Adds Perplexity to Roo-Code as a first-class provider, including handler and provider-registry wiring so users can select and call Perplexity chat-completions models (defaulting to `sonar-pro`) from existing settings and model-selection flows.
What ChangedAdds Perplexity to Roo-Code as a first-class provider, including handler and provider-registry wiring so users can select and call Perplexity chat-completions models (defaulting to `sonar-pro`) from existing settings and model-selection flows.
Why It MattersRoo-Code users can now use Perplexity directly from the existing UI and configuration flow, so teams with Perplexity access can onboard a new LLM backend without custom connectors or external integration code. The new provider path also hardcodes endpoint and model exposure (`api.perplexity.ai`, `sonar*` models) and key fallback precedence, so we should watch for misconfigured key precedence, incorrect model fallback behavior, and any regressions in stream/error handling after release.
Final score 82Confidence 961 evidence itemRoo-CodePerplexity APIPerplexityHandlerProvider registryapi.perplexity.aiperplexityApiKey
Fixes ELECTRON-19G by handling deleted-conversation 404 responses in the conversation fetch flow, so the app no longer rethrows an unhandled rejection on a stale `/conversation/<id>` route and instead recovers with a not-found warning and redirect to home.
What ChangedFixes ELECTRON-19G by handling deleted-conversation 404 responses in the conversation fetch flow, so the app no longer rethrows an unhandled rejection on a stale `/conversation/<id>` route and instead recovers with a not-found warning and redirect to home.
Why It MattersFor users who delete chats and immediately navigate with the back button, the app now returns to the welcome page instead of leaving them in an empty conversation shell or triggering an unhandled promise rejection, so workflow continuity is restored and support noise from repeated 404 navigation crashes should drop. The underlying mechanism uses `ConversationIndex` error swallowing plus `useNavigationType()` history replacement; teams should continue watching for other conversation routes that still allow stale `#/conversation/<id>` entries to trigger unhandled 404 rethrows or back-stack inconsistencies.
Final score 82Confidence 971 evidence itemConversationIndexConversationArtifactProvideruseNavigationTypeBackendHttpErrorconversation.notFound
This PR fixes a packaging regression by ensuring `plugin/` runtime dependencies are actually shipped in the marketplace bundle. It does two related steps: runs `npm install --production` when building the plugin package, and includes `plugin/node_modules` in `package.json` `files` so required modules (for example zod and parser/runtime deps) are present after install.
What ChangedThis PR fixes a packaging regression by ensuring `plugin/` runtime dependencies are actually shipped in the marketplace bundle. It does two related steps: runs `npm install --production` when building the plugin package, and includes `plugin/node_modules` in `package.json` `files` so required modules (for example zod and parser/runtime deps) are present after install.
Why It MattersUsers installing claude-mem from the marketplace are less likely to lose entire sessions silently because prompt-processing workers can start with the required runtime modules available instead of missing-dependency crashes. After this change, a startup failure path that previously left sessions un-summarized should stop happening, which removes a failure mode that could hide across platforms where local hook errors are not surfaced. Continue monitoring whether dependency updates outside this repo can reintroduce a gap between lockfile state and bundled modules, and watch installation size/upgrade behavior to ensure the bundle remains stable as dependencies change.
Final score 81Confidence 961 evidence itemclaude-memplugin/node_modulesscripts/build-hooks.jspackage.jsonnpm install --productionmarketplace bundleruntime dependencieszod
The PR updates local persistence recovery by version-gating Desktop-started Worker reuse. It now writes `appVersion` and `startedBy` into control-surface metadata, and Desktop launches only reuse a local Worker when that metadata shows the same launching app version; legacy or ownerless connection files are treated as stale and repaired before replacement Workers are started.
What ChangedThe PR updates local persistence recovery by version-gating Desktop-started Worker reuse. It now writes `appVersion` and `startedBy` into control-surface metadata, and Desktop launches only reuse a local Worker when that metadata shows the same launching app version; legacy or ownerless connection files are treated as stale and repaired before replacement Workers are started.
Why It MattersDesktop users who upgrade the installed app should see more reliable workspace recovery instead of resuming into an old background process that can leave workspace data stale or missing, because the app now blocks cross-version Desktop Worker reuse and forces a clean recovery path when metadata is stale or incompatible. Concretely, this changes runtime selection from optimistic reuse to version-matched reuse for Desktop-started Workers, with old connection records repaired before replacement. Teams should watch mixed CLI/Desktop workflows for any regressions in cross-process handoff and verify lock/connection file repair remains robust after repeated upgrades and rollbacks.
Final score 81Confidence 961 evidence itemlocal Workercontrol-surface connection fileappVersionstartedBysystem.capabilities--app-versionpersistence recovery
The PR introduces local session tabs for sidebar and editor-tab chat so multiple repository-scoped chats can stay open simultaneously instead of replacing the active session when users start parallel work.
What ChangedThe PR introduces local session tabs for sidebar and editor-tab chat so multiple repository-scoped chats can stay open simultaneously instead of replacing the active session when users start parallel work.
Why It MattersDevelopers running parallel repository tasks in Kilo can keep multiple chats open without losing an active conversation, so they can switch contexts faster and continue work without manual session handoffs. The change now centralizes local session lifecycle handling (tabs, state restoration, and shared connection reuse); operators should monitor for regressions in tab import/create races and cross-webview session synchronization, since those paths can still cause one chat session to shadow another if state conflicts appear.
PR #1210 makes GitHub Copilot login and request handling in oh-my-pi consistent by sending `Copilot-Integration-Id: vscode-chat` on non-auth Copilot API calls and by resolving a stable user `accountId` from `GET /user` during `/login` so existing credential rows can be matched and updated.
What ChangedPR #1210 makes GitHub Copilot login and request handling in oh-my-pi consistent by sending `Copilot-Integration-Id: vscode-chat` on non-auth Copilot API calls and by resolving a stable user `accountId` from `GET /user` during `/login` so existing credential rows can be matched and updated.
Why It MattersDevelopers and operators using oh-my-pi with GitHub Copilot should encounter fewer intermittent "model not available" failures and less credential clutter after login, making Copilot-backed workflows more predictable. By standardizing the integrator header and adding stable `accountId` matching in the credential store, the same user should no longer alternate between incompatible token scopes or accumulate stale rows; continue monitoring for `/user` fetch failures (which silently disable dedup for that login) and for future GitHub changes to token scoping or `/user` response fields that could reintroduce identity mismatch.
Final score 80Confidence 931 evidence itemoh-my-piGitHub CopilotCopilot-Integration-Idghu_ tokengho_ tokenGET /userOAuthCredentialsupsertAuthCredentialForProvider
Prevents automatic persistence writes from clearing durable workspace state with an empty list after restart or sync, requiring an explicit opt-in flag for intentional clears and threading that flag through main, renderer, and control-surface persistence paths.
What ChangedPrevents automatic persistence writes from clearing durable workspace state with an empty list after restart or sync, requiring an explicit opt-in flag for intentional clears and threading that flag through main, renderer, and control-surface persistence paths.
Why It MattersOperators and developers using opencove will be less likely to lose saved workspace configuration unexpectedly after restart or sync, because automatic background persistence no longer wipes durable state unless the clear action is explicitly requested. This reduces recovery incidents and accidental data loss, but teams should continue watching for any legitimate clear workflows that fail to pass the new allow flag and thereby appear as no-ops.
This change makes Copilot model selection honor entitlement and policy directly from the live `/models` response: `filterModel` now excludes policy-disabled or non-picker models, and `ModelManagerOptions.dynamicIsAuthoritative=true` forces `resolveProviderModels` to keep only IDs returned by the live API, preventing stale bundled entries from persisting in merged results and cache.
What ChangedThis change makes Copilot model selection honor entitlement and policy directly from the live `/models` response: `filterModel` now excludes policy-disabled or non-picker models, and `ModelManagerOptions.dynamicIsAuthoritative=true` forces `resolveProviderModels` to keep only IDs returned by the live API, preventing stale bundled entries from persisting in merged results and cache.
Why It MattersGitHub Copilot users and operators will no longer see or pick blocked/disabled models in the model picker, reducing the chance of selecting a model that fails immediately with `model_not_supported`, and teams should watch whether policy changes and temporary API anomalies keep the cached catalog synchronized in time. Concretely, the PR enforces `filterModel` checks for `policy.state != enabled` and `model_picker_enabled = false` and applies `dynamicIsAuthoritative` so merged results are constrained to live response IDs; testing showed a catalog reduction from 31 to 12 entries matching account settings.
Final score 80Confidence 961 evidence itemGitHub Copilot/models endpointfilterModelresolveProviderModelsmodel_picker_enableddynamicIsAuthoritative
Version 0.5.0 of notebooklm-py completed the phase-1 deprecation audit by removing remaining v0.3-era deprecated APIs and fields (including old sharing call, legacy type properties, and deprecated source/CLI params), while updating docs, changelogs, and version gates to schedule the removals for v0.6.0 and keeping `RPCError.rpc_id` and `RPCError.code` as permanent aliases.
What ChangedVersion 0.5.0 of notebooklm-py completed the phase-1 deprecation audit by removing remaining v0.3-era deprecated APIs and fields (including old sharing call, legacy type properties, and deprecated source/CLI params), while updating docs, changelogs, and version gates to schedule the removals for v0.6.0 and keeping `RPCError.rpc_id` and `RPCError.code` as permanent aliases.
Why It MattersDevelopers maintaining integrations with notebooklm-py now need to migrate callsites that still use removed APIs (for example old sharing calls, `mime_type`/`poll_interval`-style parameters, and legacy source/artifact type fields), because those paths will break when deprecations are enforced at v0.6.0, while existing code reading `RPCError.rpc_id` or `RPCError.code` should continue to work. This reduces long-term API ambiguity but shifts migration work to consumers, so the key follow-up is to monitor CI and downstream clients for hidden breakages in polling, file-source, and sharing flows before the next release.
Final score 80Confidence 951 evidence itemnotebooklm-pyversion 0.5.0client.notebooks.share()poll_intervalinitial_intervalRPCError.rpc_idRPCError.codev0.6.0 deprecation timeline
Aider now wraps `Path(...).is_dir()`/`resolve()` calls in three `main.py` validation paths with `try/except OSError`, so malformed or overlong paths (including those arising from large `--system-prompt` values) no longer crash the CLI with tracebacks and instead emit a warning and continue.
