Implements the full tool-calling path for `agents actor run --skill` so that LLM
actors can actually invoke tools through the ToolCallingRuntime loop when a skill
is attached.
Core changes:
- reactive/tool_caller.py (new): ToolCallingLLMCaller implements the LLMCaller protocol;
binds tool schemas via bind_tools(), threads SystemMessage+HumanMessage on first call
and AIMessage+ToolMessages on subsequent calls, extracts tool calls from LangChain
responses following the LangChainSessionCaller pattern.
- reactive/tool_caller.py: bidirectional tool name encoding via uppercase sentinels
(_encode_tool_name / _decode_tool_name) to make CleverAgents namespaced tool names
("builtin/file-read", "server:local/tool") compatible with Anthropic's tool name
pattern ^[a-zA-Z0-9_-]{1,128}$. Uses _C_ for ":" and _S_ for "/" — uppercase
sentinels are safe because valid CleverAgents tool names forbid uppercase letters.
Encoding applied in _resolve_llm() before bind_tools(); decoding applied in invoke()
when extracting tool calls from the LLM response.
- reactive/tool_agent.py (new): ToolCallingAgent builds a per-run local ToolRegistry from
resolved skill tool entries by looking up names in the shared builtin_registry; drives
ToolCallingRuntime.run_tool_loop(); exposes last_result for tool_calls surfacing.
- reactive/application.py: (ST-1) _make_agent_instance() now always merges skill tools
instead of silently dropping them when actor has no base tools list; routes
tools+llm→ToolCallingAgent, tools+non-llm→SimpleToolAgent, no-tools+llm→SimpleLLMAgent;
(ST-4) _builtin_registry created at startup with register_file_tools/git/subplan;
(ST-6) _tally_tool_calls() + last_run_tool_calls property.
- reactive/graph_executor.py: (ST-5) ToolCallingAgent added to isinstance check in
_invoke_agent() so context dict is forwarded for Jinja2 rendering.
- cli/commands/actor_run.py: prints "Tool Calls: {n}" when > 0.
Test fixes:
- features/steps/actor_cli_run_steps.py: _make_app() sets last_run_tool_calls=0 to avoid
MagicMock>int TypeError in Python 3.13.
- features/steps/actor_run_signature_resolve_steps.py: same fix.
- robot/helper_actor_run_signature.py: same fix.
- features/reactive_application_coverage_boost.feature: updated scenario to verify new
correct behavior (LLM+skills → ToolCallingAgent, not silently kept as SimpleLLMAgent).
BDD coverage: 34 scenarios in features/actor_run_tool_calling.feature covering
tool call success, multi-turn loop, no-skill regression, silent-drop fix, LLMCaller
internals, _build_tool_registry edge cases, last_run_tool_calls tallying,
tool name encoding/decoding, and LLM response decoding.
ISSUES CLOSED: #11211
