Shire Autocompletion
Fast inline code autocomplete for VS Code, powered by your own model endpoint — not a
hosted SaaS. Point it at any OpenAI-compatible server (vLLM, llama.cpp, Ollama, TGI, SGLang, or a
cloud endpoint) running a FIM-capable code model (Qwen2.5-Coder, Codestral, DeepSeek-Coder,
StarCoder2) and get Copilot-style ghost-text completions.
Features
- Inline FIM completions with the correct fill-in-the-middle prompt format — streaming, low
latency, debounced, request-cancelling.
- Recent-edit refactor assist — feeds your recent before→after edits into the prompt, so after
you change something in one file, it proposes the analogous change as you type in another.
- Cross-file repo context — indexes your workspace (BM25 lexical by default) and injects the
most relevant snippets. Optional semantic retrieval via your own embedding model.
- Private by default — your code only ever goes to the endpoint you configure. API key is
stored in VS Code's encrypted SecretStorage.
- Sidebar UI — configure everything and watch live index/latency status, no JSON editing.
Quick start
- Install the extension.
- Open the Shire Autocompletion panel from the Activity Bar (sparkle icon).
- Set Base URL (e.g.
http://localhost:8000/v1), API Key (if any), and Model.
- Start typing in a code file — grey ghost text appears; press Tab to accept.
Your endpoint must expose OpenAI-compatible /v1/completions (legacy text completion, not
chat) with stream and stop support. Serve a FIM/base code model for best infill.
Model compatibility (important)
Autocomplete needs a fill-in-the-middle (FIM) code model, and the FIM token
format must match the model family. Set fimTemplate accordingly:
fimTemplate |
Models |
qwen (default) |
Qwen2.5-Coder, Qwen3-Coder |
codestral |
Mistral Codestral |
deepseek |
DeepSeek-Coder |
starcoder |
StarCoder, StarCoder2 |
custom |
anything — set customFimTemplate ({prefix}/{suffix}) + customStop |
If your endpoint only exposes /v1/chat/completions (not the legacy
/v1/completions), set apiMode to chat or auto. Use Test Connection
in the panel to verify the model actually does FIM.
How it works (concepts)
When your cursor sits in the middle of code, the extension sends the model two pieces:
- prefix — the code before the cursor.
- suffix — the code after the cursor.
The model fills the gap between them. This is Fill-In-the-Middle (FIM) — different
from chat, and the reason a normal chat model gives bad completions. Each model family marks
the prefix/suffix with its own special tokens, which is what fimTemplate selects.
On top of that, it can add context to the prompt: your recent edits and relevant snippets
from other files (found by indexing your repo). More context = smarter completions, at the cost
of a slightly bigger prompt.
Configuration
The sidebar panel covers the essentials; everything is in VS Code Settings
(Cmd/Ctrl+, → search "Shire", or click ⚙ in the panel). Grouped reference:
Connection
| Setting |
Default |
Meaning |
shire.baseUrl |
http://localhost:8000/v1 |
Your endpoint, ending in /v1. The extension calls {baseUrl}/completions (or /chat/completions). |
shire.apiKey |
— (panel) |
Sent as Authorization: Bearer. Set it via the panel — stored encrypted, not in plaintext settings. |
shire.model |
Qwen2.5-Coder-7B |
Completion model name as registered on your server. |
Completion API
| Setting |
Default |
Meaning |
shire.apiMode |
completions |
completions = FIM (best). chat = use /chat/completions (for endpoints lacking the legacy route; lower quality). auto = try completions, fall back to chat. |
shire.fimTemplate |
qwen |
FIM token format — must match your model family: qwen, starcoder, deepseek, codestral, or custom. Wrong choice = garbage output. |
shire.customFimTemplate |
{prefix}{suffix} |
Only when fimTemplate=custom. A template string containing the {prefix} and {suffix} placeholders, e.g. <PRE>{prefix}<SUF>{suffix}<MID>. |
shire.customStop |
— |
Only when fimTemplate=custom. Comma-separated stop strings that end generation. |
shire.temperature |
0.1 |
Randomness. Low = deterministic, predictable completions (recommended). |
shire.maxTokens |
256 |
Max length of a single completion. |
shire.requestTimeoutMs |
8000 |
Give up on a request after this many ms (prevents a hung endpoint from freezing suggestions). |
When completions fire / where
| Setting |
Default |
Meaning |
shire.enabled |
true |
Master on/off. |
shire.debounceMs |
150 |
How long you must pause typing before a request fires. Higher = fewer requests, more lag. |
shire.multiline |
true |
Allow multi-line completions. Off = single line only. |
shire.disabledLanguages |
scminput, git-commit, plaintext |
Language ids where it stays silent. Add e.g. markdown, json. (Language id shows bottom-right of the editor.) |
shire.maxPrefixChars |
3000 |
How much code before the cursor to send. Bigger = more context, bigger/slower prompt. |
shire.maxSuffixChars |
1500 |
How much code after the cursor to send. |
Context (recent edits + cross-file)
| Setting |
Default |
Meaning |
shire.enableRecentEdits |
true |
Feed your recent before→after edits into the prompt, so a change in one file is suggested in another (refactor assist). |
shire.enableRepoContext |
true |
Index the workspace and inject the most relevant snippets from other files. |
shire.maxNeighborFiles |
4 |
How many cross-file snippets to inject per completion. |
shire.retrievalTimeoutMs |
150 |
Max time spent finding cross-file context before giving up and completing without it (keeps things fast). |
shire.enableSemanticRetrieval |
false |
Use embeddings (vector similarity) in addition to keyword search. Higher relevance, needs an embedding model + a slower index build. |
shire.embedModel |
nomic-embed-text |
Embedding model name — only used when semantic is on. |
Indexing
| Setting |
Default |
Meaning |
shire.maxIndexFiles |
30000 |
Max files indexed in lexical mode (semantic off). Covers large monorepos. |
shire.maxSemanticFiles |
3000 |
Max files indexed in semantic mode (every file is embedded, so kept lower). |
shire.embedConcurrency |
5 |
How many files to embed in parallel during a semantic build. Higher = faster build, more load on the endpoint. |
File selection: the indexer skips node_modules, .git, build output, anything in your
.gitignore / VS Code files.exclude, files over 120 KB, and non-code files. The set is
sorted and capped deterministically, so rebuilds are identical.
Commands
Cmd/Ctrl+Shift+P →
- Shire Autocompletion: Test Connection — verifies the endpoint and whether the model does FIM.
- Shire Autocompletion: Rebuild Repo Index
- Shire Autocompletion: Show Debug Output — per-completion latency + errors.
- Shire Autocompletion: Toggle Enabled
What leaves your machine
When a completion fires, the code around your cursor (prefix + suffix), your recent
edits, and any retrieved cross-file snippets are sent to your configured baseUrl
— nowhere else. No telemetry. The API key is stored in encrypted SecretStorage.
What it does / doesn't
- Does: Copilot-tier single/multi-line completion, cross-file context, refactor propagation —
all against your own model.
- Doesn't: it is in-context, not model training — the model never learns your code into its
weights. And it doesn't reproduce Cursor's custom next-edit "cursor jump" model; you get the
suggestion when you start typing at the new site.
Privacy
No telemetry. Your code is sent only to the baseUrl you set. The API key is kept in encrypted
SecretStorage, never in plaintext settings (a legacy plaintext apiKey setting is honored as a
fallback only).
License
MIT