What ChangedAider now wraps `Path(...).is_dir()`/`resolve()` calls in three `main.py` validation paths with `try/except OSError`, so malformed or overlong paths (including those arising from large `--system-prompt` values) no longer crash the CLI with tracebacks and instead emit a warning and continue.
Why It MattersAider users who pass very long or malformed path arguments—especially via command-line prompts—will avoid abrupt crashes, so their sessions keep running while the tool surfaces a warning and skips only the bad path. This is implemented by catching `OSError` in the three `Path(fname).is_dir()`/`resolve()` branches in `main.py`; follow whether the warning-based skip behavior causes any important files to be silently ignored and whether other path-handling code paths still raise uncaught filesystem exceptions.
Final score 80Confidence 971 evidence itemPath.is_dir()OSErrorENAMETOOLONGaider/main.pyresolve()
Added a new Ctrl+Tab task switcher with a visual overlay, enabling fast task navigation via keyboard, plus an MRU-based ordering path through a MobX task-switching store.
What ChangedAdded a new Ctrl+Tab task switcher with a visual overlay, enabling fast task navigation via keyboard, plus an MRU-based ordering path through a MobX task-switching store.
Why It MattersTask-heavy users can move between active tasks faster by using Ctrl+Tab instead of manual list browsing, which should reduce interruption time when jumping back to ongoing work; the implementation also changes task ranking to status+recency for faster access to likely targets. Continue watching for shortcut conflicts with existing bindings, correctness of MRU ordering in edge cases (e.g., rapid task changes), and whether overlay focus/accessibility behaves reliably across input-heavy sessions.
Final score 80Confidence 961 evidence itemCtrl+Tabtask switcheroverlay UIMobXMRU orderingkeyboard shortcuts
Superset now adds a local contributor setup path via `bun setup:local` that enables a fresh clone to boot a local web/API/desktop stack without required third-party credentials, while keeping strict validation for cloud/internal profiles and isolating each worktree’s local DB and desktop state.
What ChangedSuperset now adds a local contributor setup path via `bun setup:local` that enables a fresh clone to boot a local web/API/desktop stack without required third-party credentials, while keeping strict validation for cloud/internal profiles and isolating each worktree’s local DB and desktop state.
Why It MattersDevelopers can now bring up a local Superset environment from a fresh clone without hunting for OAuth, payment, email, or queue service credentials, which lowers setup friction and makes onboarding and parallel local worktree development faster. The `local` profile now runs with non-blocking optional integration checks and explicit worktree-scoped DB/desktop state, while production and canary behaviors remain strict; operators should watch for accidental environment-profile leakage (especially with `VERCEL` markers), and for cleanup behavior when local worktrees are removed so stale state does not accumulate.
Final score 79Confidence 941 evidence itembun setup:localSUPERSET_PROFILEdeployment profileslocal profileDocker PostgresElectricper-worktree local state
This change updates `entire trail list` to stop returning an unbounded legacy table and instead show a bounded recent view (default 10 items) of repo-wide in-progress trails, while adding explicit composable `--author` and `--status` filtering (including clear-filter behavior) and removing the implicit `--all`-style listing mode.
What ChangedThis change updates `entire trail list` to stop returning an unbounded legacy table and instead show a bounded recent view (default 10 items) of repo-wide in-progress trails, while adding explicit composable `--author` and `--status` filtering (including clear-filter behavior) and removing the implicit `--all`-style listing mode.
Why It MattersCLI operators and repo triage users now get a faster, less noisy view of active work by default, which reduces time spent finding current trails during debugging or incident coordination. The command now prioritizes recent in-progress items and requires explicit filter composition for broader searches, making results easier to control and reducing accidental broad scans in large repos. Watch for adoption friction in automation or power-user workflows that depended on the removed implicit full-list behavior, especially scripts assuming `--all` or default unfiltered visibility.
Final score 79Confidence 961 evidence itementire trail list--author--status--limit--allin_progress
The PR rewrites the task editor’s file-tree model to use hierarchical `FileNode` objects with per-folder `children`, while normalizing local filesystem relative paths to POSIX form before loading, storing, revealing, adding/removing, and processing watch events. This addresses a Windows-specific bug where backslash paths were treated as flat IDs (for example `src\routes\+page.svelte`), producing duplicated nested folder/file rows; root nodes are now expanded lazily and visible rows are derived from each directory’s actual children.
What ChangedThe PR rewrites the task editor’s file-tree model to use hierarchical `FileNode` objects with per-folder `children`, while normalizing local filesystem relative paths to POSIX form before loading, storing, revealing, adding/removing, and processing watch events. This addresses a Windows-specific bug where backslash paths were treated as flat IDs (for example `src\routes\+page.svelte`), producing duplicated nested folder/file rows; root nodes are now expanded lazily and visible rows are derived from each directory’s actual children.
Why It MattersWindows users editing tasks now get a correct file tree in the UI—no duplicated folder/file wrappers for nested paths—so file selection and navigation stay reliable during day-to-day work instead of causing confusion or wrong target selection. The implementation replaces the flat-tree identity logic with explicit per-folder children and applies POSIX path normalization in core file-tree operations (`LocalFileSystem` and renderer loading/reveal/watch flows), so the next signal to watch is whether path-sensitive events (renames, deletions, and cross-platform sync paths) continue to preserve expansion state and visible-node consistency at scale.
The PR changes `truncateToWidth` so legacy and custom TUI calls that pass string ellipsis values are handled in the TypeScript wrapper instead of being sent directly to the native addon, preventing the napi enum-conversion failure that could crash the render loop.
What ChangedThe PR changes `truncateToWidth` so legacy and custom TUI calls that pass string ellipsis values are handled in the TypeScript wrapper instead of being sent directly to the native addon, preventing the napi enum-conversion failure that could crash the render loop.
Why It MattersDevelopers using custom or older TUI renderers that pass string ellipsis values to `truncateToWidth` should see rendering no longer fail mid-loop, so UI text truncation remains stable during normal runs instead of dropping into crash states. The change routes strings through the TypeScript layer before native invocation while leaving enum/null inputs on the old native path, and teams should monitor whether additional legacy argument shapes or wrapper-to-native API changes reintroduce conversion failures.
Final score 79Confidence 981 evidence itemtruncateToWidthEllipsisTypeScript wrappernative addonnapi enum conversion
A broken import in `manager.py` referenced `get_default_skill_directories`, which does not exist in `discovery.py`; this PR changed it to `get_default_skill_dirs`, restoring normal execution for `create_skill`, `edit_skill`, `delete_skill`, `patch_skill`, and `write_skill_file`.
What ChangedA broken import in `manager.py` referenced `get_default_skill_directories`, which does not exist in `discovery.py`; this PR changed it to `get_default_skill_dirs`, restoring normal execution for `create_skill`, `edit_skill`, `delete_skill`, `patch_skill`, and `write_skill_file`.
Why It MattersDevelopers using PraisonAI skill workflows can again run core operations like creating, editing, deleting, patching, and writing skills without the process stopping on import, so automation and manual skill-management flows stay usable instead of failing early. The fix is a correctness fix in module wiring: a function-name mismatch between files had been breaking `SkillManager` entrypoints and making repeated skill operations unreliable. Continue to watch for similarly mismatched imports in adjacent utility modules, especially if discovery-related symbols are renamed again.
Final score 79Confidence 981 evidence itemMervinPraison/PraisonAImanager.pydiscovery.pyget_default_skill_dirsget_default_skill_directoriesSkillManagercreate_skilledit_skilldelete_skillpatch_skillwrite_skill_file
Adds persistent column reordering for InsForge `DataGrid`, saving drag-and-drop order in local storage per table and restoring it on reload while keeping default behavior when persistence is not enabled.
What ChangedAdds persistent column reordering for InsForge `DataGrid`, saving drag-and-drop order in local storage per table and restoring it on reload while keeping default behavior when persistence is not enabled.
Why It MattersUsers configuring table views will keep their preferred column layout after refreshing the page, so they spend less time reordering columns before doing data work, while developers can enable stable, per-view behavior by setting `storageKey` on each grid. The implementation now restores or safely resets order based on validated stored state, so operators should watch for malformed or stale localStorage values and storage failures (such as private-browser restrictions or quota limits) that can silently force a fallback to default ordering.
Final score 78Confidence 931 evidence itemDataGridlocalStorageuseColumnOrderstorageKeycolumn reorderingdrag-and-drop
Serena now adds first-class CUE language support by introducing a dedicated `CueLanguageServer` wired to the official `cue lsp`, enabling CUE-aware LSP behavior (startup + symbol/reference handling) for `.cue` projects.
What ChangedSerena now adds first-class CUE language support by introducing a dedicated `CueLanguageServer` wired to the official `cue lsp`, enabling CUE-aware LSP behavior (startup + symbol/reference handling) for `.cue` projects.
Why It MattersDevelopers editing `.cue` files in Serena can now get CUE language features such as symbol navigation and cross-file references to work immediately when opening CUE repositories, instead of manually wiring or debugging LSP setup themselves. The change uses a pinned default `cue` binary (`v0.16.1`) and verifies downloads by SHA-256 across Linux/macOS/Windows (x64 and arm64), with settings to override both path and version when teams need a custom toolchain. Operators and maintainers should watch for reliability in first-run downloads and for upstream `cue lsp` protocol changes, since either could break startup or reduce LSP completeness for CUE users.
Final score 78Confidence 961 evidence itemCUECueLanguageServercue lspLanguage.CUELanguageServerDependencyProviderSinglePathls_specific_settings.cue
A bug fix makes both the claude and cli parsers resolve relative `@` include paths against the directory of the source file, so includes work when the prompt file is outside the current working directory.
What ChangedA bug fix makes both the claude and cli parsers resolve relative `@` include paths against the directory of the source file, so includes work when the prompt file is outside the current working directory.
Why It MattersDevelopers using CodeCompanion prompt files outside their current working directory can now load referenced local files with `@` includes without missing-context failures, which reduces broken prompt behavior during multi-location workflows; watch whether any custom include workflows implicitly depend on cwd-based resolution and track cases where source_dir is not discoverable, as those will continue using the previous fallback behavior.
The PR makes NotebookLM emit a DeprecationWarning whenever decode is run with NOTEBOOKLM_STRICT_DECODE=0 and safe_index uses soft-mode fallback, turning a previously implicit compatibility path into an explicit, visible transition signal tied to a documented retirement plan.
