refactor: scrub migration scars, dead code, and unused helpers

- Strip PLAYBOOK / AUDIT / Phase-N / C-numbered references from
  module docstrings across 16 files; rename
  ``_PHASE1_PARALLEL_DEFAULT`` → ``_PARALLEL_TOOL_CALLS_DEFAULT``.
- Delete unused exception classes: ``SandboxInitializationError``,
  ``ImplementedInClientSideOnlyError``.
- Delete the no-op ``on_handoff`` hook (we don't use SDK handoffs).
- Delete the unreachable backward-compat tab-delimited fallback in
  ``_parse_git_diff_output``.
- Delete orphaned ``strix/tools/load_skill/`` (dir contained only a
  pycache) and stale pycache files.
- Rewrite ``strix/skills/__init__.py``: 168 → 56 LoC. Drop seven
  helper functions (``get_available_skills``, ``get_all_skill_names``,
  ``validate_skill_names``, ``parse_skill_list``,
  ``validate_requested_skills``, ``generate_skills_description``,
  ``_get_all_categories``) — none had external callers; only
  ``load_skills`` is used.
- Drop the stale ``strix/agents/sdk_factory.py`` per-file ruff ignore
  (file no longer exists).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pyproject.toml

@@ -299,38 +299,31 @@ ignore = [
 # CaidoCapability uses agents.tool.Tool at runtime — pydantic Field
 # annotations and the cached _CAIDO_TOOLS tuple need it eagerly.
 "strix/sandbox/caido_capability.py" = ["TC002"]
-# Agent factory: agents.tool.Tool is used at runtime in the _BASE_TOOLS
-# tuple type, not just for annotations.
-"strix/agents/sdk_factory.py" = ["TC002"]
 # Entry point: ``Path`` is used at runtime by the typing of the
 # session_manager call; importing under TYPE_CHECKING would defer
 # resolution past where mypy needs it. ``_build_root_task`` legitimately
 # walks every supported target type — splitting it into per-type
 # helpers would add indirection without simplifying anything.
 "strix/entry.py" = ["TC003", "PLR0912"]
-# Legacy tracer module — pre-existing PLR/E501 patterns; full refactor
+# Tracer carries a long event surface and a runtime ``Callable``
-# is out of scope for the harness migration. ``Callable`` is a runtime
 # annotation on ``vulnerability_found_callback``.
 "strix/telemetry/tracer.py" = ["TC003", "PLR0912", "PLR0915", "E501"]
-# Legacy interface utility with intentionally many branches per supported
+# Interface utility branches per scope-mode / target-type combination;
-# scope-mode / target-type combination; refactor would obscure the
+# splitting would obscure the decision tree without simplifying it.
-# decision tree without simplifying it.
 "strix/interface/utils.py" = ["PLR0912", "BLE001", "PLC0415"]
-# CLI / TUI / main keep extensive lazy imports + broad exception swallows
+# CLI / TUI / main keep extensive lazy imports + broad exception
-# for resilience around terminal-rendering errors. Refactor is out of
+# swallows for resilience around terminal-rendering errors.
-# scope for the harness migration.
 "strix/interface/cli.py" = ["BLE001", "PLC0415"]
 "strix/interface/tui.py" = ["BLE001", "PLC0415", "PLR0912", "PLR0915", "SIM105"]
 "strix/interface/main.py" = ["BLE001", "PLC0415", "PLR0912", "PLR0915"]
 "strix/interface/streaming_parser.py" = ["PLC0415"]
 "strix/interface/tool_components/agent_message_renderer.py" = ["PLC0415"]
-# Sandbox dispatch helper has many short-circuit error returns (auth fail,
+# Each short-circuit error return in the sandbox dispatch is a distinct,
-# size cap, decode fail, etc). Each is a distinct, documented failure mode
+# documented failure mode the model needs to see verbatim.
-# the model needs to see verbatim — collapsing them harms readability.
 "strix/tools/_sandbox_dispatch.py" = ["PLR0911"]
-# StrixSession + StrixTracingProcessor catch broad Exception intentionally:
+# StrixSession + StrixTracingProcessor catch broad Exception
-# the whole point is that compressor / sanitizer / disk failures must not
+# intentionally: compressor / sanitizer / disk failures must not tear
-# tear down the agent run (C10, C16). Calls already log at exception level.
+# down the run. Calls already log at exception level.
 "strix/llm/strix_session.py" = ["BLE001"]
 "strix/telemetry/strix_processor.py" = ["BLE001"]
 

strix/agents/factory.py

@@ -1,33 +1,21 @@
-"""build_strix_agent — assemble an ``agents.Agent`` for root or child runs.
+"""``build_strix_agent`` — assemble an ``agents.Agent`` for root or child.
 
-This is the keystone that links Phase 2's SDK function tools, Phase 3's
+Wires the SDK function tools, multi-agent graph tools,
-graph tools, Phase 4's CaidoCapability, and the rendered Jinja prompt
+``CaidoCapability``, and the rendered Jinja prompt into one
-from :mod:`strix.agents.prompt` into a single ``agents.Agent``
+``agents.Agent`` ready for ``Runner.run``.
-instance ready for ``Runner.run``.
 
 Two flavors:
 
-- **Root** (``is_root=True``): the top-level scan agent. Carries
+- **Root** (``is_root=True``): top-level scan agent. Carries
-  ``finish_scan`` (terminates the scan), no ``agent_finish`` (that's
+  ``finish_scan`` and stops there.
-  for subagents). ``tool_use_behavior`` stops on ``finish_scan`` so
-  the model can't accidentally keep talking after marking the scan
-  complete.
-
 - **Child** (``is_root=False``): subagents spawned by the
-  ``create_agent`` graph tool. Carries ``agent_finish``, no
+  ``create_agent`` graph tool. Carries ``agent_finish`` and stops
-  ``finish_scan``. ``tool_use_behavior`` stops on ``agent_finish``
+  there — without ``stop_at_tool_names`` the SDK loop would keep
-  (C4 — without this, the SDK loop would keep going to ``max_turns``
+  running to ``max_turns`` even after the child reported back.
-  even after the child reported back to its parent).
 
