Skip to content
| Marketplace
Sign in
Visual Studio Code>AI>Claude Code Config HelperNew to Visual Studio Code? Get it now.
Claude Code Config Helper

Claude Code Config Helper

liliangshan

|
76 installs
| (0) | Free
Use a built-in Chat Webview backed by the local Claude CLI, with task flows, shared prompts, diagnostics, and provider/model settings.
Installation
Launch VS Code Quick Open (Ctrl+P), paste the following command, and press enter.
Copied to clipboard
More Info

Claude Code Config Helper

Version: 3.2.99

Website: www.zhineng.dev

Overview · Getting Started · FAQ · MCP Tools

Move beyond AI chat and keep development moving. Claude Code Config Helper (LLS CCAI) connects the local Claude CLI, model providers, independent left/right conversations and MCP tools inside VS Code, bringing coding, troubleshooting, task flows and diagnostics into one workspace.

  • Guided setup: download or reuse an existing CLI, manage providers, and choose workspace models and reasoning effort.
  • Independent conversations: keep primary and supporting tasks in separate contexts with their own model and permission settings, leaving the editor area for code.
  • Tools on demand: connect databases, SSH hosts, caches and APIs; inspect request usage and cache statistics; work in seven interface languages.

Bring your own model service and credentials; model credits are not included. Usage charges and feature availability depend on your provider. This is an independent project, not an official Anthropic or Microsoft product.

The Help icons in both native chat title bars and the Help text buttons open https://www.zhineng.dev/ directly in your default browser, without a topic menu. Button labels remain localized, and the link contains no credentials, conversation content or workspace paths.

Recommended MCP Tools

The following MCP servers are maintained by the same author and can complement this extension or run independently with other compatible clients. They are optional: installing the extension does not automatically install or enable them.

Tool Use Cases npm Documentation
SSH Remote commands and SFTP transfers with password/key authentication and host-fingerprint verification. @liangshanli/mcp-server-ssh
MySQL Data queries, schema inspection and DDL/DML with read-only mode, permission controls and operation logs. @liangshanli/mcp-server-mysql
SQL Server SQL queries and schema operations with permission controls, connection pooling and logs. @liangshanli/mcp-server-mssqlserver
Redis Cache inspection and key/data management with configurable operation permissions and logs. @liangshanli/mcp-server-redis
API Debug HTTP API debugging, shared configuration, login requests and indexed execution of saved endpoints. @liangshanli/mcp-server-apidebug

See the MCP configuration guide for client-specific examples. Consult each package's npm documentation for current requirements. Start with a test environment and least-privilege accounts, verify SSH host fingerprints, and keep credentials out of Git. Application-level switches do not replace database, operating-system or network access controls.

Tool JSON error diagnostics

When the Responses stream converter detects malformed tool arguments, it automatically writes .LLSOAI/error/YYYY-MM-DD/error-tooljson-<timestamp>-<random>.json, even when relay.debugRecord is off. Each file preserves the entire upstream SSE received for that request, the converted response, tool argument diagnostics, HTTP completion flags, end reason and the actual output-budget/reasoning parameters. A truncated stream can only preserve bytes received before it stopped; it cannot recover data the upstream never sent. Invalid UTF-8 additionally includes a Base64 copy of the original bytes.

Files share the existing 20-snapshot error retention limit across date directories. Detection compares the upstream final arguments with the actual emitted argument fragments, including __unparsedToolInput wrappers and late fragments. The recorder also retains up to eight recent tool responses in memory (64 MiB total) to correlate a subsequent CLI InputValidationError by call ID. If the original response has expired or the extension was reloaded, the fallback file explicitly marks originalResponseUnavailable; it never substitutes an unrelated response.

Request bodies, URLs and authentication headers are not included in these diagnostics, but raw model responses and tool arguments can contain sensitive content; review them before sharing. Token limits remain unchanged, and arguments are never repaired by guessing missing JSON characters.

What's New in 3.2.99

In Remote Office, you can now send a photo from the Android app or WeChat mini program, with or without a text prompt. The photo button appears when the selected main or parallel conversation is online and reports that it can receive images. This is a workspace capability, not an account setting. If the button is missing after upgrading, reload the bound VS Code window after active tasks finish, then reopen that area on your phone.

The extension downloads the photo over the computer's bound route, verifies it and passes its local path to the selected CLI. Downloads can take up to 180 seconds with a matching workspace service. If the download fails, the extension does not run the prompt without the photo.

What's New in 3.2.98

With a matching Remote Office service and client, the send button follows the selected conversation: while it is generating, you can stop it; once it is online, ready and no longer generating, you can send a new message. An unknown result from an earlier command no longer requires a review before you continue. The earlier command is not retried. Main and parallel conversations remain independent.

What's New in 3.2.97

  • Localized Remote Office guide. The welcome card uses the signed-in route for the domain and automatically changes the language segment in /{language}/remote.html to match the extension. For example, switching from Simplified Chinese to Japanese changes /zh-cn/remote.html to /ja/remote.html without changing your signed-in route.
  • The link covers all seven interface languages and updates when you switch routes or sign out. Before sign-in, Simplified Chinese defaults to the mainland China route; other languages default to the international route.

After installing the update, let active tasks finish before reloading VS Code. Existing settings and workspace bindings are preserved.

What's New in 3.2.96

  • Use your workspace remotely. Choose Remote office below the chat input, sign in, and link the current folder. The web client, Android app and mainland China WeChat mini program can open the main or parallel conversation, send tasks, and stop an active response on a connected workspace. Tasks run on your computer, so keep VS Code open and online. See the mainland China guide or the international guide. The welcome card in an empty chat opens the guide for your current route and interface language.
  • Keep the main chat connected when working in a subdirectory. If you bind tools and your CLI enters tools/extension, the remote main chat still belongs to tools. You do not need to rebind it or set chat.cliCwd. A separately opened workspace or a symlink pointing outside the bound folder does not inherit access.
  • Open Remote Office on your phone. The login and bound-status dialogs offer QR buttons for the website and Android app. The mainland route also offers the WeChat mini program. Select a button to view its code, or copy the website address for that route.
  • Image requests keep the provider's default format. If no output_format is specified, the extension leaves it out so the provider can use its default (typically PNG). Both generation and editing drop stray frame settings unless GIF was explicitly requested.
  • Find models by ID. Provider settings sort models by their actual IDs and keep disabled models available to edit. Chat model pickers offer only enabled, selectable models.

Upgrade note: After installing the update, let active tasks finish before reloading VS Code. Existing directory bindings are retained. Remote stop requires a compatible workspace service and client.

What's New in 3.2.95

  • Models ordered by provider, then model ID. Normal, task-flow and compaction dropdowns now sort by provider name, keep identically named providers separated by provider ID, and sort each provider's models by their actual IDs rather than display aliases. Natural numeric ordering places model-2 before model-10. Sorting changes only the displayed copy: saved configuration order, selected values and save/restart behavior stay unchanged, and optional unconfigured rows remain first.
  • Larger tool-input display buffer. The CLI output adapter now accepts up to 1,000,000 JavaScript string units per streamed tool-input block, up from 200,000. This is a character-length guard, not a token budget or a fixed byte size. Large valid tool arguments below the limit remain available for card parsing; once a block exceeds the guard, its remaining display deltas are ignored and the warning is logged only once for that block.

Upgrade note: install 3.2.95 and reload VS Code, or restart the F5 development host, when ready. Reloading may interrupt active CLI/image tasks; a chat-CLI restart alone does not load changed code/assets. Existing settings, credentials, history and selected models are retained. The tool-input guard applies to the extension's CLI-output/card parser, not the upstream request stream or the CLI's execution limits; it cannot repair an interrupted or malformed upstream JSON response. The full Anthropic beta forwarding, history-cache and left-chat completion fixes remain included.

What's New in 3.2.94

  • Forward the CLI's complete Anthropic beta header. Anthropic providers now receive every beta declared in the CLI's nonempty anthropic-beta header, without a name/date whitelist. This includes the companion beta for thinking.display: "updates" when the CLI sends it; missing headers do not cause the extension to invent flags.
  • Retain explicit provider configuration. Without a custom beta header, the CLI value is preserved (array values are joined as an HTTP list). Custom provider beta values are merged with all CLI flags and deduplicated. Client credentials and other diagnostic-header filtering remain unchanged; OpenAI Chat/Responses paths do not gain Anthropic beta headers.
  • No hidden compatibility rewrite. Thinking display settings, history signatures, cache controls and context-management settings retain their existing behavior. Unsupported-provider errors remain visible; the relay does not silently change updates to summarized, disable thinking or retry rejected requests. The Anthropic history-cache and left-chat completion fixes from 3.2.93 are included.

Upgrade note: install 3.2.94 and reload VS Code, or restart the F5 development host, when ready. Reloading can interrupt active CLI/image tasks; restarting a chat CLI alone does not load new extension code. Existing settings, credentials and history are retained. Forwarding a beta flag does not establish provider support: gateways with older schemas may still reject it. This release supersedes the 3.2.93 single-beta and full-beta test packages; the full-beta forwarding behavior is retained.

What's New in 3.2.93

  • Stable Anthropic history for prompt caching. Signed native thinking and opaque redacted-thinking blocks are retained when earlier messages move out of the latest tool round. Appending a new tool call no longer removes blocks from an already sent history prefix. Empty visible thinking with a signature is retained too; the existing filter for unsigned, short or UUID-style synthetic thinking remains. Cache counts, breakpoints, TTL and context-management settings are unchanged, and genuine upstream signature errors remain visible.
  • No spurious local authentication error after left-chat completion. Finishing a turn no longer disables a still-running independent CLI's private relay. Late internal requests can finish without reopening the completed chat or receiving an empty-body local 403. Startup failure, cancellation, restart, close and process exit still revoke access for that instance; token, method, path and generation checks remain enforced.
  • Clear local rejection diagnostics. Rejected independent-relay requests return nonempty JSON with EDITOR_RELAY_REJECTED, EDITOR_RELAY_INACTIVE or EDITOR_RELAY_REVOKED and a no-retry header. No account-login endpoint is added, and actual provider 401/403 responses are not hidden.

Upgrade note: install 3.2.93 and reload VS Code, or restart the F5 development host, when ready. Reloading may interrupt active CLI/image tasks; restarting only a chat CLI does not load changed extension code. Existing settings, credentials and histories are retained, with no automatic replay of completed or interrupted work. The corrected Anthropic history may need an initial cache rebuild; actual hit rates still depend on provider TTL, routing and compatibility. Local regressions do not replace real-provider and real-window acceptance.

What's New in 3.2.92

  • Left chat messages appear immediately. The left/parallel conversation now appends the user message and pending assistant bubble before waiting for CLI startup. Slow CLI startup no longer delays the visible sent message; startup failures keep the user message visible and show the error.

Upgrade note: reload VS Code or restart the F5 development host to load this runtime change. Existing histories, drafts and models are retained; active CLI work may be interrupted. Known requestUsageBinding and reused-response-ID limitations remain unchanged.