What ChangedThe PR makes NotebookLM emit a DeprecationWarning whenever decode is run with NOTEBOOKLM_STRICT_DECODE=0 and safe_index uses soft-mode fallback, turning a previously implicit compatibility path into an explicit, visible transition signal tied to a documented retirement plan.
Why It MattersDevelopers running NotebookLM with strict decode disabled will now see a visible warning during inference/configuration fallbacks, which makes risky deprecation behavior explicit instead of hidden and gives teams time to migrate before the soft-mode path is retired. This should reduce surprise failures and debugging time when environments still depend on deprecated decode behavior; watchers should track warning frequency, log filtering, and whether CI/ops alerting treats the new warning as an actionable migration item before v0.6.0.
Final score 77Confidence 941 evidence itemsafe_indexNOTEBOOKLM_STRICT_DECODEDeprecationWarningstrict modesoft mode
The PR replaces free-text model/variant fields in Kilo-Org/kilocode’s VSCode Agent Behavior settings with dropdown selectors, so users can pick and persist model variants through a consistent settings UI.
ContributionImplemented a UI change that introduces reusable dropdown-based model and variant selectors for Agent Behavior configuration, replacing plain text entry and providing a persistent override path for model settings.
ImpactDevelopers configuring agents in Kilo-Org/kilocode can now choose model and variant via dropdowns and have the selection persist, reducing invalid manual input and making behavior changes easier to apply correctly; watch for option drift (new or removed model variants), persistence across extension restarts, and compatibility with existing manually edited config entries.
The HN thread around Google's Antigravity highlights a likely product reset: updates are being described as a pivot that breaks continuity for existing Antigravity IDE users, with users reporting disorientation and relying on community restore scripts to recover settings and chat history.
ContributionIdentifies a single behavioral shift signal: Antigravity’s direction appears to have moved away from preserving the existing IDE experience, creating backward-compatibility and migration pain for current users.
ImpactDevelopers currently depending on Antigravity for coding-agent work may see abrupt workflow disruption when updates land, including broken local setup and fragile recovery of chat/history state, so teams should watch for an official migration path and whether updates continue to preserve existing installations. The technical concern is that the product is being treated as a reset rather than a stable, incremental IDE evolution, which increases the risk of forced migration to alternatives if compatibility and recovery tooling are not provided.
Google’s Antigravity IDE update is being described as a product reset that changes prior behavior and breaks continuity for existing users, who now face reconfiguration instead of a smooth upgrade.
ContributionThe update introduces a significant workflow discontinuity by replacing existing Antigravity behavior and requiring users to manually rebuild environment state (settings, extensions, and local history) instead of preserving prior configuration.
ImpactDevelopers who were already using Antigravity may experience lost local context and interrupted coding workflows after the update, which can slow delivery and increase risk when switching back into active work. The first visible effect is compatibility rollback, and the signal to watch is whether Google provides an official migration path for existing installations, local agent state, and extension settings because community-led recovery scripts are not a dependable long-term option.
In v7.3.6 (pre-release), Kilocode changed VS Code local CLI behavior to prevent repeated reconnect attempts when the event stream is down, reducing an unstable reconnect loop during affected sessions.
ContributionAdded a reconnect-stability fix in the VS Code local CLI integration that suppresses reconnect churn when the event stream is unavailable, rather than continuously retriggering connection attempts.
ImpactDevelopers using Kilocode inside VS Code will get steadier sessions when the local event stream is interrupted, because the extension should stop thrashing with nonstop reconnect retries and keep the UI from becoming unstable during outages. The fix appears to gate reconnect behavior when stream availability is missing, which can lower resource waste and interruption frequency; watch for recovery paths after long disconnects to ensure the CLI resumes promptly when the stream returns and does not leave users stuck in a stale disconnected state.
Kilocode v7.3.5 pre-release includes a fix that prevents the local VS Code CLI path from repeatedly reconnecting when the event stream is unavailable.
ContributionThe release changes reconnect handling in the VS Code local CLI integration so transient loss of event stream availability no longer triggers rapid repeated reconnect loops; this is a concrete stability fix for the editor-side workflow.
ImpactLocal VS Code users of KiloCode should see more stable review/chat sessions when the backend event stream is temporarily unavailable, with fewer connection churn episodes that disrupt work. Technically, the patch reduces reconnect-loop behavior in the CLI event-stream bridge so the client does not continuously retry during outages. Teams should watch for whether legitimate reconnections are being delayed after long stream interruptions and whether retry metrics still surface real failures instead of silently masking them.
The PR introduces local session tabs for sidebar and editor-tab chat so multiple repository-scoped chats can stay open simultaneously instead of replacing the active session when users start parallel work.
ContributionImplements persistent local tab/session state per webview, reuses the existing shared extension connection, and restores tab behavior to keep multiple local chats stable across creation/import race scenarios while matching Agent Manager-style interaction.
ImpactDevelopers running parallel repository tasks in Kilo can keep multiple chats open without losing an active conversation, so they can switch contexts faster and continue work without manual session handoffs. The change now centralizes local session lifecycle handling (tabs, state restoration, and shared connection reuse); operators should monitor for regressions in tab import/create races and cross-webview session synchronization, since those paths can still cause one chat session to shadow another if state conflicts appear.
This release addresses a concrete reliability issue in Kilo’s VS Code extension where Windows users could see stale indexing status in Settings and chat while files were still being indexed. The fix is important because it removes a user-facing inconsistency that can mislead operators into believing indexing is stuck or outdated when it is actually progressing.
ContributionPatched the indexing status flow so Windows path-casing no longer breaks SSE event matching and status messages arriving before the initial configuration payload are no longer discarded. This keeps the UI status model coherent across the entire indexing session.
ImpactWindows users of Kilo in VS Code will see the indexing indicator in Settings and chat update correctly during indexing, so they can make follow-up edits or decisions based on real progress instead of stale state. The update achieves this by fixing path matching to account for Windows casing differences and by keeping early status events from being dropped before config initialization, but progress reliability should continue to be watched for similar race conditions in other startup paths and indexing backends.
The main change in this burst is a model replacement for autocomplete, moving kilocode to the mercury-edit-2 model. This is significant because it changes the core behavior users feel directly during code completion, so downstream quality and speed of suggestions can shift even when other changes are cosmetic or CI-only.
ContributionThe burst’s primary contribution is the upgrade of the autocomplete backend model from its previous variant to mercury-edit-2, which changes the completion inference behavior used by users during editing.
ImpactDevelopers using kilocode autocomplete will likely see different suggestion behavior after the switch to mercury-edit-2, so teams should validate immediately whether suggestion relevance and pace still match their workflow before treating it as a default. The model change is the control point for completion quality, so watch acceptance/rejection ratios, context mismatch cases, and any rise in manual correction edits in real projects to catch regressions early.
Kilocode merged a change that swaps the autocomplete backend to the mercury-edit-2 model, so code-completion behavior now depends on this new model’s generation characteristics instead of the previous default.
ContributionThe PR reconfigures Kilocode’s completion pipeline to use mercury-edit-2 as the active autocomplete model, creating a direct change in how suggestions are generated for user code edits.
ImpactDevelopers using Kilocode autocomplete should notice a behavioral change in suggestion quality or style, so teams should check whether this makes edits more useful or introduces more incorrect completions during real coding sessions. This model swap changes the generation backend itself, so monitor completion acceptance rates, suggestion correctness for common coding patterns, and latency so rollout risk can be caught before it affects developer productivity.
This PR bumps the repo dependency `turbo` from 2.9.8 to 2.9.14, with the primary effect of consuming upstream security hardening from the Turborepo release rather than introducing a feature change in agor code.
ContributionPulls the project’s `turbo` toolchain dependency to version 2.9.14, which includes upstream fixes for command execution and auth/workspace-detection security flaws.
ImpactDevelopers and CI operators using agor’s build tooling get a lower chance of being hit by newly disclosed Turborepo security issues, so repository automation is less likely to trigger unsafe command execution or session/auth handling failures after upgrades. The concrete technical change is the dependency-level security patch set in v2.9.14 (command execution hardening in the VS Code extension path, auth callback state validation, and safer Yarn detection behavior), and this should be monitored for any build-orchestration regressions or behavior changes in existing workflows.
The PR changes `truncateToWidth` so legacy and custom TUI calls that pass string ellipsis values are handled in the TypeScript wrapper instead of being sent directly to the native addon, preventing the napi enum-conversion failure that could crash the render loop.
ContributionAdded compatibility handling for string ellipsis arguments in `truncateToWidth`, preserving existing enum/null behavior in the native layer so older renderer calls remain functional without causing type-conversion crashes.
ImpactDevelopers using custom or older TUI renderers that pass string ellipsis values to `truncateToWidth` should see rendering no longer fail mid-loop, so UI text truncation remains stable during normal runs instead of dropping into crash states. The change routes strings through the TypeScript layer before native invocation while leaving enum/null inputs on the old native path, and teams should monitor whether additional legacy argument shapes or wrapper-to-native API changes reintroduce conversion failures.
This change makes Copilot model selection honor entitlement and policy directly from the live `/models` response: `filterModel` now excludes policy-disabled or non-picker models, and `ModelManagerOptions.dynamicIsAuthoritative=true` forces `resolveProviderModels` to keep only IDs returned by the live API, preventing stale bundled entries from persisting in merged results and cache.
ContributionImplemented entitlement-aware model curation by adding explicit policy/picker filtering and an authoritative live-catalog merge path, so disabled or deprecated Copilot models are removed before users can select them.
ImpactGitHub Copilot users and operators will no longer see or pick blocked/disabled models in the model picker, reducing the chance of selecting a model that fails immediately with `model_not_supported`, and teams should watch whether policy changes and temporary API anomalies keep the cached catalog synchronized in time. Concretely, the PR enforces `filterModel` checks for `policy.state != enabled` and `model_picker_enabled = false` and applies `dynamicIsAuthoritative` so merged results are constrained to live response IDs; testing showed a catalog reduction from 31 to 12 entries matching account settings.
PR #1210 makes GitHub Copilot login and request handling in oh-my-pi consistent by sending `Copilot-Integration-Id: vscode-chat` on non-auth Copilot API calls and by resolving a stable user `accountId` from `GET /user` during `/login` so existing credential rows can be matched and updated.
ContributionImplements a concrete authentication reliability fix: it forces the correct Copilot integrator context on Copilot requests and uses GitHub `/user`-provided stable identity data at login to drive dedup logic instead of creating duplicate credentials on each sign-in.
