Release v3.5.0

Changelog

[3.5.0] - 2026-06-17

Fixed

• Tool calling broken on LM Studio and Ollama: Local models emit tool calls as inline <tool_call>{...}</tool_call> blocks inside delta.content instead of native delta.tool_calls. The SSE parser now runs every text chunk through a new ToolCallAdapter (src/tools/toolCallAdapter.ts) that detects these XML-style blocks, generates a stable callId, and emits a LanguageModelToolCallPart. Native delta.tool_calls is still honored (OpenCode / Zen / Go unchanged).
• Tool-calls and reasoning tags from many model families: The parser recognizes the full range of XML/pipe-bar/bracket/markdown tag variants used by Qwen, DeepSeek, Hermes, Llama 3.x, Mistral (<antml:function_calls>), Phi-3.5, Granite, Cohere, Yi, and any local model that emits reasoning or tool blocks inside delta.content. Tool-call tags: <tool_call>, <toolcall>, <tool>, <invoke name="x">, <action>, <function>, <antml:function_calls>, <|tool_call|> (Llama 3), [TOOL_CALL] / [TOOL_CALLS] (Cohere/Yi). Reasoning tags: <think>, <thinking>, <reasoning>, <reflection>, <analysis>, <thought>, <solution>, <plan>, <|thinking|> / <|reasoning|>, [THINK], and markdown-style ### Reasoning / ### Thought / ### Plan blocks. Body parsing supports the common JSON shapes ({name, arguments|parameters|input|args}, {function: {name, arguments}}, Cohere {tool, parameters}) and the Hermes/Granite XML form <invoke name="x">JSON-args</invoke>.
• Reasoning blocks missing on LM Studio and Ollama: `` blocks inside delta.content are now parsed and emitted as `LanguageModelThinkingPart`, so reasoning models show their thinking as a collapsible block. Native `delta.reasoning_content` is still honored.
• LM Studio / Ollama servers invisible in the side panel: OpenCodeTreeProvider only received OpenCode servers in its state. The two local providers now expose getServerList() and refreshTreeView() aggregates them under the same Dashboard header with a per-type icon (LM Studio = $(chip), Ollama = $(zap)).
• LM Studio / Ollama servers lost on reload: Servers added via the "Add Server" command were kept only in memory. They are now persisted to SecretStorage (workspace state) under a new LocalServerConfig schema and re-hydrated on activate().

Added

• Per-device LM Studio model entries: Newer LM Studio versions expose the same model key once per device (GPU0, GPU1, CPU, MPS). The provider now reads loaded_instances[] and emits one RoutedModelInfo per (model, device) pair, with the device id suffixed in the display name and embedded in the composite id (${serverId}:${key}@${device}). Single-device installations keep the original naming.
• Unified "Remove Server" picker: opencode-zen.removeServer now lists OpenCode, LM Studio and Ollama entries together, dispatching to the right backend (MultiServerManager vs LMStudioProvider vs OllamaProvider) and persisting accordingly.
• ToolCallAdapter unit tests in test/tool-call-adapter.test.ts: 46 cases covering complete blocks, split chunks, partial opening tags, nested JSON shape, malformed bodies, interleaved reasoning + tool calls, and every supported tag family (Qwen, DeepSeek, Hermes, Llama 3.x, Mistral, Cohere, Yi, markdown headers).

Files

• Added: src/tools/toolCallAdapter.ts, test/tool-call-adapter.test.ts
• Modified: src/providers/OpenAICompatibleProvider.ts, src/providers/LMStudioProvider.ts, src/providers/OllamaProvider.ts, src/config/secretStorage.ts, src/extension.ts