What's New in 3.2.91

  • Independent image format switches. New configurations enable only PNG; existing GIF opt-in is retained. Tools advertise only enabled explicit formats and expose frame parameters only when GIF is enabled. GIF requires an explicit animation/GIF request and must never replace a requested static image with a two-frame animation. JPG uses the provider's jpeg format value. At least one format must remain enabled. The static-format fallback introduced here was replaced by the provider's default in 3.2.96.
  • Safe SVG output. SVG Base64/Data URLs and downloaded files receive bounded UTF-8/XML and static-content validation before .svg storage. Active content, DTDs and external resources are rejected. SVG opens in the default text editor rather than an active webpage. Safe provider output in a different format is retained under its actual extension with IMAGE_FORMAT_MISMATCH; the paid request is never automatically repeated.
  • Longer, concurrent image work. Image-task defaults are now 100 concurrent tasks per workspace/extension host and 1200 seconds (20 minutes) per executing task, with a separate 100-task waiting queue. Provider-side limits still apply; this is not a cross-window quota. Already accepted tasks keep their frozen settings, and interrupted tasks are not automatically resubmitted.
  • One thinking disclosure per reply. Multiple visible-thinking blocks share one expandable section while retaining ordered text, manual choices and selection protection. Tools and request usage keep their own positions and identities; different replies and conversations are not merged.
  • More compact chat controls. Model type/name remain one size smaller and sit beside usage on the left. Tool-card metadata has a 10px left inset without extra bottom spacing. The input's running border animation is restored, the separate three-dot row is removed, and reduced-motion users retain a static border.
  • Refined model selection and settings navigation. The model picker has responsive grouped cards, full wrapping identifiers, seven-language timing guidance and clearer keyboard focus. A Switch to parallel session button in the settings toolbar returns to the existing independent conversation without restarting it or creating an editor tab.

Upgrade note: install the new package and reload VS Code, or restart the F5 development host, when ready to load the changed extension code and assets. Reloading can interrupt active CLI and image tasks; restarting only a chat CLI is not sufficient. Existing provider credentials, model selections, histories and saved GIF opt-in are retained. The four format switches declare provider capabilities, not paid API verification; actual provider support and local/remote editor appearance still need manual acceptance. Saving image settings continues to refresh running chat CLI schemas at safe idle points.

What's New in 3.2.90

  • Fluid panel width. Messages, tool cards and the composer resize with the chat panel instead of staying in a centered fixed-width column. Model/source and all five usage fields share a row when space allows and wrap naturally on narrow screens without truncating counts.
  • A calmer chat layout. Main, left and editor conversations share the approved compact styling: subtle user bubbles and borders, readable Markdown spacing, monochrome tool icons, individual code language/copy toolbars and a quieter composer. Existing model, image, permission, attachment and five-field usage controls remain available. Colors follow the active VS Code theme, including high contrast, and reduced-motion settings suppress decorative animation.
  • Expandable visible thinking. Returned visible thinking opens while streaming and closes after block/turn completion unless you manually choose otherwise. Incremental updates and same-view redraws preserve your choice and selection-copy behavior; completed history starts collapsed. SDK/native-history thinking and legacy quote-formatted blocks are supported. No hidden reasoning, signatures, generated summaries or invented durations are displayed, and model context is unchanged.
  • Request usage stays with its own response. Ordinary Anthropic, Chat and Responses replies receive an opaque request-scoped downstream message ID. Duplicate provider IDs, startup ordering and late reports no longer reassign statistics or replace a newer context meter. UI reload and exact-ID native-history rereads retain existing saved summaries without inventing missing legacy token counts. Tool IDs, signatures, model content and token values are preserved.
  • Reliable stream and completion handling. Chat/Responses streams accept a single leading BOM, including split UTF-8 bytes, without losing the first event. Late empty notifications and background metadata no longer restore a completed independent chat's stop-square state.

Upgrade note: reload VS Code or stop and restart the F5 development host when ready to activate 3.2.90; active CLI tasks may be interrupted. Restarting only a chat CLI cannot load new extension code. Existing settings and histories are retained. The standalone design sample is excluded from the VSIX. Local/remote Webview appearance still requires hands-on acceptance, and this release does not establish GIF compatibility for every provider.

What's New in 3.2.89

  • Opt-in animated GIF generation and editing. Enable Model supports GIF in the global image configuration to expose animation fields to the model. GIF requests must set output_format: "gif" and choose a size whose width and height are multiples of 16. frame_count accepts 2–64 frames (default 9); frame_duration_ms accepts 20–2000 milliseconds (default 200). Invalid requests are rejected before provider submission; ordinary static requests omit the GIF fields and remain unchanged.
  • GIF files stay animated. GIF87a/GIF89a results are accepted from Base64, Data URLs or downloads and saved unchanged as .gif, with requested animation parameters retained in task/completion records. History restoration and the existing automatic/manual editor-open actions support these outputs. Signature validation does not verify the provider's actual frame count or timing.
  • Image settings refresh running CLIs safely. Successful saves refresh main, left and retained-editor CLIs at idle points, preserving their existing sessions. Busy conversations wait for their current work to settle; accepted image tasks continue without cancellation or replay. Stopped chats are not started.
  • GIF fields also reach resumed conversations. CLI history can replay old tool definitions even after MCP restart. The relay now synchronizes existing built-in image generation/edit schemas before forwarding model requests, without rewriting history, adding tools, changing external MCPs or disabling prompt snapshots. Both normal and ToolSearch deferred modes are covered by real CLI 2.1.312 tests against local fixtures.

Upgrade note: reload the installed extension or stop and restart the F5 development host to load the updated relay; restarting only a chat CLI does not load new extension code. Reloading can interrupt active CLI work. Then enable Model supports GIF only for a compatible image model and save. This is a user-declared capability, not a paid provider probe. Known requestUsageBinding and reused-response-ID limitations remain unchanged; real provider GIF support and actual editor animation display still require manual acceptance.

What's New in 3.2.88

  • Consistent compaction feedback on both sides. Left-side and independent editor chats now follow the main/right chat behavior after native CLI compaction: the input and token meter leave the busy state without showing the misleading "new session / 0 tokens / 0% saved" card. The host no longer supplies a placeholder summary when the CLI has not returned one.
  • Only explicit success counts. Failed, cancelled, unknown and error-bearing compaction results are reported as failures. A status reset without a result does not become a success notification. The existing session identity and visible conversation history are retained.

Upgrade note: reload VS Code or restart the development host when ready to load the fix; active CLI tasks may be interrupted. This corrects future status notifications, not the compaction model, token thresholds or retry policy, and does not actively remove an old card already displayed. Known requestUsageBinding and reused-response-ID limitations remain unchanged.

What's New in 3.2.87

  • Generated images open in the editor automatically. Newly completed image-generation and editing tasks open saved output images using VS Code's image editor, without waiting for a chat view or another model reply. Multiple outputs keep their own pinned tabs, and chat input focus is preserved.
  • Safe, one-time display. Only successful or partially failed tasks with a persisted terminal record and saved images trigger auto-open. Each image is validated against its task and file metadata before opening. Queries, history refreshes, repeated completion events and extension initialization do not reopen old images. An editor-open failure does not change the generation result or repeat the paid request.
  • Existing settings and remote workspaces. Auto-open follows claudeCodeConfigHelper.editor.autoOpenReadWriteFiles, independently of the code-preview toggle; manual opening remains available when auto-open is disabled. Remote outputs preserve their workspace URI authority and path rather than opening a local file: URL.

Upgrade note: reload VS Code or restart the development host when ready to load this behavior; active CLI tasks may be interrupted. This opens editor tabs, not inline chat thumbnails, and adds no Webview resource permissions or HTTP file server. Actual local/remote editor presentation still needs manual acceptance. Known requestUsageBinding and reused-response-ID limitations remain unchanged.

What's New in 3.2.86

  • Seven-language enablement guidance. Every setup step explains how to find Claude Code Config Helper in VS Code's Extensions view and choose Enable (Workspace) or use the gear menu to Enable / Enable Globally when the extension is missing in another workspace. The reminder follows the selected language, works in narrow sidebars and does not change extension enablement or intercept Next.
  • Independent image task files. Image task writers no longer use a workspace-wide runtime.lock. Each host updates only the UUID tasks it created; writes for one task remain ordered while other tasks and hosts can proceed independently. Legacy runtime locks are ignored, not removed. The separate global configuration lock remains.
  • Windows checkpoint resilience. Brief EPERM, EACCES and EBUSY rename failures receive bounded local retries without deleting or truncating the old JSON. A persistent write failure exposes its storage error and does not disable unrelated tasks; an unsaved terminal state stays visible in the owning host until it exits. This never replays a paid request.
  • Image downloads without IP filtering. Returned HTTP/HTTPS image URLs use normal host DNS/network routing without destination-IP filtering, DNS pinning or peer-IP matching, including private and Fake-IP addresses. URL credentials, unsupported protocols and redirects remain rejected; timeout, byte limits, image validation and credential-free downloads remain in place. The separate HTTP tool is unchanged.

Upgrade and safety notes: use trusted image providers, since returned URLs can now access host-reachable local/internal services. This is not new HTTP/SOCKS proxy support. Normal shutdown updates only its own tasks; after a crash, unfinished records are not automatically taken over or rewritten, since another host may still own them. Windows sharing behavior and proxy compatibility still require real-machine acceptance. Reload only when ready: active CLI tasks may be interrupted. Known requestUsageBinding and reused-response-ID limitations remain documented in CHANGELOG.md.

What's New in 3.2.85

  • Correct JSON image-edit payload. Requests to /images/edits use images: [{ "image_url": "data:<MIME>;base64,..." }] for both single and multiple sources, with explicit n: 1 matching the supplied provider example. The obsolete top-level HTTP image field is no longer sent. model, prompt and per-request size are retained.
  • Tool inputs remain local paths. The MCP tool still accepts image as one local path or a path array; the host converts these paths into the upstream images array. No public image upload, multipart fallback or automatic paid retry is introduced. Pure text-to-image requests keep their existing fields.
  • Response diagnostics remain available. Failures show HTTP status and a redacted response prefix of up to 500 characters in task details and completion messages.

Upgrade note: reload VS Code, or restart an extension development host, to load the new image client. Reloading interrupts active extension CLI tasks. The provider example uses HTTPS image URLs; acceptance of embedded Data URLs still depends on the provider and has not been verified with a paid request. Known requestUsageBinding and reused-response-ID limitations remain unchanged.

What's New in 3.2.84

  • Built-in HTTP requests. mcp__llsccaiHttp__http_fetch accepts a complete URL plus optional method, headers and string body, then executes native fetch in the extension host. No login helper, saved connection or business environment variables are needed. Trusted main and independent chats receive their own bridge access; startup disables built-in WebFetch only when the replacement is injected, leaving WebSearch unchanged.
  • Readable webpages and bounded redirects. HTML/XHTML responses become readable text without tags, scripts, styles or comments; JSON/plain text and binary results retain their existing representation. HTTP 301/302 follows up to five validated hops within a single 60-second deadline. Cross-origin headers are dropped, POST becomes GET without its body, and unsafe or unsupported redirects remain inspectable. HTTP error responses remain available; failures do not trigger automatic retries.
  • Local image paths, not inline bytes. openapi_edit_image.image now accepts one local path or up to 16 paths, including readable files outside the workspace and symlink targets. The host reads and converts each file into a complete Base64 Data URL for the JSON image API. Models must not supply HTTP URLs, Base64 or Data URIs. There is no added path-authorization popup; OS permissions, file/aggregate size limits and remote-host path semantics still apply.
  • Useful image failure details. Image task cards and completion messages include the HTTP status and a redacted response excerpt of at most 500 Unicode characters. These diagnostics survive task/history restoration without retaining full provider responses or request credentials; old failures cannot recover missing response data retroactively.