-Caido tools come from ``CaidoCapability.tools()`` automatically via
+Caido tools come from ``CaidoCapability.tools()`` via the SDK's
-the SDK's capability merge — we don't include them here. Skills are
+capability merge — we don't list them here. Skills are baked into the
-injected via the prompt at scan-bring-up time; runtime skill loading
+system prompt at scan bring-up; there's no runtime skill-loading tool.
-isn't exposed as a tool any more (the legacy implementation reached
-into a global agent registry that no longer exists).
-
-References:
-    - PLAYBOOK.md §4.3 (graph tool wiring)
-    - AUDIT.md §2.4 (C4 — stop_at_tool_names is required for subagents)
 """
 
 from __future__ import annotations
@@ -196,12 +184,12 @@ def make_child_factory(
 ) -> Any:
     """Return a callable suitable for ``ctx.context['agent_factory']``.
 
-    The Phase 3 ``create_agent`` graph tool reads
+    The ``create_agent`` graph tool reads
     ``ctx.context['agent_factory']`` and calls it with ``name=`` and
-    ``skills=`` to build a child Agent. We snapshot the run-level
+    ``skills=`` to build a child ``Agent``. Run-level arguments
-    arguments (scan_mode, is_whitebox, etc.) into a closure so each
+    (``scan_mode``, ``is_whitebox``, etc.) are captured in a closure so
-    child inherits the right scan-level configuration without the
+    each child inherits the scan-level configuration without
-    create_agent tool having to know about them.
+    ``create_agent`` having to know about them.
     """
 
     def _factory(*, name: str, skills: list[str]) -> Agent[Any]:

strix/interface/utils.py

@@ -733,18 +733,7 @@ def _parse_name_status_z(raw_output: bytes) -> list[DiffEntry]:
             index += 2
             continue
 
-        # Backward-compat fallback if output is tab-delimited unexpectedly.
+        break
-        status_fallback, has_tab, first_path = token.partition("\t")
-        if not has_tab:
-            break
-        fallback_code = status_fallback[:1]
-        fallback_similarity: int | None = None
-        if len(status_fallback) > 1 and status_fallback[1:].isdigit():
-            fallback_similarity = int(status_fallback[1:])
-        entries.append(
-            DiffEntry(status=fallback_code, path=first_path, similarity=fallback_similarity)
-        )
-        index += 1
 
     return entries
 

strix/llm/anthropic_cache_wrapper.py

@@ -1,13 +1,9 @@
-"""AnthropicCachingLitellmModel — inject cache_control on the system message.
+"""``AnthropicCachingLitellmModel`` — inject ``cache_control`` on the system message.
 
-ModelSettings.extra_body lands the field at the request top level, which
+``ModelSettings.extra_body`` lands fields at the request top level,
-Anthropic ignores. Anthropic only honors ``cache_control`` when it is on the
+which Anthropic ignores. Anthropic only honors ``cache_control`` when
-message itself. We patch the input list before delegating to the parent.
+it is on the message itself, so we patch the input list before
-
+delegating to the parent.
-References:
-    - PLAYBOOK.md §2.1
-    - AUDIT.md §2.2 (C2 — original blocker)
-    - AUDIT_R3.md F1 (signature: first 7 params positional, then *,)
 """
 
 from __future__ import annotations

strix/llm/multi_provider_setup.py

@@ -6,10 +6,6 @@ route so models named ``anthropic/<model>`` go through
 on the system message). Every other prefix
 (``openai/`` / ``gemini/`` / ``openrouter/`` / ``litellm/...``) falls
 through to the SDK's built-in litellm routing.
-
-References:
-    - PLAYBOOK.md §2.7
-    - AUDIT_R3.md C17 (model alias validation; raise UserError on bad alias)
 """
 
 from __future__ import annotations

strix/llm/strix_session.py

@@ -1,24 +1,17 @@
-"""StrixSession — Session wrapper that runs the MemoryCompressor.
+"""``StrixSession`` — Session wrapper that runs the MemoryCompressor.
 
-The SDK's `Session` (and ``SessionABC``) protocol owns conversation history
+Delegates storage to any underlying session implementation (in-memory,
-storage. We delegate the actual storage to any underlying session
+SQLite, Redis, …) and intercepts ``get_items`` so the
-implementation (in-memory, SQLite, Redis, …) and intercept ``get_items`` so
+``MemoryCompressor`` runs before the model sees the history.
-the ``MemoryCompressor`` runs before the model sees the history.
 
-Why wrap rather than reimplement:
+Wrapping (rather than reimplementing) keeps the pentest-tuned
-- ``MemoryCompressor`` already encodes the pentest-tuned summarization
+summarization prompt and 90K-token budget intact. ``get_items`` is
-  prompt and the 90K-token budget that Strix has been tuning for months.
+also the last hook before ``call_model_input_filter``, so compressing
-  Reimplementing inside a Session would lose that institutional knowledge.
+here means the filter sees a compressed history too.
-- The SDK gives us a clean seam in ``get_items``: it's the last call before
-  ``call_model_input_filter`` runs, so compressing here means the filter
-  sees a compressed history too.
 
-References:
+If compression raises, we fall back to the uncompressed history and
-    - PLAYBOOK.md §2.8
+flip a flag so future attempts skip — a permanently broken compressor
-    - AUDIT_R2.md §1.5 (C10 — compressor exception → uncompressed fallback)
+mustn't infinite-loop the agent into context starvation.
-    - AUDIT_R3.md §3 row W5/E2 — once compression has failed, set a flag and
-      skip future attempts so we don't infinite-loop on a permanently broken
-      compressor while the agent loop slowly drowns in context.
 """
 
 from __future__ import annotations

strix/orchestration/bus.py

@@ -1,14 +1,8 @@
-"""AgentMessageBus — peer-to-peer multi-agent state owned by Strix.
+"""``AgentMessageBus`` — peer-to-peer multi-agent state for one scan.
 
 A single ``asyncio.Lock``-protected dataclass that owns inboxes,
 parent edges, statuses, and per-agent stats for the lifetime of one
 Strix scan.
-
-References:
-    - PLAYBOOK.md §2.3
-    - AUDIT_R2.md §1.4 (cancel_descendants)
-    - AUDIT_R2.md §1.7 (stats snapshot under lock)
-    - AUDIT_R3.md C13 (finalize cleans up state to avoid orphaned-message leak)
 """
 
 from __future__ import annotations
@@ -69,9 +63,8 @@ class AgentMessageBus:
     async def send(self, target: str, msg: dict[str, Any]) -> None:
         """Append a message to ``target``'s inbox.
 
