Write less. Build more. Stay in control.
LineLoom is a free MIT-licensed VS Code extension for context-aware inline autocomplete with your own Ollama model. It uses native editor ghost text, and does not require a LineLoom account, cloud inference provider, or telemetry.
Latest GitHub stable release: v1.3.0 (published 2026-10-25). The verified stable-channel VSIX and SHA-256 checksum are available on GitHub. Automated CI and isolated installations passed; real Ollama, physical OS and accessibility/manual QA remain unverified. Marketplace and Open VSX have separate release processes.
Stable registry distribution (2026-10-25): Open VSX v1.3.0 Stable is independently confirmed as its public downloadable latest version (preRelease: false, verified: true); the version-pinned publishing workflow reused the existing stable GitHub VSIX. The maintainer also manually published v1.3.0 to VS Code Marketplace. The separate GitHub v1.3.1 preview was not uploaded to either registry. Real-model and physical-device QA remain unverified.
The v1.3.1 pre-release is available with a genuine VS Code preview-channel LineLoom-1.3.1.vsix and its SHA-256 checksum.
It adds privacy-safe native Control Center Next: recommendations, a manual check for the active editor's actual configured Ollama model (including language overrides), explicit last-check states for default-model readiness and local file privacy, and Unicode completion filtering. Checks remain opt-in and session-only; no passive network polling or workspace scanning was introduced.
For testers only: the GitHub publisher passed locked audit/build, mocked VS Code editor-host, pre-release packaging and checksum, clean install and pinned v1.3.0 stable-to-preview upgrade checks. Real Ollama quality, physical OS keyboard/screen-reader and latency acceptance are NOT RUN. GitHub stable remains v1.3.0, and this preview does not update either extension registry or the npm CLI. See preview notes and manual QA checklist.
The GitHub v1.4.2 testing preview is published with a genuine preview-channel VSIX and SHA-256 checksum. The version-pinned release workflow passed locked audits, Linux VS Code mocked editor-host tests, pre-release packaging, digest validation, a clean install and independently pinned upgrades from published v1.3.0 stable and v1.4.1 preview.
New in this preview: experimental HTML/CSS/Lua opt-in completions,
explicit LineLoom Code native local FIM routing, offline adapter
integrity checks, v4 synthetic benchmark scope verification and source-free
regression review. These are development features, not trained first-party
weights or measured model-quality improvements. The original nine
default-enabled languages and SanskarCodes.lineloom ID remain unchanged.
See release notes and
manual QA checklist.
Manual Ollama and physical desktop acceptance remain NOT RUN.
The GitHub stable version remains v1.3.0; external extension registries
and npm distributions were not updated.
Download v1.4.1 GitHub pre-release: verified genuine VS Code LineLoom-1.4.1.vsix and SHA-256 checksum. The version-pinned GitHub publisher passed tests, security audits, preview-channel manifest validation, clean installation and pinned stable-to-preview upgrade.
Changes from v1.4.0: explicit source-free support-report copy, editor-scoped readiness fixes, cancellable manual model discovery, synthetic benchmark comparison, offline dataset governance screening and v2 formatting signals across Rust, C, C++, Python and TypeScript. No LineLoom-trained weights or real Ollama quality results are claimed. Real physical-editor/accessibility testing is NOT RUN. See release notes and manual QA checklist. GitHub stable stays v1.3.0; publisher SanskarCodes.lineloom, Marketplace, Open VSX and npm distribution remain unchanged.
The ongoing development branch now includes Copy Sanitized Support Summary in the native Control Center, Quick Actions and Command Palette, using only allowlisted session counters and coarse last-check statuses—never model names, files, prompts or endpoint URLs (PR #126). Switching editors also invalidates the previous editor's manually checked model and file status. Obsolete user-initiated Ollama model-list requests now receive cancellation signals when a new check starts or their status is invalidated (PR #127).
These changes are included in GitHub v1.4.1 testing preview, but not in the
earlier published GitHub v1.4.0 preview or v1.3.0 stable. The extension publisher
remains SanskarCodes with ID SanskarCodes.lineloom.
