Skip to content

CHANGELOG.md

Latest commit

 

History

History
157 lines (112 loc) · 7.09 KB

CHANGELOG.md

File metadata and controls

157 lines (112 loc) · 7.09 KB
Raw

vscode-agent-platform-connector

0.4.2

Patch Changes

  • e6c2724: Show a "(⚠ no project)" warning in the model picker label when no GCP project is configured, instead of rendering empty parentheses.

0.4.1

Patch Changes

  • 813ae9a: Fix recurring Vertex 400 "tool_use ids were found without tool_result blocks immediately after" by synthesizing placeholder tool results for assistant tool calls whose results never landed in the chat history (e.g. a prior turn failed or was cancelled mid-tool-call). Previously one failed tool call wedged the conversation, causing every subsequent request to fail with the same 400.

0.4.0

Minor Changes

  • c9b2569: Add a Google AI Studio (Gemini API) backend alongside Vertex. Models can now set backend: 'gemini-api' to route through the AI Studio OpenAI-compatible endpoint, billed to the API key's account instead of the GCP project. Ships two new picker entries — Gemini 3.5 Flash (Gemini API) and Gemini 3.1 Pro Preview (Gemini API) — and lets customModels opt into the same backend.

    The API key is stored in VS Code's SecretStorage (OS keychain), not in settings.json. Set it with the new Set Gemini API Key command; when no key is stored the connector falls back to the GEMINI_API_KEY env var.

0.3.2

Patch Changes

  • ef4c7bd: Fix a second cause of the Anthropic tool-pairing 400 ("tool_use ids were found without tool_result blocks" / "unexpected tool_use_id in tool_result blocks"). fitRequestToContext's oldest-message trimming (used to keep long conversations within the model's context window) could drop an assistant tool_use turn while keeping its paired tool_result reply, or vice versa, producing an invalid sequence upstream. Trimming now treats a tool_use assistant turn and its immediately-following tool_result turn as one atomic unit that is always dropped or kept together.

0.3.1

Patch Changes

  • c7d3b80: Fix Anthropic 400 ("unexpected tool_use_id found in tool_result blocks") by dropping orphaned tool_result blocks when building the Claude Messages body. When VS Code trims chat history (or a prior turn failed mid-tool-call), a tool_result could be left with no matching tool_use in the preceding assistant turn; buildClaudeBody now filters unpaired results and skips the turn when none pair, keeping the required tool_use -> tool_result adjacency.

0.3.0

Minor Changes

  • b243398: Add Claude Fable 5 and Claude Sonnet 5 to the built-in model catalog, each with a base and extended-reasoning ("High") variant.

0.2.1

Patch Changes

  • b61ec44: Recover from context-window overflow errors by trimming and retrying. When a long session exceeds the model's context window, the connector now parses the actual token counts from the upstream 400, re-trims the oldest history with a tighter budget, and retries (up to 3 times) instead of failing the turn.

0.2.0

Minor Changes

    • Add a local Anthropic-compatible proxy (--proxy) so the GitHub Copilot CLI can use Claude on Vertex through the connector.
    • Prune older images from Claude requests to stay within the model's image limits, keeping long multimodal conversations working.
    • Re-label the high-effort model variants in the picker as "… – High" (e.g. "Claude Opus 4.8 – High") for clarity.
    • Add Prettier as the shared formatter, wired into ESLint and VS Code format-on-save so editor auto-fix and CLI output stay in sync.

0.1.5

Patch Changes

  • 2eb3eb2: Prevent "conversation is too long" failures by fitting requests to the model's context window. The Claude models now advertise their true 1,000,000-token input limit (was an overly conservative 200,000), and a new last-line safeguard trims the oldest non-system messages when the estimated request would still exceed the limit — so a runaway conversation degrades gracefully instead of failing with an upstream Vertex 400. System messages and the final user turn are always kept.

0.1.4

Patch Changes

  • fe1d0c8: Stop leaking VS Code/Copilot internal cache_control markers into tool-result text. Tool results can nest prompt-cache breakpoint data parts; the converter previously JSON-stringified them into the text sent upstream, which surfaced a raw {"$mid":...,"mimeType":"cache_control","data":"..."} blob in the model's view of the conversation. These internal control parts are now skipped when extracting tool-result text.

0.1.3

Patch Changes

  • bca8976: Help VS Code keep conversations within the model's context window and explain overflows clearly.

    • Token counting now includes image attachments (previously counted as zero) and adds a small per-message/part framing overhead, so the estimate is conservative rather than optimistic. This lets VS Code's chat client trim or summarize history before the request overflows the model, instead of after.
    • When an overflow does happen, the Vertex 400 now surfaces a clear, actionable message (including the token counts when available) telling the user to start a new chat, trim large context, or switch to a larger-context model — instead of the misleading generic "check your project, location, and model settings" hint.
    • Explicitly drop VS Code/Copilot internal control parts (e.g. cache_control prompt-cache markers) during message conversion so they are never forwarded to the upstream model. These were already ignored, but the guard is now named and intentional, hardening against accidental forwarding in future changes.

0.1.2

Patch Changes

  • 27c98bd: Fix image attachments being dropped. The provider now forwards image content parts to Vertex: Gemini receives them as image_url data URIs and Claude as base64 image source blocks, so vision-capable models can see attached images.

0.1.1

Patch Changes

  • Make the "Project" item in the status menu interactive: clicking "Project: Not set" (or the current project) now opens a validated input prompt to set the GCP project, which is saved to your VS Code settings.

0.1.0

Initial preview release.

Added

  • Native VS Code Language Model Chat Provider (vscode.lm.registerLanguageModelChatProvider) exposing Gemini and Claude on Gemini Enterprise Agent Platform (formerly Vertex AI), billed to your own GCP project.
  • Built-in models: Gemini 3.5 Flash, Claude Opus 4.8 (+ thinking), Claude Sonnet 4.5 (+ thinking).
  • gcloud-based authentication with two modes: adc (Application Default Credentials) and isolated (the connector's own private credential store).
  • Streaming responses with tool/function calling for both Gemini (OpenAI Chat Completions shape) and Claude (Anthropic Messages shape).
  • customModels setting to expose additional Model Garden models without a code change.
  • Local, per-day estimated cost tracking shown in the status bar tooltip and status menu (resets at local midnight; never queries the billing API).
  • Status bar item, output channel, and commands: Settings, Show Logs, Check Models, Sign In (isolated credentials).
  • Actionable upstream error messages (enable API / role / Model Garden links).