-        Idempotent if target was never registered: creates an empty inbox.
+        Messages addressed to a finalized agent are dropped silently —
-        Messages addressed to a finalized agent are dropped silently — the
+        :meth:`finalize` clears the inbox so they can't accumulate.
-        target's inbox was cleared in :meth:`finalize` (C13).
         """
         async with self._lock:
             if target not in self.statuses:
@@ -115,9 +108,8 @@ class AgentMessageBus:
     async def finalize(self, agent_id: str, status: str) -> None:
         """Move an agent from live to completed; clean up routing state.
 
-        C13 (AUDIT_R3): also clears ``inboxes``, ``parent_of``, ``names`` so
+        Also clears ``inboxes``, ``parent_of``, ``names`` so siblings
-        sibling agents that try to send to a finished agent don't accumulate
+        that send to a finished agent can't accumulate orphan messages.
-        orphan messages forever.
         """
         async with self._lock:
             self.statuses[agent_id] = status
@@ -127,7 +119,7 @@ class AgentMessageBus:
             self.names.pop(agent_id, None)
 
     async def total_stats(self) -> dict[str, Any]:
-        """Snapshot of live + completed stats. Lock-protected (C12)."""
+        """Snapshot of live + completed stats."""
         async with self._lock:
             agg = {"in": 0, "out": 0, "cached": 0, "cost": 0.0, "calls": 0}
             for stats in (*self.stats_live.values(), *self.stats_completed.values()):
@@ -138,9 +130,9 @@ class AgentMessageBus:
     async def cancel_descendants(self, root_agent_id: str) -> None:
         """Cancel ``root_agent_id`` and every transitive child, leaves first.
 
-        Wired into the CLI Ctrl+C handler and TUI stop button so a root cancel
+        Wired into the CLI Ctrl+C handler and TUI stop button —
-        actually propagates (C9 — SDK's ``result.cancel`` does not cascade
+        the SDK's ``result.cancel`` doesn't cascade to children spawned
-        to children spawned via ``asyncio.create_task``).
+        via ``asyncio.create_task``, so we walk the tree ourselves.
         """
         async with self._lock:
             queue = [root_agent_id]

strix/orchestration/filter.py

@@ -1,14 +1,8 @@
-"""inject_messages_filter — SDK ``call_model_input_filter`` for the message bus.
+"""``inject_messages_filter`` — SDK ``call_model_input_filter`` for the bus.
 
 The SDK runs ``call_model_input_filter`` exactly once per turn before
-the LLM call (``run_internal/turn_preparation.py:55-80``) and captures
+the LLM call and captures the output in a closure for any subsequent
-the filter's output in a lambda closure for any subsequent retries
+retries — so a single drain per turn doesn't lose messages on retry.
-(``run_internal/model_retry.py:34-35``) — so a single drain per turn
-does not lose messages on retry.
-
-References:
-    - PLAYBOOK.md §2.4
-    - AUDIT_R3.md C14 (filter must be defensive — exception → unmodified data)
 """
 
 from __future__ import annotations
@@ -32,12 +26,12 @@ async def inject_messages_filter(data: CallModelData) -> ModelInputData:
     Each drained message is wrapped in an ``<inter_agent_message>`` XML envelope
     so the system prompt's rules around inter-agent communication apply.
 
-    Messages from the literal sender ``"user"`` (a real human via TUI) skip
+    Messages from the literal sender ``"user"`` (a real human via TUI)
-    the XML wrap and are added as plain user messages.
+    skip the XML wrap and are added as plain user messages.
 
-    C14: any exception inside the filter — including a malformed message dict
+    Any exception inside the filter — malformed message dict, bug in
-    or a bug in ``bus.drain`` — is caught and the original ``data.model_data``
+    ``bus.drain``, etc. — is caught and the original ``data.model_data``
-    is returned unmodified. A bug in the filter must never tear down the run.
+    is returned unmodified. A bug here must never tear down the run.
     """
     try:
         if not isinstance(data.context, dict):

strix/orchestration/hooks.py

@@ -1,13 +1,4 @@
-"""StrixOrchestrationHooks — RunHooks subclass wiring bus + tracer + warnings.
+"""``StrixOrchestrationHooks`` — RunHooks wiring bus + tracer + warnings."""
-
-References:
-    - PLAYBOOK.md §2.5
-    - AUDIT.md §2.5  (C5 — streaming/hook bridge)
-    - AUDIT_R2.md §1.3 (C8 — subagent crash detection)
-    - AUDIT_R3.md F2 (context types: AgentHookContext for agent_*,
-      RunContextWrapper otherwise; on_tool_end result is str)
-    - AUDIT_R3.md C15 (every hook body try/except so a hook bug never tears down the run)
-"""
 
 from __future__ import annotations
 
@@ -27,16 +18,17 @@ class StrixOrchestrationHooks(RunHooks[Any]):
 
     Wires four concerns:
 
-    1. Turn-budget warnings injected into ``input_items`` at 85% and ``N - 3``
+    1. Turn-budget warnings injected into ``input_items`` at 85% and
-       of ``max_turns``.
+       ``N - 3`` of ``max_turns``.
-    2. LLM usage recording into the bus.
+    2. LLM usage recording into the bus + tracer.
-    3. Sandbox readiness: awaits the ``CaidoCapability._healthcheck_task``
+    3. Sandbox readiness: awaits the
-       on first agent start so the agent doesn't fire tools before Caido and
+       ``CaidoCapability._healthcheck_task`` on first agent start so
-       the tool server are ready.
+       the agent doesn't fire tools before Caido and the tool server
-    4. Subagent crash detection (C8): if ``on_agent_end`` fires without
+       are ready.
-       ``agent_finish_called`` being set in context, posts a synthetic
+    4. Subagent crash detection: if ``on_agent_end`` fires without
-       ``<agent_crash>`` message to the parent's inbox so the parent learns
+       ``agent_finish_called`` being set, posts a synthetic
-       on its next turn instead of polling ``wait_for_message`` forever.
+       ``<agent_crash>`` message to the parent's inbox so the parent
+       learns on its next turn instead of waiting forever.
     """
 
     async def on_llm_start(
@@ -105,7 +97,7 @@ class StrixOrchestrationHooks(RunHooks[Any]):
                     input_tokens=int(getattr(usage, "input_tokens", 0) or 0),
                     output_tokens=int(getattr(usage, "output_tokens", 0) or 0),
                     cached_tokens=cached,
-                    cost=0.0,  # litellm cost computation lives in the legacy LLM
+                    cost=0.0,
                     requests=1,
                     bucket="live",
                 )
@@ -203,12 +195,3 @@ class StrixOrchestrationHooks(RunHooks[Any]):
                 tracer.log_tool_end(ctx.get("agent_id", "?"), tool.name, result)
         except Exception:
             logger.exception("on_tool_end failed")
-
-    async def on_handoff(
-        self,
-        context: RunContextWrapper[Any],
-        from_agent: Any,
-        to_agent: Any,
-    ) -> None:
-        # Strix multi-agent goes through the bus; SDK handoffs are unused.
-        pass

strix/run_config_factory.py

@@ -1,18 +1,7 @@
-"""make_run_config — assemble a Strix-flavored ``RunConfig`` for ``Runner.run``.
+"""``make_run_config`` — assemble a Strix-flavored ``RunConfig`` for ``Runner.run``.
 