No publisher-ID or registry migration is authorized or underway.
Download GitHub v1.4.0 pre-release with its genuine VS Code preview-channel VSIX and SHA-256 checksum. The publisher completed successfully after audits, mocked editor-host testing, digest validation, clean installation and a pinned v1.3.0 stable-to-preview upgrade.
What's new: bounded transient Ollama automatic retry cooldowns, confirmation-gated session-activity reset, source-free acceptance and inference timing aggregates, clearer model-readiness wording, and dedicated model evaluation tooling. See versioned release notes and manual QA checklist. Physical Windows/macOS/Linux human accessibility and actual Ollama quality are still NOT RUN. The latest stable remains v1.3.0, and Marketplace/Open VSX/npm were not updated by this GitHub preview.
With lineloom.performanceDiagnostics enabled, LineLoom now computes rolling p50/p95 from up to 64 successful Ollama request durations, exposes them in native Session Activity and Diagnostics, and offers Reset Inference Timing Samples. Values are source-free, never persisted, erased on opt-out and protected from stale in-flight results. Suggestion counters remain unchanged after a timing reset. No real-model latency benchmark has been run, and this feature is included in the GitHub v1.4.0 preview, not in v1.3.0 stable or either independent registry. See sidebar timing/privacy documentation.
The status bar differentiates autocomplete enabled from model availability actually checked. While explicit local model or file-policy checks are running, the sidebar displays non-clickable progress wording. No passive Ollama calls, file reads or new endpoint consent are introduced. See sidebar documentation. This feature is in the GitHub v1.4.0 preview, but not the v1.3.0 stable or independent registries.
The GitHub v1.4.0 preview distinguishes suggestions offered from VS Code item acceptance callbacks, using up to 128 random one-shot tokens and aggregate in-memory counters. No code, file identities, event logs, analytics or stored acceptance history. Counts are not proof of quality and the feature is included in the v1.4.0 GitHub preview, not v1.3.0 stable or v1.3.1 preview. See sidebar privacy and status.
The following work is now included in the GitHub v1.4.2 testing preview but not in the v1.3.0 stable build or external extension registries:
- PR #144 recognizes HTML, CSS and Lua as experimental, explicit opt-in languages. The original nine default-enabled modes are unchanged; locally installed model overrides may be assigned separately.
- PR #145 recognizes an explicitly selected, locally installed
lineloom-codeFIM model tag, with bounded Ollama prefix/suffix completion. This is request routing only, not trained or shipped weights. - PR #143 adds SHA-256 local LoRA adapter artifact receipts and a read-only verification gate before research smoke runs. The checks do not prove provenance, permission to train, or model usefulness.
These changes have verified v1.4.2 GitHub pre-release packaging and automated CI evidence, not real Ollama quality, physical-editor accessibility or a stable/registry update. The publisher ID and independent registries are unchanged.
A separate, optional LineLoom Code research track now has an opt-in local Ollama evaluation harness and 30 original categorized synthetic fixtures for Rust, C, C++, Python and TypeScript. The evaluator has strict bounded stream decoding and language/category-level aggregate results. No custom model has been trained, downloaded or bundled, and no editor defaults change. Use npm run model:eval:dry to inspect the offline fixture scope or npm run model:eval -- --model qwen2.5-coder:1.5b --runs 3 to run a voluntary localhost-only comparison against an already installed model. Results report strict synthetic exact matches and latency, not proven quality, correctness or human acceptance. See model research, training milestones and data governance.
The latest GitHub preview v1.4.2 was published on 2026-10-27 with its own VSIX and checksum. This does not include a trained first-party model; v1.3.0 remains stable.
The published v1.3.0 stable build adds opt-in, session-only explicitly selected related-file context: at most two tiny supported, same-workspace source files (2,048 bytes each), subject to existing canonical-path/ignore checks and post-response validation. The feature is disabled by default, never automatically scans other files, and requires explicit consent before reference snippets reach Ollama. A reviewed dev-time VSIX packager mitigation clears the high-severity audit gate, but four moderate transitive dev-tool findings remain. Download the verified GitHub v1.3.0 stable VSIX and SHA-256 checksum. See release notes and manual acceptance. Physical-device, actual local-model, accessible UI and publisher-authentication QA remain NOT RUN, and registries are not updated by the GitHub release.
