feat(llm): add configurable thinkTagName for thinking output formats

[vinaychauhannumber]

Summary

Fixes #5992

Different LLM providers use different XML tag names for reasoning/thinking output. Previously, Continue hardcoded <think>...</think> in multiple places, which broke support for providers like vLLM that use custom tag formats.

This PR introduces a thinkTagName option on LLMOptions (defaulting to "think") so users can configure the tag name to match their provider.

Changes

core/config/types.ts

• Added thinkTagName?: string to the LLMOptions interface with full JSDoc documentation

core/llm/index.ts

• Added thinkTagName: string property to the BaseLLM class
• Wired it up in the constructor with a default of "think" (fully backward-compatible)

core/llm/llms/Ollama.ts

• Replaced hardcoded "<think>" / "</think>" strings with dynamic thinkOpenTag / thinkCloseTag values derived from this.thinkTagName

core/util/index.ts

• Updated removeCodeBlocksAndTrim to accept an optional thinkTagName parameter
• The think-block regex is now built dynamically instead of being hardcoded

core/autocomplete/postprocessing/index.ts

• Updated the Qwen3 postprocessing block to use llm.thinkTagName for regex construction instead of hardcoded <think> tags

Usage Example

{
  "provider": "openai-compatible",
  "model": "my-vllm-model",
  "apiBase": "http://localhost:8000/v1",
  "thinkTagName": "reasoning"
}

This will correctly strip <reasoning>...</reasoning> blocks instead of the default <think>...</think>.

Backward Compatibility

Fully backward compatible — thinkTagName defaults to "think", so all existing configs work unchanged.

References

• https://docs.vllm.ai/en/latest/features/reasoning_outputs.html#streaming-chat-completions

---

Summary by cubic

Make the reasoning tag name configurable via thinkTagName so providers like vLLM and Ollama work correctly; default stays "think" for backward compatibility. Fixes #5992.

• New Features
  • Added thinkTagName?: string to LLMOptions; wired through BaseLLM, ILLM, and public types (core/index.d.ts).
  • Updated Ollama streaming to use dynamic open/close tags.
  • Made regexes dynamic in Qwen3 postprocessing and removeCodeBlocksAndTrim to strip custom thinking blocks.

^{Written for commit f23b9e5. Summary will update on new commits.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[vinaychauhannumber]

I have read the CLA Document and I hereby sign the CLA

[vinaychauhannumber]

recheck

[cubic-dev-ai]

1 issue found across 5 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="core/config/types.ts">

<violation number="1" location="core/config/types.ts:583">
P2: The configurable `thinkTagName` is documented as an XML tag name, but downstream code interpolates it directly into `new RegExp()` without escaping regex metacharacters. A user setting `thinkTagName` to something like `reason[ing]` or `think.*` would hit a runtime `SyntaxError` or match unintended content. The codebase already has `escapeLiteralForRegex` in `core/util/regexValidator.ts`; it should be applied when building the think-block regexes in `core/util/index.ts` and `core/autocomplete/postprocessing/index.ts` (or the value should be escaped when assigned in `BaseLLM`).</violation>
</file>

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: The configurable thinkTagName is documented as an XML tag name, but downstream code interpolates it directly into new RegExp() without escaping regex metacharacters. A user setting thinkTagName to something like reason[ing] or think.* would hit a runtime SyntaxError or match unintended content. The codebase already has escapeLiteralForRegex in core/util/regexValidator.ts; it should be applied when building the think-block regexes in core/util/index.ts and core/autocomplete/postprocessing/index.ts (or the value should be escaped when assigned in BaseLLM).

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At core/config/types.ts, line 583:

<comment>The configurable `thinkTagName` is documented as an XML tag name, but downstream code interpolates it directly into `new RegExp()` without escaping regex metacharacters. A user setting `thinkTagName` to something like `reason[ing]` or `think.*` would hit a runtime `SyntaxError` or match unintended content. The codebase already has `escapeLiteralForRegex` in `core/util/regexValidator.ts`; it should be applied when building the think-block regexes in `core/util/index.ts` and `core/autocomplete/postprocessing/index.ts` (or the value should be escaped when assigned in `BaseLLM`).</comment>

<file context>
@@ -572,6 +572,15 @@ declare global {
+     * Set this to match your provider's reasoning output format.
+     * See: https://docs.vllm.ai/en/latest/features/reasoning_outputs.html
+     */
+    thinkTagName?: string;
   }
   
</file context>