Port from cline/cline#11514: encourage parallel tool calls

Add a universal system-prompt guidance block telling the model to batch
independent tool calls (reads, searches, web fetches, read-only commands)
into a single assistant turn instead of one call per turn. The runtime
already executes independent batches concurrently (read-only tools always;
non-overlapping path-scoped file ops); the open-source system prompt had
nothing steering the model to PRODUCE the batch. Fewer round-trips means
less resent context, which compounds over a long conversation.

- prompt_builder.py: new PARALLEL_TOOL_CALL_GUIDANCE block (short, static,
  cache-amortised) modeled on TASK_COMPLETION_GUIDANCE.
- system_prompt.py: inject right after the task-completion block, gated by
  agent.valid_tool_names + the new toggle.
- agent_init.py: read agent.parallel_tool_call_guidance (default True).
- config.py: add the default under the agent section.
- test_prompt_builder.py: behavior-contract tests (batching steer, dependent
  carve-out, length bound) — invariants, not wording snapshots.

Adapted from Cline's TypeScript tool-surface guidance to hermes-agent's
Python prompt-assembly architecture and config-over-env conventions.

agent/agent_init.py

@@ -1224,6 +1224,12 @@ def init_agent(
     # targets.
     agent._task_completion_guidance = bool(_agent_section.get("task_completion_guidance", True))
 
+    # Universal parallel-tool-call guidance toggle.  Default True.  Separate
+    # flag from task_completion_guidance because a user may want one but not
+    # the other.  Steers the model to batch independent tool calls into a
+    # single turn; the runtime already executes such batches concurrently.
+    agent._parallel_tool_call_guidance = bool(_agent_section.get("parallel_tool_call_guidance", True))
+
     # Local Python toolchain probe toggle.  Default True.  When False,
     # the probe is skipped entirely (no subprocess calls, no system-prompt
     # line).  Useful for users on exotic setups where the probe heuristics

agent/prompt_builder.py

@@ -305,6 +305,45 @@ TASK_COMPLETION_GUIDANCE = (
     "is always better than inventing a result."
 )
 
+# Universal parallel-tool-call guidance — applied to ALL models.
+#
+# Why this matters for cost: every assistant turn resends the entire
+# accumulated conversation (and, on cache-friendly providers, re-reads the
+# cached prefix and pays for the newly-appended turn). A model that issues
+# one tool call per turn multiplies the number of round-trips — and therefore
+# the resent context — for any task that needs several independent reads,
+# searches, or safe lookups. Batching independent calls into a single
+# assistant response collapses N turns into one, cutting both latency and the
+# resent-context cost that compounds over a long conversation.
+#
+# The hermes-agent runtime already executes a batch of tool calls
+# 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.
+#
+# 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
+# sessions via prefix caching. Keep it tight.
+#
+# Ported from cline/cline#11514 ("encourage parallel tool calls"), adapted
+# from Cline's TypeScript tool-surface guidance to hermes-agent's Python
+# prompt-assembly architecture.
+PARALLEL_TOOL_CALL_GUIDANCE = (
+    "# Parallel tool calls\n"
+    "When you need several pieces of information that don't depend on each "
+    "other, request them together in a single response instead of one tool "
+    "call per turn. Independent reads, searches, web fetches, and read-only "
+    "commands should be batched into the same assistant turn — the runtime "
+    "executes independent calls concurrently, and batching avoids resending "
+    "the whole conversation on every extra round-trip.\n"
+    "Only serialize calls when a later call genuinely depends on an earlier "
+    "call's result (e.g. you must read a file before you can patch it). When "
+    "in doubt and the calls are independent, batch them."
+)
+
 # OpenAI GPT/Codex-specific execution guidance.  Addresses known failure modes
 # where GPT models abandon work on partial results, skip prerequisite lookups,
 # hallucinate instead of using tools, and declare "done" without verification.

agent/system_prompt.py

@@ -33,6 +33,7 @@ from agent.prompt_builder import (
     KANBAN_GUIDANCE,
     MEMORY_GUIDANCE,
     OPENAI_MODEL_EXECUTION_GUIDANCE,
+    PARALLEL_TOOL_CALL_GUIDANCE,
     PLATFORM_HINTS,
     SESSION_SEARCH_GUIDANCE,
     SKILLS_GUIDANCE,
@@ -123,6 +124,17 @@ def build_system_prompt_parts(agent: Any, system_message: Optional[str] = None)
     if getattr(agent, "_task_completion_guidance", True) and agent.valid_tool_names:
         stable_parts.append(TASK_COMPLETION_GUIDANCE)
 
