Strix (STRIX) — agents
What this is
strix/agents is where a single AI agent gets built — its tool list assembled, its
system prompt rendered from a Jinja template and a resolved set of knowledge-pack
"skills." It doesn't run anything itself; think of it as the factory that produces the
agent object strix/core then drives.
What flows through it
Technical details
Package: strix/agents/
Depends on: strix.tools (imports every tool module to build _BASE_TOOLS),
strix.skills (skill markdown bodies via load_skills()), strix.config
(uses_chat_completions_tool_schema — decides tool schema style per model).
Depended on by: strix.core (runner.py calls build_strix_agent/make_child_factory
directly; execution.py uses the returned child factory closure).
Files at a glance
| File | Role |
|---|---|
factory.py |
build_strix_agent() — assembles one agents.sandbox.SandboxAgent: tool list, capabilities, lifecycle stop condition. make_child_factory() — returns a closure children are built from. |
prompt.py |
render_system_prompt() — Jinja2 rendering of system_prompt.jinja against a resolved skill set and scope context. |
prompts/system_prompt.jinja |
The actual prompt template: communication rules, authorization/scope framing, execution guidelines, and the root-only <root_agent_directive> block. |
Files & functions, explained
factory.py
- build_strix_agent(*, name, skills, is_root, scan_mode, is_whitebox, interactive, chat_completions_tools, system_prompt_context, extra_tools, instructions_override) — the main constructor. Assembles _BASE_TOOLS (think, load_skill, todo/notes CRUD, web_search, reporting tools, proxy read tools, agent-graph tools) + _EXTRA_TOOLS (pluggable, via register_agent_tools()) + call-specific extra_tools; appends respond_to_user only if interactive; appends the lifecycle tool (finish_scan for root, agent_finish otherwise).
- _with_coerced_arguments / _with_bounded_result — wrap every FunctionTool to repair malformed args and cap output size (delegates to strix.tools.output_store).
- _finish_tool_use_behavior — the SDK tool_use_behavior hook: only the four lifecycle tool names end an agent's Runner cycle.
- capabilities=[Filesystem(...), Shell(...)] — SDK-native sandbox capabilities; Strix's configurator closures adjust their tool schemas (e.g. renaming apply_patch's input field to patch).
- make_child_factory(*, scan_mode, is_whitebox, interactive, chat_completions_tools, system_prompt_context) — returns _factory(*, name, skills) -> SandboxAgent, so core.execution can build children without knowing run-level config internals.
prompt.py
- render_system_prompt(*, skills, scan_mode, is_whitebox, is_root, interactive, system_prompt_context) — resolves a skill list (_resolve_skills: caller-requested + always scan_modes/<mode> + always tooling/agent_browser + always tooling/python + coordination/root_agent if root + custom/source_aware_sast if whitebox), loads their markdown bodies, and renders the Jinja template. Root vs. child role is prompt-driven, not tool-driven — both get essentially the same tools; only the root's prompt includes the orchestration directive forbidding it from testing directly.
Related: STRIX — Architecture (HLD), core.md, tools.md, skills.md.