GitHub v1.2.0 stable release is published with LineLoom-1.2.0.vsix and the SHA-256 checksum.
The native Activity Bar Control Center now includes active-editor eligibility,
default/per-language model routing, model-connection guidance, and additional
privacy-safe, in-memory session counters. The stable VSIX has no pre-release
marker. Install it using VS Code → Extensions → … → Install from VSIX…
or code --install-extension LineLoom-1.2.0.vsix.
Manual QA disclosure: Real local Ollama model usefulness/latency, interactive Windows/macOS/Linux editor and accessibility checks, real privacy acceptance and upgrades from older installations are not yet verified. The owner authorized GitHub stable publication after passing automated CI; unperformed manual checks are not represented as passed. Release notes · Acceptance matrix · Follow-up issue #21.
External registry distribution is separate: Open VSX now publicly serves v1.2.0 Stable (verified on 2026-10-20). The Open VSX publishing workflow reused and verified the stable GitHub VSIX. Visual Studio Marketplace is a separate rollout; a GitHub release never updates registries automatically. The older v1.1.1 pre-release and v1.0.0 stable release remain accessible.
The v1.2.1 maintenance pre-release includes Stop Current Generation, hardened sensitive-directory exclusions, ignore-policy invalidation, more regression tests and a complete TypeScript source-inventory gate. The attached VSIX is marked for the actual VS Code pre-release channel and includes a SHA-256 checksum. Real-model/editor hardware acceptance remains outstanding, so this is not a manually verified stable update. GitHub stable v1.2.0 and independently published registry versions are unchanged.
The GitHub v1.2.2 pre-release adds stricter Ollama settings validation, fail-closed malformed path exclusions, and a preview-channel VSIX. Download the preview VSIX and SHA-256 checksum. This was a GitHub-only preview; v1.2.0 was the current GitHub stable release at the time. The latest GitHub stable is now v1.3.0 (2026-10-25), while Open VSX and Visual Studio Marketplace remain independently managed. Real model performance, full manual desktop acceptance and accessibility remain unverified.
GitHub v1.2.3 — Live Ignore Policy Safety Preview is publicly available with a genuine preview-channel VSIX and SHA-256 checksum. It improves privacy against delayed/dropped ignore-file watcher notifications: existing and previously missing ignore sources are revalidated, cached source rules use in-memory SHA-256 fingerprints of rule bytes, and up to eight concurrent policy read tasks preserve root-to-leaf rule ordering.
Publishing CI passed automated tests, mocked VS Code host integration, preview VSIX verification and isolated stable-to-preview upgrade. This does not prove real-model quality, physical workstation performance, actual watcher timing or accessibility; those remain tracked in issue #46. The GitHub stable release remains v1.2.0; registry distributions are separate and were not changed by this publication.
GitHub v1.2.4 — Privacy Eligibility Diagnostics Preview is published with the genuine preview-channel VSIX and SHA-256 checksum. It adds opt-in, in-memory privacy eligibility check timing counters (allowed/denied, last, session average and max, rolling 64-sample p95). Collection is disabled by default and contains no file paths, prompts, source code, rule text or response bodies. These measurements are not end-to-end Ollama/model latencies.
Publisher CI passed the real preview VSIX/channel/checksum and isolated installs/upgrades from SHA-256-pinned v1.2.0 stable and v1.2.3 preview VSIX packages. Manual real-model, physical Windows/macOS/Linux performance and accessibility QA remain unverified in issue #54. Use the v1.2.4 manual acceptance checklist to record reproducible, sanitized evidence before considering a stable rollout. This GitHub-only pre-release did not update either extension registry. At the time this preview was published, GitHub stable was v1.2.0; the latest stable is now v1.3.0.
