Inline autocomplete

CoopAI inline autocomplete shows ghost-text suggestions as you type in the editor. Suggestions stream from the Coop API and appear via VS Code's InlineCompletionItemProvider.

The feature ships off by default (coopAI.autocomplete.enabled: false) so GitHub Copilot and other inline providers keep working until you opt in. Coop assigns Mistral Codestral for inline completions — see Model assignments. Your choice is saved globally (User scope) so workspace folders cannot silently override it.

Turn autocomplete off (or back on)

Autocomplete is controlled from Settings → Preferences → Model & chat — there is no toggle in the Coop sidebar header.

Extension UI — Settings → Preferences → Model & chat

1. Open CoopAI Settings (gear icon in the sidebar title bar).
2. Go to Preferences → Model & chat.
3. Check or uncheck Enable inline autocomplete.
4. Click Save model settings.

The Autocomplete row in the read-only assignment list shows On or Off based on that checkbox. Chat, quick actions, and edit mode are always on. Autocomplete always uses the assigned Codestral model and does not follow the model menu in chat — see Model assignments.

Success: With Enable inline autocomplete checked and saved, typing in an eligible file (e.g. .ts) shows ghost text after a short pause. With it unchecked and saved, Coop does not request ghost text.

File — VS Code User settings

Prefer User settings (not Workspace) so the preference stays consistent across folders:

"coopAI.autocomplete.enabled": false

Coop strips workspace overrides on activate and on Save so a folder or .code-workspace file cannot keep autocomplete on or off against your Settings choice.

Extension UI — Command Palette (optional)

Run CoopAI: Toggle Autocomplete to flip the global enabled flag without opening JSON settings.

Explicit opt-out is remembered (coopAI.autocomplete.userDisabled in extension global state) so index-discovery toasts do not re-enable autocomplete after you turn it off.

Prerequisites

• Signed in under Settings → Account (Google, email, or SSO)
• File type is supported (code files; sensitive files such as .env are skipped)

How it works

flowchart LR
  A[Type in editor] --> B[Trigger detector]
  B --> C[Completion router]
  C --> D["POST /v1/completions/inline"]
  D --> E[SSE stream]
  E --> F[Ghost text in editor]
  F --> G{User action}
  G -->|Tab| H[Accept]
  G -->|Escape| I[Reject]

1. Context extraction — Prefix, suffix, indentation, and surrounding lines from the open buffer.
2. FIM (fill-in-the-middle) — When coopAI.autocomplete.useFim is true (default), the extension sends segments: { prefix, suffix }. Production routing uses Codestral (Mistral) when the server has FIM keys configured; otherwise it falls back to chat-style completion.
3. SSE streaming — The extension requests stream: true. Tokens arrive incrementally so ghost text can appear before the full completion finishes.
4. Client intelligence — Hot Streak, Smart Throttle, request recycling, and multi-line detection tune when and how requests fire.

Hot Streak

After you Tab-accept a suggestion, autocomplete stays snappy for ~8 seconds (up to 3 keystrokes). Debounce drops to 0–50 ms so the next completion feels immediate.

Smart Throttle

Debounce adapts to your typing speed and rolling p95 latency:

• Fast typing → shorter debounce
• Elevated server latency → longer debounce to avoid wasted requests

Request recycling

If you keep typing while a request is in flight, the extension reuses the in-flight request when the new prefix extends the old one, instead of firing a duplicate call.

Multi-line detection

When the cursor is after {, =>, (, or on an empty line inside a block, the client requests up to 200 tokens (vs 96 for single-line) and allows longer ghost-text spans.

Keyboard shortcuts

Action	macOS	Windows / Linux
Accept suggestion	Tab	Tab
Reject suggestion	Escape	Escape
Next suggestion	Alt+]	Alt+]
Previous suggestion	Alt+[	Alt+[

Next / previous apply only when coopAI.autocomplete.showMultipleSuggestions is true.

Run CoopAI: Show Autocomplete Help from the Command Palette for a quick reference.