Upgrade note: restart the relevant chat CLI after loading this version so it receives the new HTTP tool and path-only image schema. Separately configured external image MCP packages are not modified. This release does not change Linux working-directory permission policy. Known requestUsageBinding and reused-response-ID association limitations remain documented in CHANGELOG.md.

What's New in 3.2.83

  • Native WebSearch compatibility. The CLI's dedicated search requests now use native web_search with compatible Responses models and Chat Completions gateways, instead of an unimplemented function declaration. Streamed/JSON replies preserve real source URLs and available citation titles in CLI-compatible search blocks; ordinary function tools keep their existing behavior.
  • Chat answers retain their evidence level. Structured search metadata becomes source results. Gateways returning only answer text and Markdown links keep that content with a clear notice that live search execution is not independently verified; the adapter does not invent source records, encrypted snippets or observed search counts.
  • No unsupported call-budget parameter. Neither Chat nor Responses sends max_tool_calls, avoiding HTTP 400 failures on gateways that reject it. Streaming, supported domain allowlists and approximate location remain; blocked-domain exclusions report an explicit compatibility error rather than being silently ignored.
  • Isolated search execution. Search subrequests no longer inject task-flow context or trigger automatic task-flow continuation. Duplicate snapshots are deduplicated, while empty, failed and unfinished searches report errors instead of returning a successful-looking empty template.
  • Verification. Search acceptance passed 32/32 against local fixtures, including real CLI 2.1.312 round trips for Responses, Chat structured sources and Chat text-only answers. No paid API requests were made for these implementation checks.

Known issues and scope: the existing requestUsageBinding failure (true !== false) and reused-response-ID usage-association limitation remain unresolved; intermittent leftUsage assertions are also documented in the changelog. Linux working-directory permissions, symlink authorization and CLI permission policies are unchanged in this release. Provider-specific support still requires manual acceptance. See Native Web Search and CHANGELOG.md.

What's New in 3.2.82

  • Dedicated compaction failures return an error. Failed summary-model requests are converted to an Anthropic invalid_request_error with HTTP 400, code dedicated_compaction_failed and x-should-retry: false. HTTP errors, network failures, timeouts, interrupted streams, inline errors and empty/incomplete summaries do not fall back to the normal model or start automatic chat recovery.
  • No repeated upstream summary charges. The CLI's compaction flow can attempt another summary even after HTTP 400. A session-scoped failure latch answers those follow-up attempts locally without calling the provider again. A new explicit submission or manual /compact clears the corresponding latch; other conversations and ordinary chat retain their existing behavior.
  • Consistent failure cleanup. Main and independent conversations stop automatic recovery/task-flow continuation and release compaction waits. Invalid, missing or disabled dedicated models/providers also report terminal errors. Successful summaries still pass through; failed streaming summaries are buffered instead of publishing a misleading successful prefix.
  • Local verification. Real CLI 2.1.312 with synthetic local upstreams passed 10 acceptance cases without disabling CLI API retries; all six failure scenarios made exactly one upstream summary request. These checks do not call paid APIs or replace acceptance with your provider.

Known issues: the pre-existing requestUsageBinding assertion (true !== false) and reused-response-ID usage-association limitation remain unresolved. See CHANGELOG.md for release checks.

What's New in 3.2.81

  • Built-in image tools. The Image chip after Task flow opens a compact, seven-language global connection dialog. Choose an existing compatible provider/model or configure Base URL, API key and model independently; all chat views share the selection without replacing their ordinary models.
  • Background generation, automatic results. llsccaiImage provides six tools for generation, JSON editing, task lookup, listing, prompt search and configuration information. Models submit the image size per call. Results return to the originating conversation as a user message and CLI continuation once it is idle; routine polling is unnecessary.
  • Files stay in the workspace. Task records and images go into .LLSOAI/openapi-image-generation/tasks/ and images/. Completion events give the model the workspace path, relative/absolute file paths and complete fileUri, with instructions to use that URI verbatim in output. No thumbnail/URL conversion or dynamic Webview file access is installed; inline file:// previews are not guaranteed.
  • Direct submission with existing safeguards. Image tools add no billing confirmation popup. The global switch, configuration checks, conversation capabilities, limits and no-automatic-regeneration policy remain. Image requests and the model turn handling completion may incur charges.
  • Linux installer compatibility. CLI package detection falls back to local libc probes when VS Code/Electron does not expose process.report, distinguishing glibc/musl without blindly assuming either.

Known issues and acceptance: the existing requestUsageBinding failure (true !== false) and reused-response-ID association limitation remain unresolved. JSON editing and supported sizes depend on the provider; real Linux/remote installation and the previously reported chat-refresh behavior require manual rechecking. See CHANGELOG.md for validation details.

What's New in 3.2.80

  • Moved conversations stay selectable. The parallel/settings view no longer reopens the main chat whenever it becomes visible. Both conversations can be selected independently when moved into the same sidebar or placed on opposite sides.
  • View identity, not default placement. Chat-opening actions focus the registered view or existing WebviewView rather than its original container. Moving a view does not exchange its history, model settings or CLI identity; repeated switching does not remount the parallel runtime.
  • Settings keep their intended focus. Explicit settings opening may reveal the main chat first, but finishes at the settings view's current location. Manual visibility changes and opening the parallel conversation never force main-chat activation.

Known issue: the existing requestUsageBinding failure (true !== false) and the reused-response-ID association limitation remain unresolved. Simulated layout regressions pass; real-window drag-and-drop still requires manual acceptance.

What's New in 3.2.79

  • Selected history actually reloads. Explicit history selection replaces the visible transcript even when cached message IDs match. Reselecting the current left/editor history rereads its native file; ordinary duplicate initialization still avoids unnecessary redraws.
  • Broader native transcript support. String messages and content-block arrays are both recognized. Missing, unreadable or empty selected histories report failure without silently clearing the current chat.
  • Safe switching and earlier display. The main chat reads and validates history before stopping its previous CLI and saving the selected identity. The left/editor chat publishes loaded history before process startup, keeping it visible even if that startup fails. Stop an active left/editor task before selecting history; interrupted prompts are never automatically replayed.

Known issue: the existing requestUsageBinding failure (true !== false) and the reused-response-ID association limitation remain unresolved. The reload behavior is covered by regression tests and an isolated browser check; manual acceptance in real workspaces is still required.

What's New in 3.2.78

  • Conversation history is back on both sides. Each chat header includes a Conversation list button that opens its native CLI history drawer. Selecting a history resumes it only in the originating left or right conversation; listing or dismissing history does not reset either chat.
  • Localized and keyboard-friendly. Loading/empty states, close controls, date groups and relative timestamps follow the selected UI language. Close through the button, backdrop or Escape; focus returns to the history button and the closed drawer is inert.
  • Parallel entry stays direct. Parallel session and the native plus action still open the left conversation without an intermediate list or editor tab. This restores native CLI history, not the separate independent-session picker. Website Help continues to open the homepage directly.

Known issue: the existing requestUsageBinding failure (true !== false) and the reused-response-ID association limitation remain unresolved. Real-window history selection still requires manual acceptance.

What's New in 3.2.77

  • One-click website help. Both chat title bars have a Help icon, and both chat headers have a Help text button. Every entry opens www.zhineng.dev directly in the system browser without a dropdown or topic menu. The command palette also provides LLS CCAI: Open Help.
  • Localized controls, English documentation. Help labels, tooltips, accessible names, command titles and launch-failure notices support all seven interface languages. The README is English-only, with website links and recommended optional MCP tools near the top.
  • No interruption or private URL data. Opening Help does not wait for chat task/settings queues or CLI updates, restart a conversation, change a model or send a model request. The fixed homepage URL contains no credentials, messages or workspace paths; simultaneous clicks share the pending browser launch.

Known issue: the existing requestUsageBinding failure (true !== false) and the reused-response-ID association limitation remain unresolved. Real-window external-browser behavior still requires manual acceptance.

What's New in 3.2.76

  • Two-sided conversations. The right main chat stays in place. Parallel session and the native + open the left independent conversation directly, restoring its saved selection or first history, and creating one only when no history exists. Repeated clicks reuse the same conversation without refreshing the chat or restarting its CLI. Settings can replace the left page without stopping its task.
  • Matching title-bar actions. The left chat has Edit title, Settings and Restart CLI icons like the right chat, but no plus. Rename/restart affect only the corresponding conversation. Chat no longer exposes conversation-list controls; editor task code, histories and restoration remain retained behind hidden legacy entries.
  • Left usage and MCP fixes. Exact request/response bindings and persisted statistics restore per-request token/cache usage and the context meter. Each conversation owns its bindings. Native workspace .mcp.json discovery follows the user's strict-mode setting rather than being forcibly disabled; folder-scoped VS Code MCP settings resolve against the conversation's own workspace.
  • Automatic setup activation. Missing or incomplete global installation records automatically open and focus the shared wizard before either chat starts. The Reset Installation State and Open Setup command clears only setup records through VS Code's live API; it preserves CLI files, providers, secrets, MCP configuration and histories. Development-only one-shot reset requests are ignored by production installations.
  • Build identification. Startup output records the actual extension path, mode, version and setup-completion flag, making installed-versus-development testing easier to verify.

Known issue: the existing requestUsageBinding failure (true !== false) and reused-response-ID association limitation remain unresolved. Manual acceptance is still needed for real workspace/provider/MCP behavior.

What's New in 3.2.75

  • Setup and settings share one view. The left LLS CCAI view shows the three-step wizard until this host has an explicit global setup-completion record, then switches in place to settings. Existing CLI/model settings are prefilled, not treated as proof that setup is complete. Step 1 configures language, download source and CLI; step 2 manages providers and models directly; step 3 selects workspace Normal, Task flow and Compaction models plus Auto/Low/Medium/High/XHigh reasoning effort.
  • Managed CLI downloads and updates. Choose official npm (default) or the explicitly selected third-party China mirror, npmmirror. The extension downloads the host's native platform package, verifies its SHA512 integrity, extracts without running install scripts and validates --version. Executables are published directly as <selected directory>/claude-<version> (.exe on Windows), without overwriting an existing file. Custom download-directory choices survive reloads.
  • Updates stay in settings. The CLI card contains only the current version and Check for updates. Checking shows the result inline without switching back to onboarding or refreshing provider forms. New versions require confirmation; validated updates switch the path and restart extension-managed live sessions, preserving their identities and settings. Running tasks can be interrupted and are not replayed; unopened histories remain stopped and old executables are retained.
  • Up to three parallel task tabs. The main chat defaults to the right sidebar; settings remain on the left. The title-bar + Parallel task action reopens saved editor conversations in creation order without a picker, creating a new UUID only after all saved conversations are open. The maximum is three editor task areas, excluding the main chat; hidden, opening and closing tabs count toward the limit. Each tab owns its CLI, context, model choices, reasoning effort, task flow and draft. This is independent chat concurrency, not the separate multi-agent orchestration project.
  • Cleaner chat and provider controls. Editor tabs place Edit title, Restart CLI, Conversation list and New chat after the title. Main-chat page actions stay right-aligned, with existing native title-bar shortcuts retained. Manual restart reuses the relay-and-CLI restart flow with visible progress; editor restarts affect only their own runtime. Provider lists show * plus the last four key characters only when a hint was generated during saving; older entries retain Saved key, and full credentials remain in SecretStorage.
  • Recovery and copy fixes. Reload restores the main chat's exact saved CLI history when its message cache is empty or mismatched. Editor tabs retain UUID association and automatically recover locks only when the recorded owner process is gone. Copying selected chat output uses the host clipboard with a real success acknowledgement and preserves right-click selection.

