perf: eliminate redundant LLM calls in agentic loop (#24)

Three optimizations to the agentic loop:
1. Cache loop agent as singleton (avoid repeated warmups)
2. Sliding window for step context (last 2 results, not all)
3. Replace summary LLM call with deterministic summary

Saves 1 full LLM inference call per agentic loop invocation
(30-60s on local models) and reduces context window pressure.

Also fixes pre-existing test_cli.py repl test bugs (missing result= assignment).

src/timmy/agentic_loop.py

@@ -58,6 +58,8 @@ class AgenticResult:
 # Agent factory
 # ---------------------------------------------------------------------------
 
+_loop_agent = None
+
 
 def _get_loop_agent():
     """Create a fresh agent for the agentic loop.
@@ -65,9 +67,12 @@ def _get_loop_agent():
     Returns the same type of agent as `create_timmy()` but with a
     dedicated session so it doesn't pollute the main chat history.
     """
-    from timmy.agent import create_timmy
+    global _loop_agent
+    if _loop_agent is None:
+        from timmy.agent import create_timmy
 
-    return create_timmy()
+        _loop_agent = create_timmy()
+    return _loop_agent
 
 
 # ---------------------------------------------------------------------------
@@ -168,11 +173,11 @@ async def run_agentic_loop(
     for i, step_desc in enumerate(steps, 1):
         step_start = time.monotonic()
 
+        recent = completed_results[-2:] if completed_results else []
         context = (
             f"Task: {task}\n"
-            f"Plan: {plan_text}\n"
-            f"Completed so far: {completed_results}\n\n"
-            f"Now do step {i}: {step_desc}\n"
+            f"Step {i}/{total_steps}: {step_desc}\n"
+            f"Recent progress: {recent}\n\n"
             f"Execute this step and report what you did."
         )
 
@@ -273,27 +278,15 @@ async def run_agentic_loop(
                 completed_results.append(f"Step {i}: FAILED")
 
     # ── Phase 3: Summary ───────────────────────────────────────────────────
-    summary_prompt = (
-        f"Task: {task}\n"
-        f"Results:\n" + "\n".join(completed_results) + "\n\n"
-        "Summarise what was accomplished in 2-3 sentences."
-    )
-    try:
-        summary_run = await asyncio.to_thread(
-            agent.run,
-            summary_prompt,
-            stream=False,
-            session_id=f"{session_id}_summary",
-        )
-        result.summary = (
-            summary_run.content if hasattr(summary_run, "content") else str(summary_run)
-        )
-        from timmy.session import _clean_response
-
-        result.summary = _clean_response(result.summary)
-    except Exception as exc:
-        logger.error("Agentic loop summary failed: %s", exc)
-        result.summary = f"Completed {len(result.steps)} steps."
+    completed_count = sum(1 for s in result.steps if s.status == "completed")
+    adapted_count = sum(1 for s in result.steps if s.status == "adapted")
+    failed_count = sum(1 for s in result.steps if s.status == "failed")
+    parts = [f"Completed {completed_count}/{total_steps} steps"]
+    if adapted_count:
+        parts.append(f"{adapted_count} adapted")
+    if failed_count:
+        parts.append(f"{failed_count} failed")
+    result.summary = f"{task}: {', '.join(parts)}."
 
     # Determine final status
     if was_truncated: