Tags: continuedev/continue
Tags
fix: remove Ollama template-based tool support gate (#11905) * fix: let explicit capabilities.tools override Ollama template tool detection The Ollama provider checks the model template from /api/show for a `.Tools` placeholder to decide whether to pass tools to the API. For newer models (e.g. qwen3.5) whose templates may not yet include `.Tools`, this silently strips tools from the request — even when the user explicitly sets `capabilities: tools: true` in config.yaml. Use nullish coalescing so an explicit capabilities.tools setting always wins, with the template heuristic as a fallback. * remove verbose comments * remove templateSupportsTools check; trust upstream tool support detection * fix: remove extra blank line to pass prettier check
fix(intellij): handle square brackets in file paths for autocomplete (#… …11474) * fix(intellij): handle square brackets in file paths for autocomplete Use File.toURI() instead of URI constructor for Windows two-slash file:// URIs to properly percent-encode special characters like [ ]. Fixes #10978. * test(intellij): strengthen Windows URI assertions for bracket encoding Verify that square brackets are percent-encoded and drive letter and directory structure are preserved in parsed URI path. * fix(intellij): make Windows bracket test cross-platform The CI runs on Linux where C: is not a drive letter. Assert that brackets are encoded and structure is preserved without platform-specific path assumptions. * fix: use cross-platform URI constructor instead of File.toURI() File.toURI() is platform-dependent — on non-Windows hosts it treats Windows drive paths (e.g. C:/Users/...) as relative paths, producing corrupted URIs. The multi-arg URI(scheme, host, path, fragment) constructor is pure RFC 2396 string manipulation and works identically on all platforms. Also strengthens the Windows bracket test with exact equality assertions instead of loose contains/assertFalse checks. * chore: remove verbose comments for minimal diff --------- Co-authored-by: Dallin Romney <dallinromney@gmail.com>
fix: vLLM provider should respect user-configured contextLength and m… …odel (#11868) * fix: respect user-configured maxTokens and contextLength for autocomplete Two fixes for autocomplete's maxTokens not being respected (#6003): 1. vLLM provider: _setupCompletionOptions() unconditionally overwrote the user's contextLength and model with server-reported values. Now user-explicit settings take priority over auto-detected values. 2. Autocomplete pipeline: renderPromptWithTokenLimit() returned completionOptions without maxTokens, relying on an implicit merge deeper in the call stack. Now maxTokens is explicitly included in the returned completionOptions so the value flows through to the API request reliably. * remove verbose comments from Vllm provider * style: fix prettier formatting * fix: only set maxTokens when llm provides a value * remove defensive-only maxTokens pass-through from autocomplete pipeline
fix: remove Ollama template-based tool support gate (#11905) * fix: let explicit capabilities.tools override Ollama template tool detection The Ollama provider checks the model template from /api/show for a `.Tools` placeholder to decide whether to pass tools to the API. For newer models (e.g. qwen3.5) whose templates may not yet include `.Tools`, this silently strips tools from the request — even when the user explicitly sets `capabilities: tools: true` in config.yaml. Use nullish coalescing so an explicit capabilities.tools setting always wins, with the template heuristic as a fallback. * remove verbose comments * remove templateSupportsTools check; trust upstream tool support detection * fix: remove extra blank line to pass prettier check
PreviousNext