[3.4.0] - 2026-06-16

Added

• VS Code Proposed APIs Integration: Activadas 6 APIs propuestas de VS Code para mejorar la experiencia de chat
  • chatProvider (v5): Provider de chat con firma actualizada
  • languageModelThinkingPart (v1): Thinking blocks colapsables en el UI
  • chatInputNotification: Notificaciones en el área de input del chat
  • chatStatusItem: Items de estado en la vista de chat
  • languageModelPricing: Información de precios por token
  • languageModelSystem (v3): Mensajes de sistema nativos
• Thinking Blocks Colapsables: Emisión de LanguageModelThinkingPart para reasoning content
  • Soporte en streaming SSE OpenAI/LM Studio (delta.reasoning_content)
  • Soporte en streaming SSE Anthropic (thinking_delta)
  • Los bloques de thinking se renderizan como elementos colapsables en VS Code
• Chat Input Notifications: Sistema de notificaciones en el input del chat
  • showMissingConfigNotification(): Aviso cuando un provider no está configurado
  • showConnectionErrorNotification(): Aviso de errores de conexión
  • showConnectedNotification(): Confirmación de conexión exitosa
  • Fallback a mensajes estándar de VS Code cuando la API propuesta no está disponible
• Chat Status Items: Manager de estado para providers
  • ChatStatusItemManager: Crea items de estado usando la API propuesta
  • Fallback a StatusBarItem cuando la API no está disponible
• Model Pricing: Información de precios integrada en los modelos
  • Campo _pricing en RoutedModelInfo con inputTokenPrice, outputTokenPrice, currency
  • Precios populados desde models.dev para providers OpenCode (Free, Go)
  • Conversión automática de $/M tokens a $/token
• System Messages Nativas: Reemplazo del hack numérico por API propuesta
  • Eliminado (msg.role as number) === 2
  • Usa (vscode as any).LanguageModelChatMessageRole.System en convertMessages() y convertMessagesAnthropic()

Changed

• Refactor de Providers: Unificación de providers bajo OpenAICompatibleProvider
  • Eliminados BaseLocalProvider.ts, BaseOpenCodeProvider.ts
  • Nuevo OpenAICompatibleProvider.ts como clase base para todos los providers
  • Soporte unificado para formatos: OpenAI, Anthropic, OpenAI Responses
• Firma de método actualizada: provideLanguageModelChatResponse usa LanguageModelChatRequestMessage[] en lugar de LanguageModelChatMessage[]
• Eliminados archivos obsoletos: openCodeApiClient.ts, anthropicAdapter.ts, messageConverter.ts, openaiResponsesAdapter.ts, responseStreamer.ts, toolCallAdapter.ts

[3.3.0] - 2026-06-11

Added

• LM Studio Provider: Nuevo provider para conectar LM Studio local o remoto
  • Auto-detección de modelos vía /v1/models
  • Detección de capacidades: Reasoning, Vision, Tools
  • Streaming SSE real con chunks progresivos
  • Soporte para múltiples servidores LM Studio
  • Heurísticas de context length basadas en tamaño del modelo
• Ollama Provider: Nuevo provider para conectar Ollama local o remoto
  • Auto-detección de modelos vía /api/tags
  • Detección de capacidades: Reasoning, Vision, Tools
  • Streaming NDJSON con chunks progresivos
  • Soporte para múltiples servidores Ollama
  • Heurísticas de context length basadas en parameter_size
• Rebranding: Extensión renombrada a "+ Providers on Copilot Chat"
  • Nuevo nombre: plus-providers-copilot-chat
  • Nuevas keywords: lmstudio, ollama, local-ai, self-hosted, remote-server
  • README actualizado con documentación de todos los providers
  • Soporte para conexiones remotas en todos los providers locales

Changed

• package.json: Actualizado displayName, description, keywords
• README.md: Documentación completa de LM Studio, Ollama, OpenCode Servers, y OpenCode Cloud

[3.2.5] - 2026-06-11

Added

• Tests reales con datos del servidor: 30 tests pasando validados contra respuestas reales.
  • test/OpenCodeServerProvider.test.ts — 15 tests: extracción de texto/reasoning/tokens de parts[], manejo de errores, respuestas vacías, tool calls, validación de estructura JSON, simulación de orden de eventos streaming
  • test/OpenCodeServerProvider.streaming.test.ts — 9 tests: orden reasoning→texto, concatenación de múltiples text parts, cancelación de token, manejo de tool calls, respuesta vacía
  • test/opencode-server.integration.test.ts — 6 tests de integración REAL con servidor opencode + LMStudio:
    • Verifica health del servidor, crea sesión, envía mensajes
    • Tarea pesada: genera 900+ chars con 28K tokens en 22s
    • Verifica reasoning en respuesta compleja (matemáticas)
    • Valida estructura JSON: step-start → reasoning → text → step-finish
    • Usa provider lm-studiolocal configurado en opencode → http://localhost:1234/v1
  • test/provider-behavior.test.ts — 4 tests HONESTOS del comportamiento REAL:
    • Documenta que el servidor opencode devuelve JSON completo (NO SSE/streaming)
    • El provider acumula parts[] y emite al final con yield al event loop
    • 100 parts se procesan en ~165ms (si fuera streaming real tardaría 10-20s)
    • Valida estructura JSON real del servidor: step-start → reasoning → text → step-finish
    • Documenta la diferencia: opencode (JSON completo) vs LMStudio (SSE real)
  • Framework: node:test + tsx (sin jest/mocha)

