Available in the published GitHub v1.3.0 stable release. The sidebar was introduced in the v1.1.x preview series and stabilized in v1.2.0. You can open it using the LineLoom icon in VS Code's Activity Bar or the Command Palette's LineLoom: Open Control Center. The VS Code Marketplace and Open VSX may serve different versions; verify the installed extension before expecting this view.
This is a native VS Code Tree View, not a web app, separate chat service or background Ollama process.
The in-development Control Center adds one Next: action near Engine Status. It prioritizes disabled autocomplete, restricted workspace trust, unsupported/non-local editors, disabled languages, invalid or unapproved endpoint settings, and unavailable models before offering actions for generation or diagnostics. When no safe command can perform the required step (for example, opening a supported file or changing Workspace Trust), it displays guidance without a clickable command.
Where a native LineLoom command already exists, clicking the recommendation performs that explicit user action: Enable Autocomplete, Manage Autocomplete Languages, Select Model, Open Settings, Check Ollama Connection, Stop Current Generation, Open Diagnostics or Trigger Completion. The recommendation itself never sends a prompt, reads file contents or contacts Ollama; it does not claim the configured model is reachable or that the file passed privacy inspection. Engine Status remains independent of model availability and permission checks.
The Related-File Context entry now distinguishes setting enabled but no files chosen from session-only selected references. Enabling the setting alone is not consent to read additional files, and selection remains bounded to two user-chosen files.
Release status: This sidebar refinement is source work proposed for a future maintenance build, not part of the published v1.3.0 VSIX or either marketplace version. Physical keyboard/screen-reader testing and real Ollama acceptance remain open.
After the user explicitly invokes LineLoom: Check Ollama Connection—from the sidebar, Command Palette, or Setup Guide—the model section can show a Last Manual Model Check line. Its statuses distinguish Not checked, Checking, Default model found, Default model missing, and Connection failed.
This is a snapshot of the configured default model's presence in Ollama's returned model list at the time of that action, not automatic availability monitoring or proof that the model can generate useful code. Per-language model overrides remain independent; a successful default-model check does not verify a language-specific override. If the last check reports a missing default model, the sidebar suggests selecting an installed one; if a connection failed, it suggests an explicit retry.
- Sidebar rendering, refreshing and navigation never trigger model discovery, inference, file access or remote approval.
- A check is initiated only after a deliberate command; remote HTTPS hosts still require the existing separate opt-in and origin approval.
- Only one of five non-sensitive status values is retained in memory for the active session. No model names, endpoint URLs, server responses or timestamps are stored by the readiness tracker.
- A second manual check supersedes the first; late results from a previous check cannot overwrite the current status. Changing any LineLoom configuration or workspace trust clears the previous result. A dismissed approval is not treated as a successful check.
- The code is a future release candidate, not included in the already published v1.3.0 VSIX, Marketplace or Open VSX packages. Actual physical-device accessibility and real-model quality remain outstanding.
The sidebar now displays a Last Manual File Check row below Check Active File Privacy in the Privacy & Connection section. The row starts Not checked, changes to Checking only after a person invokes the command, and then displays the previous local path-and-ignore inspection outcome (Passed, Blocked, or No active editor). The normal Next: hint also explains an explicitly blocked file, without offering an unsafe bypass.
This is not continuous monitoring, a file-content scanner, or a guarantee of future completion eligibility. A Passed result means only that existing path and ignore checks succeeded at that instant; the completion engine independently revalidates privacy before inference. Model readiness, language settings, endpoint approval and eventual prompt safety are separate checks.
- No filename, absolute path, file content, actual ignore-rule pattern or secret enters the sidebar's stored status. The status lives only in extension memory.
- The sidebar's normal refresh, focus and rendering perform no file reads, prompt construction or Ollama requests.
- The command runs the pre-existing local inspector only after a deliberate click. A second inspection supersedes any older pending result.
- Active editor changes, active document edits/closures, extension configuration, workspace trust and ignore-file watcher changes clear the previous result. Late asynchronous results cannot restore an old approval after these events.
- The change is source-only until a future version is explicitly tested and released; it is not in the published GitHub v1.3.0 VSIX or either extension marketplace.
