YCraSy DeepSeek Agent v0.1.15

Released 2026-09-13 details.

• Added the move_path tool: moves or renames a file or a whole directory inside the workspace, creating the destination's parent folder. It never overwrites, so an existing destination fails before anything is touched, and an optional expectedBeforeHash rejects the move if the source file changed after the preview.
• Added the delete_path tool: deletes a file or, with recursive: true, a directory and everything inside it. The path goes to the operating system trash unless permanent: true is passed, and recursive is never inferred: a directory without it fails instead of taking its contents with it.
• Path changes decided from Git facts. In the automatic permission modes, move_path and delete_path skip the security review when Git proves the change can be undone: the path is inside a repository, tracked, not ignored, and free of uncommitted changes. An untracked, ignored, modified, or non-repository path keeps the confirmation it had before, so a delete that cannot be undone is never treated as a routine workspace write.
• DeepSeek V4 Pro is selectable again, with its images delegated to Flash. The model registry records capability flags, so the extension knows V4 Pro cannot read images itself: a Pro generation that carries a live attachment is offered analyze_images, which asks Flash about the attachments named in the current message, sends DeepSeek file IDs instead of Base64 or local paths, disables thinking, caps its own output at 8K tokens, rejects expired files and files uploaded to another API origin, and reports the delegated usage against Flash in the vision_analysis phase. A Flash generation keeps reading images natively, in chat and in tool rounds.
• The usage popover now shows the context window and can compact on demand. It reports measured usage once a request has been sent and an estimate before that, with the window size, the soft and hard limits, the point where auto-compaction starts, and how many compactions the conversation has had; Compact context runs a manual compaction and reports the tokens it freed, or that there was nothing to compact.
• Added the low reasoning effort, between off and high, so thinking can be trimmed for speed without turning reasoning off.
• Conversation history can be filtered and grouped by workspace, with an All workspaces entry, a header per workspace, and the conversation count of each group.
• Terminal results keep their full output viewable. The row still shows a bounded preview, and a View retained output disclosure opens the stdout or stderr that was kept, so a command whose preview looks truncated can be inspected without running it again.
• Tooltips are drawn by one portal layer. data-tooltip anchors are rendered by TooltipLayer, which flips sides and aligns when the requested position has no room, so hints survive the panels that clip their children and the narrow sidebar; the composer pickers and select now share the same control metrics, so the model menu matches the permission select instead of carrying its own border and height.
• Credentials are stripped from tool traffic. Tool results, stored tool-call arguments, and the content handed back after a confirmation pass through redaction, which replaces credential-shaped keys and patterns with [REDACTED] before the text reaches the transcript, the tool ledger, or a checkpoint; structured output stays parseable, so the paths and results a tool consumer reads are unaffected.

Yar's DeepSeek Copilot v0.1.14

Released 2026-09-10 details.