-Factory pattern: every Strix scan goes through here so the defaults are
+Every scan goes through here so defaults apply uniformly. Per-call
-applied uniformly. Per-call overrides are accepted via ``model_settings_override``
+overrides land via ``model_settings_override``.
-for the rare case a single run wants different reasoning effort or
-``tool_choice`` (C21).
-
-References:
-    - PLAYBOOK.md §2.10
-    - AUDIT.md §2.1 (C1 — parallel_tool_calls=False until Phase 6 relaxes the
-      tool server's per-agent task slot serialization)
-    - AUDIT_R2.md §1.6 (C11 — retry policy explicitly excludes 401/403/400;
-      auth and validation errors must fail fast, not waste retries)
-    - AUDIT_R3.md C21 — RunConfig override + context fields including
-      ``is_whitebox``, ``diff_scope``, ``run_id``
 """
 
 from __future__ import annotations
@@ -39,14 +28,12 @@ if TYPE_CHECKING:
     from strix.orchestration.bus import AgentMessageBus
 
 
-# Phase 6 relaxes the tool server's per-agent task-slot serialization
+# Sequential tool calls per agent — the tool server serializes one task
-# (``runtime/tool_server.py:94-97``) and flips this to ``True`` after
+# per agent at a time, so concurrent calls would queue anyway.
-# multi-agent stress tests confirm safety.
+_PARALLEL_TOOL_CALLS_DEFAULT = False
-_PHASE1_PARALLEL_DEFAULT = False
 
-# Default retry policy. Explicitly does NOT include 401/403/400 — those are
+# Retry policy. 401/403/400 are deliberately excluded — auth and
-# auth and validation errors that retrying cannot fix; they should fail fast
+# validation errors can't be fixed by retrying and should fail fast.
-# so the user sees the real error within seconds. 429/5xx is the right set.
 _RETRYABLE_HTTP_STATUSES = (429, 500, 502, 503, 504)
 
 # Default retry budget: 5 attempts with ``min(90, 2*2^n)`` backoff.
@@ -82,7 +69,7 @@ def make_run_config(
     *,
     sandbox_session: BaseSandboxSession | None,
     model: str = "anthropic/claude-sonnet-4-6",
-    parallel_tool_calls: bool = _PHASE1_PARALLEL_DEFAULT,
+    parallel_tool_calls: bool = _PARALLEL_TOOL_CALLS_DEFAULT,
     tool_choice: Literal["auto", "required", "none"] | None = "required",
     reasoning_effort: Literal["low", "medium", "high"] | None = None,
     model_settings_override: ModelSettings | None = None,
@@ -90,32 +77,27 @@ def make_run_config(
 ) -> RunConfig:
     """Build a ``RunConfig`` with Strix defaults.
 
-    Note: ``max_turns`` and ``isolate_parallel_failures`` are NOT
+    Note: ``max_turns`` is not a ``RunConfig`` field — pass it directly
-    ``RunConfig`` fields — they are passed directly to ``Runner.run``.
+    to ``Runner.run``. ``STRIX_DEFAULT_MAX_TURNS`` is the budget Strix
-    Use ``STRIX_DEFAULT_MAX_TURNS`` for the budget; pass
+    uses.
-    ``isolate_parallel_failures=False`` to ``Runner.run`` if Phase 6 has
-    not yet flipped ``parallel_tool_calls=True``.
 
     Args:
-        sandbox_session: Live sandbox session shared by every agent in this
+        sandbox_session: Live sandbox session shared by every agent in
-            scan (one container per scan; see ``strix.sandbox.session_manager``).
+            this scan (one container per scan; see
-            ``None`` is allowed for unit tests and dry runs.
+            :mod:`strix.sandbox.session_manager`). ``None`` is allowed
-        model: Model alias to pass to ``MultiProvider``. Defaults to the
+            for unit tests and dry runs.
-            current production-favored Anthropic alias.
+        model: Model alias passed to ``MultiProvider``. Defaults to the
-        parallel_tool_calls: Default ``False`` to keep behavior sequential
+            production Anthropic alias.
-            per the tool server's slot serialization (C1).
+        parallel_tool_calls: Default ``False`` — the tool server
+            serializes one task per agent.
         tool_choice: Forces tool use per turn unless explicitly relaxed.
-            Pass ``None`` to omit.
         reasoning_effort: ``"low" | "medium" | "high"``; routes to
-            ``ModelSettings.reasoning``. ``None`` defers to provider default.
+            ``ModelSettings.reasoning``.
-        model_settings_override: Optional ``ModelSettings`` to merge over
+        model_settings_override: Optional per-run ``ModelSettings``
-            the factory defaults (C21 — per-run override path).
+            merged over factory defaults.
-        sandbox_client: Optional pre-built sandbox client (e.g., the Strix
+        sandbox_client: Optional pre-built sandbox client (Strix Docker
-            Docker subclass). Defaults to ``None``; the SDK will instantiate
+            subclass). The SDK instantiates its built-in if a session is
-            its built-in if a session is supplied without a client.
+            supplied without a client.
-
-    Returns:
-        A ``RunConfig`` ready to pass to ``Runner.run``.
     """
     base_settings = ModelSettings(
         parallel_tool_calls=parallel_tool_calls,
@@ -175,14 +157,14 @@ def make_agent_context(
 ) -> dict[str, Any]:
     """Build the per-agent ``context`` dict passed to ``Runner.run(context=...)``.
 
-    The dict is the canonical place where bus, sandbox handles, identity,
+    The canonical place where bus, sandbox handles, identity, tracer
-    tracer reference, and per-agent toggles live. Tools, hooks, and the
+    reference, and per-agent toggles live. Tools, hooks, and
-    ``inject_messages_filter`` all reach in via ``ctx.context.get(...)``.
+    ``inject_messages_filter`` reach in via ``ctx.context.get(...)``.
 