ImpactDevelopers and operators using oh-my-pi with GitHub Copilot should encounter fewer intermittent "model not available" failures and less credential clutter after login, making Copilot-backed workflows more predictable. By standardizing the integrator header and adding stable `accountId` matching in the credential store, the same user should no longer alternate between incompatible token scopes or accumulate stale rows; continue monitoring for `/user` fetch failures (which silently disable dedup for that login) and for future GitHub changes to token scoping or `/user` response fields that could reintroduce identity mismatch.
Version 0.5.0 of notebooklm-py completed the phase-1 deprecation audit by removing remaining v0.3-era deprecated APIs and fields (including old sharing call, legacy type properties, and deprecated source/CLI params), while updating docs, changelogs, and version gates to schedule the removals for v0.6.0 and keeping `RPCError.rpc_id` and `RPCError.code` as permanent aliases.
ContributionExecuted a concrete API lifecycle cleanup: removed deprecated methods/fields and compatibility shims, switched waiting API usage away from `poll_interval` to `initial_interval`, and synchronized docs/version-gate coverage so the deprecation contract is explicit, with only two explicit error-field aliases preserved.
ImpactDevelopers maintaining integrations with notebooklm-py now need to migrate callsites that still use removed APIs (for example old sharing calls, `mime_type`/`poll_interval`-style parameters, and legacy source/artifact type fields), because those paths will break when deprecations are enforced at v0.6.0, while existing code reading `RPCError.rpc_id` or `RPCError.code` should continue to work. This reduces long-term API ambiguity but shifts migration work to consumers, so the key follow-up is to monitor CI and downstream clients for hidden breakages in polling, file-source, and sharing flows before the next release.
The primary change is a default-on strict RPC decode policy (`NOTEBOOKLM_STRICT_DECODE=1`) so schema mismatches in batchexecute responses now fail as `UnknownRPCMethodError` instead of silently becoming `None`, with docs/tests updated to define behavior and the legacy override.
ContributionSet strict decoding as the default for RPC response handling and standardized environment/docs guidance so drifted API payloads are surfaced as explicit errors at the decode boundary; added/updated tests and ADR/docs to lock behavior and override semantics.
ImpactDevelopers and operators using notebooklm-py RPC flows now get immediate, actionable failures when upstream response schema changes, instead of silently propagating bad responses that can corrupt downstream logic without clear blame. This happens through a default strict-mode flip (`safe_index` raising `UnknownRPCMethodError` on mismatch) while keeping `NOTEBOOKLM_STRICT_DECODE=0` as a temporary compatibility override; watch for clients that spike unknown-method errors after rollout and confirm all integrations can handle the explicit failure mode before disabling the opt-out period.
This update fixes artifact-type filtering so unknown labels and quiz/flashcard artifacts are handled explicitly, reducing mis-filtering behavior in notebook artifact processing.
ContributionAdjusted the filtering logic to explicitly tolerate unrecognized artifact types and to properly process quiz/flashcard artifacts instead of dropping or mishandling them under generic rules.
ImpactOperators and users of notebook artifact workflows should see fewer quiz and flashcard items silently dropped or wrongly treated when metadata is irregular, which makes research outputs and exports more reliable; watch for regression cases where other uncommon artifact types still misbehave and for any custom integrations that depended on the prior filtering behavior.
The PR makes NotebookLM emit a DeprecationWarning whenever decode is run with NOTEBOOKLM_STRICT_DECODE=0 and safe_index uses soft-mode fallback, turning a previously implicit compatibility path into an explicit, visible transition signal tied to a documented retirement plan.
ContributionImplemented warning logic so a fallback triggered by NOTEBOOKLM_STRICT_DECODE=0 is explicitly reported, and added unit tests verifying warning, successful soft-mode traversal, and strict-mode silence so behavior is controlled and regression-resistant.
ImpactDevelopers running NotebookLM with strict decode disabled will now see a visible warning during inference/configuration fallbacks, which makes risky deprecation behavior explicit instead of hidden and gives teams time to migrate before the soft-mode path is retired. This should reduce surprise failures and debugging time when environments still depend on deprecated decode behavior; watchers should track warning frequency, log filtering, and whether CI/ops alerting treats the new warning as an actionable migration item before v0.6.0.
This change adds explicit `__all__` declarations for four internal collaborator modules and adds CI-pinned tests for those export lists, creating a stable internal module boundary while performing the Tier 12 to Tier 13 migration docs work.
ContributionIntroduced and pinned explicit symbol-export lists for four internal collaborator modules so the documented Tier13 module map is enforced by tests, reducing risk that renamed or moved private modules silently break consumers that still import through internal paths.
ImpactDevelopers maintaining NotebookLM’s internal integrations will get earlier, automated detection of export-surface drift during Tier12→13 migration work, so internal imports are less likely to break unexpectedly after refactors. This is done by adding `__all__` in the four collaborator modules and a dedicated CI-backed test that fails on silent export-list changes, which also aligns with the migration documentation and symbol-shim mapping. We should still watch whether modules outside this pinned set start to drift and whether the migration document remains synchronized with future symbol moves.
The PR renamed the remaining lifted internal `_core_*` modules to explicit ownership-oriented names (session, transport, RPC, cache, metrics), updated imports/docs/tests to match, and explicitly retained `_core.py` as a legacy private shim to avoid public API breakage.
ContributionIntroduced a concrete internal refactor that standardizes module naming by ownership layer and updates all direct references, while preserving compatibility through the legacy `_core.py` shim.
ImpactDevelopers maintaining notebooklm-py should be able to work on session-layer and transport-layer internals with less confusion, so refactors are less likely to touch the wrong code path and onboarding into the codebase should be faster, while runtime behavior for users is expected to remain unchanged. We should continue watching for missed internal or downstream imports of retired `_core_*` names (including niche integrations) because those are the main remaining failure risk after the rename sweep.
This change moves the concrete orchestration implementation from the legacy private `ClientCore` location to `notebooklm._session.Session`, and keeps `notebooklm._core` as a compatibility shim so existing private imports keep working.
ContributionEstablishes `_session.Session` as the canonical internal session/orchestration implementation and updates package docs, tests, and fixtures to use it, while preserving `_core.ClientCore` as an alias to avoid private-import breakage.
ImpactDownstream projects that still touch internal notebooklm private paths can continue to run without a forced rewrite, so maintainers get architectural cleanup without forcing immediate breakage for current integrations. The concrete mechanism is a split where implementation moves to `src/notebooklm/_session.py` and `src/notebooklm/_core.py` is reduced to a legacy compatibility/re-export shim; watch for regressions in monkeypatch or private-import coverage and verify any integrations relying on non-public `_core` attributes over subsequent releases.
The PR adds a primary workflow path for Kiro IDE by extending ai-dev-kit installers (`install.sh`, `install.ps1`) with Kiro detection and adding a `--tool kiro` flag to `databricks-skills/install_skills.sh`; installer output is directed to `.kiro/skills/` and `.kiro/settings/mcp.json`, and related docs are updated to describe the new flow and a multi-tool setup.
ContributionImplemented built-in Kiro IDE installation support in Databricks AI Dev Kit by wiring installer-time tool detection and a `--tool kiro` option to write generated skills and MCP settings into Kiro’s expected directories, then aligned documentation to reflect the new integration.
ImpactDevelopers setting up Databricks skills in the Kiro IDE can now use the existing installers to complete IDE-specific configuration automatically, reducing manual file placement and speeding environment onboarding. This should mainly improve reliability of setup for teams using Kiro, while organizations should continue to watch for compatibility drift in Kiro directory layout and MCP config schema across future Kiro versions, plus verify that non-Kiro workflows still behave unchanged.
A bug fix makes both the claude and cli parsers resolve relative `@` include paths against the directory of the source file, so includes work when the prompt file is outside the current working directory.
ContributionImplemented source-file-relative resolution for relative @-includes in both parsers and added corresponding tests, directly fixing include lookup failures caused by relying on the process cwd and keeping explicit exceptions (absolute paths or unknown source_dir) untouched.
ImpactDevelopers using CodeCompanion prompt files outside their current working directory can now load referenced local files with `@` includes without missing-context failures, which reduces broken prompt behavior during multi-location workflows; watch whether any custom include workflows implicitly depend on cwd-based resolution and track cases where source_dir is not discoverable, as those will continue using the previous fallback behavior.
This PR updates ruvnet/ruflo’s CI dependency from pnpm/action-setup v2 to v6, moving the workflow action to a newer major line with updated install/runtime behavior.
ContributionBumps the pinned pnpm/action-setup action version in repository automation from 2 to 6, replacing older action behavior with the newer major release in setup workflows.
ImpactMaintainers and CI operators in ruvnet/ruflo can run workflows against a newer pnpm/action-setup baseline, which should reduce setup-time mismatch issues as pipelines move toward newer Node and pnpm versions. The action upgrade introduces changed internals (including pnpm 11/Node24-era behavior and revised install/bootstrap handling), so the team should watch for regressions in cache behavior, PATH/bin resolution, and Windows/self-update flow consistency after merge.
Added a new Ctrl+Tab task switcher with a visual overlay, enabling fast task navigation via keyboard, plus an MRU-based ordering path through a MobX task-switching store.
ContributionImplemented a keyboard-first workflow for task switching: pressing Ctrl+Tab opens a task-selector overlay and chooses tasks through a MobX-managed ordering model that prioritizes recent and status-appropriate tasks, reducing manual task navigation.
ImpactTask-heavy users can move between active tasks faster by using Ctrl+Tab instead of manual list browsing, which should reduce interruption time when jumping back to ongoing work; the implementation also changes task ranking to status+recency for faster access to likely targets. Continue watching for shortcut conflicts with existing bindings, correctness of MRU ordering in edge cases (e.g., rapid task changes), and whether overlay focus/accessibility behaves reliably across input-heavy sessions.
The PR rewrites the task editor’s file-tree model to use hierarchical `FileNode` objects with per-folder `children`, while normalizing local filesystem relative paths to POSIX form before loading, storing, revealing, adding/removing, and processing watch events. This addresses a Windows-specific bug where backslash paths were treated as flat IDs (for example `src\routes\+page.svelte`), producing duplicated nested folder/file rows; root nodes are now expanded lazily and visible rows are derived from each directory’s actual children.
ContributionIntroduced a concrete tree-structure fix: stable `children` relationships per folder, path normalization for Windows-style separators, and root-only initial loading with expansion-based child loading, eliminating duplicated wrapper rows caused by path-identity misparsing.