• Moved to deepseek-flash (DeepSeek-V4.1-Flash) and removed the DeepSeek V4 Pro option, the analyze_images delegation tool, the hidden Vision-to-Flash fallback, and the retired-model compatibility map.
• Updated the cost catalog to the current rates with the documented peak/off-peak split, and let usage cost be displayed in USD or CNY.
• Fixed the largest avoidable cost of a long chat: a minute-precision timestamp in the system message invalidated DeepSeek's prefix cache on the first request of every generation.
• Reviewer guidance is appended as the newest message instead of being spliced into the system message, so a progress checkpoint or a completion recovery no longer re-bills the whole transcript as a cache miss. Both reviewers ignore that guidance when they read the user's request.
• The coding prompt now budgets code comments: intent that the code does not already state, never a restatement, and matching the file's existing density, since a comment costs output once and input on every later round that re-reads the file.
• The usage popover reports the conversation-wide total computed by the extension host instead of the messages the webview happens to hold.
• Auxiliary compaction uses per-phase output ceilings, the cache-hit ratio is shown per phase, and model capability flags nobody read were removed.
• A finished turn ends with an edited-files summary whose rows open the native change diff; reverting changes is out of scope.
• Pending edit_file and apply_patch previews no longer steal focus, including on macOS, where the platform activates the shown editor even when it is asked to preserve focus.
• Long conversations render lazily, with memoized message rows.
• Added the list_workspace tool; reasoning blocks render Markdown and collapse like the tool lists; fixed the tool-call chevron; attached images open a zoomable, draggable viewer; and the composer gained spacing above its toolbar.
• Added the read_func tool, which returns only the functions, methods, or types named in a source file, chained through their enclosing declarations, so reading one function no longer pulls the whole document into the context. Symbol ranges come from the editor's own language providers, so the tool follows the language instead of a text pattern. Module-level const bindings are resolvable, and an empty answer is believed only after the file has been loaded and the provider has repeated it across a short backoff.
• read_file reads a line range through offset and limit, the fallback for content the editor cannot name as a symbol: search_content locates the name and only those lines are read, with the range bounds, hasMore, and the whole-file hash.
• The prompt and both read tools state one reading order — read_func for a named declaration, search_content plus a read_file range for what the editor cannot name, and a whole-file read only for context spanning declarations — so every chat follows it instead of one conversation's habit.
• Safety decisions start from machine-verifiable facts, so a routine workspace edit or a finite diagnostic command runs without a review round; everything unproven still goes to the independent review.
• A second incomplete verdict from the completion reviewer keeps the delivered answer instead of replacing it with an error.

Yar's DeepSeek Copilot v0.1.13

Released 2026-09-09.

• Added the compact_context tool to request a tool-cycle context compaction on demand. Its handler is a read-only no-op that forces the active tool protocol to roll over into a compacted continuation at the next round, and the model is expected to trust prior successful tool outcomes without repeating completed mutations.
• Raised the default maximum output tokens to DeepSeek's documented 384K limit instead of the conservative 8192 token allowance, so long reasoning and file-generation rounds are no longer prematurely truncated.
• Replaced the native VS Code workspace-reassign prompt with an in-webview Workspace Mismatch modal offering Open here (rebind) or Cancel, backed by a new conversationLoadRejected protocol message and English, Spanish, and Chinese localization.
• run_terminal_command now reuses one dedicated VS Code integrated terminal across successive commands so its scrollback and history remain visible; the terminal is closed only on timeout, cancellation, a working-directory mismatch, or extension shutdown, never right after a successful command.
• Fixed generation-event scoping so protocol handlers that update conversation/generation refs before the next React render accept the first user message immediately after admission.
• Reworked the chat composer footer around CSS grid and container queries: narrow sidebars collapse the model/reasoning/permission picker into a settings gear with an inline usage breakdown, and the activity and tool panels use a cleaner left-indented layout without status capsules.

Yar's DeepSeek Copilot v0.1.12

Released 2026-08-28.

• Removed the provider-specific DSML text detector, streaming buffer, hidden retry, and recovery prompt. Tool execution now depends exclusively on native API tool_calls; tool-shaped assistant text remains ordinary content.
• Split chat generation, history, settings, webview protocol, attachments, usage accounting, tool-call completion, and SearXNG installation into focused services while preserving their external behavior.
• Removed obsolete compatibility shims, forwarding modules, unused provider and FIM paths, and redundant tool-result aliases. Restored and hardened the immutable SearXNG runtime publication workflow with pre-publish metadata and checksum gates.
• Enabled TypeScript unused-code checks and a Knip production gate, then removed trivial implementation-detail tests while retaining 269 behavioral unit tests and nine VS Code integration tests.
• Simplified the interface-language automatic option to Auto in every supported locale.

Yar's DeepSeek Copilot v0.1.11

Released 2026-08-27.

