Skip to content

roadmap.md

Latest commit

 

History

History
175 lines (128 loc) · 18.6 KB

roadmap.md

File metadata and controls

175 lines (128 loc) · 18.6 KB
Raw

LineLoom roadmap

Updated 2026-11-03 · Planned milestones are not implementation or release commitments.

Current state (verified repository and GitHub release)

  • GitHub stable: v1.3.0, published 2026-10-25 with stable-channel VSIX and SHA-256. Default-off explicitly selected related-file context, native sidebar and single-file autocomplete are shipped.
  • Latest GitHub preview: v1.4.3, published 2026-11-01 with two distinct pre-release VSIX identities, SanskarCodes.lineloom and Sanskar.lineloom, each with its own SHA-256 checksum. This is a GitHub-only migration-testing preview, not an automatic publisher transfer, stable release or Marketplace/Open VSX update. v1.3.0 remains GitHub stable. v1.4.2 is an earlier preview with optional HTML/CSS/Lua support and model-research tooling. Neither preview contains trained first-party model weights or verified real-model benchmarks.
  • Shipped in preview: PR #99 added invisible-Unicode completion hardening, now included in the v1.3.1 pre-release VSIX.
  • Independent distribution: Visual Studio Marketplace and Open VSX listings use SanskarCodes.lineloom and are not automatically updated by GitHub. Open VSX v1.3.0 Stable was independently published and verified on 2026-10-25; the maintainer separately updated Marketplace to v1.3.0. Verify live registry versions before giving future install advice. GitHub Packages npm installer @sanskarin/lineloom-cli is independently published at v0.1.0; merged CLI v0.3.0 source includes offline verification and installation, but neither feature exists in the published package.
  • Unreleased source after v1.4.3: PR #239 added opt-in CUDA C++ (13 registered language IDs in source, nine enabled by default); PR #240 strengthened packager dependency validation; PRs #243 and #244 hardened completion output handling. PR #246 restored reviewed packaging and VS Code type baselines after incompatible automatic dependency upgrades. These merges are source, not retroactive changes to published VSIX files, registry listings or npm packages. The CLI v0.3.0 offline installer is also merged source, not a published CLI v0.3.0 package.
  • First-party model research tooling (not trained weights): #166 adds reserved-token and tokenizer-boundary checks, #167 adds paired before/after local LoRA validation losses, and #168 adds per-target-token weighting alongside per-example and five-language macro losses. These pass source/CI checks; no checkpoint, training corpus, actual GPU run, trained weights, human code-quality evidence or model release exists. The metrics only apply if a future authorized local experiment is performed.
  • Outstanding quality evidence: real Ollama completion quality, physical Windows/macOS/Linux editor UX, accessibility, selected-context consent and real publisher-authentication workflows are not verified by the automated release. See #66, #60, #79, and manual acceptance.

v1.4.3 GitHub-only dual-publisher pre-release — published

The v1.4.3 GitHub release contains a separately packaged SanskarCodes.lineloom preview and Sanskar.lineloom preview, with independent SHA-256 files. The tracked source identity and both registry listings were not migrated by this publication. Installation under the new publisher ID requires a deliberate uninstall/reinstall and a settings/privacy review; no user counts, extension-private storage or update history transfer automatically. See release notes and migration instructions. Real Ollama, OS accessibility and marketplace publisher checks remain NOT RUN.

v1.4.2 GitHub-only pre-release — published

The verified GitHub v1.4.2 preview includes opt-in HTML/CSS/Lua, local-only native FIM tag routing, adapter-integrity receipts and stronger source-free synthetic benchmark checks. Publication ran through the owner-authorized v1.4.2 workflow; physical editor acceptance, real-model quality and independent registry updates are not implied. See release notes and manual QA — NOT RUN.

v1.4.1 GitHub-only pre-release — published

The verified GitHub pre-release packages post-v1.4.0 improvements (#126, #127, #129, #130, #131) after an owner-authorized publisher run, with both public checksum-matching assets. Stable stays v1.3.0 and all registries/publisher SanskarCodes.lineloom remain untouched. No trained weights or verified real Ollama model benchmark. See versioned release notes and manual QA NOT RUN.

Post-v1.4.0 source work (now included in v1.4.1 preview)

  • PR #126: sanitized clipboard support summary and editor-scoped manual-status invalidation.
  • PR #127: explicit Ollama model checks now abort obsolete pending discovery requests on settings, trust and editor changes.
  • Both PRs passed mocked Linux extension-host, Windows/macOS/Linux CI, dependency audit and VSIX candidate checks. These tests are not real-model or physical keyboard/screen-reader acceptance. The source changes shipped in the GitHub v1.4.1 preview.
  • Historical note (v1.4.1): registry publisher migration was deferred at this point. The later v1.4.3 GitHub dual-publisher preview does not itself transfer registry listings or existing installations.

