NESweep autocompletion

NESweep — Next Edit autocompletion for VSCode

NESweep is a fork of Sweep Next Edit that retargets the extension at a local OpenAI-compatible /v1/completions server (e.g. llama.cpp's llama-server) running an edit-prediction model. The upstream uvx sweep-autocomplete Python child process — which falls back to CPU and is unusable for next-edit latency — is removed.

Features

Local OpenAI-compatible backend. Posts to /v1/completions on any server you bring up (llama.cpp, vLLM, sglang, Ollama with the OpenAI shim).
SweepAI + Zed Zeta-2 / Zeta-2.1 models. Format auto-detected from sweep.modelName. Zeta-2.1 uses Zed V0318-style numbered boundaries across one contiguous cursor-centered editable excerpt.
Targeted related-code context. LSP definitions/usages are preferred, including distant definitions in the active file; visible and recent buffers fill the remaining related-context slots. A compact Document Symbols outline identifies the active class/function and nearby callables.
LSP-diagnostics aware. Cursor-radius filter, cascading-error suppression below a root-cause line, and user-configurable regex rewrites on the messages (clang / clang-tidy presets included).
Per-language workspace rules. .vscode/nes-<languageId>.md is editable from the NESweep status-bar menu with a configurable soft-cap warning when the file grows large enough to bloat latency.
Cache-friendly + persistent. Stable content emitted first / volatile last for maximum prefix-cache hits; recent files, edits, and cursor positions survive window reload via workspaceState, so the model has context immediately after restart.
Status-bar menu + trace logging. Toggle, snooze, ping server, edit instructions. Set the NESweep output channel to Trace (Developer: Set Log Level… → NESweep) for full request/response visibility.

Settings

Key	Default	Purpose
`sweep.serverUrl`	`http://localhost:8080`	`/v1/completions` base URL
`sweep.modelName`	`sweepai/sweep-next-edit`	`model` field in the request body; substring-matched to pick the prompt format
`sweep.completionTimeoutMs`	`10000`	Per-request timeout (ms)
`sweep.retriggerOnContextExit`	`false`	Opt in to requesting a completion after the cursor moves at least half the last model editable window, even with no text edit
`sweep.contextExitRetriggerDebounceMs`	`1500`	Debounce for context-exit cursor retriggers, allowing navigation to settle; `0` uses the next event-loop turn
`sweep.maxContextFiles`	`5`	Related excerpt cap; Zeta prioritizes LSP retrieval, then visible/recent buffers
`sweep.outlineSymbols`	`0`	Optional nearby LSP symbols plus the active symbol path for every model. Includes methods, fields, enum members, and globals when provided; asynchronous and separate from `maxContextFiles`, a positive value opts in
`sweep.maxRecentChangesChars`	`12000`	Character budget for formatted recent-edit history; `0` disables history
`sweep.includeClipboardContext`	`true`	Include clipboard text as retrieval context; it is emitted last in retrieval
`sweep.stableRetrievalOrdering`	`false`	Sort retrieval chunks deterministically to improve prefix-cache reuse
`sweep.reuseIdenticalPromptResults`	`false`	Reuse recent temperature-0 results for byte-identical prompts
`sweep.identicalPromptCacheTtlMs`	`5000`	TTL for identical-prompt result reuse
`sweep.diagRadius`	`12`	±N lines around cursor; `0` disables
`sweep.editableTokens`	`2000`	Shared approximate active-file budget, estimated as UTF-8 bytes / 3 and snapped to whole lines. Sweep allocates 2/3 before and 1/3 after the cursor; Zeta uses it for its editable excerpt.
`sweep.zetaContextTokens`	`150`	Approximate additional Zeta-2.1 current-file context budget, using the same tokenizer-free estimate
`sweep.rulesMaxChars`	`3000`	Soft cap on per-language workspace-rules file size; overflow surfaces as a diagnostic + red background in the editor
`sweep.injectInlineDiagnostics`	`false`	Inline `BUG:` comments next to diagnosed lines in the prompt — recommended for 0.5B / 1.5B sweep checkpoints
`sweep.inlineDiagnosticsMarker`	`BUG: LSP error here`	Marker phrase used by the inline injection + response-side strip anchor
`sweep.diagnosticsMessageTransforms`	clang preset	`{regex: replacement}` rewrites applied to every diagnostic message after the built-in normalisations

