Files
strix/strix/agents/factory.py
T
0xallam 33e5e61b0c feat(orchestration): full parity with legacy harness — 8 gaps closed via SDK natives
Audit found 8 behavioral gaps between post-migration and the legacy
``BaseAgent.agent_loop``. All 8 are now closed using SDK-native
primitives — no custom workarounds, no shadow state machines.

What was broken / different:

- G1: ``inherit_context`` was dead code; children always started fresh.
- G2: TUI user message couldn't interrupt an in-flight LLM/tool turn.
- G3: ``llm_failed`` state never set; hard failures propagated as crashes.
- G4: No graceful ``stop_agent`` tool.
- G5: Parked subagents waited forever (no auto-resume timeout).
- G6: Inter-agent messages used a plain header instead of legacy XML.
- G7: Completion reports used JSON instead of legacy XML.
- G11/G12: Turn counter reset per cycle; budget warnings could re-fire.

What we did:

Bus extensions (``orchestration/bus.py``):
- ``streams`` registry + ``attach_stream`` ctx manager + ``request_interrupt``
  for SDK-native ``RunResultStreaming.cancel(mode="after_turn")``.
- ``mark_llm_failed`` + ``wait_for_user_message`` (filtered: only ``from="user"``
  satisfies; peer messages don't unstick a stuck model).
- ``stopping: set[str]`` for graceful programmatic exit.
- ``cancel_descendants_graceful`` — leaves-first via ``request_interrupt``.
- ``record_usage`` increments ``calls`` unconditionally so it doubles as the
  per-agent-lifetime turn counter (legacy ``state.iteration`` parity).
- ``warned_85`` / ``warned_final`` flags on ``stats_live`` for once-fire
  budget warnings.

Run loop rewrite (``orchestration/run_loop.py``):
- ``Runner.run`` → ``Runner.run_streamed`` with ``bus.attach_stream`` so
  cancel has a target. Catch ``(AgentsException, APIError)`` after retries
  exhaust; in interactive mode call ``mark_llm_failed`` + wait for user.
- ``UserError`` / ``MaxTurnsExceeded`` / ``CancelledError`` propagate.
- Outer loop: ``asyncio.wait_for(bus.wait_for_message, timeout=300)`` for
  interactive subagents (root waits forever). ``TimeoutError`` injects
  ``"Waiting timeout reached. Resuming execution."``.
- Honors ``bus.stopping`` at top of each iteration.

Hooks (``orchestration/hooks.py``):
- Counter source moved from per-cycle ``ctx["turn_count"]`` to
  per-lifetime ``bus.stats_live[agent_id]["calls"]``.
- Warnings guarded by once-flags — exactly-once across all cycles.

Filter (``orchestration/filter.py``):
- Restored legacy ``<inter_agent_message>`` XML envelope with the
  ``<delivery_notice>DO NOT echo back</delivery_notice>`` instruction.

Agents-graph (``tools/agents_graph/tools.py``):
- G1: ``create_agent`` reads ``ctx.turn_input`` (SDK populates it before
  tool execution at ``run_internal/turn_resolution.py:806``). Wraps as
  one ``<inherited_context_from_parent>`` block.
- G7: ``agent_finish`` emits the legacy ``<agent_completion_report>``
  XML. ``child_ctx["task"] = task`` threaded so the report echoes the
  original task.
- G4: New ``stop_agent`` tool — refuses self-stop, refuses already-
  finalized targets, ``cascade=True`` uses ``cancel_descendants_graceful``.

TUI (``interface/tui.py``):
- ``_send_user_message`` schedules ``bus.send`` AND
  ``bus.request_interrupt(target, mode="after_turn")`` — SDK finishes
  current turn cleanly, next cycle picks up the user's message.

Factory (``agents/factory.py``):
- Registered ``stop_agent`` in ``_BASE_TOOLS``.

Out of scope:
- G8 (``[ABORTED BY USER]`` marker) is auto-resolved by G2 — the SDK
  saves the full assistant message before honoring
  ``cancel(mode="after_turn")``, so partial content is preserved in the
  session.

Verified all bus behaviors with a smoke test. Lint at baseline.
2026-04-25 17:30:29 -07:00

