Release v3.2.0

Changelog

[3.2.0] - 2026-06-02

Added

• Subagent tool (opencode_subagent): Registered via vscode.lm.registerTool(), delegates to first available OpenCode provider. Accepts query and description parameters. Runs with temperature 0, no additional tools.
• Thinking blocks: Server provider and BaseOpenCodeProvider now use LanguageModelThinkingPart for collapsible reasoning content. Falls back to text markers if API unavailable.

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 must be { providerID, modelID } object (not string)
  • providerID comes from server's actual provider data (e.g., "opencode")
  • Session API only accepts model and parts — no tools, temperature, etc.
  • Filter user/assistant messages only (ignore system/tool messages)

[3.0.0] - 2026-06-02

BREAKING CHANGES

• models.dev API integration: Model capabilities (context size, pricing, vision, reasoning) now fetched live from https://models.dev/api.json instead of hardcoded metadata
• Server provider rewrite: OpenCode Server now uses session-based API (POST /session + POST /session/:id/message) instead of non-existent /chat/completions endpoint
• Single server provider: All connected servers register under one opencode-server vendor instead of one per server

Added

• Live model registry: modelRegistry.ts fetches from models.dev on activation with 30-minute cache
• 40+ Zen models + 16 Go models with real context sizes, pricing, and capabilities
• Pricing in tooltips: hover over any model shows In: $X/M · Out: $Y/M · Cache: $Z/M
• Server provider models: Local servers now appear in Language Models view with correct capabilities
• ThinkingEffort configuration: Reasoning models show configurationSchema for low/medium/high
• Partial model ID matching: Handles prefixed IDs (e.g., opencode/deepseek-v4-flash)
• Debug logging: Detailed SSE parsing logs in Output panel for troubleshooting

Fixed

• Image detection too aggressive: Was matching ANY message part with mimeType as image, causing "Image input not supported" for all models. Now only detects actual LanguageModelImagePart and LanguageModelDataPart with image/* mime types
• Tool schema validation: Filters out tools with empty/undefined schemas that cause 400 errors from strict providers (MiniMax, etc.)
• Server provider auth: buildHeaders() made public, auth headers properly sent
• Server provider streamResponse: Fixed ReadableStream vs ReadableStreamDefaultController mismatch
• Server provider model registration: Models now appear in Language Models view after registration
• models.dev data loading: Now blocking (await) before provider registration so data is available when VS Code queries models

Removed

• Anthropic Messages adapter (anthropicAdapter.ts) — all models use /chat/completions via OpenCode routing
• OpenAI Responses adapter (openaiResponsesAdapter.ts) — all models use /chat/completions via OpenCode routing
• @vscode-elements/elements dependency — not usable in webviews

[2.4.0] - 2026-06-02

Added

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

Changed

• TreeView sidebar with Dashboard + Config sections
• Model registry with fallback static data

[2.3.3] - 2026-06-02

Fixed

• /usage endpoint returns 404 — confirmed locally: GET /zen/v1/usage → 404, /zen/go/v1/usage → 404. Global tree now gracefully handles missing endpoint and falls back to showing model families instead of hanging
• Global tree now shows model families: when usage API is unavailable, the Global tree still displays available models grouped by family (minimax, kimi, glm, deepseek, qwen, mimo, hy3) with model counts — making the view useful even without billing data
• getModelFamilies() added to BaseOpenCodeProvider to expose model group info
• refreshGlobal command now refreshes models + re-fetches global data instead of a separate refresh method

[2.3.2] - 2026-06-02

Changed

• Added debug logging to getUsage: logs URL, HTTP status, response body (or empty/JSON parse error) so we can see exactly what the API returns
• Added debug logging to streamResponse: logs first chunk id/model, reasoning start/done, usage tokens

Fixed

• Config tree now uses correct command IDs (was using opencodeConfigureZen instead of opencode-zen.configureZen)
• Global tree shows pending/error/ok state per provider instead of hanging
• fetchApiUsage has 8s timeout to prevent tree from hanging forever

[2.3.0] - 2026-06-02

Added

• 3-tree sidebar layout: Session (local session stats), Global (API account data from server), Config (quick actions)
• Global tree: fetches real account data from /usage endpoint — balance, used, remaining, quota limits per time period (hour/day/week/month). Shows progress bars with percentages and reset timestamps
• Config tree: one-click buttons for Configure Zen Key, Configure Go Key, Refresh All Models, Clear Usage Stats — each as a native tree item with command binding
• ApiQuota interface: id, name, unit, limit, used, remaining, reset_at, period for granular rate limit display
• Context menus: right-click items in Config/Global trees for relevant actions

Changed

• Session tree renamed from UsageTreeProvider → SessionTreeProvider with cleaner helper methods

[2.2.0] - 2026-06-02

Changed

• Replaced WebviewView with native TreeView: usage sidebar now uses VS Code's native TreeDataProvider API (like Explorer/Timeline/Outline). No custom HTML/CSS — all rendering uses VS Code's native tree components with proper labels, icons, and collapse states
• Sidebar sections: API Keys, Balance, Session Summary (expandable), By Provider, By Model, Sessions (grouped by sessionId with nested requests), Recent Requests

[2.1.1] - 2026-06-02

Fixed

• Usage stats blank: byModel and byProvider changed from Map to plain Record<string, ...> so they serialize properly over webview postMessage (Maps were invisible to JSON.stringify)
• Thinking rendering: thinking content is now buffered and emitted as a single block with [reasoning]...[/reasoning] markers when reasoning ends, preventing interleaved text output
• Webview sessions: new "Sessions" section groups requests by sessionId with nested request details

[2.1.0] - 2026-06-02

Changed

• Sessions and request tracking: every request now has a unique requestId and belongs to a sessionId. Tokens reported for the same requestId are replaced (not accumulated), fixing the token overcount bug where the API reports total tokens per session, not per message
• Reasoner steps: streaming now tracks reasoner_step events so the provider collects reasoning steps (label, id, timestamps) per request
• 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