Settings

Setting	Default	Description
coopAI.autocomplete.enabled	false	Master switch — persisted at global scope
coopAI.autocomplete.trigger	auto	auto — debounced while typing; manual — Command Palette CoopAI: Trigger Inline Autocomplete only; off — no requests
coopAI.autocomplete.useFim	true	Send FIM segments for Codestral routing
coopAI.autocomplete.useGraphContext	false	Force indexed graph context on; when false, graph is still auto-attached when Deep-Index is ready (see below)
coopAI.autocomplete.debounceMs	300	Pause after typing before auto-trigger (0–2000)
coopAI.autocomplete.requestTimeoutMs	1500	Drop slow requests after this many ms (100–5000)
coopAI.autocomplete.maxSuggestionLength	200	Max characters in one suggestion (8–500)
coopAI.autocomplete.showMultipleSuggestions	false	Cycle distinct ranked suggestions (Alt+[ / Alt+]). Near-duplicates (`
coopAI.autocomplete.projectImports	[]	Extra import paths to bias project-style completions

Advanced coopAI.autocomplete.model presets exist for developer tuning; production inline routing uses the assigned Codestral model regardless. See Extension settings.

GitHub Copilot

When Coop autocomplete is on, Coop automatically disables Copilot inline suggestions (github.copilot.enable) and restores your previous Copilot setting when you turn Coop autocomplete off. No prompt — Copilot chat and other features stay available; only competing inline ghost text is turned off.

When Coop autocomplete is off, Copilot inline behavior is unchanged.

Graph context (indexed repos)

When your workspace repo is Deep-Indexed and index status is ready (SCIP or Zoekt available), Coop automatically attaches graph context to inline completions — even when coopAI.autocomplete.useGraphContext is false (the default).

• The extension sends useGraphContext: true with repoId and file path when the index is healthy
• The API attaches a short slice of dependents and ownership from the indexed graph (150 ms budget)
• Available on all plans when the repo is Deep-Indexed (free orgs: up to 3 repos)

Set coopAI.autocomplete.useGraphContext to true to force graph context on regardless of index health. Leave at false for auto when indexed.

Requires a connected, indexed repo in the admin portal. Set Workspace owner/repo/branch so repoId resolves correctly.

Response header x-graph-context: degraded means the graph slice timed out or was unavailable — completion still works from buffer context.

FIM (fill-in-the-middle)

Traditional completion sends only text before the cursor. FIM sends both prefix (before cursor) and suffix (after cursor) so the model can fill the gap.

Production routing (when segments.prefix is present and useFim is enabled):

1. Mistral Codestral — assigned autocomplete model (codestral-latest) when MISTRAL_API_KEY is configured
2. DeepSeek FIM — fallback when only DEEPSEEK_API_KEY is set
3. Chat fallback — when FIM providers are unavailable

Set coopAI.autocomplete.useFim to false to always use chat-style message requests.

Zero-retention routing

Inline requests use a dedicated path separate from chat:

x-use-case: code-completion-only

See Zero-retention LLM routing.

Telemetry

Event	Where	Purpose
completion.requested	Server	Usage tracking, server-side latency
completion.suggested	Extension	Ghost text actually shown (CAR denominator)
completion.accepted	Extension	Tab accept (CAR numerator)
completion.rejected	Extension	Escape or superseded suggestion
completion.performance	Extension	Batched client p50/p95 snapshots

Org admins can view org completion metrics in the admin portal → Completions tab. Members see personal usage on My Usage → Completions.

Suggested next-step pills in chat (suggest_chip.*, suggest_intent.*, chat_intent.*) follow VS Code's telemetry setting and are not used for Admin usage charts.

Troubleshooting

Problem	Fix
No ghost text	Confirm Enable inline autocomplete is checked in Settings → Preferences → Model & chat and saved; sign in under Settings → Account
Nothing on manual trigger	Enable autocomplete first; Command Palette → CoopAI: Trigger Inline Autocomplete
Slow or missing suggestions	Increase requestTimeoutMs; check network; self-hosted API needs MISTRAL_API_KEY for Codestral FIM
Completions in strings/comments	By design — trigger detector skips comment and string contexts
Graph context empty	Deep-Index the repo in admin portal; confirm index status is ready in Settings → Indexing; check Workspace owner/repo/branch; set coopAI.autocomplete.useGraphContext to true to force on
Save unchecks, then the box checks itself	A workspace file was forcing it on. Coop now clears that on Save. Reload the Extension Host, then save again.

More fixes: Troubleshooting.

API

Direct API usage: API reference — Inline completion.

Next steps

• Model assignments — per-feature models and settings UI
• Extension settings
• Getting started
• Edit mode — /edit patches with apply and undo
• Create pull request — open a PR from the patch card or by asking in chat