Fix CLI flags: remove --max-tokens, add --dangerously-skip-permissions

Claude Code CLI uses positional args for prompts, not --prompt flag.
Also enables full access mode so orchestrator runs without permission prompts.

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

symbiont/dispatcher.py

@@ -81,7 +81,6 @@ def dispatch(
     prompt: str,
     tier: ModelTier = ModelTier.SONNET,
     system_prompt: Optional[str] = None,
-    max_tokens: int = 4096,
     timeout_seconds: int = 120,
 ) -> DispatchResult:
     """
@@ -113,14 +112,15 @@ def dispatch(
         "claude",
         "--print",              # non-interactive, just print the response
         "--model", model_flag,
-        "--max-tokens", str(max_tokens),
         "--output-format", "json",
+        "--dangerously-skip-permissions",  # full access, no prompts
     ]
 
     if system_prompt:
         cmd.extend(["--system-prompt", system_prompt])
 
-    cmd.extend(["--prompt", prompt])
+    # Prompt is a positional argument in Claude Code CLI
+    cmd.append(prompt)
 
     start = time.monotonic()
 

symbiont/router.py

@@ -41,7 +41,6 @@ def classify_task(task_description: str) -> tuple[ModelTier, float, str]:
         prompt=f"Classify this task:\n\n{task_description}",
         tier=ModelTier.HAIKU,
         system_prompt=CLASSIFIER_SYSTEM_PROMPT,
-        max_tokens=256,
     )
 
     if result.success: