Skip to content

armgit / strix

Public Imported from GitHub
0

Open-source AI penetration testing tool to find and fix your app’s vulnerabilities.

docs: harness wiki + SDK migration plan + audits + playbook + testing strategy

0xallama35a4a22b175
diff --git a/AUDIT.md b/AUDIT.md
new file mode 100644
index 00000000..3b5ee331
--- /dev/null
+++ b/AUDIT.md
@@ -0,0 +1,350 @@
+# Migration Audit — Pre-Execution Verification
+
+> Source-verified against `openai-agents` v0.14.6 at `/tmp/openai-agents` and Strix at `9fb1012`. Five parallel deep-dives covering: agent loop, tool execution, sandbox, LLM/sessions/tracing, and Strix internals.
+>
+> **Verdict: GO.** No architectural blockers. Five concrete corrections to the plan must be applied before Phase 1; all are <300 LOC total. Migration today is feasible if we apply the corrections below in the listed order.
+
+---
+
+## 1. Verified bridges (no change needed)
+
+These claims from `MIGRATION_EVALUATION.md` were **confirmed against source** — proceed as planned:
+
+| Plan claim | Verification | Source |
+|---|---|---|
+| `call_model_input_filter` runs before every model call | ✓ Confirmed at `turn_preparation.py:55-80`. **Bonus: filter runs ONCE per turn — its output is captured in a lambda closure for retries (`model_retry.py:34-35`). Inbox messages will NOT be drained twice on retry.** Open question #1 in plan §11 is resolved. | `run_internal/turn_preparation.py:48-82`, `run_internal/run_loop.py:1363-1369, 1803-1809` |
+| `asyncio.create_task(Runner.run(...))` is isolation-safe | ✓ Each task gets own `RunContextWrapper`; contextvars properly isolated. No global state mutation inside `Runner.run`. | `run.py:486-615`, `run_context.py:42-51` |
+| Shared `SandboxRunConfig(session=...)` across parallel runs | ✓ SDK does NOT tear down sandbox sessions at run end; caller owns lifecycle. Safe to reuse one session across N children. | `run_config.py:115-138`, `docker.py:1372-1401` |
+| `RunContextWrapper.context` mutable across turns | ✓ Dict is by-reference; mutations persist. Bus + agent_id stash will work as designed. | `run_context.py:42-51`, `run.py:615` |
+| `RunHooks.on_agent_end` fires once per Runner.run | ✓ Single fire when `final_output` established. | `run_internal/turn_resolution.py:204-255` |
+| Custom Docker image accepted | ✓ `DockerSandboxClientOptions(image=str)` is verbatim pass-through to `containers.create(image=...)`. No assumed binaries. | `sandbox/sandboxes/docker.py:106-122, 1340, 1444-1456` |
+| `Manifest.environment` reaches container | ✓ Resolved via `await manifest.environment.resolve()` and passed to `containers.create(environment=...)`. | `sandbox/sandboxes/docker.py:1448-1450` |
+| `Manifest` entries are a strict superset (LocalDir, GitRepo, mounts) | ✓ Direct LocalDir maps to our tar-pipe with concurrency limits. | `sandbox/entries/__init__.py`, `artifacts.py:127-179` |
+| Capability lifecycle (clone, bind, tools, instructions, process_manifest) | ✓ Per-run cloned; bound after container start; can hold mutable state. CaidoCapability viable. | `sandbox/capabilities/capability.py:15-100`, `sandbox/runtime.py:180-256` |
+| MultiProvider with custom prefix routing | ✓ `MultiProviderMap.add_provider("strix", StrixModelProvider())` works exactly. | `models/multi_provider.py:16-49, 138-232` |
+| Custom `ModelProvider` interface | ✓ Just `get_model(model_name) -> Model`. Post-prefix-strip name received. | `models/interface.py:127-151` |
+| LitellmModel reasoning effort priority | ✓ Exact: `reasoning.effort` > `extra_body["reasoning_effort"]` > `extra_args["reasoning_effort"]`. | `extensions/models/litellm_model.py:162-199` |
+| LitellmModel streaming + tool-call assembly across providers | ✓ `ChatCmplStreamHandler.handle_stream()` unifies provider-native streaming (Anthropic, OpenAI, etc.) into common stream format. | `extensions/models/litellm_model.py:315-351` |
+| `add_trace_processor()` / `set_trace_processors()` | ✓ Both exist; can disable defaults entirely. | `tracing/__init__.py:94-130` |
+| `RunHooks` 7-hook surface area | ✓ All 7 hooks fire as documented. RunHooks + AgentHooks both fire (gathered). | `lifecycle.py:13-99, 102-199` |
+| Per-tool timeout default is `None` | ✓ Confirmed. Our `strix_tool()` factory will re-impose 120s. | `tool.py:337-338` |
+| `RunState.to_json()/from_json()` resumable across processes | ✓ Schema v1.9; full serialization. | `run_state.py:1-200` |
+| `tracing_disabled` per-RunConfig | ✓ Disables ALL tracing for that run. | `run_config.py:186-188` |
+| `OPENAI_AGENTS_DONT_LOG_MODEL_DATA` env | ✓ Logging only; independent from tracing. | `_debug.py:12-21` |
+| Sync function tools auto-offload via `asyncio.to_thread` | ✓ **Plan was wrong** — SDK DOES auto-thread sync `@function_tool` bodies. We can drop the manual `asyncio.to_thread` wrapping in our libtmux/IPython tools and just write sync functions. ~30 LOC saved. | `tool.py:1820-1829` |
+| `ToolGuardrailFunctionOutput.reject_content("nope")` continues run | ✓ Model sees "nope" as tool output and proceeds. Run NOT halted. | `tool_guardrails.py:79-105` |
+| Multi-agent stat aggregation via hooks | ✓ `on_llm_end` + `on_agent_end` fire on each child Runner.run; bus aggregation works. | `lifecycle.py`, `run_internal/turn_resolution.py:204-255` |
+
+---
+
+## 2. Critical corrections (must apply before / during Phase 1)
+
+Five corrections to the plan. All are concrete and small.
+
+### 2.1 [BLOCKER] Strix tool server slot serialization vs SDK parallel tool calls
+
+**The collision.** SDK fires N tool calls in one turn as N concurrent `asyncio.create_task` (`run_internal/tool_execution.py:1414`, `:1424-1430`). Strix tool server cancels the previous in-flight task for the same agent on every new request (`tool_server.py:94-97`). When SDK issues `terminal_execute` + `web_search` simultaneously for the same agent, the second cancels the first.
+
+**Fix (Phase 1 — safe default).** Add to the default RunConfig for every Strix run:
+
+```python
+RunConfig(
+ model_settings=ModelSettings(
+ parallel_tool_calls=False, # model-side hint: emit one tool call per turn
+ ...
+ ),
+ isolate_parallel_failures=False, # if model emits multiple anyway, don't cascade-cancel
+ ...
+)
+```
+
+**Caveat.** `parallel_tool_calls` is a **provider hint** (`model_settings.py:89-96`), not enforced SDK-side. The model may still emit multiple. With `isolate_parallel_failures=False`, sibling tools survive a single failure; but the tool server still cancels prev-task on same-agent collision.
+
+**Fix (Phase 2 — proper).** Relax `tool_server.py:94-97` to allow concurrent same-agent tool calls. The cancellation logic was Strix's old serialization; we don't need it under the SDK's orchestration. ~10 LOC removal. Re-test multi-agent end-to-end.
+
+**Effort:** 0.5 day (safe default in Phase 1) + 0.5 day (proper fix + tests in Phase 2).
+
+---
+
+### 2.2 [BLOCKER] Anthropic prompt cache placement is wrong in plan
+
+**The defect.** Plan §3.2 said: set `ModelSettings(extra_body={"cache_control": {"type": "ephemeral"}})`. Verified at `extensions/models/litellm_model.py:509-516` — this lands `cache_control` in the **request-level** `extra_body`, not on the system message. **Anthropic requires `cache_control` on the system message itself** (per Anthropic API spec). Plan would silently cache nothing.
+
+**Fix.** Build a thin `LitellmModel` subclass that injects `cache_control` into the message list before delegating to parent:
+
+```python
+# strix/llm/anthropic_cache_wrapper.py (~40 LOC)
+from agents.extensions.models.litellm_model import LitellmModel
+
+class AnthropicCachingLitellmModel(LitellmModel):
+ def _patch_system_message_for_cache(self, input_items: list) -> list:
+ if not _is_anthropic(self.model):
+ return input_items
+ patched = []
+ for item in input_items:
+ if isinstance(item, dict) and item.get("role") == "system":
+ content = item["content"]
+ if isinstance(content, str):
+ content = [{"type": "text", "text": content,
+ "cache_control": {"type": "ephemeral"}}]
+ patched.append({**item, "content": content})
+ else:
+ patched.append(item)
+ return patched
+
+ async def get_response(self, *, input, **kwargs):
+ return await super().get_response(input=self._patch_system_message_for_cache(input), **kwargs)
+
+ async def stream_response(self, *, input, **kwargs):
+ async for ev in super().stream_response(input=self._patch_system_message_for_cache(input), **kwargs):
+ yield ev
+```
+
+Wire into our `MultiProviderMap` so any `litellm/anthropic/...` route uses this wrapper.
+
+**Effort:** 0.5 day (~40 LOC + tests).
+
+---
+
+### 2.3 [BLOCKER] DockerSandboxClient subclass requires full method duplication
+
+**The reality.** Audit #2 verified that `_create_container()` (`sandbox/sandboxes/docker.py:1434-1477`) builds `create_kwargs` locally and **does not** expose a hook for kwarg injection. Subclass must reimplement the method body. ~100-120 LOC duplication. Plan said "~80 LOC" — bump to ~120 LOC.
+
+**Fix.** Subclass and copy the parent body verbatim, adding our injections before the final `containers.create(**create_kwargs)` line:
+
+```python
+# strix/runtime/strix_docker_client.py (~120 LOC)
+from agents.sandbox.sandboxes.docker import (
+ DockerSandboxClient, _build_docker_volume_mounts,
+ _manifest_requires_fuse, _manifest_requires_sys_admin,
+ _docker_port_key, parse_repository_tag,
+)
+
+class StrixDockerSandboxClient(DockerSandboxClient):
+ async def _create_container(self, image, *, manifest=None, exposed_ports=(), session_id=None):
+ # --- copy of parent _create_container body (lines 1442-1476) ---
+ if not self.image_exists(image):
+ repo, tag = parse_repository_tag(image)
+ self.docker_client.images.pull(repo, tag=tag or None, all_tags=False)
+
+ environment = None
+ if manifest:
+ environment = await manifest.environment.resolve()
+
+ create_kwargs = {
+ "entrypoint": ["tail"],
+ "image": image,
+ "detach": True,
+ "command": ["-f", "/dev/null"],
+ "environment": environment,
+ }
+
+ if manifest is not None:
+ mounts = _build_docker_volume_mounts(manifest, session_id=session_id)
+ if mounts:
+ create_kwargs["mounts"] = mounts
+ if _manifest_requires_fuse(manifest):
+ create_kwargs.setdefault("devices", []).append("/dev/fuse")
+ create_kwargs.setdefault("cap_add", []).append("SYS_ADMIN")
+ create_kwargs.setdefault("security_opt", []).append("apparmor:unconfined")
+ if _manifest_requires_sys_admin(manifest):
+ create_kwargs.setdefault("cap_add", []).append("SYS_ADMIN")
+
+ if exposed_ports:
+ create_kwargs["ports"] = {
+ _docker_port_key(p): ("127.0.0.1", None) for p in exposed_ports
+ }
+
+ # --- STRIX INJECTIONS ---
+ create_kwargs.setdefault("cap_add", []).extend(["NET_ADMIN", "NET_RAW"])
+ create_kwargs.setdefault("extra_hosts", {})["host.docker.internal"] = "host-gateway"
+
+ return self.docker_client.containers.create(**create_kwargs)
+```
+
+**Risk.** SDK upstream changes to `_create_container` won't propagate. Pin SDK version; track upstream in CI; consider upstream PR for `additional_create_kwargs` hook.
+
+**Effort:** 0.5 day (~120 LOC + integration test).
+
+---
+
+### 2.4 [BLOCKER] Subagent must exit cleanly via `agent_finish` — needs `tool_use_behavior` configuration
+
+**The risk.** When subagent calls `agent_finish` and we return a result string from the tool, SDK's loop checks `Agent.tool_use_behavior` (`turn_resolution.py:512-544`). If not configured to permit early exit, the loop continues until `max_turns`. Children would burn budget instead of finishing.
+
+**Fix.** On every Strix subagent's `Agent`, configure:
+
+```python
+child_agent = Agent(
+ name=name,
+ instructions=...,
+ tools=[..., agent_finish, ...],
+ tool_use_behavior={
+ "stop_at_tool_names": ["agent_finish"],
+ },
+ ...
+)
+```
+
+This tells the SDK: as soon as `agent_finish` returns, treat that as final output. Same pattern for root agent + `finish_scan`:
+
+```python
+root_agent = Agent(
+ name="strix-root",
+ tool_use_behavior={"stop_at_tool_names": ["finish_scan"]},
+ ...
+)
+```
+
+**Effort:** Trivial — one config line per agent factory.
+
+---
+
+### 2.5 [HIGH] Streaming TUI integration needs a planned shape
+
+**The reality.** Plan §10 phase 5 said "TUI re-pointed at `Runner.run_streamed().stream_events()`" — true, but Strix's current TUI polls `tracer.streaming_content` at 2 Hz with **per-chunk granularity**. SDK `stream_events()` exposes `RawResponsesStreamEvent` (raw chunks) and `RunItemStreamEvent` (semantic items) — sufficient, but we lose Strix's `update_streaming_content(agent_id, accumulated_text)` API that aggregates incremental text.
+
+**Fix.** Build a `StrixStreamAccumulator` that consumes `Runner.run_streamed().stream_events()` and synthesizes the same shape Strix's tracer used to expose:
+
+```python
+async for event in result.stream_events():
+ if event.type == "raw_response_event":
+ delta = _extract_text_delta(event.data)
+ if delta:
+ tracer.append_streaming_content(agent_id, delta)
+ elif event.type == "run_item_stream_event":
+ if event.name == "tool_called":
+ tracer.log_tool_start(agent_id, event.item.tool_name)
+ elif event.name == "tool_output":
+ tracer.log_tool_end(agent_id, event.item.tool_name, event.item.output)
+```
+
+Plus, `RunHooks.on_llm_start/on_llm_end/on_tool_start/on_tool_end` fire regardless of streaming mode, so child agents launched via `Runner.run` (not streamed) still feed the tracer through hooks. The TUI subscribes to the same tracer.
+
+**Effort:** 1.5 days for both stream accumulator + hook bridge + TUI repoint.