ImpactWindows users editing tasks now get a correct file tree in the UI—no duplicated folder/file wrappers for nested paths—so file selection and navigation stay reliable during day-to-day work instead of causing confusion or wrong target selection. The implementation replaces the flat-tree identity logic with explicit per-folder children and applies POSIX path normalization in core file-tree operations (`LocalFileSystem` and renderer loading/reveal/watch flows), so the next signal to watch is whether path-sensitive events (renames, deletions, and cross-platform sync paths) continue to preserve expansion state and visible-node consistency at scale.
The commit burst includes a concrete feature change: INT-420 adds MySQL IAM authentication support in mirrord, including schema and lint-related updates to accept IAM auth configuration, enabling identity-based database access in mirrored local workflows.
ContributionIntroduced a MySQL IAM auth path in mirrord’s MySQL integration, adding configuration schema support and lint validation changes so users can configure IAM-based authentication instead of relying only on static DB credentials.
ImpactDevelopers running local sessions through mirrord can now connect to MySQL using IAM credentials, which makes mirrored testing align with production-style access control and avoids keeping permanent database passwords for local workflows. This should also reduce secret-handling risk for operators, but rollout should be watched for IAM token lifetime handling, permission mapping accuracy, and compatibility with existing non-IAM MySQL auth setups.
This release’s main change is the addition of an Anthropic fast mode in the AI coding-agent, with built-in automatic fallback behavior when that fast mode cannot be used, making this the primary user-facing capability update in v15.1.7.
ContributionAdded an Anthropic fast path to the AI coding-agent and implemented automatic fallback handling so coding requests can continue through an alternate path if fast-mode execution is not available.
ImpactDevelopers using oh-my-pi’s coding-agent can see faster AI coding assistance by default, because the agent now tries a fast Anthropic path first and can continue via fallback instead of stopping on fast-mode failures. The practical follow-up is to watch fallback behavior under real workloads for silent degradation: check whether latency and answer quality change when fallback triggers, and whether fallback failures are clearly surfaced in logs so operators can spot provider-level regressions quickly.
In PR #4272, mirrord changed README links that pointed to `metalbear.co` over to `metalbear.com`, aligning the docs with the canonical website domain and removing the extra redirect path for README navigation.
ContributionReplaced README URL targets from `metalbear.co` to `metalbear.com` for public documentation and community links, so the repository’s documentation now points directly to the canonical domain instead of a redirecting domain.
ImpactDevelopers and operators browsing the mirrord README now reach docs and onboarding links through the canonical site domain directly, so they are less exposed to extra redirect hops and inconsistent URL behavior during evaluation or daily use. The change is a pure documentation-domain canonicalization, but it should be watched for any mixed-domain links left behind in the README and whether external environments or automations that rely on the previous `metalbear.co` host tolerate the domain move without breakage; continued monitoring should include link-health checks on Slack, docs, and /mirrord/ai targets after publication.
An open issue reports a rendering failure in Superset terminal where Claude code responses can appear as unreadable text and repeated twice or three times in the same output area.
ContributionThe issue identifies a concrete user-facing rendering defect in Superset’s terminal UI path for AI-generated (Claude) output, where streamed code output is incorrectly displayed and may be duplicated.
ImpactUsers reading Claude output in Superset terminal can lose trust in what the terminal shows, which can trigger unnecessary reruns, incorrect copy/paste, and debugging mistakes in day-to-day operations. This likely points to a terminal frontend rendering regression for AI response streams; follow whether it also affects other prompts or output sizes, whether logs become consistently reproducible across browsers, and when a fix is merged if any downstream workflow automation consuming terminal text becomes unreliable.
This PR removes the upstream `sdks/vscode` extension now that Kilo ships `packages/kilo-vscode`, updates merge automation to skip `sdks/vscode/**`, and cleans stale annotation/changelog references tied to the removed SDK.
ContributionThe change consolidates VS Code SDK management by decommissioning the old upstream `sdks/vscode` path and making upstream sync logic ignore it, while removing stale release/docs references so the old package is not reintroduced or presented as supported.
ImpactDevelopers and repo operators maintaining Kilo will spend less time resolving confusing extension path drift during merges, because the removed `sdks/vscode` package is now explicitly excluded from upstream sync and documentation/tests are aligned to the active `packages/kilo-vscode` path. This should reduce maintenance churn and accidental reappearance of obsolete files, but downstream consumers should be checked for lingering imports or configs referencing `sdks/vscode`, and CI scripts should be monitored to ensure no hidden dependency still expects that directory.
Prevents automatic persistence writes from clearing durable workspace state with an empty list after restart or sync, requiring an explicit opt-in flag for intentional clears and threading that flag through main, renderer, and control-surface persistence paths.
ContributionIntroduced an explicit ownership check in persistence flow that blocks accidental empty-workspace writes, while keeping user-requested clearing as a dedicated opt-in path; the flag is wired through the main, renderer, and control-surface layers and covered by unit, contract, IPC, and control-surface tests.
ImpactOperators and developers using opencove will be less likely to lose saved workspace configuration unexpectedly after restart or sync, because automatic background persistence no longer wipes durable state unless the clear action is explicitly requested. This reduces recovery incidents and accidental data loss, but teams should continue watching for any legitimate clear workflows that fail to pass the new allow flag and thereby appear as no-ops.
Among the large frontend dependency bump, the key change is upgrading axios from 1.15.2 to 1.16.0, which enforces fetch adapter request-size limits, preserves caller-provided Host headers through proxy requests, and updates URL/credential parsing behavior while adding QUERY method support.
ContributionUpgrades the frontend axios package in OpenHands from 1.15.2 to 1.16.0, enabling stricter fetch-side request-size enforcement (maxBodyLength/maxContentLength), safer proxy header handling (Host preservation), and updated request parsing plus QUERY method support.
ImpactIf your OpenHands frontend integrations send HTTP requests through axios, oversized payloads can now be blocked earlier and proxy-based calls are less likely to lose the intended Host header, which directly lowers the risk of unexpected upload exposure and misrouted API requests. This behavior change is tied to the axios 1.16.0 bump; teams should re-run API integration tests that rely on percent-encoded credentials, virtual-host proxy flows, or unusual protocol strings, and track any regressions in redirects, aborts, and timeout behavior.
OpenHands PR #14466 upgrades `actions/github-script` from v7 to v9, with the core change being migration to the v9 `@actions/github` runtime and the new script-context `getOctokit` pattern, which is a breaking compatibility shift compared with the older `require('@actions/github')` usage.
ContributionUpgrades the workflow dependency chain to `actions/github-script` v9 (including related package updates), replacing legacy CJS import-based client creation patterns with the injected `getOctokit` function and updating the action runtime behavior to match v9 API expectations.
ImpactWorkflow maintainers for this repository can see immediate breakage risk: CI scripts that still use `require('@actions/github')` or redeclare `getOctokit` with `const/let` may fail at runtime, so operators need to audit and migrate workflow scripts before merging to keep automation reliable. This change also enables multi-client token flows through injected `getOctokit` and adds orchestration-id visibility in request user-agent, so teams should monitor post-merge workflow runs for compatibility regressions and verify integration tests around script execution paths before rollout.
A broken import in `manager.py` referenced `get_default_skill_directories`, which does not exist in `discovery.py`; this PR changed it to `get_default_skill_dirs`, restoring normal execution for `create_skill`, `edit_skill`, `delete_skill`, `patch_skill`, and `write_skill_file`.
ContributionCorrected a concrete runtime bug by aligning the imported helper symbol in `manager.py` with its real definition in `discovery.py`, which removes the ImportError barrier for skill management execution.
ImpactDevelopers using PraisonAI skill workflows can again run core operations like creating, editing, deleting, patching, and writing skills without the process stopping on import, so automation and manual skill-management flows stay usable instead of failing early. The fix is a correctness fix in module wiring: a function-name mismatch between files had been breaking `SkillManager` entrypoints and making repeated skill operations unreliable. Continue to watch for similarly mismatched imports in adjacent utility modules, especially if discovery-related symbols are renamed again.
This PR’s strongest tracked change is upgrading `@chenglou/pretext` from `^0.0.5` to `^0.0.7`, adding concrete text-layout fixes (including `keep-all` mixed-script handling and punctuation attachment) that materially affect rendered text correctness for downstream consumers.
ContributionThis change upgrades the text layout dependency to a side-effect-free v0.0.7 build and applies concrete wrap/line-break fixes for mixed-language, punctuation-heavy, and symbol-embedded text, which matters because it directly improves rendering quality and reduces layout churn in user-visible outputs.
ImpactUsers and developers working with hyperframes text rendering (chat, markdown, or rich-text content) should see fewer awkward line wraps and dangling symbols/punctuation in mixed-script content, which means generated or displayed text remains more readable and consistent; continue monitoring multilingual and punctuation-dense templates for any visual wrap regressions after the dependency bump, since this is the first practical follow-up to verify before broad rollout.
The PR bumps frontend lint tooling, with a concrete change to eslint-plugin-i18next 6.1.4 (from 6.1.3), which includes a flat-config type fix and ESLint 10 support, directly addressing tooling compatibility and lint rule correctness during frontend checks.
ContributionUpgraded the frontend lint plugin to a version that fixes a flat-config typing issue and adds compatibility for ESLint 10, reducing friction for teams running lint validation with newer ESLint baselines.
ImpactFrontend developers on this repository should experience fewer lint setup and compatibility failures when running local or CI checks with updated ESLint tooling, which should reduce time spent on false errors during upgrades. This comes from the plugin’s flat-config type correction and explicit ESLint 10 support path, both of which matter in day-to-day dev workflows. Continue to watch for any new rule false positives/negatives after rollout, especially in existing frontend files that rely on flat config patterns.
Kilo-Org/kilocode PR #10221 moves JetBrains package guidance into a single file at `packages/kilo-jetbrains/AGENTS.md` and removes the separate UI skill document, so JetBrains agent instructions (including IntelliJ source lookup, split-mode architecture, and threading/EDT basics) are maintained from one canonical location with `UiStyle` as the shared source for UI constants.
ContributionIntroduces a single authoritative JetBrains guidance document for agents and removes duplicate UI instructions, reducing documentation fragmentation. The change explicitly centralizes IDE usage rules and UI style ownership (`UiStyle`) in `packages/kilo-jetbrains/AGENTS.md`.
