Codex/raise cb and optimize coding image

.gitignore

@@ -44,3 +44,6 @@ platform/htmlcov/
 platform/coverage.xml
 platform/**/*.cover
 platform/**/*.py,cover
+
+# Local git worktrees
+.worktrees/

platform/.env.example

@@ -10,6 +10,7 @@ LLM_API_KEY=your-api-key-here
 LLM_BASE_URL=https://api.openai.com
 LLM_MODEL=gpt-4o-mini
 LLM_TIMEOUT=120.0
+LLM_ROLE_MODEL_MAP={"orchestrator":"gpt-4o-mini","test":"gpt-4o-mini"}
 EXTRA_SKILL_DIR_WHITELIST=
 
 # SkillKit compatibility env vars.

platform/app/api/v1/tasks.py

@@ -92,9 +92,17 @@
     task = await service.cancel_task(task_id)
     if task is None:
         raise HTTPException(status_code=404, detail="Task not found")
     return task
 
 
+@router.post("/{task_id}/clone", response_model=TaskDetailResponse, status_code=201)
+async def clone_task(task_id: str, service: TaskService = Depends(get_task_service)):
+    task = await service.clone_task(task_id)
+    if task is None:
+        raise HTTPException(status_code=404, detail="Task not found")
+    return task
+
+
 @router.post("/{task_id}/retry", response_model=TaskDetailResponse)
 async def retry_task(task_id: str, service: TaskService = Depends(get_task_service)):
     task = await service.retry_task(task_id)

platform/app/config.py

@@ -16,8 +16,9 @@ class Settings(BaseSettings):
     LLM_TIMEOUT: float = 120.0
 
     # Per-role model routing (JSON string: {"coding": "gpt-4o", "review": "claude-sonnet-4-20250514"})
-    # Unspecified roles fall back to LLM_MODEL
-    LLM_ROLE_MODEL_MAP: str = "{}"
+    # Unspecified roles fall back to LLM_MODEL. Keep lightweight defaults on
+    # orchestrator/test so parse + signoff stay cheaper unless env overrides them.
+    LLM_ROLE_MODEL_MAP: str = '{"orchestrator":"gpt-4o-mini","test":"gpt-4o-mini"}'
     # Comma-separated absolute path prefixes allowed in agent config `extra_skill_dirs`.
     # Empty means only built-in platform/skills directory is allowed.
     EXTRA_SKILL_DIR_WHITELIST: str = ""
@@ -38,9 +39,12 @@ class Settings(BaseSettings):
     DB_POOL_TIMEOUT: int = 30
 
     # Circuit breaker configuration
-    CB_MAX_TOKENS_PER_TASK: int = 200000
-    CB_MAX_COST_PER_TASK_RMB: float = 50.0
+    CB_MAX_TOKENS_PER_TASK: int = 400000
+    CB_MAX_COST_PER_TASK_RMB: float = 100.0
     CB_TOKEN_PRICE_PER_1K: float = 0.01
+    # Per-stage token budgets (JSON string). Empty means disabled.
+    # Example: {"parse": 60000, "code": 250000}
+    CB_STAGE_TOKEN_BUDGETS: str = '{"parse": 60000, "code": 250000}'
 
     # Webhook secrets (empty = skip verification)
     JIRA_WEBHOOK_SECRET: str = ""
@@ -96,6 +100,12 @@ class Settings(BaseSettings):
     SANDBOX_ROLES: str = '["coding", "test"]'
     SANDBOX_DUMP_MODEL_API_RESPONSE: bool = True
     SANDBOX_MODEL_API_RAW_LOG_HOST_DIR: str = "/tmp/silicon_agent/model_api_logs"
+    SANDBOX_GRADLE_CMD_TIMEOUT_SECONDS: int = 480
+    SANDBOX_GRADLE_CACHE_HOST_DIR: str = "/var/lib/silicon_agent/gradle-cache"
+    SANDBOX_GRADLE_USER_HOME: str = "/var/lib/silicon_agent/gradle-cache"
+    SANDBOX_DEFAULT_JAVA_VERSION: int = 8
+    SANDBOX_GRADLE_WRAPPER_PREWARM: bool = True
+    SANDBOX_GRADLE_WRAPPER_PREWARM_TIMEOUT_SECONDS: int = 180
 