+
+---
+
+## 3. Medium-severity adjustments (Phase 1-2)
+
+| # | Issue | Source | Fix | Effort |
+|---|---|---|---|---|
+| M1 | Cost tracking — SDK has `Usage(input/output/cached_tokens)` but no cost field. `litellm.completion_cost()` requires raw litellm response, not SDK's `ModelResponse`. | `usage.py`, `extensions/models/litellm_model.py:254-293` | Inside our `AnthropicCachingLitellmModel` and a similar light wrapper for OpenAI, capture the litellm response and store cost in `ModelResponse.usage` via `litellm.cost_per_token(...)` (which takes tokens, not response). Then `RunHooks.on_llm_end` reads it. | 0.5 day |
+| M2 | Vision-less model image stripping — SDK has none, will pass-through and provider rejects. | None | If we end up routing to a non-vision model, build a wrapper Model that strips images. Defer; current models (Claude Sonnet 4.6, GPT-5, Gemini) are all vision-capable. | Defer (0 day) |
+| M3 | SQLiteSession uses `threading.RLock`, not `asyncio.Lock`. Concurrent async writes from parallel children may interleave. | `memory/sqlite_session.py:17-175` | Use a separate `Session` per child (history is per-agent anyway); only share `SandboxRunConfig.session`. Plan §4.7 already says this — emphasize it in code review. | 0 day (already in plan) |
+| M4 | Trace processor memory pressure on 300-turn runs. | `tracing/processor_interface.py` | Custom processor batches every 100 spans + `force_flush()` periodically. | 0.5 day |
+| M5 | Streaming events don't expose token deltas — only raw chunks. | `stream_events.py` | Parse `RawResponsesStreamEvent.data` chunks for token text manually in our accumulator. | (rolled into 2.5) |
+| M6 | `trace_include_sensitive_data` is binary, no field-level. | `run_config.py:193-199` | Custom trace processor scrubs PII via existing `TelemetrySanitizer`. Plan already says this. | 0 day (already in plan) |
+| M7 | Caido + tool server readiness check needs a place to await — Capability.bind() is sync. | `sandbox/capabilities/capability.py:29-31` | Spawn a background task in bind() (`_healthcheck_task`); await it inside `RunHooks.on_agent_start`. ~30 LOC. | 0.5 day |
+| M8 | `vulnerability_found_callback` (TUI popup trigger) — no SDK-native equivalent. | Strix `telemetry/tracer.py:89` | Wrap `create_vulnerability_report` tool with an output guardrail that fires the callback on success. | 0.5 day |
+| M9 | `<agent_delegation>` XML wrapper today contains structured identity that the system prompt has rules to ignore. | Strix `system_prompt.jinja:19-22`, `agents_graph_actions.py:238-266` | Replicate exact XML envelope when `inject_messages_filter` adds the parent's task message OR when `create_agent` builds the child's initial input. Keeps system prompt rules intact unchanged. | 0.5 day |
+| M10 | Whitebox wiki note auto-update on subagent finish (side effect on `agent_finish` tool). | Strix `agents_graph_actions.py:161-202` | Implement directly inside our `agent_finish` function tool body, just like today. | 0 day (free port) |
+| M11 | `_force_stop` mid-turn soft-interrupt has no SDK equivalent. | Strix `base_agent.py:84` | Use `result.cancel(mode="after_turn")` for cooperative cancel; for mid-turn hard cancel, `.cancel(mode="immediate")`. | 0 day (use `result.cancel`) |
+| M12 | 85% / N-3 turn warnings as user messages. | Strix `base_agent.py:186-211` | `RunHooks.on_llm_start` checks `ctx.usage` turn count; if at threshold, mutate `input_items` (passed by reference per `lifecycle.py:18-26`). Verify mutation visibility in source: hook signature shows `input_items` is the list; mutations propagate. | 0.5 day |
+
+---
+
+## 4. Pre-Phase-1 spike (1 day)
+
+Before writing production code, validate the assumptions in a tiny throwaway script:
+
+1. **Two-children messaging smoke test.** Build minimal `MessageBus` + `inject_messages_filter` + 2 child agents that exchange one message each. Run with `LitellmModel("anthropic/claude-sonnet-4-5-20250929")` (or whatever Anthropic alias is current). Verify: messages arrive, hooks fire, no deadlock, no message duplication on retry.
+2. **Anthropic cache wrapper smoke test.** Send 3 requests with identical system prompt; check Anthropic response usage `cache_creation_input_tokens` on call 1 and `cache_read_input_tokens` on calls 2-3.
+3. **`StrixDockerSandboxClient` smoke test.** Pull our Kali image, create a session, run `nmap -sS scanme.nmap.org` via `session.exec()` to verify NET_RAW works.
+4. **`tool_use_behavior={"stop_at_tool_names": [...]}` smoke test.** Toy agent with `agent_finish`-equivalent; verify SDK terminates exactly when expected.
+5. **Tool server parallel-call smoke test.** Issue two POSTs to local tool server with same `agent_id` simultaneously; observe whether second cancels first under current code.
+
+If any spike fails, fix before Phase 1. If all pass, proceed.
+
+**Effort:** 1 day.
+
+---
+
+## 5. Updated migration plan (rev 3 sequencing)
+
+Replaces `MIGRATION_EVALUATION.md` §10. Same scope, with corrections folded in.
+
+### Phase 0 — Spike & corrections (1.5 days)
+- Run the 5 spikes above.
+- Build `AnthropicCachingLitellmModel` (~40 LOC). Smoke-tested.
+- Build `StrixDockerSandboxClient` (~120 LOC). Smoke-tested.
+- Decide tool server fix: relax serialization (recommended) OR set `parallel_tool_calls=False` + `isolate_parallel_failures=False` (safe default).
+
+### Phase 1 — Foundation (4 days)
+- `MultiProvider` + `MultiProviderMap` with `StrixModelProvider` for our aliases.
+- Wire `AnthropicCachingLitellmModel` into the provider map.
+- `strix_tool` decorator (~30 LOC; just `function_tool` with default `timeout=120, timeout_behavior="error_as_result"`).
+- Custom `Session` subclass with our memory compressor strategy.
+- Custom `TracingProcessor` with JSONL + scrubadub PII scrub. `set_trace_processors([StrixProcessor()])` to disable defaults.
+- `RunConfig` factory that bakes in: `tracing_disabled=False`, `isolate_parallel_failures=False`, `model_settings.parallel_tool_calls=False` (until Phase 6 relaxes), our processors.
+- Cost tracking inside the model wrapper (M1).
+
+### Phase 2 — Tool ports (8 days)
+- Sandbox dispatcher: one helper that POSTs to FastAPI tool server with httpx (`Timeout(connect=10, total=150)`) + Bearer auth.
+- All 30+ tools as `@strix_tool`. Sync ones use `def`, SDK auto-threads them.
+- Browser as `ComputerTool` + `AsyncComputer` subclass (or as a single `@strix_tool` if `ComputerTool` semantics don't match).
+- Stateful tools key off `RunContextWrapper.context["agent_id"]` (helper: `get_agent_id(ctx)`).
+- `create_vulnerability_report` wraps `ToolOutputGuardrail` to fire the TUI popup callback (M8).
+- Verify reentrancy on browser singleton, tmux sessions, IPython kernels.
+
+### Phase 3 — Multi-agent orchestration (4 days)
+- `AgentMessageBus` + tests.
+- `inject_messages_filter` + tests including retry simulation (verified safe by audit; no de-dup needed).
+- `StrixOrchestrationHooks` for stat aggregation + tracer wiring.
+- Six graph tools (`create_agent`, `send_message_to_agent`, `wait_for_message`, `agent_status`, `view_agent_graph`, `agent_finish`).
+- Every child Agent: `tool_use_behavior={"stop_at_tool_names": ["agent_finish"]}`.
+- Root Agent: `tool_use_behavior={"stop_at_tool_names": ["finish_scan"]}`.
+- Identity injection via `<agent_delegation>` XML in the **first user message** of the child Runner (M9).
+- Wiki auto-update on whitebox `agent_finish` (M10).
+
+### Phase 4 — Sandbox + Caido capability (2 days)
+- `StrixDockerSandboxClient` (already in Phase 0).
+- `CaidoCapability` with `process_manifest` (env vars), `tools()` (7 Caido tools), `instructions()` (proxy-aware system block), `bind()` (spawn healthcheck task).
+- `RunHooks.on_agent_start` awaits Caido + tool server readiness via the capability's healthcheck task (M7).
+- Container reuse keyed by scan_id in our session map.
+
+### Phase 5 — Interface + persistence (3 days)
+- Streaming accumulator wires `Runner.run_streamed().stream_events()` → tracer (replaces today's per-chunk `update_streaming_content`).
+- TUI keeps its 2 Hz polling against the tracer; tracer is now event-driven from the accumulator.
+- 85% / N-3 turn warnings via `RunHooks.on_llm_start` mutating `input_items` (M12).
+- Run-directory layout via custom `TracingProcessor` writing `events.jsonl`, `vulnerabilities/`, etc.
+- CLI / config / argparse layer unchanged.
+
+### Phase 6 — Validation + tool server relaxation (4 days)
+- Smoke: every tool runs in sandbox.
+- Multi-agent: 2+ parallel children + messaging + cancel.
+- Bedrock + Anthropic + OpenAI parity.
+- Memory compression at 90K.
+- PII redaction.
+- Real pentest end-to-end vs Strix baseline diff.
+- **Now relax tool_server.py:94-97** (remove per-agent task cancellation), set `parallel_tool_calls=True` and `isolate_parallel_failures=True`. Re-run multi-agent test. If clean → ship parallelism.
+
+### Buffer (2 days)
+For unforeseen issues from spike feedback or test failures.
+
+**Total: ~28.5 days** (vs plan's 25–35). Within budget.
+
+---
+
+## 6. Final go/no-go
+
+✅ **GO.**
+
+**Why:**
+- All architectural assumptions validated. No showstoppers found.
+- The 5 corrections in §2 are concrete, small, and isolated to specific phases.
+- The 12 medium adjustments in §3 are sensible and most are already implicit in the plan.
+- The plan's rev-2 effort estimate (25–35 days) holds with corrections (~28.5 days).
+
+**Day-1 first commits (in order):**
+1. `strix/llm/anthropic_cache_wrapper.py` — `AnthropicCachingLitellmModel`.
+2. `strix/runtime/strix_docker_client.py` — `StrixDockerSandboxClient`.
+3. `strix/orchestration/bus.py` — `AgentMessageBus`.
+4. `strix/orchestration/filter.py` — `inject_messages_filter`.
+5. `strix/orchestration/hooks.py` — `StrixOrchestrationHooks`.
+6. `strix/tools/_decorator.py` — `strix_tool` factory.
+7. `strix/llm/multi_provider_setup.py` — `MultiProviderMap` wiring + `StrixModelProvider`.
+
+These seven files (~600 LOC total) form the migration's load-bearing foundation. Everything else is incremental ports onto this foundation.
+
+Branch is already on `harness-migration`. Ready when you are.
diff --git a/AUDIT_R2.md b/AUDIT_R2.md
new file mode 100644
index 00000000..7bb61904
--- /dev/null
+++ b/AUDIT_R2.md
@@ -0,0 +1,243 @@
+# Migration Audit — Round 2 Findings
+
+> Five-agent deep verification of areas not covered in `AUDIT.md`: SDK surface area exhaustive catalog, Strix port-readiness contracts, concurrency forensics, data-flow mapping, error-handling forensics. Source-verified against `openai-agents` v0.14.6 and Strix at `9fb1012`. Adds seven new concrete corrections to the migration plan.
+
+---
+
+## 1. New corrections discovered
+
+These are additive to the five blockers in `AUDIT.md` §2. Each was missed in earlier rounds.
+
+### 1.1 [CRITICAL] notes/notes.jsonl writes are not lock-protected
+
+**Defect.** `strix/tools/notes/notes_actions.py:40-54` (`_append_note_event`) opens the JSONL file and writes without holding `_notes_lock`. Today this is invisible because Strix daemon-thread subagents serialize on Python's GIL during the `f.write(...)` call — but **the file `open + seek + write + close` is not atomic across multiple threads**. Two simultaneous notes operations from sibling agents can interleave bytes mid-line, corrupting the JSONL file.
+
+**Post-migration risk.** Same bug. SDK runs tool calls in parallel within a turn (`run_internal/tool_execution.py:1414, 1424`), so two `create_note` invocations on different agents in the same event loop tick will hit the file simultaneously.
+
+**Fix.**
+
+```python
+# notes_actions.py:40-54 (today)
+def _append_note_event(op, note_id, note=None):
+ notes_path = _get_notes_jsonl_path()
+ if not notes_path:
+ return
+ event = {"timestamp": datetime.now(UTC).isoformat(), "op": op, "note_id": note_id}
+ if note is not None:
+ event["note"] = note
+ with _notes_lock: # <- ADD
+ with notes_path.open("a", encoding="utf-8") as f:
+ f.write(f"{json.dumps(event, ensure_ascii=True)}\n")
+```
+
+Same fix for `_persist_wiki_note()` (write to `wiki/<slug>.md`).
+
+**Apply during Phase 2** when porting notes tool.
+
+### 1.2 [CRITICAL] events.jsonl writes are not lock-protected either
+
+**Defect.** `strix/telemetry/tracer.py:162-268` (`_emit_event` → `_append_event_record`) calls `append_jsonl_record(self.events_file_path, record)` **without** acquiring the lock that `_get_events_write_lock()` (line 106-108) is designed to provide. The lock exists in the codebase but is unused at the call site.
+
+**Post-migration risk.** Even more acute. Our custom `TracingProcessor` will write SDK spans → `events.jsonl` from multiple concurrent agent tasks. JSONL corruption guaranteed under load.
+
+**Fix.**
+
+```python
+# tracer.py:_append_event_record (today)
+def _append_event_record(self, record):
+ try:
+ with self._get_events_write_lock(): # <- ADD
+ append_jsonl_record(self.events_file_path, record)
+ except OSError:
+ logger.exception("Failed to append JSONL event record")
+```
+
+In our custom processor (the migration-phase replacement), apply the same lock.
+
+**Apply in Phase 1** when wiring the custom `TracingProcessor`.
+
+### 1.3 [HIGH] Subagent crash silent — parent never learns
+
+**Defect.** `strix/tools/agents_graph/agents_graph_actions.py:281-287` catches the daemon-thread exception, sets the graph node status to `"error"`, and **re-raises** inside the thread. The thread dies. The parent agent calling `wait_for_message(timeout=600)` polls for 600s and resumes with "Timed out" — never knows the child was dead.
+
+**Post-migration risk.** Same problem in different shape. If a child `Runner.run` task raises, our `MessageBus.tasks[child_id]` is in `done` state with exception, but parent's `wait_for_message` only checks `inboxes`.
+
+**Fix.** In `StrixOrchestrationHooks.on_agent_end` (Phase 3), if exit was due to exception, push a synthetic completion report to parent's inbox so `call_model_input_filter` surfaces it on parent's next turn:
+
+```python
+class StrixOrchestrationHooks(RunHooks):
+ async def on_agent_end(self, ctx, agent, output):
+ bus = ctx.context["bus"]
+ me = ctx.context["agent_id"]
+ parent = bus.parent_of.get(me)
+ # Detect crash: did agent_finish run? if not, output is None or the run errored.
+ crashed = (output is None) or (ctx.context.get("agent_finish_called") is not True)
+ if crashed and parent is not None:
+ await bus.send(parent, {
+ "from": me,
+ "content": f"<agent_crash agent_id='{me}' name='{bus.names.get(me)}'>"
+ f"Agent terminated without calling agent_finish. "
+ f"Parent should not wait further on this child."
+ f"</agent_crash>",
+ "type": "crash",
+ })
+ await bus.finalize(me, "completed" if not crashed else "crashed")
+```
+
+The `agent_finish_called` flag is set by the `agent_finish` tool body. Also add a watchdog in the bus: any task in `tasks` whose `done()` is True but `bus.statuses` is still `running` is reaped.
+
+### 1.4 [HIGH] Cancellation cascade incomplete
+
+**Defect.** Strix's `stop_agent(agent_id)` (`agents_graph_actions.py:688-748`) requires explicit invocation. Today if the user Ctrl+C's the root, only the root agent loop is cancelled — children running in daemon threads keep executing.
+
+**Post-migration risk.** Same. SDK's `result.cancel()` cancels the root task; child `Runner.run` tasks (spawned by `asyncio.create_task` in `create_agent` tool) are NOT cancelled by SDK and continue.
+
+**Fix.** Top-level run wrapper walks `bus.parent_of` to enumerate descendants and explicitly cancels each:
+
+```python
+# strix/orchestration/cancellation.py
+async def cancel_run_with_descendants(bus: AgentMessageBus, root_agent_id: str):
+ descendants = []
+ queue = [root_agent_id]
+ while queue:
+ aid = queue.pop()
+ descendants.append(aid)
+ queue.extend(child for child, parent in bus.parent_of.items() if parent == aid)
+ for aid in reversed(descendants): # leaves first
+ task = bus.tasks.get(aid)
+ if task is not None and not task.done():
+ task.cancel()
+ # Wait briefly for cancellations to settle
+ await asyncio.gather(*(t for t in bus.tasks.values() if not t.done()),
+ return_exceptions=True)
+```
+
+Wire from CLI signal handler and TUI stop button.
+
+### 1.5 [MEDIUM] Memory compressor has no graceful fallback
+
+**Defect.** `strix/llm/memory_compressor.py:152-219` makes a separate LLM call to summarize old messages. If that call times out or fails, the exception bubbles to the agent loop and **fails the iteration** — the only purpose of the compressor (avoiding context-window overflow) is undermined by an even harsher failure.
+
+**Post-migration risk.** Same. Custom `Session` subclass calling our compressor inherits the brittleness.
+
+**Fix.** Wrap compressor invocations:
+
+```python
+# In our custom Session subclass
+async def _compress_if_needed(self, items):
+ try:
+ return await self._compressor.compress_history(items)
+ except (asyncio.TimeoutError, Exception) as e:
+ logger.warning("Compression failed (%s); returning uncompressed history", e)
+ return items # let context-window error happen later if it must
+```
+
+The downstream context-window error (if it happens) is itself retryable via SDK retry policies, so we degrade rather than fail.
+
+### 1.6 [MEDIUM] 401 retry policy mismatch between Strix and SDK
+
+**Detail.** Strix's `_should_retry` (`llm/llm.py:326-330`) treats `status_code is None` as retryable AND defers HTTP codes to `litellm._should_retry(code)` — which does NOT retry 401. So Strix fails fast on auth errors.
+
+The SDK's retry default (configurable via `ModelRetrySettings.retry_policies`) may include 401 retries depending on policy composition. We don't want to retry 401 (it wastes time and clutters traces).
