feat(code): pre-analysis optimization and MiniMax compatibility

- Add LlmPlanner::pre_analyze() for single-call intent, goal, plan, and
  input optimization (reduces LLM round-trips)
- Add ParallelStepResult structured JSON output for parallel plan steps
- Fix OpenAI client: fallback to total_characters when total_tokens=0
  (MiniMax API compatibility)
- Fix delta_content extraction for reasoning models (handle empty deltas)
- Add pre_analysis_system.md prompt for the new pre-analysis LLM call

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: hikejs

core/prompts/plan_parallel_results.md

@@ -1,4 +1,11 @@
-The following steps were executed in parallel:
-{results}
+The following plan steps completed in parallel. The results are provided as structured JSON in a USER message (not plain text) with this envelope:
 
-Continue with the next steps.
+```json
+{"type":"parallel_results","steps":[{"step_id":"...","step_number":1,"status":"completed","summary":"...","key_findings":[...],"error":null},...]}
+```
+
+- `summary`: Brief description of what this step accomplished (in the user's language).
+- `key_findings`: Optional array of notable findings or outputs.
+- `error`: Present only when `status` is `"failed"`.
+
+Use the structured step data above to continue planning and execution.

core/prompts/pre_analysis_system.md

@@ -0,0 +1,31 @@
+You are a pre-analysis assistant. Given a user message, perform all needed analysis in a single pass and return JSON. The JSON schema is:
+
+{
+  "intent": "GeneralPurpose|Plan|Explore|Verification|CodeReview",
+  "requires_planning": true|false,
+  "goal": {
+    "description": "what the user ultimately wants to achieve",
+    "success_criteria": ["criterion 1", "criterion 2"]
+  },
+  "execution_plan": {
+    "complexity": "Simple|Medium|Complex|VeryComplex",
+    "steps": [
+      {
+        "id": "step-1",
+        "description": "what to do in this step",
+        "tool": "bash|read|write|edit|grep|...",
+        "dependencies": [],
+        "success_criteria": "what counts as done for this step"
+      }
+    ],
+    "required_tools": ["tool_name"]
+  },
+  "optimized_input": "rewritten user message with ambiguous pronouns resolved, implied context made explicit, in the same language as the user message"
+}
+
+Rules:
+- intent: "Plan" only if the user explicitly asks for planning/architecture/design. "Explore" only for read-only search tasks. "Verification" only for adversarial testing. "CodeReview" only for code quality analysis. Default to "GeneralPurpose".
+- requires_planning: true if the task has multiple steps, could benefit from structured reasoning, or the user asks for a plan. false for simple single-step tasks or read-only lookups.
+- complexity: Simple (<50 chars, 1 step), Medium (50-150 chars, 2-4 steps), Complex (150-300 chars, 5-8 steps), VeryComplex (>300 chars or many steps)
+- optimized_input: fix grammatical issues, resolve "it", "that", "the file" references, make implicit assumptions explicit. Keep the same meaning and language.
+- Respond with JSON only, no markdown fences or explanation.