ImpactDevelopers working on JetBrains-integrated agents can rely on one canonical guide instead of hunting across multiple files, which should reduce onboarding confusion and inconsistent implementation choices. The practical effect is fewer setup or UI-behavior mismatches when adding or updating JetBrains guidance, since both threading/EDT expectations and reusable style references are now documented in one place. In production-like contributor workflows, watch for any stale links or automation still referencing the removed UI skill file, because those can silently reintroduce the old split documentation flow.
Adds persistent column reordering for InsForge `DataGrid`, saving drag-and-drop order in local storage per table and restoring it on reload while keeping default behavior when persistence is not enabled.
ContributionImplemented a new optional persistence path for column ordering: `DataGrid` now accepts a `storageKey` and uses a `useColumnOrder` hook to read/write ordered column keys in localStorage, validate stored keys, append missing keys after schema changes, and persist reorder events. It also fixes a reorder splice off-by-one edge case in the same flow.
ImpactUsers configuring table views will keep their preferred column layout after refreshing the page, so they spend less time reordering columns before doing data work, while developers can enable stable, per-view behavior by setting `storageKey` on each grid. The implementation now restores or safely resets order based on validated stored state, so operators should watch for malformed or stale localStorage values and storage failures (such as private-browser restrictions or quota limits) that can silently force a fallback to default ordering.
Fixes ELECTRON-19G by handling deleted-conversation 404 responses in the conversation fetch flow, so the app no longer rethrows an unhandled rejection on a stale `/conversation/<id>` route and instead recovers with a not-found warning and redirect to home.
ContributionIntroduced a recoverable missing-conversation path: the conversation fetcher now catches `BackendHttpError` 404s from `ConversationIndex`/`ConversationArtifactProvider` flows, displays the new `conversation.notFound` warning, and navigates with `replace: true` via navigation history handling so back-button navigation cannot loop to the deleted conversation route.
ImpactFor users who delete chats and immediately navigate with the back button, the app now returns to the welcome page instead of leaving them in an empty conversation shell or triggering an unhandled promise rejection, so workflow continuity is restored and support noise from repeated 404 navigation crashes should drop. The underlying mechanism uses `ConversationIndex` error swallowing plus `useNavigationType()` history replacement; teams should continue watching for other conversation routes that still allow stale `#/conversation/<id>` entries to trigger unhandled 404 rethrows or back-stack inconsistencies.
Hyperframes Studio now supports a multi-tab right inspector by adding a CSS tab and refactoring inspector state so users can work across Layers, Design, and CSS in one panel, with tab focus and layout state persisted in the URL and a two-tab limit to keep behavior stable.
ContributionIntroduced a concrete editor capability: a multi-tab right inspector flow (Layers/Design/CSS) by replacing prior single-tab behavior with `rightPanelTabs` + `rightPanelFocusTab`, wiring this state through layout/session/selection/url paths, and gating inspector rendering by feature flag so deployments can enable safely.
ImpactStudio users can keep editing Layers, Design, and CSS in the same right panel and return to the same tab arrangement via URL links, reducing manual UI setup time and context switching during styling and layout work. The implementation persists right-panel tab state in URL and enforces a two-tab guardrail, while defaulting to the Design tab and collapsing single-tab view to full height; teams should watch shared-link compatibility and flag-gated rollout behavior to catch regressions when existing projects open in old or mixed session states.
Superset now adds a local contributor setup path via `bun setup:local` that enables a fresh clone to boot a local web/API/desktop stack without required third-party credentials, while keeping strict validation for cloud/internal profiles and isolating each worktree’s local DB and desktop state.
ContributionIntroduces a local-first bootstrap flow that copies local env examples, derives worktree-specific names, starts required local services, applies migrations, and seeds a local admin account, while allowing missing optional integration credentials in `local`/`ci` and preserving strict checks for `cloud`/`internal`.
ImpactDevelopers can now bring up a local Superset environment from a fresh clone without hunting for OAuth, payment, email, or queue service credentials, which lowers setup friction and makes onboarding and parallel local worktree development faster. The `local` profile now runs with non-blocking optional integration checks and explicit worktree-scoped DB/desktop state, while production and canary behaviors remain strict; operators should watch for accidental environment-profile leakage (especially with `VERCEL` markers), and for cleanup behavior when local worktrees are removed so stale state does not accumulate.
Aider now wraps `Path(...).is_dir()`/`resolve()` calls in three `main.py` validation paths with `try/except OSError`, so malformed or overlong paths (including those arising from large `--system-prompt` values) no longer crash the CLI with tracebacks and instead emit a warning and continue.
ContributionAdded explicit OSError handling around directory checks and path resolution in `read_only_fnames`, multi-file validation, and single-directory detection, converting hard failures into graceful invalid-path warnings.
ImpactAider users who pass very long or malformed path arguments—especially via command-line prompts—will avoid abrupt crashes, so their sessions keep running while the tool surfaces a warning and skips only the bad path. This is implemented by catching `OSError` in the three `Path(fname).is_dir()`/`resolve()` branches in `main.py`; follow whether the warning-based skip behavior causes any important files to be silently ignored and whether other path-handling code paths still raise uncaught filesystem exceptions.
PR #5149 adds a `--one-shot` mode so aider can process a single prompt and terminate immediately, skipping git repository setup; the command now accepts the prompt from `--message`, `--message-file`, or the first positional argument.
ContributionImplemented a new one-shot CLI path that detects one-off message input, disables git initialization, executes the coder once, and exits, enabling lightweight one-prompt usage.
ImpactDevelopers and script operators can run quick explain/refactor tasks with aider without opening a full interactive coding session, which should reduce startup friction and simplify lightweight automation flows. The implementation in `args.py` and `main.py` adds explicit message-source resolution and enforces single-run exit behavior; monitor adoption for scripts that previously relied on repo initialization or persistent session state, and verify edge cases where users omit/ambiguate message input.
Release v2.8.1 updates the packaged skill set by excluding php_deserialization.md to align with VS Marketplace policy, addressing a packaging issue that could affect distribution compliance.
ContributionImplemented a release packaging correction by removing a specifically listed skill file from the shipped bundle so the v2.8.1 artifact no longer includes content flagged by marketplace policy.
ImpactDevelopers and users installing or publishing the latest costrict release should see fewer marketplace acceptance problems because the disallowed php_deserialization.md file is no longer included, reducing the chance that updates are blocked during validation. Continue to monitor whether any users lose discoverable php deserialization guidance in the default bundle and whether the missing file changes onboarding or troubleshooting workflows for PHP-related usage.
This change adds packaging safeguards in the build flow so `php_deserialization.md` is removed/ignored during release packaging, preventing the file from being shipped inside bundled skills.
ContributionAdded a concrete release hardening step by removing `php_deserialization.md` in the build process and applying package-ignore rules, which changes release artifact composition to exclude an unintended file.
ImpactExtension publishers can publish the CoStrict package with fewer chances of Marketplace policy rejections caused by extra bundled files, so operators avoid failed uploads and users receive cleaner install artifacts. The technical change is enforced through updated build and ignore configuration (`.vscodeignore`/`.gitignore`), so teams should monitor future skill-file additions for accidental exclusion and track recurring Marketplace validation failures after the 2.8.1 packaging change.
This change performs a repo-wide rename of CLI-facing identifiers from `aioncli`/`AionCLI` to `aioncore`/`AionCore`, including the executable name, bundled paths, prepare script name, CI step labels, and related strings/messages so internal tooling and commands align to one naming standard. The npm package `@office-ai/aioncli-core` was explicitly left unchanged.
ContributionImplemented a coordinated naming migration for AionUi’s CLI entrypoint and startup resources, replacing mixed old/new command/path identifiers with `aioncore` across scripts, CI, and repository references while explicitly scoping out the unrelated package `@office-ai/aioncli-core`.
ImpactDevelopers and operators running AionUi locally or through CI get fewer startup and automation failures because the project now uses the same CLI name and paths (`aioncore`) everywhere internally, instead of a mix that could send scripts to stale locations. Continue to watch release tooling, installers, and external wrappers that may still call `aioncli` or `bundled-aioncli`, since those external touchpoints can fail even after the in-repo migration is complete.
The PR updates cc-connect so reply footers now append the current git branch name in parentheses when the working directory is inside a git repository, with fallback to unchanged output in non-git directories.
ContributionIntroduces a separate branch-resolution path (`replyFooterGitBranch`) and integrates it into footer generation, executing `git rev-parse --abbrev-ref HEAD` with a 1-second timeout, appending the branch name for git repos, and truncating long names with multi-byte-safe ellipsis logic.
ImpactDevelopers using cc-connect in terminal workflows will now see the active git branch directly in each reply footer, so they can more reliably tell which branch context they are operating in and reduce the chance of acting on the wrong checkout; monitor whether branch lookup timeouts or truncation hide important branch information in repos with slow git metadata and very long branch names. The change splits path resolution from footer compaction, uses a direct git command to resolve the branch, and applies rune-safe truncation at 30 characters for readability.
A commit in this burst changes model selection precedence so an `OPENAI_MODEL` value in `env_json` is now applied as an override by the model-provider flow, preventing accidental use of a different model source.
ContributionAdded a precedence fix that enforces `OPENAI_MODEL` from `env_json` to override alternative model configuration paths, correcting model resolution behavior when multiple model settings are present.
ImpactDevelopers and operators using `env_json` to configure codeg now get the model they explicitly set, reducing confusion and retries from calls silently running against an unintended model. Follow-up should confirm how this override interacts with other provider-level model settings to avoid future ambiguity in mixed-configuration environments and ensure no environment-specific regressions in model binding.
A single validation fix was made to `scripts/validate-skills-schema.py` to close the gap where plugins using root-level `SKILL.md` (alongside `.claude-plugin/plugin.json`) were skipped during batch validation as `Skills: 0`. The update adds a second discovery pass from `.claude-plugin/plugin.json`, merges results with the legacy `skills/<name>/SKILL.md` walk, and deduplicates by absolute path so existing nested-layout behavior is preserved.
ContributionIntroduced root-layout skill discovery to the validator: `find_skill_files` now performs a second walk anchored at `.claude-plugin/plugin.json` and `validate_plugin` explicitly scans root-level `SKILL.md`, with absolute-path deduplication against legacy nested discovery to avoid duplicate counting.