+
+**Fix.** Explicit retry policy in our `RunConfig` factory:
+
+```python
+from agents.retry import retry_policies, ModelRetrySettings, ModelRetryBackoffSettings
+
+DEFAULT_RETRY = ModelRetrySettings(
+ max_retries=5,
+ backoff=ModelRetryBackoffSettings(
+ initial_delay=2.0, multiplier=2.0, max_delay=90.0, jitter=0.0,
+ ),
+ policy=retry_policies.any(
+ retry_policies.network_error(),
+ retry_policies.http_status([429, 500, 502, 503, 504]),
+ # explicitly NOT including 401, 403, 400
+ ),
+)
+```
+
+Bake into our `make_run_config()` factory so every Strix run gets it automatically.
+
+### 1.7 [MEDIUM] `_completed_agent_llm_totals` read without lock from tracer
+
+**Defect.** `agents_graph_actions.py:35` declares the dict; finalize writes hold `_agent_llm_stats_lock`. Tracer's `get_total_llm_stats()` (`telemetry/tracer.py:801-834`) reads it without acquiring the lock. Possible partial-update read.
+
+**Post-migration risk.** Reduced (single asyncio loop), but our `MessageBus.total_stats()` should still snapshot under the bus's own `asyncio.Lock`.
+
+**Fix.** Already in `MessageBus` design — `total_stats` acquires lock. Just confirm the implementation does this.
+
+---
+
+## 2. Round 1 verification snapshot
+
+What the five Round 1 audits actually verified:
+
+| Audit | Output | Key new finding |
+|---|---|---|
+| 1.1 SDK surface | Exhaustive catalog (~55 sections) — every `Agent` field, every `RunConfig` knob, every `ModelSettings` field, every span type, every error class, every hook, every Session impl, every Model interface method | No surprises — confirms Strix-side decisions in plan |
+| 1.2 Strix port-readiness | Per-tool exact contract reference (params, return shapes, side effects, threading) | Confirms tool-level mapping; surfaces no new blockers |
+| 1.3 Concurrency forensics | Lock-by-lock inventory both repos + post-migration topology | **Discovered the two JSONL race conditions (§1.1, §1.2 above) and cancellation cascade gap (§1.4)** |
+| 1.4 Data flow & persistence | Every artifact + every in-memory structure mapped pre/post | Confirms invariants survive migration; no data loss paths |
+| 1.5 Error handling forensics | 60+ failure modes catalogued with detection/error class/retry/fallback/visibility | **Discovered subagent-crash silence (§1.3), compressor fail-open (§1.5), 401 retry mismatch (§1.6)** |
+
+---
+
+## 3. Updated correction set (consolidated from `AUDIT.md` + Round 1)
+
+| # | Severity | Defect | Phase to apply |
+|---|---|---|---|
+| **C1** | Blocker | Strix tool-server slot serialization vs SDK parallel calls (`AUDIT.md` §2.1) | Phase 0 (set safe `parallel_tool_calls=False`/`isolate_parallel_failures=False`) → Phase 6 (relax) |
+| **C2** | Blocker | Anthropic `cache_control` placement on system message (`AUDIT.md` §2.2) | Phase 0 (`AnthropicCachingLitellmModel`) |
+| **C3** | Blocker | `DockerSandboxClient` subclass needs full method-body copy (`AUDIT.md` §2.3) | Phase 0 (`StrixDockerSandboxClient`) |
+| **C4** | Blocker | Subagent `tool_use_behavior={"stop_at_tool_names": [...]}` required (`AUDIT.md` §2.4) | Phase 3 (multi-agent) |
+| **C5** | High | Streaming TUI integration via `StrixStreamAccumulator` (`AUDIT.md` §2.5) | Phase 5 |
+| **C6** | Critical | notes JSONL write race (Round 1 §1.1) | Phase 2 (notes tool port) |
+| **C7** | Critical | events.jsonl write race (Round 1 §1.2) | Phase 1 (custom processor) |
+| **C8** | High | Subagent crash silent — synthetic completion-report on `on_agent_end` (Round 1 §1.3) | Phase 3 |
+| **C9** | High | Cancellation cascade walks `bus.parent_of` tree (Round 1 §1.4) | Phase 3 |
+| **C10** | Medium | Memory compressor try/except → degrade to uncompressed (Round 1 §1.5) | Phase 1 (custom Session) |
+| **C11** | Medium | Retry policy excludes 401/403/400 (Round 1 §1.6) | Phase 1 (RunConfig factory) |
+| **C12** | Medium | Bus stats snapshot under lock (Round 1 §1.7) | Phase 3 (already in design) |
+
+Plus the original twelve medium adjustments from `AUDIT.md` §3 (M1–M12).
+
+---
+
+## 4. Verified-safe areas (no further investigation needed)
+
+| Area | Verification |
+|---|---|
+| `call_model_input_filter` retry safety | Filter runs once per turn; output captured in lambda closure, not re-invoked on retry. Inbox drain is safe. (Round 1 §1.1 confirmed via `turn_preparation.py:55-80` + `model_retry.py:34-35`.) |
+| `asyncio.create_task(Runner.run)` isolation | Each task gets fresh `RunContextWrapper`; contextvars isolated per task; no global state mutation in `Runner.run`. |
+| Shared `SandboxRunConfig.session` across parallel runs | SDK does NOT auto-tear-down sandbox sessions; safe to reuse one session across N children. |
+| `RunHooks.on_agent_end` firing | Once per `Runner.run` invocation (verified `turn_resolution.py:204-255`). |
+| `RunContextWrapper.context` mutability | Dict by-reference; mutations persist within and across turns. |
+| Sync function tools | SDK auto-threads sync `@function_tool` bodies via `asyncio.to_thread` (`tool.py:1820-1829`) — drop manual offload. |
+| Custom Docker image | `DockerSandboxClientOptions(image=str)` pass-through; no assumed binaries. |
+| `Manifest.entries` superset of Strix needs | `LocalDir`, `LocalFile`, `GitRepo`, `Mount` types cover all Strix patterns. |
+| MultiProvider routing | `MultiProviderMap.add_provider("strix", StrixModelProvider())` works as designed. |
+| Tracing API | `set_trace_processors([...])` disables defaults; custom processors can write to JSONL/OTel. |
+| `RunState.to_json/from_json` | Serializable (`CURRENT_SCHEMA_VERSION=1.9`); cross-process resumable. |
+| Sandbox capability hooks | `process_manifest`, `tools()`, `instructions()`, `bind()` cover `CaidoCapability` needs. |
+
+---
+
+## 5. Areas flagged for monitoring during implementation
+
+These aren't blockers but warrant attention during Phase work:
+
+- **Browser singleton event-loop init race** — low risk, double-check pattern recommended in `_ensure_event_loop` (`browser_instance.py:34-48`).
+- **`agent_tasks` dict in tool server** — currently unprotected; if we ever switch uvicorn to threaded workers, needs `asyncio.Lock`.
+- **SQLiteSession async-task ordering** — `threading.RLock` doesn't serialize asyncio tasks deterministically. Mitigated by per-child Sessions (already in plan).
+- **Trace processor memory pressure on long runs** — `BatchTraceProcessor` accumulates spans; periodic `force_flush()` recommended.
+- **Bus.inboxes resize race** — asyncio.Lock around all dict mutations covers this; verify lock scope in implementation.
+
+---
+
+## 6. Round 1 outcome
+
+**No new architectural blockers.** Plan structure remains sound. Twelve corrections (five from `AUDIT.md`, seven from Round 1) all bounded, all implementable in their assigned phase.
+
+Next: **Round 2** dispatches deep-dives on file-by-file implementation specs, per-tool migration contracts, test plans, and cross-cutting concerns. Round 2 output is the actual day-1 engineering reference, not more audit findings.
diff --git a/AUDIT_R3.md b/AUDIT_R3.md
new file mode 100644
index 00000000..51aff623
--- /dev/null
+++ b/AUDIT_R3.md
@@ -0,0 +1,476 @@
+# Migration Audit — Round 3 Findings
+
+> Five new deep-dives covering scenario walkthroughs, type/signature compatibility, CLI/interface migration, pathological edge cases, and build/deps/deployment. Adds 13 more concrete corrections (C13–C25) and 3 critical type-signature fixes that must be applied to `PLAYBOOK.md` before Phase 0 starts.
+
+---
+
+## 1. Critical type/signature fixes (PLAYBOOK skeletons are wrong)
+
+These three fixes are **applied inline to `PLAYBOOK.md`** in this round. The skeletons as written would not compile against SDK v0.14.6.
+
+### F1 — `AnthropicCachingLitellmModel` method signature wrong
+
+**PLAYBOOK §2.1 was:** `async def get_response(self, *, system_instructions, input, model_settings, ...)`
+
+**SDK reality** (`models/interface.py:57-89`): The first 7 params (`system_instructions, input, model_settings, tools, output_schema, handoffs, tracing`) are **positional-first**, then `*,` then keyword-only (`previous_response_id`, `conversation_id`, `prompt`).
+
+Same fix applies to `stream_response`. Both must drop the leading `*,` and pass the first 7 params positionally to `super()`.
+
+**Status:** Inline-corrected in `PLAYBOOK.md` §2.1 below.
+
+### F2 — `RunHooks` lifecycle hook signatures wrong
+
+**PLAYBOOK §2.5 was:** `on_agent_start(self, ctx, agent)` and `on_agent_end(self, ctx, agent, output)` typed `ctx` as `RunContextWrapper`.
+
+**SDK reality** (`lifecycle.py:37-59`): These two specific hooks receive `context: AgentHookContext[TContext]`, not `RunContextWrapper`. `AgentHookContext` is a different generic wrapper with the same `.context` attribute pattern but a distinct type.
+
+Also: `on_tool_end(self, ctx, agent, tool, result)` — the `result` parameter is typed `str`, not `Any`.
+
+**Status:** Inline-corrected in `PLAYBOOK.md` §2.5 below.
+
+### F3 — `TracingProcessor` hook methods are SYNC
+
+**PLAYBOOK §2.9 was:** All hook methods (`on_trace_start`, `on_trace_end`, `on_span_start`, `on_span_end`, `force_flush`, `shutdown`) shown without explicit `async` keyword — implementation accidentally implied async.
+
+**SDK reality** (`tracing/processor_interface.py:53-129`): All hooks are **synchronous** (`def`, not `async def`). Our processor must use sync methods. JSONL writes are sync I/O which is fine; if we ever want async export, we'd need to schedule via `asyncio.run_coroutine_threadsafe()` from the sync hook.
+
+**Status:** Inline-corrected in `PLAYBOOK.md` §2.9 below.
+
+---
+
+## 2. New corrections (C13–C25)
+
+Additive to the 12 corrections in `AUDIT.md` + `AUDIT_R2.md`.
+
+### C13 [HIGH] Bus must clear inbox/parent_of/names on finalize (Round 3.4)
+
+**Defect.** When an agent finishes, `bus.finalize` only updates statuses and stats — but children whose parent already finished may still call `bus.send(parent_id, msg)`, accumulating messages in `bus.inboxes[parent_id]` forever. Memory leak bounded by agent count × messages per cycle.
+
+**Fix.** Update `AgentMessageBus.finalize()`:
+
+```python
+async def finalize(self, agent_id: str, status: str) -> None:
+ async with self._lock:
+ self.statuses[agent_id] = status
+ self.stats_completed[agent_id] = self.stats_live.pop(agent_id, {})
+ self.inboxes.pop(agent_id, None) # NEW
+ self.parent_of.pop(agent_id, None) # NEW
+ self.names.pop(agent_id, None) # NEW
+```
+
+**Apply in Phase 3** (in `strix/orchestration/bus.py`).
+
+### C14 [HIGH] `inject_messages_filter` must be defensive
+
+**Defect.** If a bug in the filter raises, SDK treats it as a model invocation failure and retries. Filter raises on every retry → run fails after `max_retries` exhausted.
+
+**Fix.** Wrap filter body in try/except; return unmodified `data.model_data` on exception:
+
+```python
+async def inject_messages_filter(data: CallModelData) -> ModelInputData:
+ try:
+ if not isinstance(data.context, dict):
+ return data.model_data
+ bus = data.context.get("bus")
+ agent_id = data.context.get("agent_id")
+ if bus is None or agent_id is None:
+ return data.model_data
+ pending = await bus.drain(agent_id)
+ if not pending:
+ return data.model_data
+ new_input = list(data.model_data.input)
+ for msg in pending:
+ # ... XML wrapping
+ return ModelInputData(input=new_input, instructions=data.model_data.instructions)
+ except Exception:
+ logger.exception("inject_messages_filter failed; proceeding without injection")
+ return data.model_data
+```
+
+**Apply in Phase 3** (in `strix/orchestration/filter.py`).
+
+### C15 [HIGH] `RunHooks` must be defensive
+
+**Defect.** If our hook bodies raise (e.g., bus operation fails, tracer disk error), exception propagates and tears down the run.
+
+**Fix.** Each hook wraps its body in try/except; logs and continues:
+
+```python
+async def on_llm_start(self, context, agent, system_prompt, input_items):
+ try:
+ # ... mutate input_items, increment turn count, etc.
+ except Exception:
+ logger.exception("on_llm_start failed")
+
+# Same for on_llm_end, on_agent_start, on_agent_end, on_tool_start, on_tool_end, on_handoff
+```
+
+**Apply in Phase 3** (in `strix/orchestration/hooks.py`).
+
+### C16 [HIGH] Custom `TracingProcessor` must catch disk errors
+
+**Defect.** PLAYBOOK §2.9 `_emit()` opens file with `"a"` mode; OSError (disk full, permission denied) propagates from sync hook. SDK's hook caller may not gracefully handle. Run dies.
+
+**Fix.** Wrap `_emit` body in try/except; log and continue:
+
+```python
+def _emit(self, event: dict[str, Any]) -> None:
+ try:
+ clean = self.sanitizer.sanitize(event)
+ with _lock_for(self.events_path):
+ with self.events_path.open("a", encoding="utf-8") as f:
+ f.write(json.dumps(clean, ensure_ascii=True) + "\n")
+ except OSError:
+ logger.exception("Failed to write event to JSONL")
+```
+
+**Apply in Phase 1** (in `strix/telemetry/strix_processor.py`).
+
+### C17 [MEDIUM] `StrixModelProvider` must validate model alias
+
+**Defect.** If `STRIX_LLM=strix/typo-model-name` (alias not in `STRIX_MODEL_MAP`), our provider falls through to `(model_name, model_name)` and the LLM call later fails with provider's "model not found" — opaque diagnostic.
+
+**Fix.** Validate at `get_model()` entry; raise `UserError` with the list of valid aliases:
+
+```python
+def get_model(self, model_name: str | None) -> Model:
+ if model_name is None:
+ raise UserError("Model name required for StrixModelProvider")
+ if model_name not in STRIX_MODEL_MAP:
+ raise UserError(
+ f"Unknown Strix alias '{model_name}'. "
+ f"Valid: {list(STRIX_MODEL_MAP.keys())}"
+ )
+ api_model, _ = STRIX_MODEL_MAP[model_name]
+ if "anthropic/" in api_model or "claude" in api_model.lower():
+ return AnthropicCachingLitellmModel(model=api_model, base_url=STRIX_API_BASE)
+ return LitellmModel(model=api_model, base_url=STRIX_API_BASE)
+```
+
+**Apply in Phase 1** (in `strix/llm/multi_provider_setup.py`).
+
+### C18 [MEDIUM] Model output size cap on sandbox tools
+
+**Defect.** A tool returning 50MB binary output (e.g., browser screenshot of a huge map) gets base64-encoded into JSON; httpx loads the response into RAM; OOM on small hosts.
+
+**Fix.** Configure `httpx.Limits(max_content_size=...)` in `post_to_sandbox`:
+
+```python
+_TIMEOUT = httpx.Timeout(connect=10.0, read=150.0, write=150.0, pool=150.0)
+_LIMITS = httpx.Limits(max_connections=10, max_keepalive_connections=5)
+
+async def post_to_sandbox(ctx, tool_name, kwargs) -> dict:
+ # ...
+ async with httpx.AsyncClient(timeout=_TIMEOUT, limits=_LIMITS) as client:
+ r = await client.post(url, json=body, headers=headers)
+ if int(r.headers.get("content-length", 0)) > 50_000_000:
+ return {"error": "Sandbox response too large (>50MB)"}
+ # ...
+```
+
+Plus: cap on the sandbox side too (tool server limits its own response payload).
+
+**Apply in Phase 2** (in `strix/tools/_sandbox_dispatch.py`).
+
+### C19 [MEDIUM] `tool_choice="required"` requires at least one enabled tool
+
+**Defect.** If `is_enabled` callbacks gate out all tools and `ModelSettings(tool_choice="required")`, model has no legal response. SDK raises `ModelBehaviorError`. Run fails opaquely.
