the Next Edits pickers offer Same as inline completions. The default effort is the server/model default; the parameter is only sent when a level is chosen, so providers that do not support it are unaffected. Commands update the setting at the scope where it is already defined (workspace folder, workspace or user).

Next Edits

1. Enable cliproxyInline.nextEdits.enabled in settings.
2. Make a small change, such as renaming a function declaration.
3. Run CLIProxy Inline: Suggest Next Edit. A valid related edit is marked in the source and status bar without changing the file.
4. Use Go To Next Edit to navigate or Preview Next Edit to compare read-only before/after snapshots.
5. Use Accept Next Edit or Reject Next Edit from the Command Palette or editor title actions. Acceptance works from the preview too and is one undoable operation. It never saves the file.

Automatic predictions and Tab navigation are separate opt-in settings. Automatic NES waits 800 ms after editing. If an inline completion is still in flight it waits for it to finish, then requests the prediction; completions and Next Edits never cancel each other. Moving the cursor does not discard a pending prediction, editing the document does. Accepting a next edit immediately requests the following one, with the accepted change as the latest edit, so related changes can be applied one after another; this does not depend on the automatic setting. When a next edit appears, any visible ghost text is hidden and no inline completion is offered until the next edit is accepted, rejected or invalidated, so Tab cannot accept a stale completion instead. With Tab navigation enabled, the first Tab navigates and the second accepts at the destination, unless IntelliSense, a snippet or another conflicting editor mode is active. Escape dismisses an available proposal when no conflicting widget is active. Use the commands if another extension owns these keys.

Next Edits can use a different model from completions: set cliproxyInline.nextEdits.model and cliproxyInline.nextEdits.reasoningEffort. Completions benefit from a small, fast model; Next Edits need a model that follows the JSON contract reliably. Each Next Edits attempt logs one line to the CLIProxy Inline Completion output channel with its outcome and latency (for example ok, null, truncated, anchor-missing, stale), never code or prompts.

This uses stable VS Code APIs, not Copilot's experimental inline-edit renderer. It predicts one contiguous insertion, replacement or deletion in the active file. Ambiguous anchors, invalid JSON, oversized changes and results for an older document version are discarded. Changing the source invalidates its proposal; request a new one instead of applying a stale diff. No recent edits can mean no suggestion.

Settings

Privacy and limits

The local proxy may forward code to remote providers. Local transport is not a guarantee of local inference. Document content is sent only for suggestions; recent edit history is bounded in memory, never stored on disk. Logs omit code, prompts, raw responses and credentials, including HTTP error bodies.

Suggestions require a trusted workspace, an eligible writable file, untitled or vscode-remote document, and one empty selection. Virtual previews, output documents, .env variants, .ssh, common private-key files and credential filenames are excluded. This is not comprehensive secret detection: do not use the extension on sensitive material. The language filter and global toggle apply to both features.

Inline completion context is limited to 24,000 UTF-16 characters including serialized data and prompt allowances, and output to maxTokens. Very long lines may result in no request.

Next Edits has no input or output token limit: it sends the whole active document plus up to 16 recent edits from the last 60 seconds, each as its line number and full changed lines, and does not send max_tokens. The model's context window and requestTimeoutMs are the effective limits; large files cost more and are slower, and a provider may reject a request that exceeds its context. NES tracking is skipped above two million characters. Proposals are limited to 64 lines and 8,000 characters per side. No extra files are sent or indexed.

The chat completion contract approximates fill-in-the-middle behavior; it is not native FIM and does not guarantee Copilot-equivalent quality. A model that answers conversationally or returns unreliable JSON may abstain frequently. A response wrapped in a single json code block is accepted; anything else around the JSON is rejected. systemPrompt customizes cursor completions only; NES has a separate fixed output contract.

Development and verification

npm run test:unit
npm run test:integration
VSCODE_TEST_VERSION=1.90.0 npm run test:integration
CLIPROXY_LIVE_URL=http://127.0.0.1:8317 CLIPROXY_API_KEY=... npm run test:live
npm run package

Unit tests include an in-memory editor adapter and synthetic HTTP server. Integration tests launch an isolated VS Code Extension Host and never send fixture code to an external model. Live tests are skipped unless CLIPROXY_LIVE_URL is set; they send only small synthetic fixtures to the configured models (CLIPROXY_LIVE_MODEL, CLIPROXY_LIVE_NES_MODEL) and allow a few attempts per scenario because model output is nondeterministic. Downloads require access to the official VS Code CDN and Electron requires a display (or Xvfb on headless Linux). See implementation and verification results for actual coverage and remaining checks.

License

MIT