The v1.2.5 implementation PR #56
has been merged into main, alongside
privacy-hardening PR #59
and CI-safety PR #58.
This v1.2.5 source milestone, now incorporated into the published v1.3.0 stable VSIX, includes LineLoom: Check Active File Privacy
(lineloom.checkActiveFilePrivacy): a user-triggered local inspection of
workspace, sensitive-path and root/nested ignore policies. It is available
from the Command Palette, Quick Actions and the Privacy & Connection sidebar,
and does not read source content or contact Ollama. Results only describe
path/ignore eligibility at the time checked, not full completion or model
readiness. No filenames or policy contents are shown.
The source includes LineLoom: Reset Privacy Timing Statistics
(lineloom.resetPrivacyTimings), an opt-in-only sidebar / Quick Actions entry,
and protection against asynchronous statistics being restored after a reset.
Malformed UTF-8 or NUL-containing .gitignore / .lineloomignore files
fail closed before inference.
Combined candidate CI passed Windows/macOS/Ubuntu builds and tests, mocked VS Code editor-host integration, preview-channel VSIX validation, isolated install and checksum-pinned upgrades from v1.2.0, v1.2.3 and v1.2.4. This is not real-model or physical-device acceptance. The v1.2.5 development plan, manual acceptance checklist and release readiness tracker #60 track the remaining gates.
No standalone v1.2.5 GitHub release was published. Its merged source changes were included in the later v1.3.0 GitHub stable VSIX. The latest GitHub testing preview is v1.4.0; the older v1.2.4 preview remains available. GitHub publication did not update either external extension marketplace.
Install the currently published version from your editor's supported registry:
- Visual Studio Marketplace — for VS Code users.
- Open VSX Registry — for compatible editors using the Open VSX extension registry.
- Both listings use publisher
SanskarCodesand extension IDSanskarCodes.lineloom. v1.3.0 Stable is now available in both extension registries; Open VSX v1.3.0 was verified independently, and the maintainer updated Marketplace separately.
In VS Code, open Extensions (Ctrl+Shift+X), search for LineLoom, verify the publisher SanskarCodes, and choose Install. You can also use the Command Palette's Extensions: Install Extensions action, or the VS Code CLI:
code --install-extension SanskarCodes.lineloomIf your editor obtains extensions from Open VSX, open its Extensions view, search for LineLoom, and verify the publisher SanskarCodes. The independently verified Open VSX publication is now stable v1.3.0; check the live listing for later updates.
The latest GitHub v1.3.0 stable release includes LineLoom-1.3.0.vsix and its SHA-256 checksum. Verify the downloaded file and choose VS Code → Extensions → ... → Install from VSIX (or code --install-extension LineLoom-1.3.0.vsix).
Important for Windows: Do not double-click the VSIX file. Windows may open Visual Studio VSIX Installer, which correctly rejects VS Code extensions and displays “One or more extensions are for Visual Studio Code.” Open Visual Studio Code and select Extensions → … → Install from VSIX… instead. Step-by-step installation and troubleshooting.
Distribution status: GitHub v1.3.0 stable is published. Open VSX v1.3.0 Stable was independently published and verified on 2026-10-25 (publishing run); Marketplace v1.3.0 was updated manually by the maintainer. GitHub releases never synchronize registries automatically. GitHub owner sanskarIN and publisher SanskarCodes are intentionally different. Publisher instructions.
The LineLoom website is live and maintained in site/, deployed through the GitHub Pages workflow. The deployment guide documents verification and maintenance.
Website content is excluded from the extension VSIX package. This does not add hosted inference, an account requirement or telemetry.
Documentation index: Browse all guides.
Development and upcoming features: forward-looking roadmap · v1.3.1 maintenance plan · proposed future features · native sidebar guide · package-manager distribution · sanitized release evidence. These plans describe unreleased ideas, not features already in the installed extension.
The v1.3.0 GitHub stable publishing workflow passed cross-platform checks, Linux mocked VS Code editor-host integration, stable VSIX identity, checksum and isolated installation/upgrade tests. These do not demonstrate real Ollama quality, physical operating-system interaction or accessibility. The v1.3.0 manual acceptance checklist and evidence protocol record remaining work.