+
+**Fix.** Assert at agent build time:
+
+```python
+def build_strix_agent(name, tools, ...) -> Agent:
+ enabled_count = len([t for t in tools if not _statically_disabled(t)])
+ if enabled_count == 0:
+ raise UserError(f"Agent {name} has no enabled tools but tool_choice='required'")
+ return Agent(name=name, tools=tools, ...)
+```
+
+**Apply in Phase 1** (in agent factory).
+
+### C20 [MEDIUM] Per-tool `timeout_behavior` discrimination
+
+**Defect.** If `timeout_behavior="error_as_result"` on a critical sandbox tool (e.g., `terminal_execute`), model sees the timeout error string and may retry the same tool with same args → infinite loop.
+
+**Fix.** For critical sandbox tools, use `timeout_behavior="raise_exception"` so the model is told via SDK's error machinery that the tool genuinely failed (not just timed out gracefully). For idempotent local tools (notes, todos), `error_as_result` is fine.
+
+**Apply in Phase 2** — when porting each tool, pick the appropriate behavior.
+
+### C21 [MEDIUM] `make_run_config` and `make_agent_context` need overrides
+
+**Defect.** Plan §H1: today there's no path for per-run override of `model_settings` (e.g., user wants `tool_choice="auto"` for a specific run). And `is_whitebox` flag isn't propagated to context — wiki auto-update on subagent finish (M10) reads `ctx.context.get("is_whitebox")` but it's never set.
+
+**Fix.**
+
+```python
+def make_run_config(*, sandbox_session, bus, model="strix/claude-sonnet-4.6",
+ max_turns=300, model_settings_override: dict | None = None) -> RunConfig:
+ base_settings = ModelSettings(parallel_tool_calls=False, tool_choice="required", retry=...)
+ if model_settings_override:
+ base_settings = base_settings.model_copy(update=model_settings_override)
+ return RunConfig(model_settings=base_settings, ...)
+
+
+def make_agent_context(*, bus, sandbox_session, sandbox_token,
+ tool_server_host_port, caido_host_port, agent_id, agent_name,
+ parent_id, tracer, model_settings, max_turns=300,
+ is_whitebox: bool = False, # NEW
+ diff_scope: dict | None = None, # NEW (J1)
+ run_id: str | None = None) -> dict: # NEW (run-id propagation)
+ return {
+ "bus": bus, "sandbox_session": sandbox_session,
+ "sandbox_token": sandbox_token,
+ "tool_server_host_port": tool_server_host_port,
+ "caido_host_port": caido_host_port,
+ "agent_id": agent_id, "agent_name": agent_name,
+ "parent_id": parent_id, "tracer": tracer,
+ "model_settings": model_settings, "max_turns": max_turns,
+ "turn_count": 0, "agent_finish_called": False,
+ "is_whitebox": is_whitebox,
+ "diff_scope": diff_scope,
+ "run_id": run_id,
+ }
+```
+
+**Apply in Phase 1** (in `strix/run_config_factory.py`).
+
+### C22 [MEDIUM] `finish_scan` must check children status before exit
+
+**Defect.** Strix today's `finish_scan` validates that all child agents are not running/stopping (`tools/finish/finish_actions.py:98`). PLAYBOOK §4.2 didn't carry this forward. Without the check, root could finish while children are still in-flight.
+
+**Fix.** Inside `finish_scan` tool body:
+
+```python
+@strix_tool(timeout=30)
+async def finish_scan(ctx, executive_summary: str, methodology: str,
+ technical_analysis: str, recommendations: str) -> str:
+ if ctx.context.get("parent_id") is not None:
+ return "Error: finish_scan is for the root agent only. Subagents must call agent_finish."
+ bus = ctx.context["bus"]
+ me = ctx.context["agent_id"]
+ async with bus._lock:
+ in_flight = [
+ child_id for child_id, parent in bus.parent_of.items()
+ if parent == me and bus.statuses.get(child_id) in ("running", "waiting")
+ ]
+ if in_flight:
+ names = [bus.names.get(c, c) for c in in_flight]
+ return (
+ f"Cannot finish: subagents still running: {names}. "
+ f"Wait for completion (or call stop_agent) before finishing the scan."
+ )
+ ctx.context["agent_finish_called"] = True
+ # ... write narrative fields, persist final report
+ return "Scan completed. Report written."
+```
+
+**Apply in Phase 2** (when porting `finish_scan`).
+
+### C23 [MEDIUM] Diff-scope context injection point
+
+**Defect.** Plan §J1: PLAYBOOK doesn't say where the diff scope context (from `resolve_diff_scope_context()`) is injected post-migration.
+
+**Fix.** Two-part:
+1. CLI parses `--scope-mode=diff` + `--diff-base=...` and computes `DiffScopeResult` (same as today).
+2. The `instruction_block` from the result is **prepended to the user's instruction** in the first message of `Runner.run`. (Same as Strix today; the agent sees it as part of its task.)
+
+```python
+# strix/interface/cli.py (or main.py)
+diff_scope = resolve_diff_scope_context(args)
+user_instruction = args.instruction or ""
+if diff_scope.instruction_block:
+ user_instruction = f"{diff_scope.instruction_block}\n\n{user_instruction}".strip()
+
+context = make_agent_context(..., diff_scope=diff_scope.metadata, ...)
+result = await Runner.run(
+ agent,
+ input=[{"role": "user", "content": user_instruction or "Conduct a thorough penetration test."}],
+ ...
+)
+```
+
+**Apply in Phase 5** (CLI/interface migration).
+
+### C24 [MEDIUM] Run-name uniqueness + Docker availability checks
+
+**Defect.** Plan §28 + §32: nothing prevents two parallel `strix` invocations colliding on `run_name` and competing for the same container name. And nothing surfaces a clear error when Docker daemon isn't running.
+
+**Fix.** Pre-flight checks at CLI startup:
+
+```python
+def main():
+ args = parse_arguments()
+ apply_config_override(args.config)
+ if args.use_sdk_harness:
+ if not _docker_daemon_available():
+ sys.exit("Docker daemon unavailable. Start Docker Desktop / dockerd and try again.")
+ run_dir = Path("strix_runs") / args.run_name
+ if run_dir.exists() and (run_dir / "events.jsonl").exists():
+ sys.exit(
+ f"Run '{args.run_name}' already exists at {run_dir}. "
+ f"Use a different --name or rm the directory."
+ )
+ # ... continue with scan
+```
+
+**Apply in Phase 5** (in `strix/interface/main.py`).
+
+### C25 [MEDIUM] Hook cancel mode mapping + cleanup
+
+**Defect.** Plan §C8: PLAYBOOK §C9 mentions `result.cancel(mode=...)` but doesn't specify which mode for which trigger.
+
+**Fix.**
+- **Ctrl+C from user** → `result.cancel(mode="immediate")` + `await bus.cancel_descendants(root_id)`.
+- **TUI "stop agent" button (graceful)** → `result.cancel(mode="after_turn")` + `await bus.cancel_descendants(root_id)`.
+- **`stop_agent(child_id)` tool called by parent** → directly `bus.tasks[child_id].cancel()`.
+- **Run finished naturally** → no cancellation needed; `on_agent_end` hooks finalize.
+
+**Apply in Phase 5** (signal handler + TUI binding).
+
+---
+
+## 3. New scenario gaps from walkthrough audit (Round 3.1)
+
+| # | Scenario | Gap | Fix |
+|---|---|---|---|
+| **W1** | Cold-start single-agent | Most gaps already in C1–C12 | Use new C13–C25 |
+| **W2** | Multi-agent parallel | `finish_scan` had no children-running check | C22 |
+| **W3** | Mid-run Ctrl+C | Cancel-mode mapping ambiguous | C25 |
+| **W4** | Subagent silent crash | Background post-invoke task exceptions don't trigger crash detection | Document — exceptions in `post_invoke_task` are logged async, not via `on_agent_end`. Bus watchdog optional. |
+| **W5** | Compressor cascade fail | After compression fails once, next iteration retries forever | Set `_compression_disabled=True` in context after first failure; subsequent calls skip. Apply in Phase 1 custom Session. |
+| **W6** | Container dies mid-run | No periodic liveness check | Optional: background asyncio task pings `/health` every N turns. Phase 5 / Phase 6 enhancement. |
+| **W7** | Whitebox wiki on finish | `is_whitebox` not propagated to context | C21 |
+| **W8** | RunConfig override | No injection point | C21 |
+| **W9** | Resume from RunState | Out of scope for MVP | Defer; document as Phase 7+ |
+| **W10** | Diff-scope mode | Injection point unspecified | C23 |
+
+---
+
+## 4. CLI/interface migration spec (Round 3.3 highlights)
+
+Full spec in the audit; key takeaways folded into PLAYBOOK and corrections above.
+
+**Survives unchanged:** All argparse flags, `Config` class, target inference, run-name generation, scope-diff resolution, helper utilities (`format_*`, `infer_target_type`, `clone_repository`, etc.), exit code logic.
+
+**Refactor needed:**
+
+- **`run_cli()` headless** — wrap `Runner.run_streamed()` with `StrixStreamAccumulator`; same Rich panels, same exit codes.
+- **`run_tui()` interactive** — Textual app subscribes to tracer reactive fields; tracer fed by accumulator + hooks.
+- **Vulnerability popup** — direct call from `create_vulnerability_report` tool body to `tracer.vulnerability_found_callback` (simpler than `ToolOutputGuardrail`; both viable).
+- **Live stats** — `build_live_stats_text(tracer, agent_config)` reads tracer; tracer fed via hooks; no change to read path.
+- **Interactive resume** — after `Runner.run()` returns, re-run with appended history + new user message; SDK `Session` makes this clean.
+
+**`infer_target_type()` → Manifest entries:**
+
+| Inferred type | Manifest action |
+|---|---|
+| `local_code` | `LocalDir(src=Path(...))` mounted under `/workspace/<subdir>` |
+| `repository` (git URL) | Pre-clone via existing `clone_repository()`; mount cloned dir as `LocalDir`. (Keep pre-clone logic; don't use SDK's `GitRepo` until after MVP.) |
+| `web_application` / `domain` / `ip_address` | No mount; agent reaches via Caido proxy |
+
+---
+
+## 5. Build/deps/deployment (Round 3.5 highlights)
+
+### pyproject.toml diff
+
+**Drop:**
+```diff
+- "litellm[proxy]>=1.81.1,<1.82.0",
+```
+
+**Add:**
+```diff
++ "openai-agents[litellm]==0.14.6",
+```
+
+**Transitive new (no action):** `griffelib`, `mcp`, `websockets`, `types-requests`, `openai>=2.26.0`. Pulled in by `openai-agents`.
+
+**Container image (`containers/Dockerfile`):** **NO changes.** `[sandbox]` extras stay in image. SDK's host-side code does not run inside the container.
+
+**`strix.spec` (PyInstaller):** Add SDK hidden imports:
+
+```python
+hiddenimports += [
+ "agents", "agents.agent", "agents.run", "agents.run_config",
+ "agents.memory.session", "agents.memory.sqlite_session",
+ "agents.sandbox.sandboxes.docker", "agents.sandbox.manifest",
+ "agents.sandbox.capabilities.capability", "agents.sandbox.entries",
+ "agents.extensions.models.litellm_model",
+ "agents.tool", "agents.tool_context", "agents.tool_guardrails",
+ "agents.lifecycle", "agents.guardrail",
+ "agents.tracing.processor_interface", "agents.tracing.spans", "agents.tracing.traces",
+ "agents.models.interface", "agents.models.multi_provider",
+ "agents.retry", "mcp", "websockets",
+]
+```
+
+### Config bridge
+
+`strix/config/config.py` adds an env-var bridge before SDK init:
+
+```python
+def bridge_to_sdk_env() -> None:
+ """Map legacy Strix env vars to SDK-native names where applicable."""
+ if Config.get("llm_api_key") and not os.environ.get("OPENAI_API_KEY"):
+ os.environ["OPENAI_API_KEY"] = Config.get("llm_api_key")
+ if Config.get("llm_api_base") and not os.environ.get("OPENAI_BASE_URL"):
+ os.environ["OPENAI_BASE_URL"] = Config.get("llm_api_base")
+```
+
+Call from `main.py` before any SDK import.
+
+### CI
+
+`.github/workflows/build-release.yml` — unchanged. Add new test workflow (`tests.yml`) for pytest + mypy + ruff on PRs (recommended but not blocking for cutover).
+
+### Feature flag
+
+`STRIX_USE_SDK_HARNESS` env var, default `0`. CLI entry checks; routes to legacy or SDK harness implementation.
+
+---
+
+## 6. Consolidated correction register (full)
+
+After Rounds 1, 2, 3 — twenty-five corrections to apply.
+
+| # | Severity | Phase | Source | Topic |
+|---|---|---|---|---|
+| C1 | Blocker | 1/6 | AUDIT.md §2.1 | Tool-server slot serialization vs SDK parallel calls |
+| C2 | Blocker | 0 | AUDIT.md §2.2 | Anthropic cache-control on system message |
+| C3 | Blocker | 0 | AUDIT.md §2.3 | DockerSandboxClient subclass |
+| C4 | Blocker | 3 | AUDIT.md §2.4 | Subagent `tool_use_behavior` |
+| C5 | High | 5 | AUDIT.md §2.5 | StrixStreamAccumulator |
+| C6 | Critical | 2 | AUDIT_R2.md §1.1 | Notes JSONL write lock |
+| C7 | Critical | 1 | AUDIT_R2.md §1.2 | events.jsonl write lock |
+| C8 | High | 3 | AUDIT_R2.md §1.3 | Subagent crash detection |
+| C9 | High | 3 | AUDIT_R2.md §1.4 | Cancellation cascade |
+| C10 | Medium | 1 | AUDIT_R2.md §1.5 | Compressor try/except |
+| C11 | Medium | 1 | AUDIT_R2.md §1.6 | Retry policy excludes 401/403/400 |
+| C12 | Medium | 3 | AUDIT_R2.md §1.7 | Stats snapshot under lock |
+| **F1** | **Critical** | **0** | **AUDIT_R3 §1** | **AnthropicCachingLitellmModel signature** |
+| **F2** | **Critical** | **3** | **AUDIT_R3 §1** | **RunHooks signature (`AgentHookContext`, `result: str`)** |
+| **F3** | **Critical** | **1** | **AUDIT_R3 §1** | **TracingProcessor methods are sync** |
+| C13 | High | 3 | AUDIT_R3 §2 | Bus.finalize cleans up stale state |
+| C14 | High | 3 | AUDIT_R3 §2 | Filter try/except |
+| C15 | High | 3 | AUDIT_R3 §2 | Hooks try/except |
+| C16 | High | 1 | AUDIT_R3 §2 | TracingProcessor catches OSError |
+| C17 | Medium | 1 | AUDIT_R3 §2 | Model alias validation |
+| C18 | Medium | 2 | AUDIT_R3 §2 | Sandbox response size cap |
+| C19 | Medium | 1 | AUDIT_R3 §2 | Assert ≥1 enabled tool when `tool_choice='required'` |
+| C20 | Medium | 2 | AUDIT_R3 §2 | Per-tool `timeout_behavior` discrimination |
+| C21 | Medium | 1 | AUDIT_R3 §2 | RunConfig override + context fields (`is_whitebox`, `diff_scope`, `run_id`) |
+| C22 | Medium | 2 | AUDIT_R3 §2 | `finish_scan` checks children running |
+| C23 | Medium | 5 | AUDIT_R3 §2 | Diff-scope injection in user message |
+| C24 | Medium | 5 | AUDIT_R3 §2 | Run-name + Docker preflight |
+| C25 | Medium | 5 | AUDIT_R3 §2 | Cancel mode mapping (immediate/after_turn) |
+
+---
+
+## 7. Outcome
+
+**No new architectural blockers.** All corrections are bounded.
+
+The three F-fixes (type/signature corrections) are inline-applied to `PLAYBOOK.md` in this round. C13–C25 are added to the playbook's correction register and the relevant phases. After this round, the playbook's load-bearing skeletons compile against SDK v0.14.6, and defensive error handling is wired through filter, hooks, processor, and bus.
+
+Ready for Phase 0.
diff --git a/HARNESS_WIKI.md b/HARNESS_WIKI.md
new file mode 100644
index 00000000..3bb717b2
--- /dev/null
+++ b/HARNESS_WIKI.md
@@ -0,0 +1,780 @@
+# Strix Harness — Internal Architecture Wiki
+
+> Internal deep-dive for the team. Maps every subsystem, every tool, every architectural decision, with `path:line` references throughout.
+>
+> Generated against `main` at `9fb1012` (`fix: --config flag now fully overrides ~/.strix/cli-config.json`). When code drifts, prefer `git log` and the source over this document.