-    ``agent_factory`` is a callable ``(name, skills) -> agents.Agent`` used by
+    ``agent_factory`` is a callable ``(name, skills) -> agents.Agent`` —
-    the ``create_agent`` graph tool to spin up children. ``sandbox_client``
+    the ``create_agent`` graph tool uses it to spin up children that
-    is the host-side Docker subclass; ``create_agent`` reuses it across
+    inherit the same wiring. ``sandbox_client`` is the host-side Docker
-    child runs.
+    subclass, reused across child runs.
     """
     return {
         "bus": bus,

strix/runtime/__init__.py

@@ -1,31 +1,13 @@
 """Strix runtime package.
 
-What lives here:
+- :class:`strix.runtime.strix_docker_client.StrixDockerSandboxClient` —
+  host-side ``DockerSandboxClient`` subclass that injects
+  ``NET_ADMIN`` / ``NET_RAW`` capabilities and ``host.docker.internal``
+  extra-hosts, used by the per-scan session manager
+  (:mod:`strix.sandbox.session_manager`).
 
-- :class:`StrixDockerSandboxClient` — host-side ``DockerSandboxClient``
+- ``tool_server.py`` — FastAPI server that runs inside the sandbox
-  subclass that injects ``NET_ADMIN`` / ``NET_RAW`` capabilities and
+  container. Sandbox-bound tools (browser, terminal, python, file_edit,
-  ``host.docker.internal`` extra-hosts, used by the per-scan session
+  proxy) POST here from the host via
-  manager (:mod:`strix.sandbox.session_manager`).
-
-- ``tool_server.py`` — the FastAPI server that runs *inside* the
-  sandbox container; sandbox-bound tools (browser, terminal, python,
-  file_edit, proxy) POST here from the host via
   :func:`strix.tools._sandbox_dispatch.post_to_sandbox`.
-
-The legacy DockerRuntime / AbstractRuntime + ``get_runtime`` /
-``cleanup_runtime`` globals were removed when the SDK harness took
-over scan lifecycle; sandbox sessions are now per-scan and managed by
-:func:`strix.sandbox.session_manager.create_or_reuse`.
 """
-
-
-class SandboxInitializationError(Exception):
-    """Raised when sandbox initialization fails (e.g., Docker issues)."""
-
-    def __init__(self, message: str, details: str | None = None):
-        super().__init__(message)
-        self.message = message
-        self.details = details
-
-
-__all__ = ["SandboxInitializationError"]

strix/runtime/strix_docker_client.py

@@ -11,13 +11,8 @@ additions before the final create call:
 These are required for raw-socket pentest tools (nmap -sS) and for letting
 the agent reach host-served apps via ``host.docker.internal``.
 
-Pinned to ``openai-agents==0.14.6``. Bumping the SDK version requires
+Pinned to ``openai-agents==0.14.6``. Bumping the SDK requires
-re-merging the parent body. Track upstream PR for an injection hook.
+re-merging the parent body. Track upstream for an injection hook.
-
-References:
-    - PLAYBOOK.md §2.2
-    - AUDIT.md §2.3 (C3 — original blocker)
-    - SDK source: ``/tmp/openai-agents/src/agents/sandbox/sandboxes/docker.py:1434-1477``
 """
 
 from __future__ import annotations

strix/sandbox/__init__.py

@@ -1,7 +1,8 @@
 """Strix sandbox layer on top of OpenAI Agents SDK SandboxAgent / Manifest.
 
-Phase 4 deliverables:
+- :mod:`.caido_capability` — Caido proxy + 7 GraphQL function tools
-- CaidoCapability: Caido proxy + 7 GraphQL function tools + system prompt block
+  + system prompt block.
-- healthcheck: wait_for_ports_ready
+- :mod:`.healthcheck` — ``wait_for_ports_ready``.
-- session_manager: create_or_reuse / cleanup keyed by scan_id
+- :mod:`.session_manager` — ``create_or_reuse`` / ``cleanup`` keyed
+  by scan id.
 """

strix/sandbox/caido_capability.py

@@ -8,10 +8,9 @@ Three concerns wired into the SDK's capability lifecycle:
    etc.) now flows through the proxy automatically.
 
 2. **Tool exposure** (``tools``): the seven Caido SDK function-tool
-   wrappers from Phase 2.5 are returned here. The SDK runtime collects
+   wrappers are returned here. The SDK runtime collects tools from
-   tools from every capability and merges them with the agent's
+   every capability and merges them with the agent's ``tools=[...]``
-   ``tools=[...]`` declaration, so individual agents don't have to
+   declaration, so agents don't have to redeclare them.
-   redeclare them.
 
 3. **Healthcheck task** (``bind``): when a session binds, we kick off
    :func:`wait_for_http_ready` against the FastAPI tool server's
@@ -21,10 +20,6 @@ Three concerns wired into the SDK's capability lifecycle:
    :class:`StrixOrchestrationHooks.on_agent_start` hook awaits before
    the first LLM call so the agent never hits a connection-refused
    on its very first tool invocation.
-
-References:
-    - PLAYBOOK.md §3.2
-    - AUDIT.md §2.5 (C5 — healthcheck wired to RunHooks)
 """
 
 from __future__ import annotations

strix/sandbox/healthcheck.py

@@ -16,9 +16,6 @@ Two helpers are exposed:
 - :func:`wait_for_tcp_ready` for Caido, which serves an HTTP forward
   proxy on its port and does *not* expose ``/health``. A TCP connect
   is the most we can probe without sending real proxy traffic.
-
-References:
-    - PLAYBOOK.md §3.1
 """
 
 from __future__ import annotations

strix/sandbox/session_manager.py

@@ -12,9 +12,6 @@ issuing multiple ``create_or_reuse`` calls (e.g., resume after a crash
 on the host side) gets the same bundle back. ``cleanup`` is best-effort
 — a leaked container is preferable to a stuck cleanup that prevents the
 next scan from starting.
-
-References:
-    - PLAYBOOK.md §3.3
 """
 
 from __future__ import annotations

strix/skills/__init__.py

@@ -1,167 +1,56 @@
+import logging
 import re
 
 from strix.utils.resource_paths import get_strix_resource_path
 
 
-_EXCLUDED_CATEGORIES = {"scan_modes", "coordination"}
+logger = logging.getLogger(__name__)
+
 _FRONTMATTER_PATTERN = re.compile(r"^---\s*\n.*?\n---\s*\n", re.DOTALL)
 