The default-model check introduced after v1.3.0 reports only the configured default tag. If a language-specific override is enabled (for example TypeScript using another local model), a successful default check does not verify the model used for suggestions in that editor.
The Last Active Model Check entry in the Active Editor section now offers an explicit Check Active Editor Model action (Command Palette: LineLoom: Check Active Editor Model). It checks which model is resolved for the active supported local editor, including a configured language override. After an authorized user click, the existing Ollama approval boundary applies and the extension requests the bounded model listing. The result is Not checked, Checking, Found, Missing, or Connection failed. It is only a historical model-list snapshot, not a guarantee of generation quality or live availability.
- Rendering or refreshing the sidebar performs no network request, prompts, source reads or path inspection.
- Only a supported, enabled local file in a trusted workspace can be used. A command on another editor reports an action hint, not a network operation.
- The status is session-only and stores no filenames, prompts, endpoint URLs or model tags. Switching editors, editing the active document, changing LineLoom settings or changing workspace trust clears the status. Obsolete concurrent network results are ignored.
- When an override is missing at the last manual check, Next: Change Language Model opens the existing explicit model assignment dialog; when the default is missing, Next: Choose Installed Model is offered instead.
- This is source development and has not been published into v1.3.0 or the editor marketplaces.
| Section | Visible information and actions |
|---|---|
| Completion | Engine state, enable/disable, pause/resume automatic suggestions, manual trigger and Stop Current Generation. The future maintenance candidate adds one contextual Next: recommendation |
| Models & Languages | Current default model, count of per-language overrides, enabled-language count, and native selection commands; an unreleased enhancement adds the last user-triggered default-model check |
| Active Editor | Language, effective model and provisional settings-level eligibility; an upcoming preview adds explicit active-editor model verification (including overrides), never a passive model request |
| Privacy & Connection | Workspace trust, endpoint category, Check Active File Privacy, selected related-file count, and manual Check Ollama Connection; a future source candidate adds the last explicit file-inspection status |
| Session Activity | In-memory counts for requests, suggestions offered and accepted, failures, cancellations, cache hits, continuation reuse and empty responses; optional measured latency, diagnostics and privacy-timing reset when enabled |
| More Actions | Settings, setup guide, Quick Actions |
Engine status = Ready does not establish model availability. The model may be missing or Ollama may be offline. Use the explicit Check Ollama Connection action instead of interpreting the sidebar's settings-derived status as a live health check.
Active Editor status = Allowed by settings is NOT a file privacy clearance. To evaluate a specific file's current path and ignore-policy status, invoke Check Active File Privacy. The command checks local rules without reading source text or requesting a completion; privacy is independently checked again just before inference.
This feature ships in GitHub stable v1.3.0, but is OFF by default.
- In LineLoom Settings, enable
lineloom.crossFileContext. The setting itself does not read additional files. - Open a supported local source file in a trusted workspace.
- Select Related-File Context in the sidebar or run LineLoom: Select Related Files for Context.
- Select one or two distinct supported source files from the same workspace folder (not the active file). Each selected file must be nonempty and no larger than 2,048 bytes.
- Read and accept the modal disclosure before extra snippets may enter Ollama prompts. If you configured and separately approved a remote HTTPS endpoint, those snippets may leave the device.
- The sidebar shows the number selected, not filenames. Choose Clear Related Files or use LineLoom: Clear Selected Related Files to revoke selection. Changing LineLoom settings, active editors or ignore rules also clears it.
Selected paths/snippets are kept only in current-session memory; hard exclusions and nested ignore policies are checked before and after use. No repository scan, automatic reference discovery or persistent code index occurs. The tool cannot detect every secret in an otherwise allowed file—choose only source you are comfortable sending to the configured Ollama endpoint. See privacy and resource boundaries.
When lineloom.performanceDiagnostics is OFF (default), the extension neither stores nor displays inference timing samples. After explicit opt-in, the native Session Activity sidebar and source-free Diagnostics include nearest-rank p50/p95 of the last 64 completed, authorized Ollama requests, alongside the existing session-wide average. The fixed-size rolling window stores only finite elapsed milliseconds—not source code, document paths, model names, prompts, responses, endpoint identities, or per-request timestamps. These measurements cover completion generation and privacy revalidation and do not establish model accuracy, user acceptance rate, or hardware-quality benchmarks.