Opt-in HTML, CSS and Lua — included in v1.4.2 preview

PR #144 introduced three editor language modes, now packaged in the GitHub v1.4.2 preview. HTML, CSS and Lua require explicit opt-in; the nine original default-enabled languages remain unchanged. Per-language local model overrides and existing endpoint/privacy gates remain in force. No auto-download, workspace scanning or trained LineLoom Code weights are included. Review language controls and manual acceptance before recommending broader distribution.

Model research included in v1.4.2 preview (no model weights)

  • PR #145: an explicitly configured lineloom-code model tag uses native bounded FIM requests when no extra references are selected. This does not establish that a first-party model was actually trained or installed.
  • PR #143: integrity receipts for local LoRA adapter experiment artifacts, strict offline verification and matching-receipt gates before smoke comparison. Digest matching alone is not signature or provenance verification.
  • Cross-platform and editor-host automated checks passed for the merged source. Actual model inference, generated-code quality, physical accessibility and distribution approval remain outstanding.

Forward-looking milestones

Target State Priority and proposed deliverables Explicit gates
v1.3.1 — reliability preview GitHub pre-release published Unicode completion safety, native sidebar next-action guidance, manually checked default and effective editor models, local file privacy Public preview and notes, manual QA still outstanding; not a stable or registry release
CLI v0.3.0 Merged source; not published Already-downloaded pinned VSIX verification and offline staged-copy installation Merged PR #92; manual acceptance NOT RUN; separate package version/pinset/security/authorization gates
v1.4.0 — usability & quality GitHub pre-release published 2026-10-26; not stable 30-fixture, five-language synthetic evaluator and repeatability aggregates; native status clarity, accepted-item metrics, bounded p50/p95, automatic transient retry pause and explicit session reset Candidate notes, manual QA NOT RUN, locked audit/VSIX/upgrade and separate GitHub-only publication
v1.4.x+ — language evaluation Research Investigate C# and Dart opt-in support and model-specific prompting only if real samples justify them VS Code language mapping, fixtures, bounded prompts, manual local model quality across OSs
v1.5.x+ — limited symbol intelligence RFC only Explore optional developer-selected symbol/import hints with bounded parser effort New opt-in privacy RFC, no automatic workspace scan or persistent indexing by default, independent I/O and security evaluation
v2.0.0 — additional local providers Research, not scheduled Abstract backend transport without losing Ollama defaults Explicit endpoint permissions, no cloud inference by default, strict stream decoding, cancel and budget equivalence

The roadmap is not release evidence. GitHub's latest v1.4.3 pre-release has two publisher-specific preview VSIX identities; separate registry publishing, the npm CLI and GitHub stable v1.3.0 are independent. Verify live registry versions and actual artifacts before describing them as updated. Future releases remain channel-specific.

Suggested next work order

  1. Resolve or explicitly record remaining manual v1.3.0 acceptance tasks with sanitized dated evidence. Keep security #79 open while compatible upstream packaging options and inherited moderate advisories are assessed.
  2. Fix the highest-impact small runtime regressions in focused PRs with mocked VS Code host tests; improve feedback without recording user code or filenames.
  3. Finish CLI v0.3.0 manual acceptance on physical Windows/macOS/Linux and review current VSIX pinsets; the merged development source must not be confused with GitHub Packages v0.1.0 or an authorized new package release.
  4. Make readiness status precise: the native status bar distinguishes enabled from manually verified models; pending user-initiated checks show non-clickable Next notices. No new network or source access on refresh. Physical accessibility QA remains NOT RUN.
  5. Measure suggestions offered separately from the session-only VS Code accepted-item callbacks in development; do not infer usefulness or human model quality. Manual acceptance and accessibility on physical OS devices remain NOT RUN. See sidebar notes.
  6. Maintain the opt-in, bounded 64-request inference latency window with source-free p50/p95 readings and user-initiated timing reset. This is a developer feature only: no real Ollama machine benchmark has yet been recorded.
  7. Use the dedicated-model synthetic evaluation harness and reproducible local baseline protocol to collect genuine optional local Ollama evidence. PR #119 added aggregate repeated-fixture outcomes to unreleased v1.4 development source on main, with six passing automated checks, but no real-model results. Review training and licensing gates before any fine-tuning. No benchmark results or model weights are claimed yet.
  8. Use v1.3.1 preview feedback and unresolved manual acceptance to plan the next stable candidate separately; then investigate v1.4.0 usability and model-quality work.