 
-def get_available_skills() -> dict[str, list[str]]:
-    skills_dir = get_strix_resource_path("skills")
-    available_skills: dict[str, list[str]] = {}
-
-    if not skills_dir.exists():
-        return available_skills
-
-    for category_dir in skills_dir.iterdir():
-        if category_dir.is_dir() and not category_dir.name.startswith("__"):
-            category_name = category_dir.name
-
-            if category_name in _EXCLUDED_CATEGORIES:
-                continue
-
-            skills = []
-
-            for file_path in category_dir.glob("*.md"):
-                skill_name = file_path.stem
-                skills.append(skill_name)
-
-            if skills:
-                available_skills[category_name] = sorted(skills)
-
-    return available_skills
-
-
-def get_all_skill_names() -> set[str]:
-    all_skills = set()
-    for category_skills in get_available_skills().values():
-        all_skills.update(category_skills)
-    return all_skills
-
-
-def validate_skill_names(skill_names: list[str]) -> dict[str, list[str]]:
-    available_skills = get_all_skill_names()
-    valid_skills = []
-    invalid_skills = []
-
-    for skill_name in skill_names:
-        if skill_name in available_skills:
-            valid_skills.append(skill_name)
-        else:
-            invalid_skills.append(skill_name)
-
-    return {"valid": valid_skills, "invalid": invalid_skills}
-
-
-def parse_skill_list(skills: str | None) -> list[str]:
-    if not skills:
-        return []
-    return [s.strip() for s in skills.split(",") if s.strip()]
-
-
-def validate_requested_skills(skill_list: list[str], max_skills: int = 5) -> str | None:
-    if len(skill_list) > max_skills:
-        return "Cannot specify more than 5 skills for an agent (use comma-separated format)"
-
-    if not skill_list:
-        return None
-
-    validation = validate_skill_names(skill_list)
-    if validation["invalid"]:
-        available_skills = list(get_all_skill_names())
-        return (
-            f"Invalid skills: {validation['invalid']}. "
-            f"Available skills: {', '.join(available_skills)}"
-        )
-
-    return None
-
-
-def generate_skills_description() -> str:
-    available_skills = get_available_skills()
-
-    if not available_skills:
-        return "No skills available"
-
-    all_skill_names = get_all_skill_names()
-
-    if not all_skill_names:
-        return "No skills available"
-
-    sorted_skills = sorted(all_skill_names)
-    skills_str = ", ".join(sorted_skills)
-
-    description = f"List of skills to load for this agent (max 5). Available skills: {skills_str}. "
-
-    example_skills = sorted_skills[:2]
-    if example_skills:
-        example = f"Example: {', '.join(example_skills)} for specialized agent"
-        description += example
-
-    return description
-
-
-def _get_all_categories() -> dict[str, list[str]]:
-    """Get all skill categories including internal ones (scan_modes, coordination)."""
-    skills_dir = get_strix_resource_path("skills")
-    all_categories: dict[str, list[str]] = {}
-
-    if not skills_dir.exists():
-        return all_categories
-
-    for category_dir in skills_dir.iterdir():
-        if category_dir.is_dir() and not category_dir.name.startswith("__"):
-            category_name = category_dir.name
-            skills = []
-
-            for file_path in category_dir.glob("*.md"):
-                skill_name = file_path.stem
-                skills.append(skill_name)
-
-            if skills:
-                all_categories[category_name] = sorted(skills)
-
-    return all_categories
-
-
 def load_skills(skill_names: list[str]) -> dict[str, str]:
-    import logging
+    """Load skill markdown bodies (frontmatter stripped) by name.
 
-    logger = logging.getLogger(__name__)
+    Skill files live at ``strix/skills/<category>/<name>.md``. Names
-    skill_content = {}
+    can be ``"name"`` (any category), ``"category/name"``, or a bare
+    file at the skills root. Missing skills are logged and skipped.
+    """
     skills_dir = get_strix_resource_path("skills")
+    if not skills_dir.exists():
+        return {}
 
-    all_categories = _get_all_categories()
+    by_category: dict[str, str] = {}
+    for category_dir in skills_dir.iterdir():
+        if not category_dir.is_dir() or category_dir.name.startswith("__"):
+            continue
+        for file_path in category_dir.glob("*.md"):
+            by_category[file_path.stem] = f"{category_dir.name}/{file_path.stem}.md"
 
+    skill_content: dict[str, str] = {}
     for skill_name in skill_names:
+        rel_path: str | None
+        if "/" in skill_name:
+            rel_path = f"{skill_name}.md"
+        elif skill_name in by_category:
+            rel_path = by_category[skill_name]
+        elif (skills_dir / f"{skill_name}.md").exists():
+            rel_path = f"{skill_name}.md"
+        else:
+            rel_path = None
+
+        if rel_path is None or not (skills_dir / rel_path).exists():
+            logger.warning("Skill not found: %s", skill_name)
+            continue
+
         try:
-            skill_path = None
+            content = (skills_dir / rel_path).read_text(encoding="utf-8")
+        except (OSError, ValueError) as e:
+            logger.warning("Failed to load skill %s: %s", skill_name, e)
+            continue
 
-            if "/" in skill_name:
+        var_name = skill_name.split("/")[-1]
-                skill_path = f"{skill_name}.md"
+        skill_content[var_name] = _FRONTMATTER_PATTERN.sub("", content).lstrip()
-            else:
+        logger.info("Loaded skill: %s -> %s", skill_name, var_name)
-                for category, skills in all_categories.items():
-                    if skill_name in skills:
-                        skill_path = f"{category}/{skill_name}.md"
-                        break
-
-                if not skill_path:
-                    root_candidate = f"{skill_name}.md"
-                    if (skills_dir / root_candidate).exists():
-                        skill_path = root_candidate
-
-            if skill_path and (skills_dir / skill_path).exists():
-                full_path = skills_dir / skill_path
-                var_name = skill_name.split("/")[-1]
-                content = full_path.read_text(encoding="utf-8")
-                content = _FRONTMATTER_PATTERN.sub("", content).lstrip()
-                skill_content[var_name] = content
-                logger.info(f"Loaded skill: {skill_name} -> {var_name}")
-            else:
-                logger.warning(f"Skill not found: {skill_name}")
-
-        except (FileNotFoundError, OSError, ValueError) as e:
-            logger.warning(f"Failed to load skill {skill_name}: {e}")
 
     return skill_content

strix/telemetry/strix_processor.py

@@ -1,14 +1,8 @@
-"""StrixTracingProcessor — SDK trace processor that writes events.jsonl.
+"""``StrixTracingProcessor`` — SDK trace processor that writes ``events.jsonl``.
 
