Support multiple inference engines and model providers via a strategy
enum so each receives its own thinking-disabling field in
chat-completion request bodies.
Strategies:
- "vllm" → chat_template_kwargs: { enable_thinking: false } (vLLM / SGLang)
- "deepseek" → enable_thinking: false (DeepSeek official API)
- "dashscope" → enable_thinking: false (Alibaba DashScope / Qwen)
- "openai" → reasoning_effort: "low" (OpenAI o-series, cannot fully disable)
- "anthropic" → thinking: { type: "disabled" } (Anthropic Claude)
- "kimi" → thinking: { type: "disabled" } (Kimi / Moonshot)
- "gemini" → thinking_config: { thinking_budget: 0 } (Google Gemini)
Defaults to false (no wrapper). Also accepts boolean true as a shorthand
for "vllm" for convenience.
Covered paths:
- StandaloneLLMRunner (OpenClaw plugin + gateway): llm.disableThinking,
wired through parseConfig, tdai-core, seed-runtime, and gateway config
(TDAI_LLM_DISABLE_THINKING env also accepts strategy names).
- Offload local mode (L1/L1.5/L2): separate offload.disableThinking.
The fetch wrapper lives in src/utils/no-think-fetch.ts with a
STRATEGY_TRANSFORMERS map for clean dispatch. StandaloneLLMRunner builds
it once in the constructor; LocalLlmClient caches it at construction
time and passes it through to callLlm().
Add vitest unit tests for all 7 strategies, normalization, validation,
embedding skip, and non-JSON tolerance (20 tests total).
Signed-off-by: jackson.jia <jiazhenghua0@gmail.com>
- Add `recall.maxCharsPerMemory` and `recall.maxTotalRecallChars` with defaults of `0`, which do not alter existing behavior. Users can opt in by setting positive values to cap injected memory context.
- Apply the budget after L1 search and before `<relevant-memories>` injection, preserving score order while truncating oversized entries and dropping overflow.
- Document the new guards in README, README_CN, and `openclaw.plugin.json`.
jackson-jia-914