LineLoom v1.4.0 — Reliability & Model Evaluation Preview

LineLoom v1.4.0 — GitHub reliability and model-evaluation preview

Release state: candidate only until the v1.4.0 GitHub pre-release and BOTH exact public assets are verified. This is not a stable release. The verified stable remains v1.3.0. No automatic publishing to Visual Studio Marketplace, Open VSX, npmjs.com, or GitHub Packages.

Major changes compared with GitHub preview v1.3.1

• Safe automatic retry cooldown (PR #121) — temporary HTTP 429, HTTP 5xx, transport and request-timeout failures cause a bounded 1–30s per hashed endpoint/model profile delay for automatic requests only; explicit manual completion bypasses it. No background polling, model downloads, source or endpoint identity is retained in the cooldown.
• Session Activity reset (PR #122) — confirmation-gated native sidebar, Quick Actions and Command Palette control resets completion counters, in-memory timings, one-shot acceptance callbacks and transient cooldown; stops old generations while preserving language/model/privacy preferences and explicit related-file choices.
• Acceptance callback accounting (PR #116) — count one-shot VS Code accepted-inline-item command callbacks separately from offers (including cached suggestions). A callback does not prove actual user usefulness or provide a meaningful acceptance rate.
• Opt-in inference p50/p95 (PR #118) — rolling 64 completed-request duration samples, source-free native sidebar and diagnostics display, explicit timing reset and stale-response protection. No timing samples when diagnostics disabled.
• Native status and recovery clarity (PR #117) — “autocomplete enabled” is not misrepresented as verified model availability; pending explicit checks are non-clickable and preserve consent boundaries.
• First-party completion model research (PR #119 and #120) — stricter aggregate repeatability outcomes for 30 original synthetic fixtures across Rust, C, C++, Python and TypeScript, plus a documented voluntary local benchmark protocol. No model was trained, weights packaged or real inference results claimed.

Default inference remains on user-configured Ollama; remote HTTPS still requires opt-in and per-origin approval. No silent workspace scanning, account, telemetry or automatic related-file context.

Verification and limitations

Required automated release gates: Node 22 locked npm ci, npm run check, Linux VS Code extension-host with a mocked loopback Ollama server, cross-platform CI, npm audit --audit-level=high, production npm audit --omit=dev --audit-level=moderate, real vsce package --pre-release marker, VSIX SHA-256 verification, clean-profile installation, and an isolated upgrade from the independently digest-verified v1.3.0 public stable archive. Prevent duplicate existing tag, release or unexpected asset metadata. This is a checklist, not a claim that the v1.4.0 publishing workflow has already passed.

Manual QA outstanding — NOT RUN: actual local Ollama model response quality, dedicated-model hardware p50/p95 and 30-case reports, physical Windows/macOS/Linux editor behavior, keyboard-only/high-contrast/screen-reader acceptance, consented related-file behavior with real models, and real hosted remote endpoint behavior. A green mocked editor-host suite is not a human quality study. See v1.4.0 manual QA and release evidence policy.

Four moderate development-only transitive dependency advisories were previously documented in security issue #79. The high-severity VSIX packager MSAL mitigation and explicit audit gates remain mandatory. Do not infer a zero-advisory dependency tree from a passing high-severity gate.

Installation after public confirmation

Only after GitHub v1.4.0 is tagged Pre-release, publicly has exactly LineLoom-1.4.0.vsix and LineLoom-1.4.0.sha256, and their digest is verified, download both assets and use VS Code → Extensions → ... → Install from VSIX, or code --install-extension LineLoom-1.4.0.vsix. A VSIX is not a Windows .exe. Prefer stable v1.3.0 when a conservative release is needed. For rollback, verify and reinstall the independently published v1.3.0 stable VSIX.

This GitHub-only preview must not change stable/latest tagging, extension marketplaces or the independently versioned @sanskarin/lineloom-cli package. No release URL or asset should be represented as publicly available before verification.