 Hooks into the SDK's tracing pipeline and writes events to
-``strix_runs/<run-name>/events.jsonl``. PII scrubbing via the existing
+``strix_runs/<run-name>/events.jsonl``. PII scrubbing runs through
-``TelemetrySanitizer``.
+:class:`TelemetrySanitizer`.
-
-References:
-    - PLAYBOOK.md §2.9
-    - AUDIT_R2.md §1.2 (C7 — JSONL writes must be lock-protected)
-    - AUDIT_R3.md C16 (writes must catch OSError; never tear down the run)
-    - AUDIT_R3.md F3 (every TracingProcessor hook is SYNC, not async)
 """
 
 from __future__ import annotations
@@ -47,17 +41,11 @@ def _lock_for(path: Path) -> threading.Lock:
 class StrixTracingProcessor(TracingProcessor):
     """Append trace + span events as JSONL into ``run_dir/events.jsonl``.
 
-    Every hook is synchronous — required by ``TracingProcessor`` ABC.
+    Every hook is synchronous — required by the ``TracingProcessor``
-    Every write is protected by a per-path ``threading.Lock`` so concurrent
+    ABC. Each write is protected by a per-path ``threading.Lock`` so
-    spans (e.g., from parallel agent tasks) cannot interleave bytes
+    concurrent spans can't interleave bytes mid-line. ``OSError`` is
-    mid-line and corrupt the JSONL (C7).
+    swallowed so a full disk or permission error doesn't tear the run
-
+    down. PII scrubbing runs on every event before it hits the file.
-    Every write is wrapped in ``try/except OSError`` so a full disk or a
-    permission error during the run does NOT propagate up the hook chain
-    and tear down the agent (C16).
-
-    PII scrubbing via :class:`TelemetrySanitizer` runs on every event
-    before it hits the file.
     """
 
     def __init__(
@@ -80,8 +68,8 @@ class StrixTracingProcessor(TracingProcessor):
     def _emit(self, event: dict[str, Any]) -> None:
         """Sanitize ``event`` and append it as one JSONL line.
 
-        Failures are swallowed — we'd rather lose a trace event than fail
+        Failures are swallowed — we'd rather lose a trace event than
-        the run. Errors are logged at WARNING (C16).
+        fail the run.
         """
         try:
             clean = self.sanitizer.sanitize(event)

strix/tools/__init__.py

@@ -17,13 +17,7 @@ from .finish import *  # noqa: F403
 from .notes import *  # noqa: F403
 from .proxy import *  # noqa: F403
 from .python import *  # noqa: F403
-from .registry import (
+from .registry import get_tool_by_name, get_tool_names, register_tool, tools
-    ImplementedInClientSideOnlyError,
-    get_tool_by_name,
-    get_tool_names,
-    register_tool,
-    tools,
-)
 from .reporting import *  # noqa: F403
 from .terminal import *  # noqa: F403
 from .thinking import *  # noqa: F403
@@ -32,7 +26,6 @@ from .web_search import *  # noqa: F403
 
 
 __all__ = [
-    "ImplementedInClientSideOnlyError",
     "get_tool_by_name",
     "get_tool_names",
     "register_tool",

strix/tools/_decorator.py

@@ -1,24 +1,15 @@
-"""strix_tool — function_tool factory with Strix defaults.
+"""``strix_tool`` — ``function_tool`` factory with Strix defaults.
 
-Every tool in the migrated harness should be decorated with ``@strix_tool``
+Every tool uses ``@strix_tool`` instead of bare ``@function_tool`` so
-instead of bare ``@function_tool`` so the team's defaults stay consistent
+defaults stay consistent across the suite. Override per call when
-without per-tool boilerplate. Override per call when needed.
+needed.
 
 Defaults:
-    - ``timeout``: 120s (matches the legacy tool server's
+    - ``timeout``: 120s.
-      ``STRIX_SANDBOX_EXECUTION_TIMEOUT``).
     - ``timeout_behavior``: ``"error_as_result"`` for idempotent tools.
-      Critical sandbox tools (terminal, browser, python) should pass
+      Critical sandbox tools (terminal, browser, python) opt into
-      ``timeout_behavior="raise_exception"`` explicitly so the SDK can fail
+      ``timeout_behavior="raise_exception"`` explicitly so the SDK
-      the run rather than letting the model retry the same hung call (C20).
+      fails the run rather than letting the model retry a hung call.
-
-The SDK auto-threads sync function bodies via ``asyncio.to_thread``
-(``tool.py:1820-1829``), so libtmux / IPython / blocking httpx code can be
-written as plain ``def`` and the decorator will not block the event loop.
-
-References:
-    - PLAYBOOK.md §2.6
-    - AUDIT_R3.md C20 (per-tool timeout_behavior discrimination)
 """
 
 from __future__ import annotations
@@ -44,17 +35,10 @@ def strix_tool(
 ) -> Callable[[_ToolFn], FunctionTool]:
     """Wrap ``agents.function_tool`` with Strix defaults.
 
-    The SDK's ``FunctionTool`` requires ``async def`` for ``timeout_seconds``
+    Strict mode is on by default (forbids free-form ``dict[str, X]``
-    to apply (sync handlers cannot be cleanly cancelled). All Strix tools are
+    parameters because the strict JSON schema needs
-    ``async def``; sync libraries (libtmux, IPython) get wrapped in
+    ``additionalProperties: false``). A few tools that take arbitrary
-    ``asyncio.to_thread`` inside the async tool body.
+    header / modification dicts opt out via ``strict_mode=False``.
-
-    The SDK enforces ``strict_mode=True`` by default, which forbids
-    free-form ``dict[str, X]`` parameters (the strict JSON schema needs
-    ``additionalProperties: false``). A handful of legacy tools
-    (``send_request``, ``repeat_request``) take arbitrary header /
-    modification dicts whose keys can't be enumerated, so they must
-    opt out of strict mode to preserve parity with the XML schema.
 
     Usage::
 

strix/tools/_sandbox_dispatch.py

@@ -1,24 +1,14 @@
-"""post_to_sandbox — host-to-container HTTP transport for sandbox tools.
+"""``post_to_sandbox`` — host-to-container HTTP transport for sandbox tools.
 