Zeta always uses the suffix-first FIM layout and chronological context order from its Zed training template. This prioritises edit quality over llama.cpp prefix-cache reuse: moving the cursor can change the leading suffix and invalidate the server cache.

Setup

Run any supported edit-prediction GGUF behind an OpenAI-compatible /v1/completions server. Examples with llama.cpp:

# Sweep Next Edit V2 7B
llama-server -hf henrik3/sweep-next-edit-v2-7B-GGUF --ctx-size 32768

# Sweep 1.5B (smaller, faster — turn on sweep.injectInlineDiagnostics)
llama-server -hf sweepai/sweep-next-edit-1.5B --ctx-size 32768

# Zeta-2 (Zed's SeedCoder-8B, single-region)
llama-server -hf bartowski/zed-industries_zeta-2-GGUF --ctx-size 16384

# Zeta-2.1 (Zed's SeedCoder-8B, multi-region)
llama-server -hf mradermacher/zeta-2.1-GGUF --ctx-size 16384

Then point sweep.modelName at the right name. Detection rules:

zeta-2.1 / zeta2.1 / zeta-2-1 / zeta_2_1 → Zeta-2.1 multi-region
zeta2 / zeta-2 / seedcoder → Zeta-2 single-region
everything else → Sweep layout (default)

Sweep's GGUF advertises 32k natively; the full prompt routinely runs 15–20k tokens for non-trivial files, so a smaller --ctx-size truncates real prompts. Zeta-2.1 follows the provider profile of roughly 350 editable and 150 surrounding-context tokens. NESweep estimates those budgets as one token per three UTF-8 bytes and expands to whole lines around the cursor, avoiding a real tokenizer in the extension-host hot path.

Build & install the extension:

bun install
bun run build
bunx @vscode/vsce package --no-dependencies --skip-license
code --install-extension nesweep-*.vsix --force

Credits

Original Sweep Next Edit by SweepAI.
Sweep prompt format ported from cursortab.nvim.
Zeta-2 / Zeta-2.1 model card: zed-industries on Hugging Face.

License

GNU Affero General Public License v3.0 or later — see LICENSE.

The upstream repository sweepai/vscode-nes does not ship a LICENSE file, but its initial commit (fcdfb50 — init: Base vscode foundation based on zed impl) is a line-for-line TypeScript translation of zed-industries/zed/crates/zeta/src/sweep_ai.rs — the wire-protocol structs, the ActionType enum with its SCREAMING_SNAKE_CASE serde rename, the brotli (quality=11, lgwin=22) params, the hardcoded https://autocomplete.sweep.dev/... endpoint, even the // TODO-fenced privacy_mode_enabled: false were carried over verbatim. The Rust file was removed from Zed in commit 42583c1 on 2025-12-04, but at the time of the initial commit it was AGPL-3.0 as part of the Zed editor. Translating an AGPL work into another language produces a derivative work covered by the same license, so AGPL-3.0 attaches to the entire combined codebase regardless of whether the upstream author shipped a LICENSE file. This fork makes that licensing explicit.

Zed Industries, Inc. — original sweep_ai.rs (AGPL-3.0), ported in src/api/schemas.ts, src/core/constants.ts, and parts of src/api/client.ts.
SweepAI and the upstream sweepai/vscode-nes contributors — VS Code-side glue (extension activation, inline-edit provider, document tracker, telemetry plumbing), itself a combined work covered by the same AGPL terms.
This fork's authors — all subsequent commits.