This section describes the current development source. Some features may not yet be in the published stable/preview VSIX; see the version-specific notes above.
- Stricter send-time privacy controls: rechecks active model, file exclusions, workspace trust and remote-endpoint permission before sending completion context. Pending model-picker choices are also discarded if the Ollama endpoint or remote approval changes.
- Per-language Ollama models: assign a different installed model to each supported language using LineLoom: Assign Language Model or the status-bar Quick Actions picker. Unassigned languages use the default
lineloom.model; JSX/TSX use the JavaScript/TypeScript override. No automatic model downloads. - Native VS Code inline suggestions for 12 recognized language IDs in current source. The original nine defaults are Python, JavaScript/JSX, TypeScript/TSX, Rust, Kotlin, Go, Java, C and C++. HTML, CSS and Lua are additional experimental opt-in modes, not enabled by default and not yet part of a published VSIX. See language controls.
- Smarter automatic request timing: suppresses background inference directly before existing identifier characters (including Unicode letters, combining marks, connector punctuation and joiners). Configure
lineloom.suppressMidTokenAuto(default on) to opt out; manual requests always work, and skip the typing debounce. - Typed suggestion continuation: when you type the beginning of an inline suggestion, LineLoom can offer the remaining text immediately from short-lived memory without another Ollama generation. Context, settings, active editor, workspace privacy and the actual single-insertion edit event must all match. Multi-edit changes, selection replacements and edits elsewhere invalidate reuse. No cross-file reads.
- Multi-line continuation, bounded prefix/suffix context, and optional FIM using a recognized Qwen2.5 Coder model family or, in unreleased source, an explicitly configured
lineloom-codetag for an independently installed FIM-capable model. Improved cleanup removes common model control tokens and explanations outside leading fenced code blocks. Ollama streams have total-byte, event-count and per-event safety limits. - Quick Actions from the status bar: enable/disable completion, pause/resume automatic suggestions, trigger manually, choose a default or per-language model, manage languages, check connection, open diagnostics or settings. Model discovery and connection checks run only when selected by the user.
- User-selected installed model, endpoint, timeouts, output limits, temperature, automatic-trigger controls and per-language opt-in/out settings.
- Cancellation on newer requests, active cursor/editor changes and timeouts; a short-lived backoff stops repeatedly sending identical automatic requests after an empty model result. Explicit user-triggered requests bypass this backoff. Additional checks run before sending code or reusing a cached suggestion. No insertion is attempted for selected text or multiple cursors, and automatic suggestions defer to active native IntelliSense completions.
- The active trusted-workspace document is used by default. Optional v1.3.0 related-file context remains OFF until explicitly selected and consented (maximum two tiny eligible same-workspace source files). Symlink escapes, sensitive names and nested
.gitignore,.lineloomignoreand user exclusions are blocked. - Local loopback by default; remote HTTPS needs both a setting and per-origin interactive approval. HTTP redirects are rejected to prevent transferring code to an unapproved destination. No background model downloads.
- Automatic completion requires two non-whitespace characters on the active line by default (configurable 0–32); manual triggers bypass this threshold. Active-document context is read through bounded editor ranges, even in large files.
- Install VS Code and Ollama for Windows, Linux or macOS.
- Start Ollama according to its platform instructions. Confirm that
http://127.0.0.1:11434/api/tagsis reachable locally. - Install a coding model yourself:
ollama pull qwen2.5-coder:1.5b. Downloads consume storage; see model licensing and system requirements. - Install LineLoom from Visual Studio Marketplace (recommended), or use the GitHub VSIX as described above. For local development from source instead, run
npm ci,npm run build, open the repo in VS Code and press F5. - In VS Code (or the Extension Development Host), open a trusted folder containing a supported source file (e.g.
.py,.js,.ts,.rs,.kt,.go,.java,.c, or.cpp). - Type some code, or run LineLoom: Trigger Completion. Accept native inline text with your VS Code inline-suggest keybinding.