+
+---
+
+## Table of Contents
+
+1. [What Strix Is](#1-what-strix-is)
+2. [Architecture at a Glance](#2-architecture-at-a-glance)
+3. [Repository Layout](#3-repository-layout)
+4. [Lifecycle of One Run](#4-lifecycle-of-one-run)
+5. [Agent System](#5-agent-system)
+6. [LLM Layer](#6-llm-layer)
+7. [Tool System](#7-tool-system)
+8. [Runtime & Sandbox](#8-runtime--sandbox)
+9. [Interface (CLI / TUI / Headless)](#9-interface-cli--tui--headless)
+10. [Prompts](#10-prompts)
+11. [Skills](#11-skills)
+12. [Config](#12-config)
+13. [Telemetry & Persistence](#13-telemetry--persistence)
+14. [Cross-Cutting Design Decisions](#14-cross-cutting-design-decisions)
+15. [Recent Evolution (Notable Commits)](#15-recent-evolution-notable-commits)
+16. [Quick File Index](#16-quick-file-index)
+
+---
+
+## 1. What Strix Is
+
+`strix-agent` (PyPI: `strix-agent`, version `0.8.3` per `pyproject.toml:3`) is an **autonomous AI hacker harness**. The agent dynamically pentests apps — runs targets in a sandboxed Kali container, finds vulns, validates them with PoCs, and writes a report.
+
+**Mission shape:** loop an LLM against a Kali-loaded sandbox until it produces a vulnerability report. Provider-agnostic via litellm. Multi-agent orchestration. Whitebox + blackbox + greybox modes.
+
+**Core deps** (`pyproject.toml:25-39`): `litellm[proxy]`, `pydantic`, `rich`, `docker`, `textual`, `tenacity`, `cvss`, `traceloop-sdk`, `opentelemetry-exporter-otlp-proto-http`, `scrubadub`. Optional `sandbox` extra adds `fastapi`, `uvicorn`, `ipython`, `playwright`, `pyte`, `libtmux`, `gql`, `openhands-aci`.
+
+**Entry point:** `strix.interface.main:main` (`pyproject.toml:43`).
+
+---
+
+## 2. Architecture at a Glance
+
+```
+ ┌──────────────────────────────────────────────────────┐
+ │ HOST PROCESS │
+ │ │
+ user CLI ──▶ │ interface/main.py │
+ │ │ │
+ │ ▼ │
+ │ StrixAgent (root) ──spawns──▶ StrixAgent (subagent) │
+ │ │ │ thread │
+ │ │ agent_loop() │ │
+ │ ▼ ▼ │
+ │ ┌──────────┐ ┌─────────────────────────┐ │
+ │ │ LLM │ │ Tool Executor │ │
+ │ │ litellm │ │ ──local────▶ in-process │ │
+ │ │ stream │ │ ──sandbox──▶ HTTP POST │ │
+ │ └──────────┘ └─────────────┬───────────┘ │
+ │ │ │
+ │ Tracer ──▶ events.jsonl │ │
+ │ OTel/Traceloop │ │
+ │ PostHog │ │
+ └──────────────────────────────┼───────────────────────┘
+ │ Bearer-auth HTTP
+ ▼
+ ┌──────────────────────────────────────────────────────┐
+ │ SANDBOX CONTAINER (Kali, one per scan) │
+ │ │
+ │ FastAPI tool server :48081 │
+ │ POST /execute → registry → tool fn │
+ │ │
+ │ Caido HTTP proxy :48080 (CA-injected) │
+ │ Playwright Chromium (headless, no-sandbox) │
+ │ tmux + IPython (terminal, python) │
+ │ /workspace (shared FS) │
+ │ nmap, sqlmap, nuclei, semgrep, trivy, ... │
+ └──────────────────────────────────────────────────────┘
+```
+
+**Two key boundaries:**
+
+1. **Host ↔ sandbox**: HTTP/JSON over Bearer-token auth. Host owns LLM + telemetry; sandbox owns dangerous tools.
+2. **Provider-agnostic LLM**: everything goes through `litellm.acompletion`; tool calls are **custom XML in text**, not native function-calling — provides multi-provider compatibility at the cost of native streaming schemas.
+
+---
+
+## 3. Repository Layout
+
+```
+strix/
+├── agents/ # Agent loop, state, multi-agent graph orchestration
+│ ├── base_agent.py # Core loop, ~620 lines
+│ ├── state.py # AgentState pydantic model
+│ └── StrixAgent/
+│ ├── strix_agent.py # execute_scan() entry, task assembly
+│ └── system_prompt.jinja # 32 KB Jinja system prompt
+├── llm/ # Completion wrapper, provider routing, memory mgmt
+│ ├── llm.py # acompletion wrapper, streaming, retries
+│ ├── config.py # LLMConfig dataclass
+│ ├── memory_compressor.py # Token-budget pruning + LLM summarization
+│ ├── utils.py # Tool-format normalization, XML parsing
+│ └── dedupe.py # Vulnerability dedup via LLM similarity
+├── tools/ # Every tool the agent can call
+│ ├── registry.py # @register_tool decorator, schema loader
+│ ├── executor.py # Dual local/sandbox dispatch, result fmt
+│ ├── context.py # contextvar agent_id propagation
+│ ├── argument_parser.py # Type coercion of XML string args
+│ ├── browser/ # Playwright Chromium (24 actions)
+│ ├── terminal/ # tmux + libtmux interactive shell
+│ ├── python/ # IPython kernel, persistent
+│ ├── proxy/ # Caido GraphQL client
+│ ├── notes/ # Wiki + categorized notes (JSONL persisted)
+│ ├── todo/ # In-memory todo list
+│ ├── reporting/ # create_vulnerability_report w/ CVSS
+│ ├── web_search/ # Perplexity sonar-reasoning-pro
+│ ├── file_edit/ # openhands-aci str_replace_editor + rg
+│ ├── finish/ # finish_scan (root only)
+│ ├── thinking/ # think tool (planning escape hatch)
+│ ├── load_skill/ # Runtime skill injection
+│ └── agents_graph/ # create_agent, send_message_to_agent, ...
+├── runtime/ # Docker-side container management
+│ ├── docker_runtime.py # Host-side: launch/healthcheck/cleanup
+│ └── tool_server.py # Sandbox-side FastAPI tool server
+├── interface/ # CLI + Textual TUI + headless
+│ ├── main.py # argparse, validation, mode dispatch
+│ ├── cli.py # Headless mode (Rich)
+│ ├── tui.py # Textual app, modal screens
+│ └── utils.py # Helpers (target inference, run-dir, ...)
+├── prompts/ # Vulnerability-specific Jinja prompts (e.g. NoSQLi)
+├── skills/ # Markdown playbooks (vulns, frameworks, scan modes)
+├── config/ # Config class, env layering, file persistence
+├── telemetry/ # Tracer, OTel, Scrubadub PII redaction, PostHog
+└── utils/ # resource_paths.py for frozen-vs-dev path resolution
+
+containers/
+├── Dockerfile # Kali rolling, all the pentest tools
+└── docker-entrypoint.sh # Caido boot, CA install, tool server start
+```
+
+---
+
+## 4. Lifecycle of One Run
+
+End-to-end trace of `strix --target ./app`:
+
+1. **CLI parse** (`interface/main.py:267-426`): parse args, validate, infer target types via `infer_target_type()` (`interface/utils.py`). Resolve diff scope if `--scope-mode=diff`.
+2. **Config layering** (`config/config.py`): apply `~/.strix/cli-config.json` (or `--config <path>` override) into `os.environ`; resolve `STRIX_LLM`, `LLM_API_KEY` via `resolve_llm_config()` at `config/config.py:199-224`.
+3. **LLM warm**: validate model reachable.
+4. **Docker pull** if needed; image pin `ghcr.io/usestrix/strix-sandbox:0.1.13` (`config/config.py:43`).
+5. **Run name + run dir**: `strix_runs/<run-name>/`. Tracer init (`telemetry/tracer.py:50+`).
+6. **Mode dispatch**: TUI (`tui.py`) or CLI (`cli.py`). Both end up calling `StrixAgent.execute_scan(scan_config)`.
+7. **Sandbox launch** (`runtime/docker_runtime.py:111-173`): container created with name `strix-scan-{scan_id}`, two random host ports mapped to container ports `48080` (Caido) and `48081` (tool server), 32-byte bearer token generated, `local_sources` tar-copied into `/workspace`.
+8. **Container boot** (`containers/docker-entrypoint.sh`): start Caido → fetch GraphQL token via `loginAsGuest` → create temp project → install CA cert into NSS + system trust → set system-wide proxy env vars → spawn tool server as `pentester` user → wait for `/health` ready.
+9. **Agent loop** (`agents/base_agent.py:152-260`): see §5.
+10. **Termination**: root agent calls `finish_scan` (`tools/finish/finish_actions.py`) when work is done; tracer writes final report `penetration_test_report.md` and per-finding JSONs under `vulnerabilities/`.
+11. **Cleanup**: `docker rm -f` async-spawned (`runtime/docker_runtime.py:334-352`).
+12. **Exit code**: `0` clean, `2` if vulns found in headless mode (per `cli.py`).
+
+---
+
+## 5. Agent System
+
+### 5.1 Single-Agent Loop
+
+`agents/base_agent.py:152-260` (`agent_loop`). Each iteration:
+
+1. `_initialize_sandbox_and_state()` once at start (`base_agent.py:158`).
+2. Check messages: `_check_agent_messages()` (`base_agent.py:448-531`) drains the inter-agent message queue, wrapping each in `<inter_agent_message>` XML and appending to history.
+3. Iteration counter bump and warning watchdog: at 85% of `max_iterations` (default 300) emit warning; at `max-3` emit critical "next message MUST be finish" warning (`base_agent.py:186-211`).
+4. `_process_iteration()` (`base_agent.py:214-216`):
+ - Compress history (memory compressor, see §6.4).
+ - Build messages, call LLM via async generator (`llm.py:156-218`).
+ - Parse tool invocations from streamed response (custom XML parser, see §6.5).
+ - `_execute_actions()` → `process_tool_invocations()` (`tools/executor.py:313-342`) — sequential per-action dispatch.
+ - Append observation XML to history.
+5. Loop until `state.should_stop()` (`state.py`): `completed | stop_requested | iteration >= max_iterations`.
+
+In **interactive mode**, after `completed=True` the loop pauses in `_enter_waiting_state()` (`base_agent.py:287-329`) instead of exiting — user can send more input. `_wait_for_input()` resumes on message arrival or `waiting_timeout` (default 600 s for subagents, 0 = forever for root, `base_agent.py:265-266`).
+
+### 5.2 State Model
+
+`agents/state.py:12-173` (Pydantic `AgentState`):
+
+| Field | Purpose |
+|---|---|
+| `agent_id`, `agent_name`, `parent_id` | Identity. `parent_id is None` ⇔ root agent. |
+| `task` | Initial task string. |
+| `messages` | Conversation history list (role/content tuples; multimodal-capable). |
+| `iteration`, `max_iterations` | Hard budget (default 300). |
+| `waiting_for_input`, `waiting_start_time`, `waiting_timeout` | Interactive-mode pause state. |
+| `completed`, `stop_requested`, `final_result` | Termination signals. |
+| `sandbox_id`, `sandbox_token`, `sandbox_info` | Sandbox handles (port, ID). |
+| `actions_taken`, `observations`, `errors` | Local audit trail. |
+| `start_time`, `last_updated` | ISO timestamps. |
+
+Snapshots (`state.model_dump()`) are stored verbatim in the agent-graph node when the agent is registered (`base_agent.py:122-134`).
+
+### 5.3 Multi-Agent Graph
+
+`tools/agents_graph/agents_graph_actions.py` (839 lines) is the orchestration plane. Globals at `:9-37`:
+
+- `_agent_graph = {"nodes": {agent_id: node}, "edges": [...]}` — node = full agent metadata; edges = `delegation` or `message`.
+- `_agent_messages: dict[agent_id, list[msg]]` — per-agent inbox.
+- `_agent_instances: dict[agent_id, agent_obj]` — live in-process instances (for stat snapshots).
+- `_agent_states: dict[agent_id, AgentState]`.
+- `_running_agents: dict[agent_id, threading.Thread]` — daemon threads.
+- `_completed_agent_llm_totals` + `_agent_llm_stats_lock` — accumulated stats from finished subagents.
+
+**Spawning** (`create_agent`, `:383-492`): validate skills → build child `LLMConfig` inheriting parent flags → `StrixAgent(config)` → optionally copy parent history (`inherit_context=True`) → spawn daemon thread running `_run_agent_in_thread()` (`:205-298`). The thread creates a fresh asyncio event loop and runs `agent.agent_loop(state.task)`. On finish, status flips to `completed | stopped | failed`, `_finalize_agent_llm_stats()` is called.
+
+**Identity injection**: parent task is wrapped in `<agent_delegation>...<agent_identity>` so the child knows its name/ID and is told to never echo it (`:238-266`).
+
+**Finish from subagent** (`agent_finish`, `:567-685`): only callable when `parent_id != None`. Builds `<agent_completion_report>` XML with summary/findings/recommendations and pushes it onto the parent's inbox. If the agent is whitebox, the wiki note is updated with a delta.
+
+**Finish from root** (`finish_scan`, `tools/finish/finish_actions.py:86-149`): only callable from root (`parent_id is None`); blocks if any sibling/child agent is still `running`/`stopping`. Triggers `tracer.update_scan_final_fields()` which writes `penetration_test_report.md`.
+
+**Inter-agent messaging** (`send_message_to_agent`, `:495-563`): synchronous append to `_agent_messages[target]` with edge metadata. No broker, no durability, single-process only.
+
+**Waiting** (`wait_for_message`, `:796-839`): subagent calls this to pause; `_check_agent_messages` resumes it on arrival.
+
+**Graph view** (`view_agent_graph`, `:302-380`): traversal printout, root-first.
+
+### 5.4 Stats Aggregation Across the Tree
+
+Recent fix (`15c9571`). `_finalize_agent_llm_stats()` (`:54-68`) snapshots a finished subagent's `llm._total_stats` and adds it under lock to `_completed_agent_llm_totals`. The root's reported totals (via `tracer.get_total_llm_stats()` at `telemetry/tracer.py:801-834`) are: `sum(_completed_agent_llm_totals) + sum(live agent _total_stats)`. Before the fix, finalized children were dropped on cleanup — root undercounted.
+
+### 5.5 Termination, Interrupts, and Cancellation
+
+- **Hard limit**: `iteration >= max_iterations` → loop exits (`base_agent.py:174`). 85% / N-3 warnings give the model time to wrap up.
+- **User Ctrl+C / parent kill**: `stop_agent(agent_id)` (`agents_graph_actions.py:688-748`) sets `state.request_stop()` + calls `agent_instance.cancel_current_execution()` (`base_agent.py:615-623`) which cancels the running asyncio task. `asyncio.CancelledError` is caught (`base_agent.py:232-243`) — interactive mode enters waiting state; non-interactive re-raises.
+- **Finish tools**: see §5.3.
+
+### 5.6 Streaming to TUI
+
+The agent loop consumes the LLM async generator and after each chunk calls `tracer.update_streaming_content(agent_id, accumulated_text)` (`base_agent.py:373-375`). The TUI polls the tracer at 2 Hz and re-renders. On finalize, `tracer.clear_streaming_content()` and `tracer.log_chat_message()` snapshot the full message into the events log.
+
+---
+
+## 6. LLM Layer
+
+### 6.1 Completion Wrapper
+
+`strix/llm/llm.py:156-218` — `LLM.generate()` is an async generator yielding `LLMResponse` objects. Pipeline:
+
+1. **Retry loop** (`:162-171`): max `STRIX_LLM_MAX_RETRIES` (default 5) attempts. Backoff `min(90, 2 * 2**attempt)` seconds. `_should_retry()` (`:326-330`) is True for network errors (no status_code) or `litellm._should_retry(code)` for HTTP statuses.
+2. **Build args** (`_build_completion_args`, `:265-274`): model resolution, reasoning effort, drop unsupported params (`litellm.drop_params = True`, `litellm.modify_params = True`, set at `:25-26`).
+3. **Stream** (`_stream`, `:173-218`):
+ - `acompletion(...)` wrapped in `asyncio.wait_for(timeout=self.config.timeout)` — commit `60abc09`.
+ - Each `await it.__anext__()` *also* wrapped in `asyncio.wait_for(timeout)` — needed because litellm's own timeout doesn't propagate to httpx for Bedrock streaming, which can accept TCP and then send no data forever.
+ - Accumulate text; when a closing `</function>` tag is seen, set `done_streaming=1` and continue 5 more chunks for trailers, then stop.
+4. **Stats** (`_update_usage_stats`, `:287-312`): extract `response.usage` (input, completion, cached tokens). Cost via `_extract_cost()` (`:314-324`) — prefers `response.usage.cost`, else `litellm.completion_cost(...)` with provider stripped.
