fix(prompt): dedupe parallel-tool-call steer; correct its rationale

The universal PARALLEL_TOOL_CALL_GUIDANCE block already lives on main, but it
shipped with two rough edges this change cleans up:

- It duplicated the batching steer for Google models. The
  GOOGLE_MODEL_OPERATIONAL_GUIDANCE block still carried its own
  "Parallel tool calls" bullet, so Gemini/Gemma received the instruction
  twice in one prompt. Drop the redundant bullet — the universal block is now
  the single source.
- Its comment claimed "nothing in the open-source system prompt encouraged
  batching," which was wrong: the steer existed for Google models only. Reword
  to say the gap was that every *other* model got nothing.
- Tighten the test that asserts the steer (precedence-correct), and add an
  invariant guarding against re-introducing the Google duplicate.

agent/prompt_builder.py

@@ -320,9 +320,11 @@ TASK_COMPLETION_GUIDANCE = (
 # concurrently when they are independent (read-only tools always; path-scoped
 # file ops when their targets don't overlap — see
 # run_agent._execute_tool_calls / tool_dispatch_helpers). The missing piece
-# was telling the *model* to emit those calls together in the first place;
-# nothing in the open-source system prompt encouraged batching. This block
-# closes that gap.
+# was telling the *model* to emit those calls together in the first place.
+# Until now the only batching steer in the prompt lived in
+# GOOGLE_MODEL_OPERATIONAL_GUIDANCE — Gemini/Gemma got it, every other model
+# got nothing. This block makes the steer universal; the now-redundant
+# Google-only bullet has been dropped so no model receives it twice.
 #
 # Short on purpose — shipped in the cached system prompt to every user, every
 # session. Token cost is paid once at install and amortised across all
@@ -425,9 +427,10 @@ GOOGLE_MODEL_OPERATIONAL_GUIDANCE = (
     "package.json, requirements.txt, Cargo.toml, etc. before importing.\n"
     "- **Conciseness:** Keep explanatory text brief — a few sentences, not "
     "paragraphs. Focus on actions and results over narration.\n"
-    "- **Parallel tool calls:** When you need to perform multiple independent "
-    "operations (e.g. reading several files), make all the tool calls in a "
-    "single response rather than sequentially.\n"
+    # Parallel-tool-call steering now lives in the universal
+    # PARALLEL_TOOL_CALL_GUIDANCE block (injected for all models), so it is no
+    # longer duplicated here — keeping it would send Gemini/Gemma the same
+    # instruction twice.
     "- **Non-interactive commands:** Use flags like -y, --yes, --non-interactive "
     "to prevent CLI tools from hanging on prompts.\n"
     "- **Keep going:** Work autonomously until the task is fully resolved. "

tests/agent/test_prompt_builder.py

@@ -28,6 +28,7 @@ from agent.prompt_builder import (
     TOOL_USE_ENFORCEMENT_MODELS,
     OPENAI_MODEL_EXECUTION_GUIDANCE,
     PARALLEL_TOOL_CALL_GUIDANCE,
+    GOOGLE_MODEL_OPERATIONAL_GUIDANCE,
     MEMORY_GUIDANCE,
     SESSION_SEARCH_GUIDANCE,
     PLATFORM_HINTS,
@@ -1512,8 +1513,9 @@ class TestParallelToolCallGuidance:
 
     def test_steers_batching_into_one_response(self):
         text = PARALLEL_TOOL_CALL_GUIDANCE.lower()
-        # Must tell the model to group independent calls together.
-        assert "single response" in text or "same" in text and "turn" in text
+        # Must tell the model to group independent calls together — accept any
+        # phrasing that means "one turn" without freezing exact wording.
+        assert "single response" in text or ("same" in text and "turn" in text)
         assert "independent" in text
 
     def test_carves_out_dependent_calls(self):
@@ -1533,6 +1535,11 @@ class TestParallelToolCallGuidance:
         # Heading delimits it as its own section in the assembled prompt.
         assert PARALLEL_TOOL_CALL_GUIDANCE.lstrip().startswith("#")
 
+    def test_not_duplicated_in_google_guidance(self):
+        # The universal block is now the single source of parallel-batching
+        # steer. The Google-only block must NOT carry its own copy, otherwise
+        # Gemini/Gemma would receive the instruction twice in one prompt.
+        assert "parallel tool call" not in GOOGLE_MODEL_OPERATIONAL_GUIDANCE.lower()
 
 
 # =========================================================================