Known issue: the existing requestUsageBinding regression (true !== false) and the separately documented reused-response-ID usage-association problem remain unresolved. Synthetic/local tests do not replace manual acceptance on every OS, mirror and upstream provider.

What's New in 3.2.74

  • Accurate cache-write accounting. Chat and Responses JSON/SSE responses share one normalizer. Gateway-provided cache writes are displayed separately from ordinary input, with ordinary input + cache reads + cache writes matching total input. Root-level write aliases take precedence over detail fields, including explicit zero; invalid counts are ignored and oversized details cannot inflate totals.
  • Late usage preserved. Cache-write details arriving after Chat finish_reason remain available in the final usage event. Missing write details keep the existing read-cache compatibility behavior; no cache writes are inferred without evidence.
  • Previous test-package fixes included. Retains removal of the local 10 MiB Relay request limit, portable CLI version detection and actionable compatibility notices. The independently diagnosed duplicate-response-ID usage-display problem is not fixed by cache accounting changes.

What's New in 3.2.73

  • Large-context Relay requests. Removed the local 10 MiB chat/compaction request-body limit that could terminate long-session requests and cause CLI retries. Request bytes are preserved across chunks; interrupted reads release buffers without reporting a false size error. Host memory, Node runtime and upstream/model limits still apply. Image-upload and MCP limits are unchanged.
  • Portable CLI version detection. Stable versions such as v2.1.303 and Claude Code 2.1.303, startup notices and ANSI colors are recognized. Windows script probes use CMD or PowerShell as appropriate, with the configured environment and working directory. Failed probes are not cached; the detection timeout is 15 seconds.
  • Actionable compatibility notices. Deduplicated notifications distinguish an old CLI, failed detection and an unsupported context window, showing the actual executable path with recheck/selection actions. Changes continue to use the safe restart flow without interrupting an active task. Windows command construction is covered by tests; real Windows-machine acceptance remains pending.

What's New in 3.2.72

  • Responses thinking summaries stream live. With model reasoning passthrough enabled, Responses requests include reasoning.summary=auto, while effort selection remains unchanged. Visible summary deltas use the existing thinking-block typewriter rather than waiting for the full response.
  • Deduplicated completion. Summary text/part events and item/response snapshots share per-item, per-summary state. Only a matching, unsent suffix is appended; repeated completion events do not duplicate text or close blocks twice. Encrypted reasoning is never displayed. Summaries first appearing in snapshots after ordinary output are skipped to preserve ordering.
  • Existing behavior preserved. Turning reasoning passthrough off retains previous request/response output. Tool argument streaming, prefix validation and error snapshots remain intact. Local converter/HTTP tests and a real CLI 2.1.280 probe with a synthetic Responses upstream verified early summary delivery without paid API requests.
  • Known issue not fixed in this release. Some gateways reuse response IDs such as resp_stream, which can prevent per-message usage association even when token totals are logged. This release addresses thinking-summary display, not that separate usage-binding issue.

What's New in 3.2.71

  • Native auto-compaction during long tool loops. Auto compact delegates execution to the CLI, no longer waiting for the full-turn result before sending an extension-generated command. Summaries are still generated by the selected model from the complete CLI conversation; the extension does not extract/truncate local chat text.
  • Model-window policy. A valid configured contextLength takes precedence over the existing model table, then a 200,000-token default. Windows ≤200K use 90%; larger windows use 80%. The full window is passed once to CLAUDE_CODE_MAX_CONTEXT_TOKENS and CLAUDE_CODE_AUTO_COMPACT_WINDOW, with the usage ratio in CLAUDE_AUTOCOMPACT_PCT_OVERRIDE. The CLI additionally reserves output/summary space, so the meter labels the trigger as an estimate rather than exactly 90%/80% of total capacity.
  • Important off-switch change. chat.autoCompact.enabled remains off by default, but off now disables CLI automatic compaction with subprocess-only DISABLE_AUTO_COMPACT. Manual /compact remains available unless separately disabled by your own native settings. Upgrade users who relied on the previous native fallback while this switch was off should enable Auto compact.
  • Safe application. Idle changes resume-restart the same session. During generation, tools, permissions, compaction or recovery, changes display Pending until a safe restart; current work is not killed. Normal/task-flow model switches and recovery recompute the strategy from raw configuration each start.
  • Compatibility and diagnosis. Managed mode requires the selected CLI executable to report 2.1.280+ and a 100K–1M window. Unsupported ranges are not silently enlarged. Explicit environment/native-setting conflicts are shown without overriding user disables. The seven-language UI exposes configured capacity, usage ratio, estimated trigger, pending state and smaller summary-model capacity.
  • Verification. Real CLI 2.1.280 against local synthetic upstreams completed native compaction before full-turn result at 200K/500K/1M, routed the summary through the selected model, saved dated snapshots and continued the task. No paid provider requests were used. Opt-in acceptance tests use CCAI_NATIVE_TEST_CLI; default regression runs skip real-CLI tests. See the implementation plan.

What's New in 3.2.70

  • Workspace Code preview switch. The switch sits after Auto compact, defaults to off, and supports all seven interface languages. It writes claudeCodeConfigHelper.editor.codePreview.enabled to workspace settings and takes effect without restarting the CLI, including while a task is running.
  • Less editor switching. Off prevents file tools from automatically displaying non-Markdown files; .md / .markdown keep their existing behavior. Browser tools still reveal browser tabs, and manual file-link clicks are unchanged. On restores normal automatic file display.
  • Existing safeguards preserved. The master editor.autoOpenReadWriteFiles setting can still disable all automatic file opening, including Markdown. Only workspace files are eligible; toggling does not close existing tabs or restrict the tools' actual read/write operations.

What's New in 3.2.69

  • Paste webpage text and images together. Cmd/Ctrl+V preserves native text insertion and adds image attachments separately. If the clipboard has no image files, HTML img sources are imported by the extension host instead of being lost in the plain-text textarea.
  • Bounded image import. Supports embedded PNG/JPEG/GIF/WebP/BMP/TIFF data and public HTTP(S) images, with five HTML images per paste, 20 MiB per image and a 15-second download deadline. Duplicate URLs are skipped. Relative URLs need a clipboard source URL or HTML base; blob:/local/private images, unsupported formats and login-protected downloads may require copying the image directly or using a screenshot. Downloads do not send browser cookies or authentication headers.
  • Visible progress and failures. Image placeholders block sending until ready. Failed imports show warnings in all seven interface languages and are removed from attachments. The Webview CSP remains unchanged.

What's New in 3.2.68

  • Browser automation reveals its target tab. Page tools bring the tracked browser page into view in the editor area before running, while preserving chat keyboard focus. Opening or reusing a page also reveals that page's tab. This does not navigate or reload the page just to display it.
  • Best-effort compatibility. The reveal path targets VS Code's integrated-browser editor by page ID. If that editor is unavailable, the automation still runs and a warning is logged. Background login-state saves do not switch tabs again.

What's New in 3.2.67

  • UTF-8-safe streaming. Anthropic, OpenAI Chat and Responses now preserve multibyte characters across network chunks. Chinese text and emoji no longer become replacement characters in tool arguments; the typewriter effect remains enabled.
  • Validated argument suffixes. Responses only appends the untransmitted suffix when a final argument snapshot matches the exact prefix already sent. Conflicting snapshots and post-close argument fragments fail explicitly and retain diagnostic evidence instead of corrupting JSON with extra trailing characters.

What's New in 3.2.66

  • Extension identity on upstream requests. Chat, model-list and standalone compaction requests now include claude-code-config-helper/3.2.66 in User-Agent. Existing CLI or custom UA values are retained, with the extension identifier appended once; requests without a UA receive the extension identifier alone.
  • Version follows the installed package. The identifier reads its name/version from the root package.json, including inside the VSIX. Future version bumps no longer require a separate UA change.

What's New in 3.2.65

  • Workspace reasoning effort, without CLI restarts. A selector beside the task-flow model offers Auto, Low, Medium, High and XHigh. It saves claudeCodeConfigHelper.chat.reasoningEffort to workspace settings and is read on each request. Auto adds no override, preserving existing client/model handling; explicit selections apply to subsequent requests only and take priority over the model-level reasoning policy.
  • Protocol-aware parameters. OpenAI Chat sends reasoning_effort; Responses sends reasoning.effort; Anthropic sends output_config.effort, with XHigh mapped to max. The selected level must be supported by the upstream model or the upstream may reject the request.
  • Seven-language controls. Options, tooltips, accessibility labels and save-failure notifications follow English, Simplified Chinese, Traditional Chinese, Korean, Japanese, French or German. Protocol values remain auto, low, medium, high and xhigh regardless of display language.
  • Date-organized debug records. Compaction request snapshots and optional daily message logs are written to .LLSOAI/chat/YYYY-MM-DD/; upstream error snapshots go to .LLSOAI/error/YYYY-MM-DD/. Directories are created automatically using the local date. The error snapshot limit remains 20 files across all date directories. Existing files are not migrated, and session/configuration files stay in their original locations.

What's New in 3.2.59

  • Auto-compaction waits for the CLI turn to finish. Crossing the token threshold queues compaction without writing /compact during a tool loop. The command is sent after an explicit CLI result, preventing it from being absorbed as a mid-turn prompt. Pending compaction continues to hold new sends; relay request endings and message_stop do not release the command.

What's New in 3.2.58

  • Task-flow repeat-update guidance. The service remembers the latest task ID and status across continuation turns. No-op updates now report that nothing changed instead of claiming progress. Consecutive identical updates add a continuation reminder naming the next actionable task and requiring actual work before another status update. Legitimate transitions for the same task remain allowed, unfinished work is not skipped, and creating or clearing a workflow resets the in-memory record.

What's New in 3.2.57

  • Accurate compaction status. Automatic compaction no longer shows “compacting” merely because /compact was queued. The indicator starts when the CLI reports compaction or the relay detects the summary request. Resetting compaction state also clears the UI and releases waiting senders. The existing 30-minute stale-marker window and 60-minute send-wait cap are unchanged.

What's New in 3.2.56

  • Patient compaction wait. Queued sends now keep waiting while a compaction is genuinely running (re-checked every minute, 60-minute cap) instead of giving up after 3 minutes, so slow compaction models no longer get raced by task-flow continuations.

What's New in 3.2.55

  • Sends wait for compaction. Task-flow continuation, self-healing resends and timeout Continues now hold until an in-flight compaction finishes instead of racing the summary request. Stale in-flight markers are reset automatically so nothing gets stuck.

What's New in 3.2.54

  • Fewer spurious restarts. The self-healing watchdog now waits 120 seconds (was 20) for the CLI to reach the relay after you send a message before restarting and resending.

