Inline Chat & Completions

Inline chat (Cmd+I)

Press Cmd+I (Mac) or Ctrl+I (Windows/Linux) to open an inline prompt directly in the editor.

1. Select code (optional) — if you have a selection, SideCar uses it as context
2. Press Cmd+I and describe the change: “convert to async/await”, “add error handling”, “translate to TypeScript”
3. The proposed replacement streams into a side-by-side diff preview — original on the left, proposal on the right
4. An Accept / Dismiss dialog gates the final apply — nothing lands in your file until you accept

Inline chat uses the same model as the sidebar chat. It supports thinking models (reasoning blocks are shown collapsed). All tool use is disabled — inline chat is a focused edit, not an agent run.

Lightbulb integration

When VS Code shows a diagnostic (red/yellow squiggle), pressing Cmd+. / Ctrl+. surfaces two SideCar actions alongside the built-in Quick Fixes:

• Fix with SideCar — explains the error and proposes a corrective edit via inline chat
• Explain this error with SideCar — sends the diagnostic to the sidebar chat for a detailed explanation

Refactor with SideCar appears in the Refactor submenu (Cmd+Shift+R / right-click → Refactor) for any selection.

Inline completions (FIM)

Copilot-style autocomplete as you type. Opt-in — enable it:

"sidecar.enableInlineCompletions": true

Press Tab to accept a suggestion, Escape to dismiss. Completions are debounced (default 300 ms) with in-flight cancellation so they don’t pile up while you type.

Backend paths

SideCar routes completions differently per backend:

Not all Ollama models support FIM — models without a <|fim_prefix|> / <|fim_suffix|> / <|fim_middle|> token set fall back to the chat API and produce lower-quality suggestions.

Speculative decoding (v0.110+)

When enabled, a small draft model proposes tokens that the main model verifies — delivering main-model quality at draft-model latency.

"sidecar.speculativeDecoding.enabled": true,
"sidecar.completionDraftModel": "qwen2.5-coder:0.5b"

Leave completionDraftModel empty to auto-discover the smallest available model on the active backend. SideCar tracks the rolling accept rate and auto-disables speculative decoding if it falls below sidecar.speculativeDecoding.minAcceptRateToKeepEnabled (default 0.4).

Settings

Tips

• Point sidecar.completionModel at a small fast model (qwen2.5-coder:1.5b, deepseek-coder:1.3b) while keeping a larger model for the sidebar agent — completions are latency-sensitive, chat turns are not
• For Kickstand, sidecar.kickstand.flashAttn: true gives a 2–4× speed boost on long contexts with Metal or CUDA, which helps completion latency
• Lower completionDebounceMs for snappier suggestions at the cost of more GPU pressure

Next Edit Suggestions (v0.72+)

After you edit a symbol, SideCar walks the symbol graph to find callers and dependents that likely need the same update. Candidates appear as ghost-text badge decorations (①②③) at the affected lines; cross-file hits show in the status bar.

Enable it:

"sidecar.nextEdit.enabled": true

Settings: sidecar.nextEdit.{debounceMs, crossFileEnabled, maxHops, topK, autoTriggerOnSave} — see Configuration → Next Edit Suggestions.

Adaptive Paste (v0.72+)

When you paste foreign content into a file, SideCar detects the content type and offers a lightbulb code action to transform it to fit the target language. Detection is heuristic — no LLM call fires until you accept.

Built-in transforms: JSON → TypeScript type, SQL → ORM query, curl → fetch(), CSS → Tailwind classes, Python → TypeScript, shell → child_process, .env → Zod schema.

Enable/disable via sidecar.adaptivePaste.enabled (default true). Set a minimum paste length with sidecar.adaptivePaste.minPasteLength (default 50 chars) to avoid triggering on short snippets.

Code actions

Right-click on selected code (or press Cmd+. / Ctrl+.) to access SideCar code actions:

All actions send to the sidebar agent which has its full tool set available — it can read related files, run tests, and iterate.

@ references

Use @ syntax in chat messages to include specific context:

References are resolved before sending the message to the model. Use them when the automatic workspace context doesn’t include what you need.

Image support

Attach images to chat messages for vision-capable models:

• Click the paperclip button to attach the active file or browse for an image
• Paste a screenshot directly into the chat input
• Supported formats: PNG, JPG, GIF, BMP, WebP, SVG

Useful for describing a UI to build, showing an error screenshot for debugging, or comparing a mockup to current code. Requires a vision-capable model — Claude 3+, GPT-4o, or a local vision model like llava.