Reset Inference Timing Samples, accessible from the sidebar, Quick Actions and Command Palette when enabled, erases the rolling samples, current/average durations and sample count without clearing request, offered-suggestion or accepted-callback counters, changing settings, stopping current inference, or touching the separate privacy-rule timing statistics. Outstanding requests from the prior timing epoch cannot restore cleared measurements; disabling diagnostics also clears them. The reset command can be invoked from the Command Palette while diagnostics are disabled and still performs a safe erase.
Pure unit tests cover bounds, invalid timings, nearest-rank math and reset. Mocked VS Code editor-host tests cover the opt-out boundary and late results. Manual actual Ollama local-model performance, Windows/macOS/Linux hardware and accessibility acceptance remain NOT RUN. The feature is development-only, not in the published v1.3.0 stable or v1.3.1 GitHub preview.
The native status bar no longer claims Ready for local code completion merely because autocomplete is enabled. Its tooltip explicitly says that the Ollama model and active-file rules are not automatically verified. Manual-only mode identifies paused automatic suggestions while allowing manual completion. Unavailable and error states direct users to manual checks and source-free diagnostics.
The sidebar Next: recommendation distinguishes user-invoked checks currently running (default model, effective active-editor model, local file privacy) from successful outcomes. These in-progress notices have no command, do not trigger duplicate requests, and remain lower priority than blocked editor/workspace states or actively generating suggestions. Refresh is still offline and does not inspect source or approve a remote origin.
Pure unit tests cover status wording and pending-state priorities. Actual Windows/macOS/Linux keyboard-only, high-contrast, screen-reader and human model acceptance remain NOT RUN. These are unreleased source changes; published v1.3.0 Stable and v1.3.1 preview are unchanged.
Suggestions Offered counts ghost-text items returned by LineLoom, including cached results. Accepted Suggestions counts one-shot VS Code item acceptance callbacks. No text, filename, document identifier, model output, or individual acceptance event is stored or logged; only an in-memory aggregate counter and up to 128 opaque random pending callback tokens are retained. Duplicate and invalid tokens do not count. Editor refresh does not issue inference or send network requests.
These counts are not proof of quality, human usefulness, or an accurate acceptance rate: the editor can show or discard items repeatedly, and some interaction paths may not invoke the callback. Unit and mocked VS Code-host tests exercise the callback wiring, but manual acceptance and actual Windows/macOS/Linux accessibility tests remain NOT RUN. This source feature is not in the already published v1.3.0 stable or GitHub v1.3.1 preview.
Rendering/refreshing the tree does not contact Ollama, issue a prompt, read related file contents, or automatically request remote approval. Its labels omit raw document paths, code, prompts, response bodies and full endpoint URLs. Model discovery or an Ollama connection check is a separate action initiated by the user.
Diagnostics are aggregate and in-memory by default. Optional performance timing collection is controlled by lineloom.performanceDiagnostics; disabling it clears collected latency measurements. The reset command erases privacy eligibility timing statistics without changing model or completion settings.
The status and consent interactions are designed for native VS Code accessibility; however physical keyboard-only, high-contrast and screen-reader acceptance remains unverified and is tracked in v1.3.0 manual QA.
- No suggestions: confirm the extension is enabled; open a supported local file in a trusted workspace; check configured language/model and use the explicit Ollama connection command.
- Editor says allowed but nothing appears: settings-only eligibility does not bypass ignore rules, unsupported editors or missing models. Run Check Active File Privacy and inspect the non-sensitive diagnostics.
- Related files count returns to zero: settings changes, switching active editors or ignore-rule edits intentionally revoke the earlier selection.
- Remote inference unavailable: remote HTTPS origins require both the relevant setting and explicit per-origin approval. Opening the sidebar does not grant permission.
- Suggestion feels stale: use Stop Current Generation; do not assume an already-dispatched prompt can be recalled from a server.
Run npm ci, npm run check and (on a supported VS Code test setup) npm run test:integration. Verify keyboard interactions and consent with an actual editor; observe that a sidebar refresh does not cause a model request. For release/OS-specific evidence and installed VSIX checks, use the evidence protocol and manual acceptance checklist. Automated tests are not a substitute for physical platform acceptance.