206 lines
6.4 KiB
Python

"""``build_strix_agent`` — assemble an ``agents.Agent`` for root or child.
Wires the SDK function tools, multi-agent graph tools, and the rendered
Jinja prompt into one ``agents.Agent`` ready for ``Runner.run``.
Two flavors:
- **Root** (``is_root=True``): top-level scan agent. Carries
``finish_scan`` and stops there.
- **Child** (``is_root=False``): subagents spawned by the
``create_agent`` graph tool. Carries ``agent_finish`` and stops
there — without ``stop_at_tool_names`` the SDK loop would keep
running to ``max_turns`` even after the child reported back.
Skills are baked into the system prompt at scan bring-up; there's no
runtime skill-loading tool.
"""
from __future__ import annotations
import logging
from typing import Any
from agents.agent import StopAtTools
from agents.sandbox import SandboxAgent
from agents.sandbox.capabilities import Filesystem, Shell
from agents.tool import Tool
from strix.agents.prompt import render_system_prompt
from strix.tools.agents_graph.tools import (
agent_finish,
agent_status,
create_agent,
send_message_to_agent,
stop_agent,
view_agent_graph,
wait_for_message,
)
from strix.tools.finish.tool import finish_scan
from strix.tools.notes.tools import (
create_note,
delete_note,
get_note,
list_notes,
update_note,
)
from strix.tools.proxy.tools import (
list_requests,
repeat_request,
scope_rules,
send_request,
view_request,
)
from strix.tools.reporting.tool import create_vulnerability_report
from strix.tools.thinking.tool import think
from strix.tools.todo.tools import (
create_todo,
delete_todo,
list_todos,
mark_todo_done,
mark_todo_pending,
update_todo,
)
from strix.tools.web_search.tool import web_search
logger = logging.getLogger(__name__)
# Host-side Strix tools. Sandbox shell + filesystem are added per-run
# by the SDK via the ``Shell`` and ``Filesystem`` capabilities below
# (they bind to the live sandbox session and emit ``exec_command`` /
# ``write_stdin`` / ``apply_patch`` / ``view_image`` function tools).
_BASE_TOOLS: tuple[Tool, ...] = (
# Thinking + planning
think,
# Per-agent todos
create_todo,
list_todos,
update_todo,
mark_todo_done,
mark_todo_pending,
delete_todo,
# Shared notes (per-run JSONL store)
create_note,
list_notes,
get_note,
update_note,
delete_note,
# Web search (only registered if PERPLEXITY_API_KEY is set; the
# tool itself returns a structured error when not configured, so
# always exposing it is safe)
web_search,
# Reporting
create_vulnerability_report,
# Caido HTTP/HTTPS proxy
list_requests,
view_request,
send_request,
repeat_request,
scope_rules,
# Multi-agent graph tools (the bus is in ctx.context)
view_agent_graph,
agent_status,
send_message_to_agent,
wait_for_message,
create_agent,
stop_agent,
)
def build_strix_agent(
*,
name: str = "strix",
skills: list[str] | None = None,
is_root: bool,
scan_mode: str = "deep",
is_whitebox: bool = False,
interactive: bool = False,
system_prompt_context: dict[str, Any] | None = None,
) -> SandboxAgent[Any]:
"""Build a ``SandboxAgent`` configured for either root or child use.
The ``Shell`` and ``Filesystem`` capabilities are added unbound; the
SDK's runtime binds them per-run against the live sandbox session
set on ``RunConfig.sandbox`` and merges their tools (``exec_command``,
``write_stdin``, ``apply_patch``, ``view_image``) into the agent's
final tool list. We deliberately exclude ``Compaction`` (OpenAI
Responses API only).
Args:
name: Agent name. Surfaces in traces and the bus's ``names`` map.
Defaults to ``"strix"`` for the root; create_agent passes
distinct names per child.
skills: Skills to preload into the system prompt.
is_root: Selects the tool list and ``tool_use_behavior``.
Root carries ``finish_scan`` and stops there; child carries
``agent_finish`` and stops there.
scan_mode: ``"deep"`` etc.; routes the scan-mode skill section
of the prompt template.
is_whitebox: Whitebox source-aware mode toggle. Adds two extra
skills to the prompt and gates whitebox-only behavior in
the create_agent / wiki integration.
interactive: Renders the interactive-mode communication block
in the system prompt.
system_prompt_context: Free-form dict the prompt template
renders into the ``system_prompt_context`` variable —
today carries the scan scope / authorization block.
"""
instructions = render_system_prompt(
skills=skills,
scan_mode=scan_mode,
is_whitebox=is_whitebox,
interactive=interactive,
system_prompt_context=system_prompt_context,
)
if is_root:
tools: list[Tool] = [*_BASE_TOOLS, finish_scan]
stop_at = ("finish_scan",)
else:
tools = [*_BASE_TOOLS, agent_finish]
stop_at = ("agent_finish",)
return SandboxAgent(
name=name,
instructions=instructions,
tools=tools,
tool_use_behavior=StopAtTools(stop_at_tool_names=list(stop_at)),
# model=None so ``RunConfig.model`` drives provider selection
# via :func:`build_multi_provider` rather than the SDK's default.
model=None,
capabilities=[Filesystem(), Shell()],
)
def make_child_factory(
*,
scan_mode: str = "deep",
is_whitebox: bool = False,
interactive: bool = False,
system_prompt_context: dict[str, Any] | None = None,
) -> Any:
"""Return a callable suitable for ``ctx.context['agent_factory']``.
The ``create_agent`` graph tool reads
``ctx.context['agent_factory']`` and calls it with ``name=`` and
``skills=`` to build a child ``Agent``. Run-level arguments
(``scan_mode``, ``is_whitebox``, etc.) are captured in a closure so
each child inherits the scan-level configuration without
``create_agent`` having to know about them.
"""
def _factory(*, name: str, skills: list[str]) -> SandboxAgent[Any]:
return build_strix_agent(
name=name,
skills=skills,
is_root=False,
scan_mode=scan_mode,
is_whitebox=is_whitebox,
interactive=interactive,
system_prompt_context=system_prompt_context,
)
return _factory