+    # Universal parallel-tool-call guidance.  Tells the model to batch
+    # independent tool calls into one assistant turn rather than emitting one
+    # call per turn — the runtime already runs independent calls concurrently
+    # (read-only tools always; non-overlapping path-scoped file ops), so the
+    # only thing missing was steering the model to produce the batch.  Cuts
+    # round-trips and the resent-context cost that compounds over a long
+    # conversation.  Gated by config.yaml ``agent.parallel_tool_call_guidance``
+    # (default True) and only injected when tools are actually loaded.
+    if getattr(agent, "_parallel_tool_call_guidance", True) and agent.valid_tool_names:
+        stable_parts.append(PARALLEL_TOOL_CALL_GUIDANCE)
+
     # Tool-aware behavioral guidance: only inject when the tools are loaded
     tool_guidance = []
     if "memory" in agent.valid_tool_names:

hermes_cli/config.py

@@ -853,6 +853,15 @@ DEFAULT_CONFIG = {
         # plausible-looking output when a real path is blocked.  Costs ~80
         # tokens in the cached system prompt.  Set False to disable globally.
         "task_completion_guidance": True,
+        # Universal parallel-tool-call guidance — short prompt block applied to
+        # all models that tells the model to batch independent tool calls
+        # (reads, searches, web fetches, read-only commands) into one turn
+        # instead of one call per turn.  The runtime already runs independent
+        # calls concurrently, so this just steers the model to produce the
+        # batch — cutting round-trips and the resent-context cost that
+        # compounds over a long conversation.  Costs ~70 tokens in the cached
+        # system prompt.  Set False to disable globally.
+        "parallel_tool_call_guidance": True,
         # Local-environment toolchain probe — surfaces Python/pip/uv/PEP-668
         # state in the system prompt when something non-default is detected
         # (e.g. python3 has no pip module, pip→python version mismatch, PEP

tests/agent/test_prompt_builder.py

@@ -27,6 +27,7 @@ from agent.prompt_builder import (
     TOOL_USE_ENFORCEMENT_GUIDANCE,
     TOOL_USE_ENFORCEMENT_MODELS,
     OPENAI_MODEL_EXECUTION_GUIDANCE,
+    PARALLEL_TOOL_CALL_GUIDANCE,
     MEMORY_GUIDANCE,
     SESSION_SEARCH_GUIDANCE,
     PLATFORM_HINTS,
@@ -1497,6 +1498,43 @@ class TestOpenAIModelExecutionGuidance:
         assert len(OPENAI_MODEL_EXECUTION_GUIDANCE) > 100
 
 
+class TestParallelToolCallGuidance:
+    """Behavior contracts for the universal parallel-tool-call guidance block.
+
+    Asserts the invariants the block must satisfy (steer batching, scope to
+    independent calls, stay short for the cached prompt) rather than freezing
+    its exact wording.
+    """
+
+    def test_is_nonempty_string(self):
+        assert isinstance(PARALLEL_TOOL_CALL_GUIDANCE, str)
+        assert PARALLEL_TOOL_CALL_GUIDANCE.strip()
+
+    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
+        assert "independent" in text
+
+    def test_carves_out_dependent_calls(self):
+        # Must NOT tell the model to batch dependent calls — that would break
+        # ordering (read-before-patch). The block has to acknowledge the
+        # serialize-when-dependent case.
+        text = PARALLEL_TOOL_CALL_GUIDANCE.lower()
+        assert "depend" in text
+
+    def test_stays_short_for_cached_prompt(self):
+        # Shipped in every cached system prompt — keep it tight. The existing
+        # task-completion block is ~600 chars; allow generous headroom but
+        # guard against accidental essay growth.
+        assert len(PARALLEL_TOOL_CALL_GUIDANCE) < 900
+
+    def test_has_a_heading(self):
+        # Heading delimits it as its own section in the assembled prompt.
+        assert PARALLEL_TOOL_CALL_GUIDANCE.lstrip().startswith("#")
+
+
+
 # =========================================================================
 # Budget warning history stripping
 # =========================================================================