• Replaced Chromium-driven search with an extension-managed local SearXNG runtime that starts on demand, exposes the available engine catalog, supports custom engine selection, and requires no system Python, Docker, Podman, or browser installation.
• Reject DeepSeek DSML serialized inside assistant text, suppress protocol markup across streaming chunk boundaries, and retry once through the native tool-call path instead of accepting malformed assistant content.
• Run agent terminal commands in a visible, per-command VS Code integrated terminal through shell integration, closing it after completion while retaining captured output, exit codes, timeouts, and cancellation.
• Reject detached process launchers and disable .NET build-server/node reuse inside agent terminals so synthetic-test workspaces are not left locked by orphaned dotnet processes.
• Scope duplicate tool-call suppression to the current workspace mutation epoch, and review progress every five rounds after the initial 20-round soft limit so legitimate rebuilds can run without encouraging verification detours.
• Replaced language-specific stalled-response matching with a bounded, tool-free DeepSeek completion review. Premature stops retry once; a second incomplete stop is retained as a failed/incomplete generation instead of becoming a false final answer.
• Removed the configurable tool-round checkpoint, per-block tool-call budget, continuation modal, and related protocol messages. Tool cycles now continue until a real final response, cancellation, context/output boundary, or error.
• Added a progress review every 20 completed tool rounds. Its decision guides the next normal round without disabling tools or forcing a separate final response; uncertain or unavailable reviews fail open.
• Progress reviews now receive a compact cumulative activity history and explicitly stop self-initiated test/debug loops once the requested deliverables already build, instead of treating optional verification failures as unfinished product work.
• Moved usage observability out of individual messages into a compact conversation-level popover beside the chat permission selector, including per-model totals when a conversation switches models. When DeepSeek omits usage for some requests, the popover shows the calculable reported-request cost as a lower bound instead of hiding all cost data.
• Removed legacy conversation and checkpoint compatibility tracked by issue #61. History now accepts only schema-v2 conversations, checkpoints require schema 3, and incompatible files are deleted during activation instead of migrated or quarantined.

SearXNG runtime 2026.08.22-9fea41204-v2

Pinned SearXNG sidecar runtimes used by DeepSeek Copilot. Published assets are immutable; the extension trusts only checksums embedded in its VSIX.

Yar's DeepSeek Copilot v0.1.10

Released 2026-08-22.

• Simplified permissions to default, auto-approve, and full-access; removed the per-tool matrix, read-only/custom profiles, and session trust shortcuts.
• Replaced local terminal danger analysis with an independent DeepSeek review that classifies mutations as routine, elevated, or critical. Auto-approve confirms elevated and critical actions; full-access confirms only critical actions that could make the computer unusable or cause broad irreversible loss.
• Moved Web search into Tools below Permission mode and added a global switch that removes search_web and read_web from model requests while disabled.
• Migrated retired settings and checkpoints to the new three-mode contract and advanced the webview protocol to version 5.
• Unified the chat textarea, compact model/reasoning picker, attachments, permission mode, and send actions into one rounded composer, and removed the redundant settings-saved notification. During generation, one contextual button now switches between Stop when the draft is empty, Interrupt and guide for Enter, and Queue message while Ctrl is held (Ctrl+Enter).
• Kept tool calling available in non-thinking mode while honoring thinking: disabled throughout every tool round, matching the current DeepSeek API contract.
• Replaced the non-visual Flash option with DeepSeek V4 Vision (Flash) (deepseek-v4-flash-vision-exp). Vision receives image file IDs directly, while Pro can request the new analyze_images tool to delegate image interpretation to Vision and consume its text description.
• Added a provider-local fallback from unavailable experimental Vision to stable deepseek-v4-flash. Text requests and connection tests retry once; direct Vision chats omit images and disclose the limitation, while Pro's delegated analyzer fails explicitly instead of returning invented visual analysis. Authentication, rate-limit, generic provider, and custom-endpoint failures never trigger the fallback.
• Unified image and context-file attachment behind one native picker. Binary signatures, rather than file extensions, identify JPEG, PNG, GIF, and WebP images; pasted clipboard images are supported with Ctrl+V/Cmd+V.
• Added DeepSeek Files API upload and deletion, 30-day remote expiry, bounded local preview caching, attachment limits, and deferred cleanup so history Undo can restore deleted conversations without losing their images. Base64 is used only transiently for clipboard IPC and is never persisted or sent in chat requests.
• Redesigned explicit cancellation as a stable terminal outcome: Stop preserves the original prompt, partial assistant timeline, and completed tool results as cancelled, emits one terminal event, never rolls back completed side effects, and cannot leak stale cancellation state into the next message. Steering remains internally interrupted, but now carries a verified link to the source generation, explicitly continues the original task under the latest guidance, and hides the misleading interruption warning.