+5. **Tool parsing** (`:212-217`): `normalize_tool_format` → `fix_incomplete_tool_call` → `parse_tool_invocations` (all in `llm/utils.py`).
+
+### 6.2 Provider Routing
+
+`strix/llm/utils.py:34-61` defines `STRIX_MODEL_MAP` — `strix/<short>` aliases (e.g. `strix/claude-sonnet-4.6`) resolve to a tuple `(api_model, canonical_model)`:
+- `api_model` is what gets passed to `acompletion` (typically OpenAI-compatible against the Strix proxy `https://models.strix.ai/api/v1` set at `config/config.py:8`).
+- `canonical_model` is what's used for litellm capability checks like `supports_prompt_caching()` and `supports_reasoning()`.
+
+For non-`strix/` prefixes, the model string is passed straight through. `LLMConfig.__init__` (`llm/config.py:8-41`) reads `LLM_API_BASE` → `OPENAI_API_BASE` → `LITELLM_BASE_URL` → `OLLAMA_API_BASE` from env in priority order.
+
+Per-provider quirks:
+- **Anthropic**: `_is_anthropic()` (`llm.py:338-341`) detects `anthropic/` or `claude` substrings and adds an ephemeral cache control block to the system message via `_add_cache_control()` (`:371-387`).
+- **OpenAI reasoning**: if `supports_reasoning()` is true, set `reasoning_effort` (`:265-266`) — env `STRIX_REASONING_EFFORT` > config > scan-mode default (`medium` for quick, `high` otherwise).
+- **Vision-less models**: `_strip_images()` (`:343-369`) replaces image content with `"[Image removed - model doesn't support vision]"`.
+- **Vertex AI**: optional extra in `pyproject.toml:48`. Documented in `docs/llm-providers/vertex.mdx`.
+
+### 6.3 Retries & Timeouts (the `60abc09` story)
+
+Symptom: Bedrock converse-stream calls would TCP-connect, send no chunks, and hang the agent loop indefinitely. faulthandler showed the loop blocked in `selectors.select()`.
+
+Fix: wrap **both** the initial `acompletion` call **and** every per-chunk read in `asyncio.wait_for`. Timeouts surface as `TimeoutError` (no `status_code` attr) which `_should_retry()` treats as retryable, kicking the backoff loop.
+
+Default `LLM_TIMEOUT = 300` (`config/config.py:24`).
+
+### 6.4 Memory Compression
+
+`strix/llm/memory_compressor.py:152-219`. Hard ceiling `MAX_TOTAL_TOKENS = 100_000`; compression triggers above ~90 K.
+
+- **Image budget**: keep last 3 images, replace older ones with `[Previously attached image removed]` (`:134-149`).
+- **System messages**: never compressed.
+- **Recent floor**: keep last 15 messages intact (`MIN_RECENT_MESSAGES = 15`).
+- **Older messages**: chunk in groups of 10, summarize via LLM call (separate `acompletion`, 120 s timeout). The summary prompt (`:15-43`) emphasizes preserving exact technical details — URLs, payloads, credentials, failed attempts (so the agent doesn't repeat them) — wrapped as `<context_summary message_count='N'>...</context_summary>`.
+- Token counting via `litellm.token_counter()` with a fallback of `len(text) / 4`.
+
+This runs every iteration. Anthropic prompt cache helps with system-prompt cost but not history (which mutates).
+
+### 6.5 Tool-Call Format (custom XML, NOT native function-calling)
+
+Tools are injected into the system prompt as XML descriptions via `get_tools_prompt()` (`tools/registry.py:280-300`). The model is instructed to emit:
+
+```xml
+<function=tool_name>
+ <parameter=key>value</parameter>
+ <parameter=other>value</parameter>
+</function>
+```
+
+Parser (`llm/utils.py`):
+- `normalize_tool_format` (`:12-31`): converts Anthropic-style `<invoke name="X">` and other variants to the canonical form.
+- `fix_incomplete_tool_call` (`:110-121`): auto-closes unclosed tags when streaming is truncated.
+- `parse_tool_invocations` (`:80-133`): regex extraction; HTML-entity-decodes values.
+- `clean_content` (`:135-160`): strips tool XML and inter-agent control tags before logging to telemetry.
+
+**Why XML over native tool use?** Multi-provider compatibility (works on any text LLM), graceful streaming truncation (early-exit at `</function>`), and full client-side control of formatting. Cost: tool descriptions are re-injected as text every call (no native streaming of schemas), and content has to be parsed by hand.
+
+### 6.6 Prompt Assembly
+
+`_prepare_messages` (`llm.py:220-248`):
+1. Render system prompt from Jinja template (`agents/StrixAgent/system_prompt.jinja`) with: `interactive` flag, `system_prompt_context` (authorized targets), `loaded_skill_names`, tools prompt.
+2. Insert agent-identity user message (hidden control block) — see §5.3.
+3. Run `MemoryCompressor.compress_history()` over conversation.
+4. If last message is assistant, append a `<meta>Continue the task.</meta>` continuation prompt (autonomous mode).
+5. Apply Anthropic cache control on the system message if applicable.
+
+### 6.7 Stats / Cost
+
+Per-call: extracted into `RequestStats` dataclass and accumulated in `LLM._total_stats`. Across the tree: see §5.4. Tracer renders these into the live TUI status panel and the final summary text (`utils/format_token_count`, etc.).
+
+### 6.8 Vulnerability Deduplication
+
+`strix/llm/dedupe.py` (~213 lines) — separate LLM call to compare a new finding against existing ones. Wired into `tools/reporting/reporting_actions.py` so duplicate vuln reports get rejected on submission.
+
+---
+
+## 7. Tool System
+
+### 7.1 Registry & Schema Loading
+
+`strix/tools/registry.py`:
+
+- Decorator: `@register_tool(sandbox_execution: bool, requires_browser_mode: bool = False, requires_web_search_mode: bool = False)` (`:190-250`).
+- Conditional registration via `_should_register_tool()` (`:175-187`): in sandbox mode, only `sandbox_execution=True` tools register. If `STRIX_DISABLE_BROWSER=true`, browser tools are skipped. If `PERPLEXITY_API_KEY` is missing, `web_search` is skipped.
+- Schema: each tool group has `<group>_actions_schema.xml` next to the implementation. Parsed via `_parse_param_schema()` (`:131-149`) → `_tool_param_schemas[name] = {params, required, has_params}`.
+- `get_tools_prompt()` (`:280-300`) emits the XML descriptions injected into the system prompt. Includes a `{{DYNAMIC_SKILLS_DESCRIPTION}}` placeholder filled by the load-skill subsystem.
+
+### 7.2 Executor (Local vs Sandbox Dispatch)
+
+`strix/tools/executor.py`:
+
+- `execute_tool()` is the single entrypoint. `should_execute_in_sandbox(tool_name)` (`:29-37`) decides routing.
+- **Sandbox path** (`_execute_tool_in_sandbox`, `:39-99`): POST to `http://{host}:{tool_server_port}/execute` with `{agent_id, tool_name, kwargs}` and `Authorization: Bearer {token}`. Connect timeout `STRIX_SANDBOX_CONNECT_TIMEOUT=10`, request timeout `STRIX_SANDBOX_EXECUTION_TIMEOUT + 30 = 150` s.
+- **Local path** (`_execute_tool_locally`, `:101-115`): look up function, coerce arg types via `argument_parser.convert_arguments()`, inject `agent_state` if the function requests it (introspection over signature), `await` if async.
+- **Argument validation** (`_validate_tool_arguments`, `:130-186`): unknown params and missing required params → formatted error string returned to the LLM (no exception).
+- **Result formatting** (`_format_tool_result`, `:227-256`): truncate >10 KB to 4 KB head + `... [middle content truncated] ...` + 4 KB tail. Wrap in `<tool_result><tool_name>...</tool_name><result>...</result></tool_result>`. Extract any `screenshot` key into a separate base64 image attachment.
+- **Process orchestrator** (`process_tool_invocations`, `:313-342`): iterate actions sequentially, aggregate result text, attach images as multimodal content blocks, return `should_agent_finish` boolean.
+
+**Sequential, not parallel.** No `asyncio.gather` over tools — could be a future optimization.
+
+### 7.3 Tool Catalog
+
+#### Browser (`strix/tools/browser/`) — `browser_action`
+Playwright Chromium singleton per container, shared event loop in a daemon thread (`browser_instance.py:34-48`). 24 actions:
+- Navigation: `launch`, `goto`, `back`, `forward`
+- Interaction: `click`, `double_click`, `hover`, `type`, `press_key`, `scroll_up`, `scroll_down`
+- Tabs: `new_tab`, `switch_tab`, `close_tab`, `list_tabs`
+- Misc: `wait`, `execute_js`, `save_pdf`, `get_console_logs`, `view_source`, `close`
+
+Returns `{screenshot: base64-png, url, title, tab_id, all_tabs, js_result, console_logs, page_source}`. Viewport 1280×720, screenshot is viewport-only (not full-page) by default. Page source truncated to 20 KB; JS result to 5 KB; console logs capped at 200 entries / 30 KB total / 1 KB each. Browser launched with `--no-sandbox --disable-web-security` — intentional for XSS/CORS pentesting.
+
+State keyed by `get_current_agent_id()` (contextvar). Tabs persist with sequential IDs (`tab_1`, `tab_2`, ...). `atexit` registered via `_register_cleanup_handlers()`.
+
+#### Terminal (`strix/tools/terminal/`) — `terminal_execute`
+libtmux-backed (`>=0.46.2`). One tmux session per `(agent_id, terminal_id)`, default `terminal_id="default"`. PS1 customized to `[STRIX_$?]$ ` so the **exit code can be regex-extracted** from the prompt (`terminal_session.py:49-54`). Pane history limited to 10 K lines.
+
+Params: `command`, `is_input` (false=new command, true=feed input to running process), `timeout` (default 30 s), `no_enter`, `terminal_id`. Returns `{content, status, exit_code, working_dir}` where status is `completed | running | error` (with sub-states `CONTINUE`, `NO_CHANGE_TIMEOUT`, `HARD_TIMEOUT`).
+
+Special-key sequences supported: `C-c`, `^X`, `S-X`, `M-X`, `F1-F12`, arrow keys, `Enter`, `Tab`, `BSpace`. Output deduplication strips previous output prefix to show only new bytes.
+
+#### Python (`strix/tools/python/`) — `python_action`
+Persistent IPython kernel (`>=9.3.0`) per `(agent_id, session_id)`. Actions: `new_session`, `execute`, `close`, `list_sessions`. cwd `/workspace`. Stdout truncated to 10 K, stderr to 5 K, repr to 10 K. Execution timeout enforced via thread join + cancellation flag (default 30 s).
+
+Pre-injects proxy helper functions into the user namespace so the agent can `send_request(...)` directly (`python_instance.py:30-47`).
+
+`KeyboardInterrupt` and `SystemExit` are caught and returned as errors rather than propagating.
+
+#### HTTP Proxy (`strix/tools/proxy/`) — 7 tools
+GraphQL client against Caido at `http://127.0.0.1:48080/graphql` (`proxy_manager.py:25`). Bearer token from env `CAIDO_API_TOKEN`. Tools: `list_requests`, `view_request`, `send_request`, `repeat_request`, `scope_rules`, `list_sitemap`, `view_sitemap_entry`.
+
+Supports HTTPQL filter syntax for request queries. Pagination (`offset`, `limit`). `view_request` supports regex search through captured request/response pairs. Caido all-traffic capture is enabled because `/etc/profile.d/proxy.sh` sets `http_proxy`/`https_proxy` system-wide and the Caido CA cert is installed into the system + NSS trust stores.
+
+Hardcoded port `48080`. Caido v0.48.0 pinned in `containers/Dockerfile`.
+
+#### Notes (`strix/tools/notes/`) — `create_note`, `list_notes`, `get_note`, `update_note`, `delete_note` (+ internal `append_note_content`)
+In-memory dict + JSONL persistence at `{run_dir}/notes/notes.jsonl`. Wiki-category notes additionally rendered as Markdown to `{run_dir}/wiki/{note_id}-{title}.md` so they're human-readable artifacts of the run.
+
+Categories: `general | findings | methodology | questions | plan | wiki`. IDs are 5-char UUID hex (collision-retry up to 20 attempts). Thread-safe via RLock. List preview defaults to 280 chars per note.
+
+The wiki note in particular is **the shared whitebox knowledge base** between root and subagents; `agent_finish` for whitebox subagents auto-appends a delta (see §5.3).
+
+#### Todos (`strix/tools/todo/`) — 6 tools
+Per-agent in-memory storage; **not persisted**. Priorities `critical | high | normal | low`; statuses `pending | in_progress | done`. Bulk create/update via JSON list. IDs are 6-char UUID hex.
+
+#### Reporting (`strix/tools/reporting/`) — `create_vulnerability_report`
+Saves a CVSS-scored vulnerability to the run. Required fields: title, description, impact, target, technical_analysis, poc_description, **poc_script_code** (mandatory), remediation_steps, cvss_breakdown (XML with AV/AC/PR/UI/S/C/I/A enums).
+
+Optional fields: endpoint, method, cve (`CVE-\d{4}-\d{4,}`), cwe (`CWE-\d+`), code_locations (XML with file/start_line/end_line/snippet/label/fix_before/fix_after — relative paths only, no `..`).
+
+CVSS computed via the `cvss` library. Deduplication via LLM similarity check (`llm/dedupe.py`). Persisted via tracer to `{run_dir}/vulnerabilities/vuln_{id}.json`.
+
+#### Web Search (`strix/tools/web_search/`) — `web_search`
+Perplexity API (`sonar-reasoning-pro` model). 300 s timeout. System prompt tailored for security professionals — vuln details, CVEs, OWASP, exploit info. Skipped from registry if `PERPLEXITY_API_KEY` not set.
+
+#### File Edit (`strix/tools/file_edit/`) — `str_replace_editor`, `list_files`, `search_files`
+Wraps openhands-aci's file_editor (`>=0.3.0`). Commands `create | str_replace | view | insert`. Relative paths auto-prefixed with `/workspace/`. `search_files` uses ripgrep (`rg`); recursive listings capped at 500 results.
+
+#### Finish (`strix/tools/finish/`) — `finish_scan`
+Root-only. Validates: caller is root, all child agents not running/stopping, all four narrative fields non-empty (executive_summary, methodology, technical_analysis, recommendations). On success: tracer writes the final report.
+
+#### Thinking (`strix/tools/thinking/`) — `think`
+Minimal — records a thought string, validates non-empty, returns char count. Acts as the "free turn" escape hatch for planning so the agent can satisfy the per-message tool-call requirement (see §10) without doing real work.
+
+#### Agents Graph (`strix/tools/agents_graph/`) — 6 tools
+`view_agent_graph`, `create_agent`, `agent_status`, `agent_finish`, `wait_for_message`, `send_message_to_agent`. Mechanics covered in §5.3.
+
+#### Load Skill (`strix/tools/load_skill/`) — `load_skill`
+Runtime injection of additional skill content into the agent's context. Validates names against the registry. Replaces the `{{DYNAMIC_SKILLS_DESCRIPTION}}` placeholder. Max 5 skills per agent context.
+
+### 7.4 Result Sanitization (commit `4934bb8`)
+
+Three layers of defense before tool output reaches the model or telemetry:
+
+1. **Screenshot extraction** (`executor.py:345-353`): if a result dict has key `screenshot` whose value is a base64 string, pull it out into a separate image attachment, replace its dict value with `"[Image data extracted - see attached image]"`.
+2. **Length truncation** (`:246-249`): >10 KB results are split head + truncation marker + tail, 4 KB each side.
+3. **Error truncation** (`:182-183`): error strings capped at 500 chars with `[truncated]` suffix.
+4. **Telemetry sanitization** (`telemetry/utils.py:67-150`): scrubadub + regex on dict/list/tuple keys/values, redacting `screenshot`, sensitive-key patterns (`api[-_]?key|token|secret|password|...`), and bearer-style tokens (`ghp_`, `ghs_`, `xox*`).
+5. **Content cleaning before logging** (`llm/utils.py:135-160`, `clean_content`): strips tool XML, inter-agent control blocks, agent-identity blocks before they hit the JSONL events log — keeps the audit log readable and prevents log-injection of tool schemas.
+
+### 7.5 Agent Context Propagation
+
+`strix/tools/context.py` defines a single `ContextVar` `current_agent_id`. Set on every tool invocation in `tool_server._run_tool` (`runtime/tool_server.py:71-83`) and used by stateful tools (browser, terminal, python, todos) to silo per-agent state without explicit threading.
+
+---
+
+## 8. Runtime & Sandbox
+
+### 8.1 Image (`containers/Dockerfile`)
+
+- Base: `kalilinux/kali-rolling:latest` (line 1).