What's New in 3.2.53

  • Auto compact switch. A new toggle next to Subagents under the input box (backed by chat.autoCompact.enabled, default off) makes the extension send /compact automatically when the session's real context usage reaches the model's context length minus 50K tokens. Estimates never trigger it, in-flight compactions are respected, and a 60-second debounce avoids repeat sends. Off by default, so existing behaviour is unchanged unless you opt in.
  • No more flash-reload of the chat area. Long sessions that exceeded the 160-message in-memory window caused every panel open (task-flow continuation, sending a message) to rebuild the whole message list and scroll top-to-bottom. The webview now stays in sync with the trimmed list and appends incrementally, so the view keeps its position.

What's New in 3.2.52

  • Compaction model routing works with Claude CLI 2.1.260+. Newer CLI builds append a role: "system" reminder after the /compact summary request, so the summary prompt is no longer the last message in the request. The relay now skips trailing system messages when detecting compaction, so /compact reaches the configured compaction model again, compaction snapshots are written and the token meter's in-flight state arms as expected. Older CLI versions are unaffected.

What's New in 3.2.51

  • Model-level Explicit Prompt Cache is available in the Add/Edit Model dialog for compatible OpenAI Chat and Responses gateways. It defaults to off, survives model refreshes and config import/export, and takes priority over the saved cache mode while enabled. Anthropic providers retain the saved value but cannot activate it.
  • Protocol-specific caching: Chat uses the session ID as prompt_cache_key, explicit 30-minute options and a system-message breakpoint; Responses uses instructions plus the key/options, without cache_control breakpoints. Upstream support is required, and missing session IDs or static prefixes skip cache injection.
  • Low cache-hit guidance: usage footers below 80% show an underlined Solution link. The localized dialog explains how to enable caching and includes an Open extension settings link directly to provider/model configuration. No model-specific gateway warning is shown.
  • Provider settings guidance now explains that fetching models or changing provider/model settings takes effect immediately in the current workspace; other workspaces need a window reload or close/reopen. All new guidance supports seven languages.
  • Responses usage normalization now shares one non-negative calculation for JSON and SSE responses. Cache-write mapping remains deferred until its accounting semantics can be verified; cache-read tokens are counted only once.

What's New in 3.2.50

  • A localized Subagents switch now appears after the footer token meter. Its label and tooltip support all seven interface languages, including Simplified Chinese.
  • The switch defaults to off and is saved per workspace using extension workspaceState, not global settings. Each workspace keeps its own choice across reloads; legacy global values are not inherited.
  • Off removes Agent, SendMessage and ListAgents from every subsequent Relay request's tools list; on preserves them. This applies to ordinary chat, task-flow creation/execution and title-generation side requests across all three protocol adapters. The existing task-flow AskUserQuestion rule is unchanged. Changing the switch does not stop agents already running or disable other orchestration entry points such as Workflow.

What's New in 3.2.49

  • The footer token meter now reflects the latest completed response without changing when a new message is sent. The previous confirmed snapshot stays visible while the response is pending, then input, output, cache-write and cache-read values are replaced atomically when new usage arrives. Cache-read tokens are included in context usage, and stale cache-write values can no longer leak across responses.
  • Long-running tool heartbeats no longer create empty waiting cards. tool_progress events marked as heartbeats are ignored for both Bash and Agent tools.
  • Nested Agent results no longer create rows of orphaned tool_result success cards. Results that have no matching tool call in the main conversation are hidden, while paired and top-level results still render normally.

What's New in 3.2.48

  • Switching sessions can no longer race the next message send. VS Code does not await one Webview message handler before dispatching the next, so session/resume could still be loading history and restarting the CLI when user/send wrote to the old or temporarily-cleared process. Session-changing Webview operations now run in arrival order; cancel and question-answer messages remain immediate.
  • Concurrent CLI start/restart requests are serialized. Session switching, model switching, auto-continue and self-healing can all request a restart. A shared startup queue now prevents two start flows from terminating or replacing each other's newly-created process.

What's New in 3.2.47

  • Fixed the false "Chat CLI process is not running; cannot write to stdin" error while the CLI is actually running. After a CLI restart (model switch, auto-continue model swap, manual restart), the old child's late exit event could flip the shared childExited flag to true even though the new process was already running — so the next message was rejected with "process not running". Exit events now only affect shared state when they come from the current process.
  • send() now waits up to 1s for a briefly-unwritable stdin instead of failing instantly right after startup, and the rejection message now distinguishes no process handle, process exited, and stdin never became writable.

What's New in 3.2.46

  • The [claude-code:unrecognized_model] notice no longer shows up as a red error before every request. When you run the built-in Chat on a custom gateway (a model id the CLI does not officially know), the Claude CLI writes an internal debug line like [claude-code:unrecognized_model] {"model":"...","query_source":"sdk"} to stderr — and the adapter used to surface every stderr line as an error segment. Lines starting with [claude-code: are now treated as internal CLI diagnostics and dropped; real stderr errors still show up as before.

What's New in 3.2.45

This release is a hardening pass over the Chat CLI, the relay server, and shutdown — no behavior changes unless something was leaking or hanging.

  • A hung Chat CLI is really killed now. Liveness used to be judged by child.killed, which only reports "a signal was sent" — so after cancelling a wedged process, later input was silently written to a dead child's stdin. An explicit exit flag now gates send() / cancel(), and shutdown escalates SIGTERM → SIGKILL after 1.5s.
  • Reloading or closing VS Code frees the relay port immediately. Stopping the relay waited on server.close(), which never returns while an SSE stream is open; existing connections are now disconnected right away.
  • Cancelling a request stops the upstream call too. Closing the connection mid-stream used to leave the upstream request running to completion (and billing); the client disconnect is now propagated and destroy the upstream request. The deprecated req.on('aborted') listener, which never fired under keep-alive, was removed.
  • deactivate() flushes the chat session to disk and clears auto-continue timers / pending question waiters instead of leaking them.
  • The task-flow CLI exit notice is localized in 7 languages and degrades to a toast with a silent restart while a task flow is running, so it can no longer block the flow with a modal.
  • New claudeCodeConfigHelper.relay.debugRecord setting (off by default). Relay request recording used to run unconditionally, rewriting the whole daily log per request and storing base64 images; it is now opt-in, append-only, image-stripped, and upstream error snapshots are capped at 20 files.
  • The test suite is wired to a glob instead of a hardcoded file list, which brought three previously-skipped test files back and grew the suite from 283 to 308 tests.

What's New in 3.2.44

  • Fixed the startup "resume task flow?" dialog popping up mid-flow. The dialog is meant for a fresh window with an unfinished workflow on disk — but the Chat webview is lazy-loaded, so when the first thing to open it was a mid-flow auto-continue, the leftover restore flag fired the modal right on top of the running flow. The dialog is now skipped whenever the auto-continue scheduler still has pending work, and any manual continue clears the flag for good.
  • The restore dialog now has a 10-second countdown. If nobody clicks, the Continue button counts down and auto-selects continue, so even a mis-timed dialog can no longer park a task flow indefinitely.
  • Switching to the task flow model resets the missing-tool strike count. Strikes accumulated on the main model no longer carry over, so the circuit breaker can't trip right after a successful switch and silently stall auto-continue.

What's New in 3.2.43

  • The task flow model is now used for continuation, not for creation. Previously, starting a task flow switched the main model permanently and it was never switched back. The main model now always creates the workflow; before each auto-continue the extension checks the current model and, only when it differs, switches the main CLI to the configured task flow model and restarts it (the session is resumed with --resume, so context and session id are preserved). When the workflow completes or is cleared, the original main model is restored.
  • The check is idempotent: no task flow model configured, or already on it, means no restart at all — a whole run restarts the CLI twice at most, and only the switching restart waits a short settle delay before the continue prompt is sent. The original model is stored in workspaceState, so a crash mid-flow still restores on the next activation.

What's New in 3.2.42

  • Fixed task flow auto-continuing forever on a blocked or failed task. The loop only stopped when every task was completed, yet a continue prompt was still generated when nothing was pending or in_progress, so a stuck workflow was re-pushed indefinitely. No actionable task now means no continue prompt.
  • blocked is gone as a writable status. Tasks can only be pending, in_progress or completed. When something truly cannot be done, the model is instructed to append the reason to .LLSOAI/task_error.md, mark the task completed and move on — so failures are recorded instead of stalling the workflow. Existing .LLSOAI/task-flow.json files containing blocked still load fine.

What's New in 3.2.41

  • Assistant body text now renders as a whole block instead of line by line. The streaming chunker used to emit every buffered line as its own markdown segment, so multi-line structures — tables, nested lists, multi-line quotes — were split across separate render roots and each row was parsed on its own (a table came out as a stack of one-cell boxes). Text deltas are now accumulated and parsed in a single pass when the content block ends.

What's New in 3.2.40

  • Task Flow onboarding (earlier release). This release introduced the empty-chat guide for the plan-first workflow. Starting in 3.2.96, the welcome card introduces Remote Office instead. The CC task flow button and the full Task Flow usage guide remain available.
  • Fixed streaming long text stacking duplicate 「Long text output」 blocks. Thinking blocks stream as full accumulated text under one stable segment id, but the collapsible long-text branch dropped its DOM node instead of returning it, so the id was never stamped and each delta appended yet another collapsed block. Long text is now patched in place, and a block you expanded stays open while it keeps growing.
  • Task flow model replaces the expert / plan / review modes. The model picker is now a three-way Normal / Task flow / Compaction choice. A task-flow prompt can switch the main model to a dedicated model right before sending, and internal routing collapses to just normal and taskFlow.