Yar's DeepSeek Copilot v0.1.9

Released 2026-08-11.

• Publish Marketplace and GitHub versions through their normal release channels while retaining the extension's preview: true product status.
• Start a new chat automatically when restored conversation state belongs to another VS Code workspace or window.
• Isolated concurrent chat generations with protocol-v3 conversation/generation correlation, navigation request IDs, background activity status, and snapshot-based restoration when returning to a running chat.
• Changed explicit Stop to atomically remove the complete cancelled turn from UI and persisted context and restore its prompt as a per-conversation draft; steering and lifecycle interruptions retain their distinct recovery behavior.
• Centralized terminal generation ownership so each run emits exactly one completion or error, added cancelling/cancelled states and typed stop reasons, and made repeated or stale cancellation idempotent.
• Propagated cancellation through context discovery, project instructions, compaction, provider streaming, browser work, confirmations, mutation locks, tools, and descendant processes.
• Removed tool-round checkpoints and per-block tool-call budgets from auto-approve and full-access; the configured limit now applies only to default, read-only, and custom modes.
• Reworked automatic context compaction around calibrated generation budgets, with a valid starting -> compacting -> streaming lifecycle and safe rollover after closed tool rounds even when usage jumps directly to the hard limit.
• Compaction now consumes its quota and emits one localized, persisted marker only when it actually reduces the request; provider-reported prompt usage calibrates primary, auxiliary, and tool-request estimates.
• Centralized Unicode-safe UTF-8 text bounding, merged overlapping file ranges, bounded large references and summaries, and persisted only newly covered generation IDs per compaction boundary to avoid quadratic history growth.
• Added preventive output-overflow recovery for reasoning-heavy responses while keeping incomplete completion states visible, and removed redundant budget events and checkpoint fields without breaking schema 1 or 2 recovery.
• Added regression coverage for initial compaction, cancellation, calibrated hard limits, tool-cycle rollover, compaction quotas, Unicode byte boundaries, range normalization, incremental markers, legacy checkpoints, and extension-host state transitions.

Yar's DeepSeek Copilot v0.1.7

Released 2026-08-07.

• Reworked integrated-browser search around one reusable VS Code page: searches type into the active engine, registered results open by click, navigation returns through browser history, and a compatibility mode remains available when newer browser tools are missing.
• Added localized DuckDuckGo, Bing, Google, and Yahoo fallback; semantic organic-result parsing; Bing redirect decoding; public-HTTPS validation; bounded search/document caches; compact 8 KiB responses; semantic page extraction; and sanitized browser diagnostics.
• Replaced model-visible page IDs, DOM references, arbitrary navigation, and generic link following with opaque search/document IDs and only two constrained tools: search_web and the multi-mode read_web.
• Compact completed conversation context to user/final-answer pairs, retain full provider transcripts only for active or incomplete recovery, and lazily compact duplicate legacy web output when history is saved again.

v0.1.6 Preview

Preview release focused on privacy, production reliability, and provider-reported usage observability.

• Adds Incognito mode for ephemeral chats with explicit save-or-discard transitions.
• Adds per-request, generation, and conversation token usage, reasoning, cache-hit/cache-miss, and official DeepSeek V4 cost estimates.
• Keeps missing provider usage explicitly unavailable and avoids guessed pricing for custom endpoints.
• Adds configurable usage warnings and redacted diagnostics without persisting prompts, commands, paths, or response content.
• Hardens tool calls, unsaved buffers, storage concurrency, partial streams, process shutdown, webview protocol negotiation, and provider validation.
• Adds cross-platform production gates, dependency audits, deterministic VSIX checks, Extension Host coverage, and installed-package smoke tests.

Validation: Production gates passed on Linux, Windows, and macOS with VS Code 1.125.0 and stable. The pre-release VSIX contains 16 intentional entries.