feat: add support for thinking level 'max' (#283)

docs/content/docs/subagents.mdx

@@ -137,7 +137,7 @@ section of the Subagents card) or under `subagent_summary` in `config.yml`:
 | `enabled` | `true` | Off falls back to a crude first-line truncation (no LLM call) |
 | `provider` | `deepseek` | Inference provider for the summary |
 | `model` | `deepseek-v4-flash` | Fast + cheap is ideal for this distillation |
-| `thinking_level` | `""` | `""` (off) / `low` / `medium` / `high` for reasoning models |
+| `thinking_level` | `""` | `""` (off) / `low` / `medium` / `high` / `max` for reasoning models |
 
 Synchronous spawns are unaffected — the agent gets the full result inline as the
 tool result and weaves it into its own reply.

humux/api/templates/agent_editor.html

@@ -151,9 +151,10 @@ <h3 class="text-xs text-muted uppercase tracking-wider">Model &amp; inference</h
               <option value="low">Low</option>
               <option value="medium">Medium</option>
               <option value="high">High</option>
+              <option value="max">Max</option>
             </select>
             <p class="text-amber-600 text-xs mt-1"
-               x-show="['low','medium','high'].includes(llmThinking) && !window.llmModelSupportsThinking(llmModel || llmDefaults.model)">
+               x-show="['low','medium','high','max'].includes(llmThinking) && !window.llmModelSupportsThinking(llmModel || llmDefaults.model)">
               Heads up: “<span x-text="llmModel || llmDefaults.model"></span>” isn't a recognized reasoning model — only set a level if you know it supports thinking.
             </p>
           </div>

humux/api/templates/base.html

@@ -435,7 +435,7 @@
         levelOptions(prov) {
           const fetched = this.fetchedLevels[prov];
           if (fetched && fetched.length) return fetched;
-          return ['low', 'medium', 'high'];
+          return ['low', 'medium', 'high', 'max'];
         },
         fetchThinkingLevels(service, model) {
           const apiKey = this.keyFor(service);

humux/core/agents.py

@@ -373,7 +373,7 @@ def _as_group_chat(value: object) -> dict:
     }
 
 
-_LLM_THINKING_LEVELS = ("", "low", "medium", "high")
+_LLM_THINKING_LEVELS = ("", "low", "medium", "high", "max")
 
 
 def _as_llm_config(value: object) -> dict:

humux/core/config.py

@@ -93,7 +93,7 @@ class AgentConfig(BaseModel):
     openrouter_api_key: str = ""
     openrouter_base_url: str = ""
     model: str = "deepseek-v4-flash"
-    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" — only for reasoning models
+    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
     # Hard ceiling on tokens the model may emit per response. The agentic loop
     # truncates mid-tool-call when this is too small for the output (e.g. a large
     # file written via the write tool), so keep it generous. 16384 is safe across providers;
@@ -278,8 +278,8 @@ class MemoryConfig(BaseModel):
     extraction_model: str = "deepseek-v4-flash"
     consolidation_provider: str = "deepseek"
     consolidation_model: str = "deepseek-v4-flash"
-    extraction_thinking_level: str = ""  # "" (off) | "low" | "medium" | "high"
-    consolidation_thinking_level: str = ""  # "" (off) | "low" | "medium" | "high"
+    extraction_thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
+    consolidation_thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
     extraction_cooldown_seconds: int = 120  # minimum seconds between extractions
 
     embedding: EmbeddingConfig = EmbeddingConfig()
@@ -299,14 +299,14 @@ class GoalDecompositionConfig(BaseModel):
     enabled: bool = True
     provider: str = "deepseek"
     model: str = "deepseek-v4-flash"
-    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high"
+    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
 
 
 class TaskReflectionConfig(BaseModel):
     enabled: bool = True
     provider: str = "deepseek"
     model: str = "deepseek-v4-flash"
-    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high"
+    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
     db_path: str = "data/reflections.db"
     max_reflections: int = 12  # injected per turn; kept small to cut prompt bloat (#5, was 50)
 
@@ -323,7 +323,7 @@ class ReplyDecisionConfig(BaseModel):
     enabled: bool = False
     provider: str = "deepseek"
     model: str = "deepseek-v4-flash"  # fast + cheap is ideal for the yes/no reply call
-    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high"
+    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
     group_only: bool = True  # only gate group chats; DMs always get a reply
     # Hard backstop: never send more than this many auto-replies into one chat
     # per rolling window — guarantees a runaway loop terminates even if the LLM
@@ -342,7 +342,7 @@ class CompactionConfig(BaseModel):
     enabled: bool = True
     provider: str = "deepseek"
     model: str = "deepseek-v4-flash"
-    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high"
+    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
     threshold_type: str = "percent"  # "percent" (of context window) or "tokens" (absolute)
     threshold_percent: int = 80  # trigger at this % of the model's context window
     threshold_tokens: int = 150000  # absolute trigger when threshold_type == "tokens"
@@ -519,8 +519,7 @@ class SubagentSummaryConfig(BaseModel):
     enabled: bool = True
     provider: str = "deepseek"  # fast + cheap is ideal for this distillation
     model: str = "deepseek-v4-flash"
-    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high"
-
+    thinking_level: str = ""  # "" (off) | "low" | "medium" | "high" | "max"
 
 class Config(BaseModel):
     agent: AgentConfig = AgentConfig()

humux/core/llm.py

@@ -297,7 +297,7 @@ def _reasoning_kwargs(self) -> dict[str, Any]:
         Empty when no level is set, so non-reasoning calls are untouched.
         """
         level = self.thinking_level
-        if level not in ("low", "medium", "high"):
+        if level not in ("low", "medium", "high", "max"):
             return {}
         if self.provider == "anthropic":
             return {"thinking": {"type": "adaptive"}, "output_config": {"effort": level}}