Fix dispatcher: pipe via stdin, use real cost from CLI JSON

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

symbiont/dispatcher.py

@@ -113,39 +113,43 @@ def dispatch(
         "--print",              # non-interactive, just print the response
         "--model", model_flag,
         "--output-format", "json",
-        "--dangerously-skip-permissions",  # full access, no prompts
     ]
 
     if system_prompt:
         cmd.extend(["--system-prompt", system_prompt])
 
-    # Prompt is a positional argument in Claude Code CLI
-    cmd.append(prompt)
-
     start = time.monotonic()
 
     try:
+        # Pipe the prompt via stdin to avoid shell escaping issues
         result = subprocess.run(
             cmd,
+            input=prompt,
             capture_output=True,
             text=True,
             timeout=timeout_seconds,
         )
         elapsed = time.monotonic() - start
 
-        # Try to parse JSON output for token counts
+        # Parse the rich JSON output from Claude Code CLI
         output_text = result.stdout.strip()
         input_tokens = 0
         output_tokens = 0
+        actual_cost_usd = 0.0
 
         try:
             parsed = json.loads(output_text)
-            # Claude Code JSON output includes result and usage
             if isinstance(parsed, dict):
                 output_text = parsed.get("result", output_text)
+                # Claude Code gives us exact usage data
                 usage = parsed.get("usage", {})
                 input_tokens = usage.get("input_tokens", 0)
                 output_tokens = usage.get("output_tokens", 0)
+                # Real cost from the CLI (even on Pro, this tracks the equivalent)
+                actual_cost_usd = parsed.get("total_cost_usd", 0.0)
+                # Also capture cache stats for efficiency tracking
+                cache_read = usage.get("cache_read_input_tokens", 0)
+                cache_create = usage.get("cache_creation_input_tokens", 0)
         except json.JSONDecodeError:
             pass  # Plain text output, that's fine
 
@@ -171,12 +175,15 @@ def dispatch(
                 error=f"Exit code {result.returncode}: {stderr}",
             )
 
-        # Estimate cost for the ledger
-        costs = MODEL_COSTS[tier]
-        estimated_cost = (
-            (input_tokens / 1_000_000) * costs["input"]
-            + (output_tokens / 1_000_000) * costs["output"]
-        )
+        # Use real cost from CLI if available, otherwise estimate
+        if actual_cost_usd > 0:
+            estimated_cost = actual_cost_usd
+        else:
+            costs = MODEL_COSTS[tier]
+            estimated_cost = (
+                (input_tokens / 1_000_000) * costs["input"]
+                + (output_tokens / 1_000_000) * costs["output"]
+            )
 
         dispatch_result = DispatchResult(
             success=True,