Highlights

  • Plan-driven Task Flow mode: the model writes a plan document first, then drives it to completion step by step with automatic continuation, live status-bar progress, restart-safe persistence, and a circuit breaker that pauses the loop if the model stops calling tools. See docs/taskflow-usage-guide.md.
  • Fixed compaction still firing at around 166k tokens after it was made manual: the Claude CLI was compacting on its own because it never saw the model's configured context length and fell back to its built-in 200k default. The context length from the model config panel is now passed to the CLI as CLAUDE_CODE_MAX_CONTEXT_TOKENS, so the CLI and the token meter share one limit, and nothing is injected when the field is empty.
  • Made context compaction manual only: the token budget no longer fires /compact by itself once a session crosses its threshold, since that competed with the CLI's own compaction and could compact a conversation still in progress. Token metering is unchanged — the meter, threshold, and per-session accounting all keep working; compaction now happens only via the compact button on the token meter or a /compact you type yourself.
  • Fixed the built-in browser_* MCP tools disappearing entirely: the MCP server runs as a standalone Node child process without the vscode module, but it statically imported browserToolHost (which chains into require('vscode')), so it crashed on startup and the whole browser tool group silently vanished from the model's tool list. BrowserToolHost is now a type-only import, required lazily on the extension-host path.
  • Added click-to-copy inline code in Chat: a single-backtick span (typically a one-line shell command) now shows a copy icon and copies the whole command on click, with a green confirmation flash. Also fixed multiple code blocks in one message stacking all their copy buttons in the message's top-right corner.
  • Fixed context compaction running twice in a row: the compaction summary request is no longer re-measured against the threshold (its body is the whole conversation, so it always tripped), compaction started by the CLI or by a manually typed /compact now registers in-flight state so the in-progress check and 60-second debounce apply to it, and a stale in-progress flag is only reset instead of being treated as a reason to compact.
  • Added browser session persistence: browser tools now keep you logged in across page closes and VS Code restarts. Cookies (including HttpOnly), localStorage, and sessionStorage are captured automatically once the page state settles and stored per origin in VS Code SecretStorage; browser_open re-injects them before navigating so the first screen's API calls are already authenticated. Cookies travel over raw CDP because Playwright's storageState/addCookies are blocked inside VS Code. Logging out on the site is recorded faithfully — the empty state overwrites the snapshot, so the next open stays logged out.
  • Fixed a browser_open deadlock when VS Code replies "At least one similar page is already open": the page id in that listing is now parsed, instead of every following call failing with "Page not found".
  • Fixed Mac Chinese (IME) Enter-to-select sending the message by mistake: when picking a candidate, macOS Pinyin commits per character and the confirming Enter arrives with isComposing=false, so the composer and resend editor now use an arrow-key state machine — an arrow key (paging through candidates) marks the next Enter as a candidate confirmation that does not send, and any other key resets the state so a plain Enter sends normally.
  • Fixed compaction routing: /compact summary requests now reach the configured compaction model even when the Claude CLI omits the <command-name>/compact</command-name> marker (as happens with token-budget-triggered compaction and newer CLI versions), instead of falling back to the main model.
  • Fixed the Anthropic prompt-cache 400 error (a ttl='1h' cache_control block must not come after a ttl='5m' cache_control block) in the built-in Chat CLI: the relay no longer force-rewrites outbound cache_control breakpoints to 1h, which conflicted with 5m breakpoints injected by an upstream gateway. The cache TTL now defaults to Default (follow client), configurable through a new selector in the model-picker dialog (Default / 5 minutes / 1 hour) with a "switch back to Default on errors" hint, persisted in global state and applied to the relay without a reload.
  • Added an in-Chat "Skip browser confirmation" hint after the CC task-flow button: when browser tools are usable but VS Code still prompts "Open Browser Page?" on every page, the hint offers a one-click, in-webview confirmation that enables workbench.browser.enableChatTools and chat.tools.global.autoApprove together — no blocking activation popup.
  • Reworked the browser tool suite to call VS Code's built-in language-model browser tools via vscode.lm.invokeTool with page-id threading (browser_open, browser_navigate, browser_get_content, browser_screenshot, browser_console, browser_eval), exposed through the in-process browser MCP server.
  • Added task-flow persistence: the CC task workflow snapshot is now cached to .LLSOAI/task-flow.json on every create/update, and is restored on the next VS Code launch with a "Resume unfinished task flow?" prompt so progress survives restarts.
  • Added a past conversations panel in the Chat header: list previous sessions, resume one to reload its full message history into the webview, and the header now shows the resumed session's title.
  • Replaced the Chat header Copy source and Clear buttons with a single New chat button that starts a fresh empty session.
  • Made session list/content retrieval Windows-compatible: project directory names are encoded exactly like the official Claude CLI (no truncation), CLAUDE_CONFIG_DIR is honored, and the read path matches the CLI's write cwd.
  • Fixed Chat model-picker refresh after adding or editing provider models: the normal, task-flow and compaction model dropdowns now update without requiring a VS Code restart.
  • Anthropic direct providers no longer receive the OpenAI-style stream_options.include_usage request field; OpenAI-compatible streaming requests still keep usage options where supported.
  • Added token budget context compression for the built-in Chat: the token meter can trigger compression, large contexts auto-trigger compression near the configured threshold, tool call/tool result blocks are removed from the summary input, and the compressed summary is injected into a fresh hidden CLI session.
  • Native Claude TodoWrite todos now appear in a separate footer panel, independent from the CC task-flow Todo panel, so both can be shown and collapsed independently.
  • Tool call cards are collapsed by default: only the summary row (icon + name + status badge + chevron) is shown; clicking the row toggles the body. Collapse state is preserved across tool status updates.
  • Running state lockdown: while the chat is responding, the bottom composer controls (model select, permission mode, task-flow model select, CC task flow button) are disabled, and the comet-beam border animation stays visible even when the textarea is focused.
  • Upstream CLI system/taskstarted and system/tasknotification JSON events are rendered as a compact task card (status icon + description + task type + status badge) instead of leaking raw JSON into the chat.
  • Chat footer Todo card for LLS CCAI task workflows, with live status refresh and animated in-progress indicators.
  • Improved LLS CCAI task menu behavior: completed workflows are cleared silently, while running workflows still require confirmation before replacement.
  • Windows Claude CLI setup hints in the configuration panel, including the npm mirror install command and %APPDATA%-based executable path with copy buttons.
  • Built-in Chat Webview backed by a long-running local Claude CLI process.
  • Bidirectional stdio stream-json transport using --output-format stream-json --verbose --input-format stream-json.
  • Markdown, code block, diff, tool card, error, permission fallback, and workspace file-reference rendering.
  • Chat session restoration through VS Code workspaceState, with a one-time privacy notice and clear-session cleanup.
  • Visual provider/model configuration utilities for Claude Code settings migration and compatibility.
  • Secure API key storage through VS Code SecretStorage before activation.
  • Task workflow support for planning, progress tracking, and automatic continuation.
  • Import/export of provider configuration and shared prompts.
  • Multi-language UI support.

Setup and Parallel Tasks

  1. Open LLS CCAI from the Activity Bar. When no explicit setup-completion record exists for this extension host, the same left-side view opens the setup wizard—even if an older installation already has a CLI path and providers.
  2. Language and CLI: choose the language and installation source. Download a managed CLI or select an existing executable under Advanced options. Next verifies the selected path. Changing the download directory updates both its displayed destination and future installations; it does not move or overwrite an external CLI.
  3. Providers: add/edit providers, save credentials, fetch models or add them manually. No paid inference test is required to proceed. Saved credentials stay in VS Code SecretStorage; the optional *ABCD key hint is generated only when a key longer than four characters is saved. Existing providers without a hint still show Saved key.
  4. Workspace models: choose the Normal model, optional Task flow/Compaction models and reasoning effort. Finish saves the explicit completion record, replaces the wizard with settings in the same view and opens the right-side main chat. It does not send a message automatically.
  5. Use Check for updates on the CLI version card for later updates. Merely checking or cancelling does not restart sessions. Accepting an update authorizes switching to its verified executable and restarting active extension-owned CLIs; interrupted work must be reviewed and resumed explicitly. Detached/remote processes can keep running.
  6. Use the main chat's Parallel session button or native title-bar + to open the left independent conversation directly. It reuses the active left session, otherwise restores its saved selection or earliest history; without history it creates one conversation. No intermediate list or editor tab is opened.
  7. The left conversation initially copies the main chat's saved settings and then keeps its own models, reasoning, permissions, history and usage. Opening settings or hiding its view does not stop its task. The left native toolbar provides Edit title, Settings and Restart CLI, without a plus; restarting one side does not restart the other.

Independent conversation checkpoints remain under <workspace>/.LLSOAI/editor-sessions/<conversationId>/; the main chat retains its existing .LLSOAI/chat-session.json CLI identity. Each chat's Conversation list button opens native CLI histories for that conversation's working directory and CLI configuration; selecting an entry resumes it only on that side. Parallel session and the native plus action still open the left chat directly. The former editor multi-session implementation, saved histories and serializer are preserved, but its creation/history commands remain hidden from the usual UI. Restoring a conversation never resends its previous prompt.

Full development notes: onboarding and CLI management and independent editor sessions.

Task Flow Mode

Task Flow turns the built-in Chat into a plan-driven, self-continuing agent. Instead of one-shot prompts, the main model writes a plan document first, and the task-flow model then drives that plan to completion.

The recommended loop:

  1. Ask the main model to write an implementation plan into a Markdown file (for example docs/plan-x.md). That file is the contract for everything that follows.
  2. Add the plan file to context with the + button in the file area above the input box (or drag it onto the chat).
  3. Click CC task flow below the input box to insert the start prompt.
  4. Send. From here the model executes each pending step and reports it back through the update_llsccai_task_workflow tool, while the extension keeps the loop going.

Progress is visible in the status bar (completed/total), the full task list opens from a QuickPick panel, and an unfinished workflow is persisted to .LLSOAI/task-flow.json so a VS Code restart offers to resume it. If the model keeps replying with plain text without calling any tool, a circuit breaker pauses auto-continuation instead of spamming the CLI.

A dedicated task-flow model can be picked by clicking either model chip in the composer or in setup step 3. It is saved only for the current workspace under claudeCodeConfigHelper.chat.taskFlow.model; leave it unset to reuse the normal model. claudeCodeConfigHelper.taskFlow.target chooses where task-flow prompts go — builtinChat or externalClaudeCode.

Full walkthrough: docs/taskflow-usage-guide.md.

Model Roles

The model picker exposes three conversational roles, accessible from the Normal or Task flow model chip in the composer; setup step 3 also configures these roles. The separate Image chip configures the global image connection and does not select a conversational role:

Role Setting Used for
Normal claudeCodeConfigHelper.chat.currentModel Everyday chat.
Task flow claudeCodeConfigHelper.chat.taskFlow.model Swapped in right before a task-flow prompt is sent; falls back to Normal when unset.
Compaction claudeCodeConfigHelper.chat.compactionMode.* The /compact summary request.

This replaces the older on-demand expert / plan / review routing. The chat.expertMode, chat.planMode, chat.reviewMode and chat.expert* keys are no longer registered; if they linger in your settings.json they simply show up as unknown settings and can be deleted.

Dedicated Compaction Failure Policy

When a dedicated compaction model is selected, a failed summary request returns HTTP 400 with invalid_request_error, code dedicated_compaction_failed and x-should-retry: false. The extension preserves the original upstream status for diagnostics, releases the compaction wait and stops automatic recovery/task-flow continuation; it does not switch to the normal model or automatically resend /compact.

HTTP 400 alone does not prevent every internal CLI compaction attempt. After a failure, the relay therefore rejects subsequent summary attempts in the same session locally, without another provider request. Submit a new message or run /compact manually to allow a fresh attempt. Main and independent conversations keep separate ownership. Successful dedicated summaries and ordinary chat remain unchanged; without a dedicated model, the existing CLI compaction behavior is retained.

Native Web Search

The CLI's built-in WebSearch uses a separate model request with a server-side search tool. For Responses models and compatible Chat Completions gateways that support native web_search, the extension sends the actual native tool declaration rather than an unimplemented function. Streaming stays enabled when requested, and ordinary function tools remain unchanged. This does not make every OpenAI-compatible model or provider search-capable.

Responses requests include web_search_call.action.sources; Chat requests use tools: [{"type":"web_search"}] and tool_choice: "required" for the CLI's explicitly selected search tool, not web_search_options. Neither protocol sends max_tool_calls: compatible gateways may reject it, and the adapter does not claim an equivalent provider-side cap for Anthropic max_uses. Allowed domains and approximate location are preserved. blocked_domains remains unsupported and returns an explicit error instead of silently dropping the restriction. Search subrequests do not inject or continue task flows.

Structured source URLs and citation titles are returned as CLI-compatible server-tool results. Some Chat gateways return only answer text containing Markdown links: that text is retained with an explicit warning that the provider returned no structured search records and live search execution is not independently verified. Those links are not parsed into verified source blocks or counted as observed server searches. Empty, failed and unfinished replies report errors instead of a successful-looking empty template.