ImpactRepo operators and quality auditors now get visibility into previously hidden Anthropic-spec plugins, so batch checks and audits stop silently missing valid skills and can enforce quality gates on more real packages instead of ignoring them. In practical terms, this should reduce false negatives in marketplace reporting and lower the chance that valid plugins are treated as broken or absent. The implementation also keeps backward compatibility by deduplicating overlapping discoveries; the immediate behavior has already shown a +26 skill count lift with legacy layouts still passing. Watch the expected rise in orphan-count metrics and verify no duplicate counting appears in mixed-layout repositories.
The PR changes default Clawd runtime SVG assets to render with `shape-rendering="geometricPrecision"` at the root `<svg>` level (for all 35 assets listed in `themes/clawd/theme.json`), replacing `crispEdges` to smooth jagged diagonal and rotated edges in Clawd animations while keeping other themes and animation/state logic unchanged.
ContributionChanged the default Clawd runtime SVG rendering path by updating all theme-referenced runtime SVG roots to `geometricPrecision`, which directly improves visual quality for diagonal/rotated strokes, and added an automated Node test to enforce the setting for the 35 referenced files.
ImpactClawd users who use the default theme will see less jagged, stair-stepped edges on rotated or diagonal animation parts, so the avatar motion looks materially smoother without altering gameplay behavior. The PR achieves this by replacing `crispEdges` with `geometricPrecision` on root SVG elements for the theme's runtime assets while leaving state mappings, timing, renderer logic, mini mode, and non-Clawd themes untouched; continue watching for any browser/GPU-specific visual differences and potential performance changes on lower-end Electron setups, especially across working/reaction/idle animation states.
Emdash v1.1.19 fixes a Windows-specific terminal regression by changing PTY command construction so Claude can launch correctly when project directories contain spaces in their paths.
ContributionAdjusted the Windows terminal command handling to double-wrap cmd.exe command lines, preventing path-quoted command corruption and making project-path launches with spaces succeed.
ImpactWindows users running Emdash in folders with spaces can start Claude from the terminal without the startup command breaking, so daily coding and agent workflows avoid unexpected launch failures during local and remote work. The change appears to be scoped to cmd.exe quoting behavior, so teams should keep monitoring other shell modes and command patterns (for example, punctuation-heavy paths) to catch any remaining launch regressions before relying on this in fully standardized Windows automation.
In this burst, the dominant meaningful change is the addition of CUE language support through `cue lsp`, giving Serena direct language-server handling for CUE files rather than only handling release/version maintenance updates.
ContributionAdded CUE language-server integration (`cue lsp`) to Serena, introducing language-aware editing support for CUE files as a concrete new capability.
ImpactDevelopers editing CUE config files can work directly in Serena with built-in language support, reducing tool-switching and making CUE syntax/schema issues easier to catch during normal development. The change is implemented through cue LSP integration, so operators should watch for LSP startup stability, compatibility across cue versions, and whether diagnostics correctly handle complex CUE syntax patterns before treating this as fully stable.
Serena now adds first-class CUE language support by introducing a dedicated `CueLanguageServer` wired to the official `cue lsp`, enabling CUE-aware LSP behavior (startup + symbol/reference handling) for `.cue` projects.
ContributionAdded a concrete CUE LSP integration path in Serena: new `Language.CUE` registration in the LSP factory and test suite, plus automatic retrieval of the cue binary from cue-lang GitHub releases with SHA-256 verification and per-platform support, while allowing explicit override of binary path or pinned cue version through settings.
ImpactDevelopers editing `.cue` files in Serena can now get CUE language features such as symbol navigation and cross-file references to work immediately when opening CUE repositories, instead of manually wiring or debugging LSP setup themselves. The change uses a pinned default `cue` binary (`v0.16.1`) and verifies downloads by SHA-256 across Linux/macOS/Windows (x64 and arm64), with settings to override both path and version when teams need a custom toolchain. Operators and maintainers should watch for reliability in first-run downloads and for upstream `cue lsp` protocol changes, since either could break startup or reduce LSP completeness for CUE users.
This PR fixes a packaging regression by ensuring `plugin/` runtime dependencies are actually shipped in the marketplace bundle. It does two related steps: runs `npm install --production` when building the plugin package, and includes `plugin/node_modules` in `package.json` `files` so required modules (for example zod and parser/runtime deps) are present after install.
ContributionIntroduces an explicit production dependency install step in the plugin build flow and publishes `plugin/node_modules` with the extension package, so installed users receive all required runtime packages instead of a dependency-missing bundle.
ImpactUsers installing claude-mem from the marketplace are less likely to lose entire sessions silently because prompt-processing workers can start with the required runtime modules available instead of missing-dependency crashes. After this change, a startup failure path that previously left sessions un-summarized should stop happening, which removes a failure mode that could hide across platforms where local hook errors are not surfaced. Continue monitoring whether dependency updates outside this repo can reintroduce a gap between lockfile state and bundled modules, and watch installation size/upgrade behavior to ensure the bundle remains stable as dependencies change.
Loopmaster appears to be positioned as a standalone web IDE for livecoding music, highlighting inline piano-roll and waveform visualizers plus playback controls so creators can code and audition music patterns in one interface.
ContributionThe release introduces an integrated, code-first music prototyping environment in-browser, replacing a multi-tool workflow with immediate visual feedback (piano-roll and waveform) tied to the edited code during playback.
ImpactMusicians and educators using browser-based music tooling can now prototype and iterate on code-driven compositions faster because they can see and hear changes in one place, reducing context switching between editors and audio visualizers, but operators should watch iOS Safari support closely after reports of initial playback not starting on some devices. Continued observation should focus on mobile audio initialization reliability, cross-browser behavior, and whether the service remains a cohesive standalone experience versus an embedded host dependency.
GitHub made remote control of GitHub Copilot sessions generally available, allowing users to start a session in VS Code or GitHub CLI and continue controlling it from a phone through GitHub Mobile.
ContributionIntroduced a cross-device session handoff path for Copilot workflows, so one active local session can be taken over and managed from a mobile client without recreating context.
ImpactDevelopers can keep an active Copilot-assisted coding session going from laptop or CLI to phone, reducing context loss during interruptions or context switches, so productivity is less likely to drop when moving between devices; operators should monitor session sync stability, auth/session ownership enforcement, and reconnect behavior on mobile network changes.
This change updates `entire trail list` to stop returning an unbounded legacy table and instead show a bounded recent view (default 10 items) of repo-wide in-progress trails, while adding explicit composable `--author` and `--status` filtering (including clear-filter behavior) and removing the implicit `--all`-style listing mode.
ContributionIntroduces a new default listing behavior for trail discovery: bounded recent results, explicit filter composition by author/status, and a simplified branch-author-age output format, replacing the prior unbounded all-trails table flow.
ImpactCLI operators and repo triage users now get a faster, less noisy view of active work by default, which reduces time spent finding current trails during debugging or incident coordination. The command now prioritizes recent in-progress items and requires explicit filter composition for broader searches, making results easier to control and reducing accidental broad scans in large repos. Watch for adoption friction in automation or power-user workflows that depended on the removed implicit full-list behavior, especially scripts assuming `--all` or default unfiltered visibility.
Adds a new `CLAUDE_MEM_OPENROUTER_BASE_URL` setting so OpenRouter requests can be routed to any OpenAI-compatible `/v1` endpoint while preserving the default OpenRouter endpoint when unset, and exposes that setting through worker validation plus the viewer settings UI.
ContributionIntroduces a configurable base-URL override in the OpenRouter provider path, allowing deployments to redirect calls to a custom compatible endpoint without modifying generated worker bundles. The PR also keeps backward compatibility by defaulting to `https://openrouter.ai/api/v1/chat/completions` when the variable is empty, and wires the new option into settings validation and the settings viewer for easier configuration and early error surfacing.
ImpactDevelopers and operators can now point claude-mem’s OpenRouter traffic to an internal or custom gateway, so integrations can match local networking, compliance, or proxy requirements without repackaging worker code. The feature is implemented via `CLAUDE_MEM_OPENROUTER_BASE_URL` with backward-compatible fallback behavior plus UI/validation wiring, and teams should monitor compatibility of custom targets with the expected `/chat/completions` contract, credential handling, and whether validation catches malformed URLs before runtime.
Athas migrated its in-app editor from a legacy custom stack to Monaco in the v0.7.0 release, including command/settings wiring and cleanup of the old overlay editor path.
ContributionIntroduced Monaco as the primary editor surface and connected editor command and settings flows to it, while updating interaction paths (vim/theme/command routing) and removing legacy editor overlays.
ImpactDevelopers editing code in Athas should notice a more consistent and responsive editing flow, especially on larger files, because the IDE now depends on Monaco rather than the previous custom overlay editor, which directly affects day-to-day coding stability and usability. Continue monitoring for regressions in selection/cursor behavior, shortcut mapping, and workflow compatibility, since engine migrations often change edge-case behavior before full parity is reached.
The PR adds core LSP-style tools (completion, hover, inlay_hints, get_diagnostics) to oraios/serena so AI agents can inspect symbols, documentation, and diagnostics through the editor language-service layer instead of relying mostly on source browsing, improving the speed and safety of automated code-assist loops.
ContributionIntroduced four language-assisted tools in Serena’s AI-IDE path and validated them in multi-language tests: completion, hover docs, inlay hints, and diagnostics retrieval, with explicit behavior differences captured per language server. This materially changes agent capability from passive source reading toward active symbol- and error-aware assistance.
ImpactDevelopers and AI operator workflows using Serena can get faster, more actionable feedback while editing Java, TypeScript, and Rust code because symbol lookup, docs, and diagnostics are exposed directly in the assistant tooling, reducing trial-and-error edits; operators should monitor inlay-hints behavior (especially in Java/TypeScript defaults) before depending on it for production agent workflows. The change also exposes uneven LSP support—manual tests show Pyright still lacks inlay hints while Rust returns non-empty inlay responses—so teams should track parity before rolling this as a general contract across language servers and validate JDTLS Maven/Gradle discovery toggles do not break project setup assumptions.
This open pull request proposes changing Serena’s Terraform/Python language-server startup flow so initialization failures are reported as explicit issue-style diagnostics, improving how setup breakages are surfaced.
ContributionProposes a new developer-facing behavior where language-server initialization errors for Terraform/Python become explicit issue items, replacing less actionable startup failure handling.
ImpactDevelopers starting Serena for Terraform or Python would get clearer, trackable startup feedback when the language server fails, so recovery and triage are faster instead of relying on ambiguous launch behavior. If implemented, watch whether issue creation is correctly scoped to genuine initialization failures to avoid false positives and noisy issue spam during transient environment/toolchain hiccups.