Fixed

• Chat se queda "working" infinitamente: Simplificado OpenCodeServerProvider para procesar la respuesta JSON del servidor de forma directa y limpia. Ahora: fetch() → await response.json() → acumular parts[] → emitir progress.report() con yield al event loop entre cada chunk para que VS Code actualice el UI.
• Eliminada simulación por delays: Removido todo el código de parsing multi-formato con delays artificiales de 50ms (readAllStreamData, processAnyFormatResponse, parseSSE, parseNDJSON, processParsedEvents, inferEventType, safeParseJson).
• Eliminada dependencia @opencode-ai/sdk: El intento de usar el SDK para SSE streaming no funcionó porque el servidor no emite eventos de contenido en tiempo real vía /global/event. El SDK fue removido del bundle (tree-shaking). Bundle bajó de 177KB a 105KB.
• Manejo de respuesta limpio: Acumula reasoning, tool calls y texto por separado, luego los emite en orden (reasoning → tools → text) con await new Promise(r => setTimeout(r, 0)) entre cada uno para ceder al event loop.
• Manejo de errores: Verificación de messageData.error a nivel top-level antes de procesar parts[].

[3.2.4] - 2026-06-11

Changed

• Intento de integración con SDK @opencode-ai/sdk para SSE streaming. Descartado — el servidor no emite eventos de contenido en tiempo real. Solución final en v3.2.5.

[3.2.3] - 2026-06-11

Fixed

• Parsing multi-formato de respuestas del servidor (SSE, NDJSON, JSON parts[], OpenAI delta). Obsoleto en v3.2.5 — reemplazado por SDK nativo. Ver v3.2.4 para contexto.

[3.2.2] - 2026-06-11

Fixed

• 11 TypeScript compilation errors que rompían tsc --noEmit en CI. Ver release v3.2.2 para el listado completo de archivos corregidos.

[3.2.1] - 2026-06-11

Changed

• Server launch: Reemplazado exec() con spawn() para lanzar servidores locales como proceso background sin ventana cmd.exe emergente
• Launch UX: El usuario puede elegir entre lanzar servidor en terminal VS Code o como proceso background invisible
• Activation: Ya no se auto-lanzan servidores locales al activar la extensión

Added

• SSE streaming inicial: Server provider lee respuestas como text/event-stream con progreso incremental
• Server launch command: Interfaz interactiva con showQuickPick para elegir modo de lanzamiento

[3.2.0] - 2026-06-02

Added

• Subagent tool (opencode_subagent): Registered via vscode.lm.registerTool(), delegates to first available OpenCode provider
• Thinking blocks: Server provider and BaseOpenCodeProvider now use LanguageModelThinkingPart for collapsible reasoning content

Fixed

• Tool registration in package.json — requires modelDescription, displayName, and inputSchema fields

[3.1.0] - 2026-06-02

Fixed

• Server provider session API: Correct request format for OpenCode server (model como { providerID, modelID }, solo model y parts en el body)

[3.0.0] - 2026-06-02

BREAKING CHANGES

• models.dev API integration: Model capabilities fetched live from https://models.dev/api.json
• Server provider rewrite: Usa session-based API (POST /session + POST /session/:id/message)
• Single server provider: All connected servers register under one opencode-server vendor

Added

• Live model registry: 40+ Zen models + 16 Go models con context sizes, pricing, y capabilities reales
• Pricing in tooltips: hover muestra In: $X/M · Out: $Y/M · Cache: $Z/M
• ThinkingEffort configuration: Reasoning models show configurationSchema para low/medium/high

Fixed

• Image detection, tool schema validation, server auth, model registration

Removed

• Anthropic Messages adapter, OpenAI Responses adapter, @vscode-elements/elements dependency