- Run LineLoom: Check Ollama Connection or LineLoom: Select Model if suggestions do not appear.
Open the Command Palette and run LineLoom: Show Setup Guide for the short interactive guide. Click LineLoom in the VS Code status bar to open Quick Actions. Pausing automatic suggestions keeps manual LineLoom: Trigger Completion available; the pause choice remains in VS Code settings until you resume.
Quick Actions · Pause or Resume Automatic Suggestions · Enable Autocomplete · Disable Autocomplete · Trigger Completion · Select Model · Manage Autocomplete Languages · Check Ollama Connection · Open Settings · Open Diagnostics · Show Setup Guide · Show Completion Performance.
Settings are under lineloom.* in VS Code Settings. lineloom.enabledLanguages lists the enabled editor modes (12 recognized in unreleased source; original nine selected by default). HTML, CSS and Lua must be explicitly enabled; JSX and TSX follow JavaScript and TypeScript choices. An empty list disables suggestions for all languages. Defaults: model qwen2.5-coder:1.5b, endpoint http://127.0.0.1:11434, debounce 250 ms, timeout 12 s, 8,000 context characters, 128 generated tokens, eight maximum suggestion lines (configurable 1–20), two minimum automatic-trigger characters (configurable 0–32), and mid-token auto-suppression enabled. Cross-file context is not implemented in v1.0.0 even if the reserved setting is enabled.
The extension transmits a bounded snippet from the current document to the configured Ollama endpoint to generate suggestions. Localhost requests stay on the local network stack; using a remote endpoint sends snippets to that server. No telemetry or repository indexing exists. A local model is not a guarantee of safe, correct, or licensed output; review code before use.
- Privacy policy and ignore rules
- Architecture
- Development and testing
- Troubleshooting
- v1.0.0 stable release notes and ongoing manual QA checklist
Published on GitHub Packages (2026-10-24): @sanskarin/lineloom-cli@0.1.0 is a separate, dependency-free npm-format installer. Its lineloom install command downloads and SHA-256 verifies the published v1.2.0 stable VSIX; --preview selects published v1.2.4. Installing this package alone does not install the extension or run install-time hooks. GitHub's npm registry requires package-read authentication. Verified publishing workflow · installation and security guide. This does not mean it is published on npmjs.com; a real authenticated end-user install remains to be manually verified.
npm install
npm run check
npm run packageThe GitHub v1.0.0 stable release ships an installable VSIX for community testing, including nine language IDs, per-language selection, nested ignore protection, bounded editor context, foreground-editor preflight checks, a short-lived empty-result retry backoff, typed-prefix continuation reuse, bounded NDJSON parsing and redirect-safe Ollama requests. It includes a mocked-Ollama editor-level integration check. Real Ollama model quality, offline behavior and manual cross-platform verification remain unverified; see the release notes. Both the VS Code Marketplace listing and Open VSX listing are public; each future update must be published separately. Roadmap.
GitHub: sanskarIN · Website: sanskarin.github.io
LineLoom is free and has no subscription or paywall. If useful, you may support maintenance via Buy Me a Coffee, Razorpay, or PayPal.Me.
MIT License. Licensing history and third-party notices. LineLoom is an independent project; no affiliation or endorsement by Microsoft, GitHub, Ollama or model providers is implied.
The v1.3.0 GitHub stable release includes a default-off, explicitly user-selected related-file context feature, originally prepared in PR #93. It permits at most two small (2 KiB each), policy-eligible source files to be used as anonymous untrusted reference snippets in a completion prompt. There is no repository scanning, persistent index or hidden background file selection. The user must enable the setting, pick files and confirm the disclosure each session; a separately approved remote Ollama host could receive those extra snippets.
Published on GitHub, but not manually accepted across physical OSs. See the privacy/threat model, release notes, and manual QA matrix. v1.2.4 remains the most recent separate GitHub preview. A compatibility mitigation for high-severity issue #79 is included; four moderate transitive publisher-tool findings remain for review.