-Every Strix tool that runs inside the Kali container (browser, terminal,
+Every Strix tool that runs inside the Kali container (browser,
-python, the seven Caido tools) has the same wire shape: POST a JSON body
+terminal, python, file_edit, the seven Caido tools) has the same wire
-to ``http://localhost:{tool_server_host_port}/execute`` with a Bearer
+shape: POST JSON to ``http://localhost:{tool_server_host_port}/execute``
-token header and ``{"agent_id", "tool_name", "kwargs"}`` as the body.
+with a Bearer token and ``{"agent_id", "tool_name", "kwargs"}`` body.
 
-This helper centralizes that transport so:
+The helper centralizes timeouts (``connect=10s`` / ``read=150s``), a
-
+50 MB response-size cap so a runaway tool can't OOM the host, and
-- Every sandbox tool gets the same timeout policy
+predictable error-string shaping so transport failures don't tear
-  (``connect=10s`` / ``read=150s``).
+down the run.
-- Every sandbox tool inherits the same response-size cap (50 MB) so a
-  runaway tool body cannot OOM the host (C18).
-- Auth/transport errors surface as predictable error strings instead of
-  exceptions, so the model can retry / pick a different tool without the
-  run dying.
-
-References:
-    - PLAYBOOK.md §3.4
-    - AUDIT_R3.md C18 (sandbox response size cap)
-    - HARNESS_WIKI.md §7.2 (legacy executor.py wire format we mirror)
 """
 
 from __future__ import annotations
@@ -36,14 +26,9 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 
-# Connect: how long to wait for the TCP handshake to complete.
-# Read: how long the tool may spend executing before we abandon the call.
-# Mirrors the legacy executor.py (``SANDBOX_EXECUTION_TIMEOUT = 120 + 30``).
 _SANDBOX_TIMEOUT = httpx.Timeout(connect=10.0, read=150.0, write=150.0, pool=150.0)
 
-#: Cap on response body size from the tool server. Anything bigger is
+# Cap so a runaway tool body never blows up the host heap.
-#: replaced by an error string so the model sees something coherent and
-#: the host doesn't OOM trying to allocate the buffer (C18).
 _MAX_RESPONSE_BYTES = 50 * 1024 * 1024  # 50 MB
 
 

strix/tools/agents_graph/tools.py

@@ -1,29 +1,16 @@
-"""SDK function-tool wrappers for the multi-agent graph tools.
+"""Multi-agent graph tools — read/write the :class:`AgentMessageBus`.
 
-Six tools that read/write the :class:`AgentMessageBus` (built in Phase 0,
+- ``view_agent_graph``: render the parent/child tree.
-``strix.orchestration.bus``):
-
-- ``view_agent_graph``: render the parent/child tree from ``bus.parent_of``.
 - ``agent_status``: per-agent status + pending message count.
-- ``send_message_to_agent``: peer-to-peer message into a child/sibling inbox.
+- ``send_message_to_agent``: queue a message in another agent's inbox.
-- ``wait_for_message``: poll our own inbox until a message arrives or the
+- ``wait_for_message``: pause this agent until a message arrives or
-  timeout expires (the legacy harness's "I'm idle, wake me on inbox").
+  ``timeout_seconds`` elapses.
-- ``create_agent``: spawn a child via ``asyncio.create_task(Runner.run(...))``;
+- ``create_agent``: spawn a child via
-  registers the child with the bus and stores its task handle so root cancels
+  ``asyncio.create_task(Runner.run(...))``; the task handle is stored
-  cascade (C9, ``bus.cancel_descendants``).
+  so a root-level cancel cascades to descendants.
-- ``agent_finish``: subagents only — flips ``agent_finish_called`` so the
+- ``agent_finish``: subagents only — flips ``agent_finish_called`` so
-  on_agent_end hook records "completed" rather than "crashed" (C8), and
+  the ``on_agent_end`` hook records "completed" rather than "crashed",
-  posts a structured completion report to the parent's inbox.
+  and posts a structured completion report to the parent's inbox.
-
-The legacy ``strix.tools.agents_graph.agents_graph_actions`` is left
-untouched — it still drives the legacy harness. These wrappers only
-target the bus and don't touch the legacy ``_agent_graph`` dict.
-
-References:
-    - PLAYBOOK.md §4.3
-    - AUDIT_R2.md §1.4 (cancel_descendants — Runner.run task handle stored
-      in bus.tasks so a root cancel walks the tree)
-    - AUDIT_R3.md C8 (crash detection via on_agent_end + agent_finish_called)
 """
 
 from __future__ import annotations
@@ -223,8 +210,7 @@ async def send_message_to_agent(
     )
 
 
-# Polling cadence for ``wait_for_message``. 1s matches the PLAYBOOK
+# Tighter would burn CPU; slacker would feel laggy when a sibling
-# skeleton; tighter would burn CPU, slacker would feel laggy when a sibling
 # delivers a message right after the wait starts.
 _WAIT_POLL_SECONDS = 1.0
 

strix/tools/registry.py

@@ -1,14 +1,14 @@
 """Minimal in-container tool registry.
 
 Used inside the sandbox container by ``strix.runtime.tool_server`` to
-look up `@register_tool`-decorated functions by name. Sandbox-bound
+look up ``@register_tool``-decorated functions by name. Sandbox-bound
-tools (browser, terminal, python, file_edit, proxy) live as legacy
+tools (browser, terminal, python, file_edit, proxy) live as
 ``*_actions.py`` modules with this decoration; the host POSTs to
 :func:`tool_server.execute_tool` which dispatches via
 :func:`get_tool_by_name`.
 
-Host-side tools are pure SDK function tools wired through
+Host-side SDK function tools are wired through
-:mod:`strix.agents.factory` and don't touch this registry at all.
+:mod:`strix.agents.factory` and don't touch this registry.
 """
 
 import logging
@@ -25,17 +25,6 @@ tools: list[dict[str, Any]] = []
 _tools_by_name: dict[str, Callable[..., Any]] = {}
 
 
-class ImplementedInClientSideOnlyError(Exception):
-    """Raised by sandbox-side stubs whose real implementation lives host-side."""
-
-    def __init__(
-        self,
-        message: str = "This tool is implemented in the client side only",
-    ) -> None:
-        self.message = message
-        super().__init__(self.message)
-
-
 def _is_sandbox_mode() -> bool:
     return os.getenv("STRIX_SANDBOX_MODE", "false").lower() == "true"