[2.4.0] - 2026-06-02

Added

• SDD task specs for multi-provider architecture (8 tasks)
• .specs/ directory with task definitions y dependency graph

[2.3.3] - 2026-06-02

Fixed

• /usage endpoint returns 404 — Global tree ahora muestra model families como fallback

[2.3.2] - 2026-06-02

Changed

• Debug logging en getUsage y streamResponse

Fixed

• Config tree command IDs, Global tree state, fetchApiUsage timeout de 8s

[2.3.0] - 2026-06-02

Added

• 3-tree sidebar layout: Session, Global, Config
• Global tree: fetches real account data from /usage endpoint con progress bars
• Config tree: one-click buttons para Configure Zen Key, Configure Go Key, Refresh All Models, Clear Usage Stats

[2.2.0] - 2026-06-02

Changed

• Replaced WebviewView with native TreeView: usa VS Code's native TreeDataProvider API

[2.1.1] - 2026-06-02

Fixed

• Usage stats blank (Map → Record), thinking rendering interleaved, webview sessions

[2.1.0] - 2026-06-02

Changed

• Sessions and request tracking: unique requestId y sessionId, tokens reemplazados en lugar de acumulados
• Reasoner steps: streaming tracks reasoner_step events
• Usage tracker keyed by requestId: recordRequest(requestId, ...) finds existing records by requestId and replaces them instead of appending duplicates

Fixed

• Thinking content no longer renders as HTML — reportThinking sends LanguageModelTextPart directly to progress (no more literal tags)
• Tokens no longer summed incorrectly across requests — each request's total is reported once and replaced if the same requestId fires again

[2.0.1] - 2026-06-01

Fixed

• Strip <think> tags from streaming content so they no longer render as literal HTML
• Auto-refresh models when API key is configured (set, changed, or cleared)
• Status bar button now opens the usage sidebar (not the output channel)
• Models endpoint queried even without API key (fails gracefully if auth required)

Changed

• Webview redesigned with collapsible tree structure and VSCode-style aesthetics
• Each model entry expands to show requests, prompt, completion breakdown
• Each recent request expands to show full details
• Usage stats bars with percentages
• New command OpenCode Zen: Show Output Log for terminal-style stats view

[2.0.0] - 2026-06-01

BREAKING CHANGES

• 3 Independent Providers: Now registers opencode-free, opencode-go, and opencode-zen as separate vendors
• No More Hardcoded Models: All models fetched dynamically from OpenCode APIs
• Separate API Keys: zenKey and goKey stored independently in SecretStorage
• Removed Files: modelMetadata.ts, registry.ts, modelsDevClient.ts, zenClient.ts, provider.ts deleted

Added

• BaseOpenCodeProvider abstract class for shared logic
• OpenCodeFreeProvider - free models only (uses Zen key)
• OpenCodeGoProvider - Go models with /go/v1/ endpoint
• OpenCodeZenProvider - paid Zen models only
• OpenCodeClient - generic HTTP client with endpoint parameter
• FileSystemWatcher on local auth.json for new key detection
• New usageWebview with API keys status, balance, and detailed stats
• Commands: configureZen, configureGo, refreshAll
• Auto-detection of local OpenCode installation at activation

Changed

• SecretStorage now manages 2 keys: zenKey and goKey
• Model caching with 5-minute TTL
• API usage caching with 1-minute TTL
• Models endpoint: GET /v1/models for Zen/Free, GET /go/v1/models for Go
• No SDK dependency (direct fetch to HTTP APIs)

[1.0.2] - 2026-06-01

Added

• New project logo combining OpenCode and Copilot branding

Fixed

• PNG icon instead of SVG (VSCode requirement)
• readonly sessionStats TypeScript error

[1.0.1] - 2026-06-01

Fixed

• PNG icon instead of SVG (VSCode requirement)
• readonly sessionStats TypeScript error

[1.0.0] - 2026-06-01

Added

• Initial release
• 45+ OpenCode Zen models registered as Copilot Language Model Provider
• 4 free models enabled by default
• Auto-detection of OpenCode installation and API key
• Tool calling support with automatic JSON repair
• SSE streaming with reasoning content display
• Vision/image input for multimodal models
• Status bar indicator with connection state
• Settings: timeout, tool calling, image input, temperature, verbose logging
• Model catalog from builtin metadata, models.dev, and Zen API