Search depends on provider/model capabilities and can incur provider charges. Source titles fall back to their URLs when absent; no Anthropic encrypted snippets or source URLs are fabricated. Local fixture tests cover conversion, stream completion, failures and real CLI 2.1.312 round trips for Responses, Chat structured sources and Chat text-only replies without paid API calls.

Built-in HTTP Requests

The llsccaiHttp MCP server exposes http_fetch (mcp__llsccaiHttp__http_fetch) in trusted workspaces. The model supplies the complete url plus optional options.method, options.headers and string options.body, matching the JSON-serializable subset of native fetch(url, options). Query parameters belong in the URL; JSON/form bodies must already be serialized. GET/HEAD cannot include a body. No login helper, saved configuration, Base URL or business environment variables are required.

Requests execute in the extension host, including remote hosts. The result contains url, status, statusText, ok, response headers, body, encoding, bytes, truncated and redirectFollowed. HTTP 4xx/5xx responses retain their status and response content rather than becoming transport exceptions. HTML/XHTML responses are parsed into readable text: tags, comments, scripts, styles, templates and explicitly hidden content are removed; paragraph/list/table boundaries and preformatted code text are retained, and HTML entities are decoded. HTML results add bodyFormat: "html-text" and an optional page title; the original response headers and pre-extraction byte count are preserved. JSON/plain text and other non-HTML content remain unchanged; binary content is explicitly Base64. No webpage scripts, stylesheets or subresources are executed/fetched, and no extra summarization-model request is made. JavaScript-rendered pages may consequently have an empty body. This is static text extraction, not a prompt-injection filter or a browser rendering engine.

  • A 60-second deadline covers request and response reading; request bodies are limited to 1 MiB and returned response bytes to 2 MiB, with truncation indicated. Runtime fetch/network availability still applies.
  • 301/302 responses follow up to five hops within the same 60-second deadline, accepting relative or absolute Location values and revalidating each destination. POST follows as GET without its body/content headers; HEAD remains HEAD. Same-origin headers are preserved. Across origins all explicitly supplied headers are dropped, and methods/body that could mutate a different origin are not forwarded. HTTPS-to-HTTP downgrade, invalid/blocked targets, hop-limit exhaustion and 303/307/308 remain inspectable 3xx responses. Results report the final url, redirectFollowed, redirectCount and optional redirectStopped; following a redirect is not an error retry.
  • No ambient browser cookies or provider keys are added, response cookies are not reused, and no requests are automatically retried. Cancellation/timeout does not prove a remote mutation was rolled back. Explicit headers and response data can appear in CLI history; do not pass secrets without permission.
  • Only HTTP/HTTPS is accepted, URL-embedded credentials and transport-control headers are rejected, and known literal link-local/metadata endpoints are blocked. Local development API URLs remain supported. This is not a general DNS-based SSRF firewall or a bypass for enterprise egress policies; use only authorized targets.
  • Main chat uses a capability-protected relay; independent chats use their own private capability URLs and owned cancellation. When the replacement MCP is injected, CLI startup excludes only built-in WebFetch, leaving WebSearch and existing deny rules intact. Untrusted workspaces or user-supplied --mcp-config overrides do not trigger that automatic exclusion. Restart the relevant CLI after installing an updated extension to refresh its tool list.

Built-in Image Generation

Click Image, immediately after the Task flow model chip, to edit the global image connection. The dialog supports English, Simplified Chinese, Traditional Chinese, Japanese, Korean, French and German. It is disabled until explicitly enabled with a valid connection.

  • Existing provider: select an enabled model from an OpenAI-compatible or Responses provider. Reusing its connection does not prove that the provider/model supports the Images API; add a missing image model in provider settings if necessary.

  • Independent connection: enter an Images API Base URL, model ID and API key. The model field accepts manual input and existing model suggestions. The key is stored in SecretStorage, never filled back into the dialog; leaving it blank preserves the saved key. Local unauthenticated services require the explicit no-authentication choice and must satisfy the current connection policy.

  • Global selection: main, parallel and retained editor chats use the same configuration within the current VS Code profile/extension-host storage environment. Saving updates subsequent tool calls, not already accepted tasks, and automatically refreshes running chat CLIs at a safe idle point to reload tool schemas. Busy conversations wait until the current turn, permissions, compaction and recovery settle; the same session is resumed without replaying prompts. Stopped chats are not started. This is not automatic cross-device or local/remote synchronization.

  • Per-request dimensions: the model supplies size as widthxheight, for example 1024x1024, according to your request and provider capabilities. There is no global size or advanced-settings form.

  • Output format: generation and editing accept an optional output_format from the enabled PNG/JPG/GIF/SVG choices. Leave it unset to use the provider's default (typically PNG). JPG is sent as jpeg. The extension ignores frame_count and frame_duration_ms unless the request explicitly selects GIF.

  • Optional animated GIF: enable Model supports GIF only if the selected image model supports it. For animation, explicitly choose output_format: "gif" and a size whose width and height are positive multiples of 16. frame_count accepts integers 2–64 (default 9); frame_duration_ms accepts integers 20–2000 (default 200). Invalid GIF parameters or requests for a disabled explicit format fail before provider submission. Local configuration checks do not verify actual provider GIF support.

  • Resumed conversations: newer CLIs can replay tool schemas saved in native history even after restart. Before forwarding a model request, the relay refreshes only existing built-in image generation/edit definitions to the current format settings. Conversation messages, external MCP tools and native history files remain unchanged; no new conversation or paid regeneration is needed.

The bundled MCP server is llsccaiImage; installing a separate image MCP package is unnecessary.

Tool Purpose
openapi_generate_image Submit one asynchronous image-generation request with prompt and size.
openapi_edit_image Submit a provider-compatible JSON edit with prompt, size and image containing one local file path or 1–16 paths; the host reads and encodes them. Not multipart.
openapi_get_image_task Inspect an owned task on explicit request or for diagnostics, not routine polling.
openapi_list_images List this conversation's tasks, including pending and failed tasks.
openapi_search_images Search original task prompts using a case-insensitive literal substring.
openapi_image_generation_info Read non-sensitive model, enablement and limit information.

Editing inputs are paths only: pass image as an existing local path string, or an array of up to 16 paths. Absolute paths anywhere readable by the extension host, workspace-relative paths, ../ paths, symlinks, ~/ and local file:// URIs are supported; no extra workspace-boundary authorization dialog is added. Remote sessions read files on the remote extension-host machine. Do not submit HTTP URLs, Base64 or Data URIs, and do not encode the image yourself. Missing files return IMAGE_SOURCE_NOT_FOUND; directories and unreadable files are rejected. Single-file limits remain, with a 100 MiB aggregate source limit. Source-format acceptance is still left to the provider. The host reads each source once and freezes a complete data:<MIME>;base64,<encoded bytes> URL in memory before accepting the task, rather than sending bare Base64. PNG/JPEG/WebP/GIF signatures determine their MIME type even when the filename is misleading; other sources use an extension-based MIME hint or application/octet-stream without adding a new format rejection. The MCP tool keeps its path-only image string/array input, but the upstream JSON edit request always uses images: [{ "image_url": "data:<MIME>;base64,..." }] and n: 1, including for a single source. It never sends the legacy top-level image field. Static text-only generation still sends only model, prompt and size; explicit GIF requests additionally send the validated format and frame parameters. No encoded source crosses the MCP request or is returned to the model.

Failure diagnostics: task details and completion messages show the HTTP status and the first 500 Unicode characters of the response body after credential/image-data redaction. This also covers HTTP-success responses containing API errors, HTML, invalid JSON or missing image data. The bounded excerpt survives task/chat restoration; full responses and request credentials are not saved. Error-body capture is capped at 64 KiB and never triggers a retry. Existing failures without a saved response cannot recover that body retroactively.

After acceptance, continue other work or finish the turn: completion is automatically delivered to the originating conversation. Delivery waits for that conversation to become idle, adds a user bubble and sends a structured result through CLI stdin. Persistent receipts prevent automatic resending after an attempted/uncertain delivery; returning to history supplements the display without regenerating images. A busy dialog is not required, and unrelated conversations do not receive the result.

Output is stored under the conversation's workspace:

  • .LLSOAI/openapi-image-generation/tasks/<taskId>.json
  • .LLSOAI/openapi-image-generation/images/<imageId>.png (or .jpg / .webp / .gif)

GIF87a/GIF89a output is accepted from Base64, Data URLs or downloads and saved unchanged, preserving its animation bytes. GIF task/completion records retain the requested format, frame count and frame duration; the existing automatic/manual image-open actions also accept GIF files. Signature validation is not a full animation decoder and does not independently verify the provider's frame count or timing.

The model receives workspacePath, each image's relative path, absolutePath and complete fileUri. Completion guidance asks for that URI verbatim when returning an image or open-original link, rather than rewriting it as a relative path. The chat keeps a compact localized summary with a checked open-image action; it does not expose a public download URL, rewrite Markdown resource addresses or guarantee inline file previews.

Open generated images automatically: newly completed generation/editing tasks open their saved images in the VS Code editor using pinned tabs (preview: false) while preserving chat focus. Partial failures still open successfully saved images; failures without saved output do not open anything. This follows the existing claudeCodeConfigHelper.editor.autoOpenReadWriteFiles master switch, not the code-preview toggle. History refreshes, task queries and reopening the extension do not reopen old images. Local and remote workspaces use their own resource URIs, and a display failure does not turn a successful image task into a failure or resend it.

Costs and data: generation/editing are submitted without an extra extension billing/path-authorization popup after the global configuration and request checks pass; CLI-native permission settings still apply. Provider image requests and the completion chat-model turn may be charged. Editing reads only the explicitly supplied paths, including paths outside the workspace, and sends their bytes directly to the configured provider. Paths may appear in CLI history; internally encoded originals are neither returned through MCP nor stored in task metadata. Prompts and generated results remain in the workspace; no third-party image host is used. Timeouts and local stops cannot guarantee remote cancellation or refunds. No failed generation is automatically retried. Each host writes only its own UUID task files without a workspace-wide runtime.lock; old lock files are ignored, not removed. Another host's unfinished records are never automatically marked interrupted. Normal shutdown marks only this host's tasks interrupted; after a crash, old queued/running records remain for manual diagnosis rather than being taken over.

Image downloads: returned HTTP/HTTPS image URLs use the extension host's normal DNS/network routing, without destination IP filtering, DNS pinning or peer-IP comparison. This permits loopback/private addresses and proxy Fake-IP ranges. It also allows a configured provider to direct requests to host-reachable local/internal services, so use trusted providers and networks; this path is no longer an internal-network/metadata-address security boundary. Downloads still reject URL credentials and non-HTTP protocols, send no provider credentials or cookies, refuse redirects, enforce timeout/byte limits and validate image signatures. This change does not add HTTP/SOCKS proxy support or relax the separate HTTP tool's policy.

Windows task storage: closed temporary JSON files are atomically renamed over the prior checkpoint. Transient Windows EPERM/EACCES/EBUSY replacement failures receive bounded local retries; these never repeat an image API request or delete/truncate the old JSON. Persistent failures report IMAGE_STORAGE_RENAME_<errno>, remain visible in the running host's task view and do not disable unrelated tasks. Pending diagnostics that cannot be persisted survive only until that host exits. Global configuration still uses its separate revision/write lock because all windows edit the same configuration file.