Serena’s behavior changes to report syntax and linting errors when users modify TypeScript or JavaScript files, so invalid edits are surfaced directly during editing.
ContributionAdds edit-time diagnostics for TS/JS file modifications by making Serena report any syntax or linting error encountered during those edits.
ImpactDevelopers using Serena to edit TypeScript or JavaScript will see immediate error feedback while working, which helps prevent silently introduced bad code from slipping into later build, lint, or runtime checks. The PR wires in reporting for syntax and lint violations during TS/JS file changes, enabling earlier correction before review or execution; teams should watch for false positives or missed rule coverage that could either block valid edits or leave some quality issues undiscovered.
A new CLI command is introduced to fetch dependencies needed by Serena language servers, replacing manual dependency installation steps and giving users a single bootstrap path.
ContributionThe change adds a concrete dependency-bootstrap utility that automates package downloads for language-server setup and reports clear feedback when a download fails, which reduces setup ambiguity for users.
ImpactDevelopers setting up Serena language servers can now install required dependencies through one command, so first-run setup and CI provisioning are less likely to fail silently due to missing components and easier to recover when failures occur. Track whether the downloader consistently handles all supported environments and whether failed downloads return actionable messages for root causes such as network outages, permission issues, or repository source mismatches.
Release v0.13.5 adds native support in web and remote workspaces for uploading and downloading files and folders directly from the file tree using streamed transfers, removing size limits and adding a unified progress dialog with cancel support plus stricter checks for uploads and ZIP downloads.
ContributionIntroduced a concrete workspace file-transfer workflow change: users can transfer files and folders directly from the file tree in both web and remote modes through streaming pipelines, with bounded UI control (progress + cancel) and hardening of upload and ZIP download validation.
ImpactWeb and remote workspace users can move large files and folders directly in-place, so large data workflows become faster to execute and easier to control when a transfer needs to be stopped, instead of relying on external/manual workarounds. The feature now streams transfer data end-to-end, removes prior size restrictions, and adds extra validation around uploads and ZIP downloads, which should lower transfer failures and unsafe archive handling; monitor for cancellation cleanup correctness, partial-transfer state consistency, and false-rejection edge cases in safety checks during very large or unusual archives.
The PR updates local persistence recovery by version-gating Desktop-started Worker reuse. It now writes `appVersion` and `startedBy` into control-surface metadata, and Desktop launches only reuse a local Worker when that metadata shows the same launching app version; legacy or ownerless connection files are treated as stale and repaired before replacement Workers are started.
ContributionIntroduced explicit Worker ownership/version checks for Desktop launches (`appVersion` + `startedBy`) and added connection-file repair logic for stale or owner-missing records, so stale Desktop workers from prior app versions are no longer reused for the current session; CLI path is preserved via protocol-compatible reuse.
ImpactDesktop users who upgrade the installed app should see more reliable workspace recovery instead of resuming into an old background process that can leave workspace data stale or missing, because the app now blocks cross-version Desktop Worker reuse and forces a clean recovery path when metadata is stale or incompatible. Concretely, this changes runtime selection from optimistic reuse to version-matched reuse for Desktop-started Workers, with old connection records repaired before replacement. Teams should watch mixed CLI/Desktop workflows for any regressions in cross-process handoff and verify lock/connection file repair remains robust after repeated upgrades and rollbacks.
Adds Perplexity to Roo-Code as a first-class provider, including handler and provider-registry wiring so users can select and call Perplexity chat-completions models (defaulting to `sonar-pro`) from existing settings and model-selection flows.
ContributionImplemented an end-to-end Perplexity integration: added provider types/settings and schema wiring, registered the new provider and model mapping, added a dedicated OpenAI-compatible handler with API-key resolution (`perplexityApiKey`, `PERPLEXITY_API_KEY`, `PPLX_API_KEY`), and updated settings/UI validation so Perplexity models become selectable in-product.
ImpactRoo-Code users can now use Perplexity directly from the existing UI and configuration flow, so teams with Perplexity access can onboard a new LLM backend without custom connectors or external integration code. The new provider path also hardcodes endpoint and model exposure (`api.perplexity.ai`, `sonar*` models) and key fallback precedence, so we should watch for misconfigured key precedence, incorrect model fallback behavior, and any regressions in stream/error handling after release.
The editor input flow now forcibly clears React selection state for deletion-related input events (backspace/delete/cut variants), preventing the selection from being visually restored to an uncollapsed range after text removal.
ContributionIn `use-editor-textarea-input.ts`, deletion InputEvents are now handled by unconditionally clearing the editor’s React selection state, replacing the previous restore-from-DOM selection behavior that could reapply stale bounds and keep the highlight visible.
ImpactEditors using athas on affected browsers will no longer see selected text remain highlighted after Backspace/Delete/Cut actions, so user edits continue from the expected cursor position and reduces accidental overwrite or confusing UI behavior. The handler now treats deletion events as authoritative state-clearing operations instead of trusting potentially stale DOM bounds, so React/UI selection and browser selection are less likely to diverge; continue to watch for regressions in IME/composition workflows and other non-deletion selection updates.
Release @intentsolutionsio/[email protected] mainly delivers the new zero-tech-debt skill-enhancer update, synchronizing CLAUDE.md-based metadata handling to schema 3.6.0 so the plugin package exposes a defined and versioned zero-tech-debt capability.
ContributionIntroduces a concrete plugin feature in v1.0.2: a zero-tech-debt skill-enhancer that is synchronized to schema 3.6.0 metadata conventions, updating the package to a defined runtime manifest shape for skill checks.
ImpactDevelopers and repository operators can now adopt the zero-tech-debt plugin from a released package version with updated metadata alignment, which should reduce silent configuration mismatch when running zero-technical-debt checks and lower the chance of inconsistent rule behavior across environments. The package change is tied to a schema 3.6.0 sync, so teams should monitor upgrades for older manifests that still rely on pre-3.6.0 fields and watch whether existing CI jobs report new validation or behavioral differences.
Release v0.3.6 introduces a task completion function with UI updates, so users can mark tasks as finished directly in the interface instead of relying on external or manual tracking steps.
ContributionAdded a task completion flow in the running-task UI path and updated task views to reflect completion state changes.
ImpactOperators and developers using Nezha can now close tasks inside the app, which can make multi-task workflows less error-prone by clearly separating active versus completed work; watch next for cases where completion state might be inconsistent if multiple tasks are updated at the same time. The update appears to add task lifecycle state handling in the UI layer, so persistence and synchronization behavior across refreshes or state reloads should be monitored.
This release’s primary change is a focused hardening of remote SSH project workflows, adding explicit SSH connection settings and fixes so remote workspaces stay connected and synchronized more reliably.
ContributionIntroduced a remote SSH settings flow and remote-project hardening plus remote file-change detection fixes, directly improving the core remote workspace capability rather than isolated UI tweaks.
ImpactDevelopers using Emdash on remote servers can now rely more on long-running SSH workflows without losing track of upstream file changes, which reduces stale task context and accidental edits against old code when switching between local and remote work. The release adds SSH connection configuration, strengthens remote project handling, and fixes remote file-change misses, so operators should keep watching for sync regressions in large/active repos and edge cases in fork-based workflows before treating it as fully stable for critical remote collaboration.
In v1.26.1, @nanocollective/get-md is bumped and reconfigured so nanocoder no longer pulls node-llama-cpp as a required transitive dependency, removing a large platform-specific binary payload (CUDA/Vulkan/Metal/ARM) from pnpm installs and the Nix closure while preserving fetch_url behavior.
ContributionThe release changes the dependency model so get-md does not force node-llama-cpp into the install graph; this converts a heavyweight transitive dependency chain into an optional peer dependency path and keeps node-llama-cpp binaries out of the package graph/closure when they are not needed.
ImpactAnyone installing nanocoder, especially CI jobs and Nix consumers, should get significantly lighter and faster installs because roughly 500 MB of platform-native binaries are no longer fetched, and teams should monitor downstream scripts that silently assumed node-llama-cpp was already present. The HTML→Markdown conversion path used by fetch_url is unchanged, so normal usage remains functionally stable, but users relying on node-llama-cpp behavior through transitive reach-through may need to add explicit installation where required.
The release updates critical and high-severity dependencies to patched versions to address security vulnerabilities in ai-dev-kit.
ContributionUpgraded the dependency set in v0.1.11 to newer patched versions specifically for critical and high-severity issues, directly changing what bundled libraries and transitive components are pulled into install/runtime environments.
ImpactDevelopers and operators using ai-dev-kit can lower immediate exposure to known dependency vulnerabilities, making production environments safer without changing application-level APIs. Follow-on watch points should include whether any pinned packages become incompatible with existing plugins or custom scripts, whether installs still reproduce reliably after the bump, and whether new CVE disclosures appear in the updated stack before broader rollout.
CodeCompanion.nvim v19.13.0 extends its DeepSeek adapter to support both DeepSeek v4 models and the thinking-mode workflow, giving plugin users access to newer model behavior directly from chat.
ContributionUpdated the DeepSeek adapter logic so chat requests can target the v4 model family and opt into thinking-mode responses through existing plugin paths, enabling advanced model usage without external client switches or manual adapter customization.
ImpactNeovim users relying on CodeCompanion can now do complex reasoning chats with DeepSeek v4 directly inside their editor, so they can keep their coding workflow consolidated and avoid switching tooling for stronger model responses. After rollout, teams should watch for endpoint compatibility issues with DeepSeek v4, latency or token-cost increases from thinking-mode calls, and any request/response format mismatches that could break existing automations.
OpenCowork 3.3.0 introduces explicit zip-slip and path traversal protections in its security-hardening work, blocking archive entries from escaping their intended output paths during extraction-related workflows. This release positions the change as a key step in making the 3.3.x line production-grade.
ContributionAdded path-safety checks in the 3.3.0 release path to validate extracted package paths and reject traversal patterns, preventing unsafe writes outside the target directory during install/update flows.
ImpactOperators and users deploying OpenCowork 3.3.0 can be less exposed to compromised release packages, so a tampered update is less likely to alter files outside the app’s expected directories or damage the host environment during extraction. After this change, watch whether any legitimate enterprise or legacy archives are incorrectly rejected by the stricter path checks, and monitor for path-handling regressions across macOS and Windows packaging paths.