Repeated automatic completion requests after a temporary Ollama error no longer hammer an overloaded or unreachable server on every keystroke. For HTTP 429, HTTP 5xx, transport failures or request timeouts, an in-memory cooldown grows from 1 to 2, 4, 8 seconds up to 30 seconds for the same hashed endpoint/model profile. Other endpoint/model profiles remain independent. Only 16 hashed profiles can be retained in memory. A successful generation clears the profile's streak; configuration or authorization changes reset the cooldown. No content, file paths, source snippets, raw endpoint URLs or model names are recorded by the cooldown.
The guard applies only to automatic Ollama network generations, after
normal editor eligibility and privacy checks. It never bypasses exclusions,
workspace trust, cancellation or remote approval. Manual Trigger Completion
is allowed even during the cooldown; no timer or background retry is started.
A session-only Automatic Requests Deferred counter appears in the native
Session Activity section and source-free Diagnostics. This counter is not a
measure of model quality, completion acceptance or server recovery.
Unit tests cover TTL, escalating delay, capacity, error classification and reset. The mocked VS Code extension-host checks temporary HTTP 503 suppression and immediate manual recovery. Actual physical-editor, local-model and accessibility acceptance remain NOT RUN. This source change has no published VSIX until a separately verified release.
The explicit Reset Session Activity command is available from Quick Actions and Session Activity in the native Control Center. It presents a confirmation dialog, then stops queued/in-flight generation, invalidates pending one-shot acceptance callbacks, clears all in-memory completion counts, recent inference durations and automatic failure cooldowns. It does not change the user's model, language enablement, endpoint/consent preferences, or explicitly chosen related-file selection. Local privacy timing statistics are independent and remain controlled by Reset Privacy Timing Statistics.
The reset never contacts Ollama, scans files or uploads data. Old cancelled generations cannot repopulate counters after reset. The development tests cover registration, counter erasure, stale accepted-item tokens and untouched preferences with a mocked VS Code extension host. Physical accessibility and real Ollama validation remain NOT RUN.
The features described above as post-v1.3.1 development—including accepted-item counters, status clarification, opt-in latency percentiles, transient retry cooldown and session-activity reset—are now bundled in the verified GitHub v1.4.0 pre-release. Earlier historical notes retain their original implementation-stage wording. They do not mean these features are absent from v1.4.0. The latest GitHub stable is still v1.3.0; the extension marketplaces were not updated. Physical keyboard/screen-reader testing and real model-quality acceptance remain NOT RUN.
A new explicit Copy Sanitized Support Summary action in the native Session Activity panel, Quick Actions, and Command Palette copies only session-level counters, bounded configuration counts, endpoint category (not URL), and the previous user-requested model/file-check status. It never reads document text, includes filenames or model tags, contacts Ollama, or writes a permanent report. Optional latency p50/p95 appears only while performance diagnostics are enabled. A successful copy uses the operating-system clipboard: review the report before sharing it, and be aware that other local applications may be able to read clipboard contents.
The editor-scoped last manual model/file statuses are now reset together on editor switch, edit or close. In-flight checks from the prior editor cannot restore an obsolete result. This fixes stale manual readiness information after moving between files with different language overrides. Unit and mocked editor-host coverage guard these privacy boundaries. Real desktop and screen-reader acceptance remains NOT RUN.
Check Ollama Connection and Check Active Editor Model now send a cancellable signal with each explicitly requested model-list check. Starting another check, changing editor, changing settings, or invalidating the manual status aborts the outdated request instead of leaving unnecessary discovery work running. Previously returned status remains session-only; a late check cannot reintroduce a prior editor's model result. Cancellation does not undo bytes a remote service has already received. There are no automatic network checks, downloads, or telemetry. Automated tests verify abort and stale-token behavior; real physical-device/model acceptance remains NOT RUN.
The Active Editor section provides a Pause/Resume Suggestions for This File control when a local file is active, plus Resume All Session-Paused Files when any files are paused. This is a local-only switch; opening the Control Center never requests an Ollama model. Pauses block both manual and automatic completions and immediately cancel queued or in-flight generation. The sidebar shows an aggregate count and state, not filenames, URIs or source. Pauses are non-persistent, independent of global auto-trigger settings, and limited to 128 SHA-256 URI digests in memory. This source-only feature does not change the published v1.4.3 VSIX.