Built-in Chat and Claude CLI Transport

The extension now uses an independent Chat Webview backed directly by the local Claude CLI. The detailed implementation plan and migration notes live in CHAT_WEBVIEW_PLAN.md.

The CLI communication findings used for this implementation are:

  • The current verified Homebrew cask is claude-code 2.1.141.

  • The cask exposes the executable as /opt/homebrew/bin/claude.

  • If the direct Homebrew download is slow or blocked, the CLI can be installed through a local proxy, for example:

    HTTP_PROXY=http://127.0.0.1:3080 \
    HTTPS_PROXY=http://127.0.0.1:3080 \
    ALL_PROXY=http://127.0.0.1:3080 \
    http_proxy=http://127.0.0.1:3080 \
    https_proxy=http://127.0.0.1:3080 \
    all_proxy=http://127.0.0.1:3080 \
    HOMEBREW_NO_AUTO_UPDATE=1 \
    brew install --cask --verbose claude-code
    
  • claude --help confirms -p/--print, --output-format text|json|stream-json, --input-format text|stream-json, --replay-user-messages, and --include-partial-messages are available.

  • After local Claude CLI authentication, real model-output probes succeeded for text, json, stream-json, and stdin text input.

  • The official Claude Code VS Code extension 2.1.144 starts its native UI CLI process with --output-format stream-json --verbose --input-format stream-json over stdio pipes.

  • Based on that finding, the Chat Webview primary transport is long-running stdio with bidirectional stream-json JSON Lines. -p/--print is retained only for single-shot probes and fallback behavior, while PTY remains an experimental last resort.

  • The installed CLI exposes the same stream-json input/output flags needed for that long-running process, so this extension follows the official extension package's SDK-style stdio stream design.

  • Do not enter tokens, passwords, or API keys through automation; complete authentication directly in the user's terminal.

The previous built-in HTTP relay module has been removed. During activation, the extension performs best-effort cleanup of legacy managed Claude Code settings such as local ANTHROPIC_BASE_URL=http://127.0.0.1:<port> entries that were created by older versions.

Managed Claude Code Environment Variables

When you activate a provider/model configuration, the extension writes the required Claude Code settings through the official VS Code Configuration API:

  • ANTHROPIC_BASE_URL
  • ANTHROPIC_AUTH_TOKEN
  • ANTHROPIC_API_KEY
  • ANTHROPIC_CUSTOM_HEADERS
  • CLAUDE_CODE_SKIP_AUTH_LOGIN

The extension marks managed entries with __CLAUDE_ROUTER_MANAGED__ so that it can safely replace only its own environment variables during future activations.

Commands

Available commands include:

  • LLS CCAI: Open Help (claudeRouter.help.open) — opens https://www.zhineng.dev/ directly in the system browser.
  • Claude Code Config: Open Config Panel
  • Claude Code Config: Open settings.json
  • Claude Code Config: Open Global Shared Settings
  • Claude Code Config: Open Workspace Shared Settings
  • Claude Code Config: Open Built-in Chat
  • Claude Code Config: Select Claude CLI for Built-in Chat
  • Claude Code Config: Restart Built-in Chat CLI
  • Claude Code Config: Reload Window
  • Claude Code Config: New Provider
  • Claude Code Config: Refresh Providers
  • Claude Code Config: Set Current Model
  • Claude Code Config: Clear Current Model
  • Claude Code Config: Export Config
  • Claude Code Config: Import Config
  • Claude Code Config: Paste Task Flow to Claude
  • Claude Code Config: LLS CCAI Task Menu
  • Claude Code Config: Show Task Progress
  • Claude Code Config: Continue Task
  • Claude Code Config: Clear Task

Command titles may be localized according to your configured UI language.

Installation

Option 1: Install from a local VSIX

Install the packaged claude-code-config-helper-3.2.91.vsix using VS Code's Extensions → … → Install from VSIX… action. Reload the window when ready to activate the installed build; reload interrupts currently running extension CLIs. To build a package yourself, install the repository dependencies, run npm run compile, then package with @vscode/vsce.

CLI installation through setup

The first setup step can download and verify the native CLI without a global npm install. Official npm is the default; China mirror · npmmirror is an opt-in third-party source whose releases may lag the official registry. The selected extension host determines the platform and architecture, including Remote SSH/WSL/container hosts—not the desktop UI machine.

Managed executables use <download directory>/claude-<version> or claude-<version>.exe. The installer never overwrites the old executable or an external installation. You can instead select an existing executable in Advanced options and validate it. Source URLs, redirects, archive entries, package integrity and the reported CLI version are checked; SHA512 integrity is not an independent publisher signature.

Option 2: Run in development mode

  1. Open this repository in VS Code.
  2. Press F5 to start the Extension Development Host.
  3. In the new VS Code window, open the Command Palette and run the extension commands.

Basic Usage

  1. Open LLS CCAI from the Activity Bar, or run LLS CCAI: Open LLS CCAI Setting from the Command Palette.
  2. If installation is incomplete, finish Language and CLI → Providers → Workspace models in the same view. Existing users without a completion record also follow this flow; an already-installed CLI can be reused without downloading it again.
  3. Add or edit a provider with its name, Base URL, API type (anthropic, openai-compatible or v1-response), authentication mode (auth_token, api_key or none) and key/token. Fetch the model list where supported, or add models manually.
  4. Choose the normal model, optional task-flow and compaction models, and reasoning effort in step 3 or through the model chips in chat. Provider credentials remain shared; independent conversations retain their own saved model selections.
  5. Send a message from the right main chat, or open the independent left conversation with Parallel session or the title-bar +. Use Restart CLI for that conversation's restart-with-progress flow; it keeps history and never replays the last task.

The left settings view contains a compact CLI version/update card and full provider/model management. It does not require a reload for every edit; reload other workspaces when their shared-provider data needs refreshing. Upstream authentication and feature support still depend on the configured provider.

Fetching models keeps your local settings

Fetch Models calls the provider's GET {baseUrl}/models, which normally returns nothing but model ids. The upstream list decides which models exist, but it no longer resets how they are configured:

  • Models still offered upstream keep every locally configured field — display name, context length, max tokens, vision, tool calling, temperature, top_p, sampling mode, selectable, transform-think and preserve-reasoning are never reset by a fetch. The upstream display name is adopted only if you never renamed the model.
  • Models the upstream stopped returning are removed, and if the removed model was your active selection the current model is cleared.
  • New ids are appended, sorted by id, after your existing entries.

The success toast reports the outcome as Fetched N models: added A, preserved K existing configurations, removed R (localized to the selected interface language).

Built-in Chat Flow

When the built-in Chat is enabled and opened:

  1. The extension resolves the configured or selected Claude-compatible CLI executable.
  2. The extension starts a long-running process with --output-format stream-json --verbose --input-format stream-json.
  3. User messages from the Webview are sent to CLI stdin as JSON Lines.
  4. CLI stdout/stderr JSON Lines are parsed into Chat segments.
  5. The Webview renders streaming markdown, code, diff, file references, tool cards, and errors.
  6. Recent Chat messages are restored from VS Code workspaceState for the current workspace until the user clears the session.

Context Compression

The built-in Chat tracks token usage against the selected model's configured context length. When the context approaches the compression threshold, or when the user clicks the token meter compression action, the extension sends an internal @llsccai-summ trigger through the CLI so relay receives the current conversation context.

The relay intercepts that trigger and runs a non-streaming summary request with a compacted single user message. Tool call and tool result blocks are removed from the summary input, keeping the generated summary focused on user goals, assistant decisions, touched files, constraints, and remaining work. After summary generation, the extension clears .LLSOAI/chat-session.json, restarts the CLI into a fresh session, and injects the compressed context internally without showing that seed message in the Chat transcript.

Task Workflow Support

The extension includes an LLS CCAI task workflow system designed for long-running development tasks.

Task workflows help the model:

  • Create a structured task plan from a larger user request.
  • Track progress with task states such as pending, in progress, completed, and blocked.
  • Continue unfinished work through the built-in Chat send chain or the legacy external-Claude clipboard path.
  • Avoid exposing internal workflow tools directly to Claude Code as normal user-facing tools.
  • Intercept local workflow tool calls such as creating or updating the workflow.
  • Schedule automatic continuation when the model needs another turn to finish the task.

Typical workflow commands include:

  • Claude Code Config: Paste Task Flow to Claude
  • Claude Code Config: LLS CCAI Task Menu
  • Claude Code Config: Show Task Progress
  • Claude Code Config: Continue Task
  • Claude Code Config: Clear Task

The Chat host injects workflow instructions and tools only when a workflow is active or when task creation is explicitly triggered. Workflow updates can be intercepted locally so that task progress is reflected in VS Code without requiring the model to execute external side effects.

Shared Prompts

The extension supports global and workspace-level shared prompt settings. Shared prompts can be used to provide consistent system-level guidance across Claude Code sessions.

You can open these settings with:

  • Claude Code Config: Open Global Shared Settings
  • Claude Code Config: Open Workspace Shared Settings

Import and Export

Use the import/export commands to move provider configuration and shared prompt data between environments:

  • Claude Code Config: Export Config
  • Claude Code Config: Import Config

Secrets are handled separately through VS Code SecretStorage and should be reviewed carefully when moving between machines.

Security Notes

Activated tokens appear in VS Code settings.

Claude Code reads configuration from environment variables. Once a provider is activated, the selected token or API key must be written into Claude Code environment settings so that Claude Code can use it.

Inactive provider secrets are stored securely.

Before activation, provider tokens are stored in VS Code SecretStorage, backed by macOS Keychain, Windows Credential Manager, or Linux libsecret depending on your platform.

Managed settings should not be edited manually.

Environment variables placed after the __CLAUDE_ROUTER_MANAGED__ marker are managed by this extension and may be replaced on the next activation. Put manually maintained variables outside the managed block.

Reload may be required after switching models or providers.

Claude Code reads environment variables when its process starts. Reloading the VS Code window ensures the new configuration is used immediately.

Development

Useful development commands:

npm install
npm run typecheck
npm run compile

Run selected lightweight tests after compiling:

npm test

Built-in MCP bridges

The extension ships four built-in MCP servers — llsccaiBrowser, llsccaiVscode, llsccaiWakeup and llsccaiImage — that the Claude CLI launches as Node subprocesses. Those subprocesses have no vscode runtime, so tool calls are forwarded over HTTP to a relay inside the extension host. The image bridge uses a conversation-bound capability; independent conversations use their own private relay rather than the main chat's state.

Their shared machinery (stdio JSON-RPC server, HTTP relay handler, tool-name guards, CLI injection and startup logging) lives in src/mcpKit/. Each bridge only declares a descriptor — server name, HTTP path, relay port environment variable and tool schemas — in its own bridge.ts.

Anything under src/mcpKit/ is loaded by those subprocesses, so it may only statically import Node built-ins and import type. Importing vscode or any extension-host module there makes the subprocess crash at startup and the whole tool group vanish from the model's view with no visible error; src/mcpKit/__tests__/noHostImports.test.ts fails the build if that happens.

Repository

GitHub: https://github.com/liliangshan/claude-code-config-helper.git

License

MIT

  • Contact us
  • Jobs
  • Privacy
  • Manage cookies
  • Terms of use
  • Trademarks
  • Your Privacy Choices
  • Consumer Health Privacy
© 2026 Microsoft