+- Non-root `pentester` user with NOPASSWD sudo (lines 10-13) — needed for raw-socket pentest tools.
+- Pre-installed: `nmap`, `nuclei`, `subfinder`, `naabu`, `ffuf`, `sqlmap`, `zaproxy`, `wapiti`, `caido-cli` (v0.48.0); Go tools `httpx`, `katana`, `gospider`, `interactsh`; Python `arjun`, `dirsearch`, `wafw00f`, `semgrep`, `bandit`, `trufflehog`; JS `retire`, `eslint`, `js-beautify`, `jshint`, `@ast-grep/cli`, `tree-sitter-cli`; tree-sitter parsers for Java/JS/Python/Go/Bash/JSON/YAML/TS; `gitleaks`, `trivy`.
+- Sandbox-extra Python deps: `fastapi`, `uvicorn`, `ipython`, `playwright`, `pyte`, `libtmux`, `gql`, `openhands-aci`.
+- Self-signed CA chain at `/app/certs/{ca.key,ca.crt,ca.p12}`, 3650-day validity (lines 52-71).
+- Workspace `/workspace` owned by pentester (line 194).
+- Ports `48080` (Caido), `48081` (tool server) exposed.
+- Image pin: `ghcr.io/usestrix/strix-sandbox:0.1.13` (`strix/config/config.py:43`, bumped in `640bd67`).
+
+### 8.2 Boot Sequence (`containers/docker-entrypoint.sh`)
+
+1. **Caido start** (lines 12-17): `caido-cli --listen 0.0.0.0:48080 --allow-guests --no-logging --import-ca-cert /app/certs/ca.p12`.
+2. **Caido readiness poll** (24-38): GET `/graphql/` until 200/400, 30 attempts × 1 s.
+3. **Login** (50-74): GraphQL `loginAsGuest` mutation → bearer token, 5 retries with backoff. Exported as `CAIDO_API_TOKEN`.
+4. **Project setup** (79-109): create + select temp Caido project for capture.
+5. **System-wide proxy** (113-146): write `/etc/profile.d/proxy.sh` setting `http_proxy/https_proxy/HTTP_PROXY/HTTPS_PROXY/ALL_PROXY=127.0.0.1:48080`; mirror into `/etc/environment`, `/etc/wgetrc`. Import CA into NSS db so Chromium trusts it.
+6. **Tool server start** (154-180): `sudo -u pentester python -m strix.runtime.tool_server` with token, port, timeout. Wait for `/health` 200, 10 retries × 1 s.
+7. **Ready** (182): emit "✅ Container ready" and `exec` the trailing args (typically `sleep infinity`).
+
+### 8.3 Host-Side Runtime (`strix/runtime/docker_runtime.py`)
+
+- **One container per scan** (not per agent). All agents in a scan share the same container, the same `/workspace`, the same Caido proxy capture, the same browser/terminal/python sessions (keyed by `agent_id` contextvar).
+- **Port allocation** (`:43-46`): `socket.bind(("", 0))` to grab two free host ports, mapped to container 48080 and 48081.
+- **Token** (`:131`): `secrets.token_urlsafe(32)` per container.
+- **Container creation** (`:111-173`): name `strix-scan-{scan_id}`, label `strix-scan-id={scan_id}`, capabilities `NET_ADMIN | NET_RAW`, `extra_hosts={"host.docker.internal": "host-gateway"}`, env passthrough including `TOOL_SERVER_PORT` / `TOOL_SERVER_TOKEN` / `STRIX_SANDBOX_EXECUTION_TIMEOUT` / `HOST_GATEWAY`. Reuses a running container if one exists for the same scan_id.
+- **Healthcheck** (`:87-109`): poll `/health` for 30 s with backoff before declaring ready.
+- **Local source mount** (`:222-269`): tar-pipe local sources into `/workspace`. (Not a Docker bind mount — copy on init.)
+- **Reattach** (`:72-85`): on existing container, re-extract token+ports from `docker inspect` env.
+- **Cleanup** (`:322-352`): `docker stop` + `docker rm` spawned as detached subprocess so the host doesn't block on shutdown.
+
+**No CPU/memory/network egress limits configured.** Container has full host outbound access. Kill switches are: tool server `asyncio.wait_for` request timeout (default 120 s), and host-driven `docker stop`. There is no seccomp/AppArmor profile beyond Docker defaults.
+
+### 8.4 Tool Server (`strix/runtime/tool_server.py`)
+
+FastAPI app, served via Uvicorn on `0.0.0.0:{TOOL_SERVER_PORT}`. Auth via `HTTPBearer` (`:36-37, 42-57`); the `/health` endpoint is unauth.
+
+- `POST /execute` (`:86-127`): JSON `{agent_id, tool_name, kwargs}` → `_run_tool` (`:71-83`). Sets `current_agent_id` contextvar, looks up registry, calls function via `asyncio.to_thread()`. Per-agent task tracking in `agent_tasks: dict[agent_id, asyncio.Task]` (`:39`) — a new request for the same agent **cancels the previous task** (`:94-97`). Hard timeout `asyncio.wait_for(REQUEST_TIMEOUT)` default 120 s.
+- `POST /register_agent` (`:130-135`): registers an agent_id (used by host to pre-allocate state).
+- `GET /health` (`:138-147`): readiness/liveness.
+- Signal handling (`:150-162`): SIGTERM/SIGINT cancel all in-flight tasks; SIGPIPE ignored.
+- Returns `{"result": ..., "error": ...}` shape; HTTP 401 on bad token.
+
+### 8.5 openhands-aci
+
+Listed as sandbox dep (`pyproject.toml:54`). Used by `strix/tools/file_edit/` to back `str_replace_editor` (the same primitive as in OpenHands / Claude Code's `Edit` semantics — view/create/str_replace/insert with strict matching).
+
+### 8.6 Multi-Agent in One Sandbox
+
+Subagents inherit `sandbox_id`/`sandbox_token`/`sandbox_info` via the parent's state (passed implicitly through the LLMConfig copy in `create_agent`). They share `/workspace`, the Caido proxy capture, and stateful tools (each agent gets its own browser tab manager / terminal session / python kernel keyed by `agent_id`). **No per-agent process or container isolation.**
+
+---
+
+## 9. Interface (CLI / TUI / Headless)
+
+### 9.1 CLI Args (`strix/interface/main.py:267-426`)
+
+| Flag | Purpose |
+|---|---|
+| `-t / --target` (multi) | Target — URL / repo / local dir / domain / IP. Type inferred via `infer_target_type()`. |
+| `--instruction` | Inline directive. Mutex with `--instruction-file`. |
+| `--instruction-file` | File path. Read into `args.instruction`. |
+| `-n / --non-interactive` | Headless mode (`cli.py`) instead of TUI. |
+| `-m / --scan-mode {quick,standard,deep}` | Default `deep`. Controls breadth/depth via prompt-injected skill. |
+| `--scope-mode {auto,diff,full}` | Default `auto`. In CI/headless, `auto`→`diff`. |
+| `--diff-base` | Branch/commit to diff against (e.g. `origin/main`). Auto-detected if missing. |
+| `--config` | Path to custom `cli-config.json` — **fully overrides** `~/.strix/cli-config.json` (commit `9fb1012`). |
+
+`localhost` targets are rewritten to `host.docker.internal` so the sandbox can reach host-served apps.
+
+### 9.2 Scan Modes
+
+Implemented as **prompt content**, not control-flow branching. `LLMConfig.scan_mode` flows through to the agent's loaded skill set:
+
+- **quick** (`strix/skills/scan_modes/quick.md`): time-boxed, prioritize high-impact (auth, IDOR, RCE, SQLi, SSRF, secrets), skip exhaustive enumeration, breadth>depth, minimal PoC validation.
+- **standard** (`strix/skills/scan_modes/standard.md`): balanced. Whitebox = repo map → semgrep → AST → secrets/deps. Blackbox = crawl, fingerprint, capture proxy traffic. Phase 2 = business logic; Phase 3 = systematic input/auth/access tests.
+- **deep** (default): exhaustive — every file, every endpoint, every parameter, every edge case, every user role, complete state-machine and trust-boundary modeling, maximum chaining.
+
+Reasoning effort defaults flow off scan mode (`llm/llm.py:74-82`): quick→`medium`, else→`high`.
+
+### 9.3 Scope Modes
+
+`resolve_diff_scope_context()` (`interface/utils.py:40+`): computes `DiffScopeResult` from `git diff <base>...HEAD`. The result's `instruction_block` is **injected into the user instruction** rather than filtering files on disk — the agent decides prioritization. Used heavily for CI/PR workflows where you only want to test what changed.
+
+### 9.4 Textual TUI (`strix/interface/tui.py`)
+
+`StrixTUIApp` with modal screens: `SplashScreen`, `HelpScreen`, `StopAgentScreen`, `VulnerabilityDetailScreen`, plus the main agent-tree + log widgets. Multi-line `ChatTextArea` (Shift+Enter = newline, Enter = send). Keys: F1 help, Ctrl+Q/C quit, ESC stop, Tab cycle panels.
+
+Live updates: an updater thread polls the tracer at 2 Hz and refreshes `reactive` widgets. Vulnerability discoveries trigger the modal popup via `tracer.vulnerability_found_callback`.
+
+### 9.5 Headless / CLI (`strix/interface/cli.py`)
+
+Async run loop. Rich panels for vuln-found events, live stats panel updated every 2 s. Exit codes: `0` clean, `2` if `tracer.vulnerability_reports` is non-empty (used to fail CI). Final completion panel includes target, duration, stats, output path.
+
+### 9.6 Run Directory Layout (`strix_runs/<run_name>/`)
+
+Created and managed by `telemetry/tracer.py`. Contents:
+- `events.jsonl` — every span/event in append-only JSONL (thread-safe writes).
+- `vulnerabilities/vuln_{id}.json` — one file per finding, sorted by severity, dedup-checked.
+- `penetration_test_report.md` — final markdown report (executive summary + methodology + technical analysis + recommendations).
+- `notes/notes.jsonl` — note ops audit log.
+- `wiki/{note_id}-{slug}.md` — human-readable wiki notes.
+- `<target_subdir>/` — local source clones, per-target.
+
+There is **no execution checkpointing** — if the process crashes mid-run, the agent restarts from scratch on retry. Resumability is limited to the interactive-mode wait/resume on inbound messages.
+
+---
+
+## 10. Prompts
+
+### 10.1 The System Prompt
+
+`strix/agents/StrixAgent/system_prompt.jinja` (~32 KB). Single template, two behavioral modes via `{% if interactive %} ... {% else %} ... {% endif %}`. Sections:
+
+1. **Core capabilities** (4-9): security assessment, validation, remediation.
+2. **Communication rules** (11-44): CLI markdown only; never echo control XML (`inter_agent_message`, `agent_completion_report`, `agent_identity`).
+3. **Tool-call requirement** (24-44, hardened in `4f90a56`):
+ - Interactive: a message **without** a tool call **immediately stops the entire execution**. "Planning..." and "I'll now scan..." both halt. The only exceptions are genuinely done or explicitly asking the user. The `think` tool is the planning escape hatch.
+ - Autonomous: minimize chatter, never empty messages, use `wait_for_message` or finish when idle.
+4. **Execution guidelines** (47+):
+ - System-verified scope block (injected at runtime with the authorized targets).
+ - Authorization status: full authz for in-scope targets, **don't question permission**.
+ - Refusal-avoidance language: treat as internal security engineering, not generic offensive activity.
+ - Validation mandate: persist, iterate, assume more issues are hidden.
+ - Multi-target coordination: build target map, correlate findings, reuse secrets/endpoints.
+ - Testing modes: black-box, white-box, combined.
+ - Methodology: scope → recon → automated scanning → targeted validation → continuous iteration → impact documentation.
+ - Efficiency tactics: automate via Python, batch operations, parallel scans, fuzzers (ffuf/sqlmap/nuclei/semgrep) before custom payloads. For SQLi/XSS/RCE, spray via python/terminal not manual browser.
+5. **Vulnerability methodology** (230+): per-class attack surface, detection channels, chaining strategies, WAF bypasses for IDOR / SQLi / SSRF / RCE / XSS / XXE / path-traversal / race conditions / auth bypass / CSRF.
+
+### 10.2 Vulnerability-Specific Prompts (`strix/prompts/`)
+
+Currently only `vulnerabilities/nosql_injection.jinja` (~266 lines) — operator injection (`$ne`, `$gt`), boolean/timing/error oracles, auth bypass, regex-based extraction, JS execution, WAF bypass, deduplication. Covers MongoDB, CouchDB, Redis, Cassandra, Neo4j, GraphQL.
+
+### 10.3 Persona System
+
+There **is no separate persona file**. All agents are `StrixAgent` instances using the same Jinja system prompt. Differentiation comes from:
+- `parent_id` (root agent vs subagent → different finish tools, different prompts injected).
+- Loaded skills (root gets `root_agent`; whitebox gets `coordination/source_aware_whitebox` + `custom/source_aware_sast`).
+- `system_prompt_context.authorized_targets` (only set on root).
+- `is_whitebox` flag (toggles wiki-note auto-update on subagent finish).
+
+---
+
+## 11. Skills
+
+`strix/skills/` — Markdown playbooks loaded into the agent's system prompt. Categories:
+
+| Dir | Contents |
+|---|---|
+| `vulnerabilities/` | Auth/JWT, IDOR, SQLi, NoSQL, XSS, XXE, SSRF, CSRF, business logic, race conditions, path traversal, RCE, auth bypass, info disclosure, mass assignment, open redirect, insecure uploads, subdomain takeover. |
+| `frameworks/` | FastAPI, NestJS, Next.js. |
+| `technologies/` | Firebase/Firestore, Supabase. |
+| `protocols/` | GraphQL. |
+| `tooling/` | ffuf, httpx, katana, naabu, nmap, nuclei, semgrep, sqlmap, subfinder. |
+| `cloud/` | Kubernetes (RBAC, container escapes, etcd, supply chain — added in `#394`). |
+| `reconnaissance/` | placeholder. |
+| `custom/` | `source_aware_whitebox` (whitebox coordination), `source_aware_sast` (triage). |
+| `scan_modes/` | quick, standard, deep. |
+
+**Loading**: skills passed to `LLMConfig(skills=[...])` are rendered into the system prompt via the Jinja `get_skill(name)` macro (`llm/llm.py:96`). Whitebox automatically pulls in `coordination/source_aware_whitebox` + `custom/source_aware_sast`. **Max 5 skills per agent** (per `skills/README.md`). Mid-run skill loading via the `load_skill` tool replaces the `{{DYNAMIC_SKILLS_DESCRIPTION}}` placeholder.
+
+**Recent additions:**
+- NoSQL injection guide (#168) — see §10.2.
+- Kubernetes security testing (#394).
+
+---
+
+## 12. Config
+
+`strix/config/config.py` — hand-rolled `Config` class (no Pydantic). All knobs are class attributes; `Config.get(name)` resolves `os.environ[name.upper()]` first, then the default.
+
+| Knob | Default | Purpose |
+|---|---|---|
+| `strix_llm` | `None` | Model string; required. |
+| `llm_api_key` | `None` | Provider API key. |
+| `llm_api_base` / `openai_api_base` / `litellm_base_url` / `ollama_api_base` | `None` | Base URL fallbacks (resolved in priority order). |
+| `strix_reasoning_effort` | `"high"` | `low`/`medium`/`high`. |
+| `strix_llm_max_retries` | `"5"` | LLM retry count. |
+| `strix_memory_compressor_timeout` | `"30"` | Compressor LLM timeout. |
+| `llm_timeout` | `"300"` | Outer LLM timeout. |
+| `perplexity_api_key` | `None` | Web search. |
+| `strix_disable_browser` | `"false"` | Skip browser tool registration. |
+| `strix_image` | `"ghcr.io/usestrix/strix-sandbox:0.1.13"` | Sandbox image pin. |
+| `strix_runtime_backend` | `"docker"` | Only `docker` supported. |
+| `strix_sandbox_execution_timeout` | `"120"` | Tool exec timeout (s). |
+| `strix_sandbox_connect_timeout` | `"10"` | Tool server connect timeout. |
+| `strix_telemetry` | `"1"` | Master telemetry switch. |
+| `strix_otel_telemetry` / `strix_posthog_telemetry` | `None` | Per-stream override. |
+| `traceloop_base_url` / `traceloop_api_key` / `traceloop_headers` | `None` | OTel/Traceloop endpoint config. |
+
+### 12.1 Layering
+
+Order (highest first): `os.environ` → class default. Config files apply by writing into `os.environ`.
+
+Two file locations:
+- `~/.strix/cli-config.json` (default) auto-applied at `Config.apply_saved()`.
+- `--config <path>` (CLI flag) overrides via `apply_config_override()` (`interface/main.py:531-539`).