     # Memory & compression configuration
     MEMORY_ENABLED: bool = True

platform/app/services/agent_service.py

@@ -154,11 +154,11 @@
         for role, _ in AGENT_ROLES:
             model = (
                 role_model_map.get(role)
-                or FALLBACK_ROLE_DEFAULT_MODELS.get(role)
                 or settings.LLM_MODEL
+                or FALLBACK_ROLE_DEFAULT_MODELS.get(role)
             )
             if model not in available_models and available_models:
                 if settings.LLM_MODEL in available_models:
                     model = settings.LLM_MODEL
                 else:
                     model = available_models[0]

platform/app/services/project_service.py

@@ -7,6 +7,7 @@
 from sqlalchemy import func, or_, select
 from sqlalchemy.ext.asyncio import AsyncSession
 
+from app.config import settings
 from app.models.project import ProjectModel
 from app.schemas.project import (
     ProjectCreateRequest,
@@ -75,7 +76,7 @@ async def create_project(self, request: ProjectCreateRequest) -> ProjectResponse
             repo_local_path=request.repo_local_path,
             branch=request.branch,
             description=request.description,
-            sandbox_image=request.sandbox_image,
+            sandbox_image=request.sandbox_image or settings.SANDBOX_IMAGE,
         )
         self.session.add(project)
         await self.session.commit()

platform/app/services/task_service.py

@@ -129,6 +129,24 @@ async def create_task(self, request: TaskCreateRequest) -> TaskDetailResponse:
         task = result.scalar_one()
         return self._task_to_response(task)
 
+    async def clone_task(self, task_id: str) -> Optional[TaskDetailResponse]:
+        """Create a new task by copying only safe creation fields from a source task."""
+        source_task = await self._load_task_with_relations_optional(task_id)
+        if source_task is None:
+            return None
+
+        return await self.create_task(
+            TaskCreateRequest(
+                jira_id=source_task.jira_id,
+                title=source_task.title,
+                description=source_task.description,
+                template_id=source_task.template_id,
+                project_id=source_task.project_id,
+                yunxiao_task_id=source_task.yunxiao_task_id,
+                github_issue_number=getattr(source_task, "github_issue_number", None),
+            )
+        )
+
     async def get_task(self, task_id: str) -> Optional[TaskDetailResponse]:
         result = await self.session.execute(
             select(TaskModel)

platform/app/worker/agents.py

@@ -40,8 +40,8 @@
 ROLE_TOOLS: dict[str, set[str]] = {
     "orchestrator": {"read", "execute", "skill"},
     "spec":         {"read", "write", "edit", "skill"},
-    "coding":       {"read", "write", "edit", "execute", "execute_script", "skill"},
-    "test":         {"read", "write", "edit", "execute", "execute_script", "skill"},
+    "coding":       {"read", "write", "edit", "execute", "execute_script"},
+    "test":         {"read", "write", "edit", "execute", "execute_script"},
     "review":       {"read", "execute", "skill"},
     "smoke":        {"read", "execute", "skill"},
     "doc":          {"read", "write", "edit", "skill"},
@@ -57,8 +57,8 @@
 _ROLE_SKILL_DIRS: dict[str, list[str]] = {
     "orchestrator": ["shared", "orchestrator"],
     "spec":         ["shared", "spec"],
-    "coding":       ["shared", "coding"],
-    "test":         ["shared", "test"],
+    "coding":       [],
+    "test":         [],
     "review":       ["shared", "review"],
     "smoke":        ["shared", "smoke"],
     "doc":          ["shared", "doc"],

platform/app/worker/compressor.py

@@ -17,7 +17,7 @@
 # Fallback truncation limits when LLM is unavailable
 _L0_FALLBACK_CHARS = 200
 _L1_FALLBACK_CHARS = 1500
-_L2_MAX_CHARS = 20_000  # Hard cap on full-text prior output to prevent token explosion
+_L2_MAX_CHARS = 4_000  # Hard cap on full-text prior output to prevent token explosion
 
 
 @dataclass