Permanent non-goals for current releases

No account or subscription requirement, passive model downloads, telemetry, arbitrary external code uploads, automatic whole-repository indexing, persistent private-source embeddings or silent remote origin changes.

Technical plans and historical release notes

After v1.4.0 — working notes (unreleased)

  • Active-editor readiness correctness: switching/closing an editor must clear both manual last-check model and file statuses, and pending results cannot reappear after the editor changes. This does not cause passive checks.

  • Copy sanitized support summary: opt-in clipboard action with aggregate session counts and finite-latency summaries only, no source, filenames, model IDs, endpoint URLs or stored analytics. Clipboard privacy is disclosed.

  • Publisher ID migration: the v1.4.3 GitHub pre-release contains a Sanskar.lineloom testing VSIX alongside the legacy SanskarCodes.lineloom VSIX. Independent Marketplace/Open VSX listing migration and user migration remain incomplete; follow the manual migration and rollback guide and verify both registries independently.

  • Manual model check efficiency (post-v1.4.0 development): abort obsolete, user-invoked model-list requests when the editor or configuration changes, without adding passive probes or persistent network state.

  • First-party FIM model evaluation (post-v1.4.0 development): new source-free offline baseline/candidate comparator enforces synthetic fixture content fingerprint, mode, repetitions and numeric summary invariants. Per-language deltas help avoid hiding weaker C or C++ samples behind one total. No Ollama real-model measurements or trained model weights exist.

  • Model data governance (post-v1.4.0 source): opt-in, read-only model:data-audit preflight for explicitly approved local manifest files; fail-closed rights declarations, repo-group holdout splits, duplicate/heuristic benchmark contamination, and obvious secrets. No corpus was gathered, rights have not been approved and no training run has been performed. Human review is required after a mechanical PASS.

  • LineLoom Code synthetic evaluation (post-v1.4.0): v2 protocol includes seven conservative, source-free response-format indicators; offline baseline comparator refuses v1/v2 mixture and highlights aggregate response-format changes alongside exact matches and latency. This infrastructure has no real-Ollama quality measurements yet.

  • LineLoom Code M1 data preflight hardening (development after v1.4.1): refuse missing train/validation/test coverage for any of Rust, C, C++, Python or TypeScript and screen lightly altered public benchmark text with additional character-shingle overlap checks. Report only bounded language-by-split counts. No corpus was gathered and no actual model training or editor-quality benchmark took place.

  • M2 synthetic request reproducibility (post-v1.4.1 development): protocol v3 adds source-free fingerprints for generated synthetic prompts, FIM suffixes and explicit Ollama settings; offline comparison requires identical request spec and refuses legacy v1/v2 reports. Hardware, server defaults and model revisions still require independent review.

  • M2 local synthetic evaluator hardening (post-v1.4.1 development): isolated in-memory transport tests now cover real request construction, full evaluation reporting, aborts, HTTP status/redirect failures and untrusted stream handling. A green test is not a real model benchmark.

  • LineLoom Code M3 runnable LoRA experiment (development after v1.4.1): independently approved local Qwen2.5-Coder-0.5B base checkpoint and private human-reviewed FIM dataset may be used with an opt-in Python/PyTorch+PEFT adapter-training command. Prefix/suffix labels are masked, training and validation splits remain distinct, the test split is never used, and outputs remain excluded from Git/VSIX. The default command does only offline planning, with pure Python regression tests in CI. No local fine-tune was executed, no approved rights-checked corpus or weights exist, and no model release was published. See training instructions.

  • M3 early-adapter smoke (development after v1.4.1): separate local-only Python base/PEFT-adapter smoke against 30 original synthetic probes, with offline planning, explicit reviewed-checkpoint gating and fixed aggregate per-language reporting. No prompt/source/output logs, downloads, registry publishing or automatic Ollama model migration. Requires an actual approved model/adapter for real measurements; no weights or measured quality are currently available.

Queued autocomplete responsiveness (unreleased source)

The automatic typing debounce now settles superseded waits immediately using a single cancellable latest-request gate. Manual requests bypass the timer; Stop, privacy revocation, file pausing and editor switches cancel it. No duplicate automatic model generation is authorized by this optimization. No source, file URI or prompt text is stored by the gate. Only the verified existing request path can reach Ollama after a delay. Unit and mocked VS Code editor-host tests protect timer cleanup and burst behavior; actual real-model latency measurements and physical editor QA remain pending. Publisher migration and new releases remain paused.