+
+The `--config` override fix (commit `9fb1012`):
+- Track applied vars in `Config._applied_from_default: ClassVar[dict]` (`config.py:59-61`).
+- On override, **clear those tracked vars from `os.environ` first**, then load the custom file.
+- Test: `tests/test_config_override.py` validates the leak doesn't happen.
+
+### 12.2 LLM Config Resolution
+
+`resolve_llm_config()` (`config.py:199-224`): if model starts with `strix/`, force `api_base = STRIX_API_BASE`; else cascade through `llm_api_base` → `openai_api_base` → `litellm_base_url` → `ollama_api_base`.
+
+### 12.3 Telemetry Opt-Out
+
+- `STRIX_TELEMETRY=0` kills both streams.
+- `STRIX_OTEL_TELEMETRY=0` kills OTel only.
+- `STRIX_POSTHOG_TELEMETRY=0` kills PostHog only.
+- Checked via `strix/telemetry/flags.py`.
+
+---
+
+## 13. Telemetry & Persistence
+
+### 13.1 Tracer (`strix/telemetry/tracer.py`)
+
+Holds `run_id`, `start_time`, agents map, tool_executions, chat_messages, vulnerability_reports, scan_results, run_metadata. Emits structured events into `events.jsonl` (thread-safe with `_get_events_write_lock`) and OTel spans.
+
+Event types include: `run.started`, `run.configured`, `agent.created`, `agent.status.updated`, `tool.execution.started`, `tool.execution.updated`, `chat.message`, `finding.created`.
+
+### 13.2 OpenTelemetry / Traceloop
+
+`bootstrap_otel()` wires an OTLP HTTP exporter (`opentelemetry-exporter-otlp-proto-http`). If Traceloop SDK is installed, spans also stream to `TRACELOOP_BASE_URL` with `TRACELOOP_API_KEY` headers. `TRACELOOP_HEADERS` (JSON string) allows custom headers. Graceful degradation if Traceloop SDK is missing.
+
+### 13.3 Scrubadub PII Redaction (`strix/telemetry/utils.py:67-150`)
+
+`TelemetrySanitizer`:
+- Recurses dict/list/tuple structures.
+- `_SCREENSHOT_KEY_PATTERN`: keys matching `screenshot` → `[SCREENSHOT_OMITTED]`.
+- `_SENSITIVE_KEY_PATTERN`: `api[-_]?key|token|secret|password|...` → `[REDACTED]`.
+- `_SENSITIVE_TOKEN_PATTERN`: bearer tokens, GitHub `ghp_`/`ghs_`, Slack `xox*`.
+- Strings additionally run through `scrubadub.Scrubber()` for emails, IPs, phone numbers, names; placeholders `{{...}}` replaced with `[REDACTED]`.
+
+Applied on every event payload before write/export (`tracer.py:159-160, 223-227`).
+
+### 13.4 PostHog (`strix/telemetry/posthog.py`)
+
+Anonymous usage telemetry. Per-process `_SESSION_ID = uuid4().hex[:16]` — no user identifier. Public API key embedded; events go to `https://us.i.posthog.com/capture/`. Events include: scan start/end, finding reported, error. Payload metadata: OS, arch, Python version, Strix version, scan mode, finding count, LLM tokens.
+
+Disabled by `STRIX_POSTHOG_TELEMETRY=0` or `STRIX_TELEMETRY=0`.
+
+---
+
+## 14. Cross-Cutting Design Decisions
+
+| Decision | Rationale | Tradeoff |
+|---|---|---|
+| **XML tool calls in text, not native function-calling** | Multi-provider compatibility, streaming truncation control, client-side parsing, partial-tag recovery. | Tool descriptions re-injected as text every call (no schema cache); custom parser to maintain. |
+| **Thread-based subagents (daemon threads, own event loops)** | Simple parent-child coordination, shared sandbox, true `asyncio.create_task` not enough because some tools are blocking. | GIL-bound; no real CPU parallelism; daemon threads die on process exit (acceptable for CLI). |
+| **Direct dict messaging, no broker** | Simple, fast, in-process. | Single-process only; no durability; zero distribution. |
+| **One container per scan, not per agent** | Shared `/workspace` + Caido capture is the common case for security work; less Docker overhead. | No per-agent isolation — a buggy/exploited tool can affect other agents in the same scan. |
+| **No CPU/memory/network limits on sandbox container** | Pentest tools (nmap, etc.) need raw socket access and can be heavy. `NET_ADMIN`+`NET_RAW` capabilities granted. | Container could exhaust host resources or DoS targets if the agent goes wrong. Operator's responsibility. |
+| **`--disable-web-security` on Chromium** | Required for XSS / CORS testing. | Browser isn't a realistic UA mirror — can yield findings that don't repro in a normal browser. |
+| **System-wide proxy via `/etc/profile.d/proxy.sh` + CA installed in NSS + system trust** | Caido captures all HTTP/HTTPS from the container by default — agents don't need to configure proxies per-tool. | Anything in the container talking off-box is observable to Caido; don't run untrusted secrets through there. |
+| **Custom PS1 for tmux exit-code extraction** | Reliable exit code detection across arbitrary shells (`[STRIX_$?]$ ` → regex). | Breaks if user-supplied scripts mess with PS1. |
+| **Memory compressor summarizes via *separate* LLM call** | Preserves operationally-critical findings instead of dropping them. | Extra cost per compression cycle; compression timeout 120 s can stall the loop. |
+| **Ephemeral Anthropic prompt cache only** | Simple — system prompt cache resets per request. | Lost cache benefit between calls if conversation history mutates (which it always does). |
+| **Subagent stats finalized into a global dict on exit** | Avoids race conditions when the root queries during child execution. Required for accurate root-level totals. | Slight complexity in `_finalize_agent_llm_stats`. |
+| **Tool result truncation at 10 KB head/tail** | Keeps context spend bounded. | Loses information; the model sometimes can't see the part it needs. |
+| **Vulnerability dedup via LLM similarity** | Catches semantic duplicates that hash-based dedup would miss. | Costs another LLM call per finding submission. |
+| **Hard tool-call requirement enforced by prompt** | Models prone to outputting "Planning..." with no tool call, which the loop interprets as "wait for user" and halts. | Strong language in system prompt; need `think` as escape hatch. |
+| **Diff scope as prompt-injected metadata, not filesystem filter** | Agent can still read related files (imports, helpers) when investigating diffed code. | Larger context spend; relies on model self-restraint to actually focus on the diff. |
+| **Skills are Markdown files, not Python plugins** | Lower contributor friction, no import system to maintain, easy to ship. | No programmatic logic in skills — they're always pure prompt content. |
+
+---
+
+## 15. Recent Evolution (Notable Commits)
+
+| Commit | Subject | What Changed |
+|---|---|---|
+| `9fb1012` | `--config` flag now fully overrides `~/.strix/cli-config.json` (#457) | Track default-applied vars in `Config._applied_from_default`; clear them before loading custom config. New test in `tests/test_config_override.py`. |
+| `60abc09` | wrap acompletion in asyncio.wait_for to prevent indefinite hangs (#453) | Wrap **both** `acompletion()` and per-chunk `__anext__()` in `asyncio.wait_for`. Bedrock TCP-accept-then-silence stalls now raise `TimeoutError` (status_code=None) → retried via `_should_retry`. |
+| `8841294` | feat(skills): add Kubernetes security testing skill (#394) | New `strix/skills/cloud/kubernetes.md`. RBAC, container escapes, etcd, supply chain. |
+| `5c13348` | feat: Add NoSQL injection vulnerability guide (#168) | New `strix/prompts/vulnerabilities/nosql_injection.jinja` covering Mongo/Couch/Redis/Cassandra/Neo4j/GraphQL. |
+| `15c9571` | fix: ensure LLM stats tracking is accurate by including completed subagents (#441) | `_finalize_agent_llm_stats()` snapshots subagent stats into `_completed_agent_llm_totals` under lock; tracer aggregates live + completed. Root no longer undercounts. |
+| `62e9af3` | Add Strix GitHub Actions integration tip | README addition only. |
+| `38b2700` | feat: Migrate from Poetry to uv (#379) | Build system now `hatchling`; deps managed by `uv`; lockfile `uv.lock`. |
+| `e78c931` | feat: Better source-aware testing (#391) | Whitebox skill set hardened (`coordination/source_aware_whitebox`, `custom/source_aware_sast`). |
+| `4934bb8` | chore: upgrade litellm and sanitize tool result text | Bump litellm to `>=1.81.1,<1.82.0`; tool result sanitization (screenshot extraction, length truncation, error truncation). |
+| `7d5a45d` | chore: bump version to 0.8.3 | PyPI version bump. |
+| `dec2c47` | fix: use anthropic model in anthropic provider docs example | Doc only. |
+| `4f90a56` | fix: strengthen tool-call requirement in interactive and autonomous modes | Hardens `system_prompt.jinja:24-44`. Explicit "message without tool call IMMEDIATELY STOPS execution" and named exceptions. Adds the `think` tool escape hatch. |
+| `640bd67` | chore: bump sandbox image to 0.1.13 | `strix/config/config.py:43`. |
+
+---
+
+## 16. Quick File Index
+
+| Path | Role |
+|---|---|
+| `strix/interface/main.py` | CLI entrypoint. |
+| `strix/interface/cli.py` | Headless mode. |
+| `strix/interface/tui.py` | Textual TUI app. |
+| `strix/interface/utils.py` | Run-name, target inference, diff-scope helpers. |
+| `strix/agents/base_agent.py` | Core agent loop. |
+| `strix/agents/state.py` | `AgentState` model. |
+| `strix/agents/StrixAgent/strix_agent.py` | Root agent + `execute_scan`. |
+| `strix/agents/StrixAgent/system_prompt.jinja` | System prompt. |
+| `strix/llm/llm.py` | LLM wrapper. |
+| `strix/llm/config.py` | `LLMConfig`. |
+| `strix/llm/memory_compressor.py` | History compaction. |
+| `strix/llm/utils.py` | Tool-format normalization, XML parser. |
+| `strix/llm/dedupe.py` | Vulnerability dedup. |
+| `strix/tools/registry.py` | Tool registry + decorator. |
+| `strix/tools/executor.py` | Local + sandbox dispatcher. |
+| `strix/tools/context.py` | Agent ID contextvar. |
+| `strix/tools/argument_parser.py` | XML arg type coercion. |
+| `strix/tools/agents_graph/agents_graph_actions.py` | Multi-agent orchestration. |
+| `strix/tools/finish/finish_actions.py` | `finish_scan` (root only). |
+| `strix/tools/browser/` | Playwright Chromium tool. |
+| `strix/tools/terminal/` | tmux/libtmux tool. |
+| `strix/tools/python/` | IPython tool. |
+| `strix/tools/proxy/` | Caido GraphQL client. |
+| `strix/tools/notes/` | Notes + wiki. |
+| `strix/tools/todo/` | In-memory todos. |
+| `strix/tools/reporting/` | Vulnerability reports + CVSS. |
+| `strix/tools/web_search/` | Perplexity. |
+| `strix/tools/file_edit/` | openhands-aci editor. |
+| `strix/tools/thinking/` | `think` tool. |
+| `strix/tools/load_skill/` | Runtime skill injection. |
+| `strix/runtime/docker_runtime.py` | Host-side container orchestration. |
+| `strix/runtime/tool_server.py` | Sandbox-side FastAPI tool server. |
+| `containers/Dockerfile` | Sandbox image. |
+| `containers/docker-entrypoint.sh` | Container boot sequence. |
+| `strix/config/config.py` | Config + env layering. |
+| `strix/telemetry/tracer.py` | Run tracer + JSONL events. |
+| `strix/telemetry/utils.py` | Scrubadub redaction. |
+| `strix/telemetry/posthog.py` | Anonymous usage telemetry. |
+| `strix/telemetry/flags.py` | Telemetry opt-out resolution. |
+| `strix/utils/resource_paths.py` | Frozen-vs-dev path resolution. |
+| `strix/skills/**/*.md` | Vulnerability + tooling + scan-mode playbooks. |
+| `strix/prompts/vulnerabilities/nosql_injection.jinja` | NoSQLi prompt. |
+| `pyproject.toml` | Deps, entry point, lint config. |
+
+---
+
+*This wiki captures the harness as of `9fb1012`. When in doubt, source wins.*
diff --git a/MIGRATION_EVALUATION.md b/MIGRATION_EVALUATION.md
new file mode 100644
index 00000000..7cfa82bf
--- /dev/null
+++ b/MIGRATION_EVALUATION.md
@@ -0,0 +1,766 @@
+# Migration Evaluation: Strix Custom Harness → OpenAI Agents SDK
+
+> Evaluated against `openai/openai-agents-python` v0.14.6 (`/tmp/openai-agents`). Maps every Strix subsystem from `HARNESS_WIKI.md` (Strix at `9fb1012`) onto SDK primitives.
+>
+> **Revision 2** — incorporates: (a) confirmed multi-agent + messaging is bridgeable via `call_model_input_filter`; (b) accepted tradeoffs on XML tool format, skills-as-tool-output, sandbox subclass; (c) tool-execution threading & timeout deltas (sequential→parallel, no default timeouts, no auto sync offload).
+
+---
+
+## Table of Contents
+
+1. [TL;DR](#1-tldr)
+2. [SDK overview](#2-sdk-overview)
+3. [Per-subsystem mapping (revised)](#3-per-subsystem-mapping-revised)
+4. [Multi-agent design — concrete bridge](#4-multi-agent-design--concrete-bridge)
+5. [Tool execution semantics — what changes](#5-tool-execution-semantics--what-changes)
+6. [Sandbox bridge](#6-sandbox-bridge)
+7. [What we still lose control over](#7-what-we-still-lose-control-over)
+8. [What we gain](#8-what-we-gain)
+9. [Effort estimate (revised)](#9-effort-estimate-revised)
+10. [Migration plan (step-by-step)](#10-migration-plan-step-by-step)
+11. [Risks & open questions](#11-risks--open-questions)
+
+---
+
+## 1. TL;DR
+
+**Verdict: full migration is feasible.** ~25–35 engineer-days for full parity, including parallel multi-agent + messaging, with three accepted tradeoffs and one custom Docker-client subclass.
+
+| Concern | Original status | Revised status |
+|---|---|---|
+| Concurrent multi-agent graph | "Critical / not bridgeable" | **Bridgeable.** `call_model_input_filter` + `asyncio.create_task` + shared `Session` + a `MessageBus` we own. Architecturally identical to today's `_agent_messages` injection in `_check_agent_messages`. ~400 LOC, full parity (true parallel, peer-to-peer messaging, wait_for_message, view_agent_graph, identity injection, stat aggregation). |
+| XML tool-call format | "Critical" | **Accepted tradeoff.** SDK is JSON-native (provider-native via LiteLLM extension for non-OpenAI). No real loss — provider-native tool use is cleaner. Multi-provider survives. |
+| `load_skill` mid-run prompt mutation | "High loss" | **Accepted tradeoff.** Skills returned as tool output; model sees them in conversation history. Slightly more memory-compressor-eviction-prone, but cleaner semantics. |
+| Sandbox `cap_add` / `extra_hosts` | "High" | **Solvable.** Subclass `DockerSandboxClient` and inject the kwargs. ~50–80 LOC. |
+| Tool execution semantics | not addressed | **Net upgrade.** SDK runs tool calls in **parallel** within a turn (Strix is sequential). No default per-tool timeout (Strix has 120s) — we add a `strix_tool()` factory to re-impose defaults. No auto sync→thread offload (Strix's tool server `asyncio.to_thread`s every call) — we wrap sync code ourselves. |
+
+**No remaining showstoppers.** All gaps now have concrete bridges.
+
+---
+
+## 2. SDK overview
+
+`openai-agents` v0.14.6, MIT, Python 3.10+. Core abstractions:
+
+| Concept | Purpose | File |
+|---|---|---|
+| `Agent` | LLM + instructions + tools + handoffs + guardrails | `src/agents/agent.py` |
+| `Runner` / `AgentRunner` | Run loop, max_turns, streaming | `src/agents/run.py`, `run_internal/` |
+| `RunState` / `RunResult` | Run state + result, resumable serialization | `run_state.py`, `result.py` |
+| `Session` | Conversation history persistence (8+ backends) | `memory/`, `extensions/memory/` |
+| `function_tool` / `FunctionTool` | Tool decorator (native function-calling) | `tool.py:1725` |
+| `Handoff` | Linear delegation to another agent | `handoffs/` |
+| `Agent.as_tool()` | Nested agent invocation (blocking) | `tool.py` (`_is_agent_tool`) |
+| `RunHooks` / `AgentHooks` | 7 lifecycle hooks | `lifecycle.py` |
+| Guardrails (input/output/tool) | Three-layer validation | `guardrail.py`, `tool_guardrails.py` |
+| Tracing | Built-in spans, processors, OpenAI dashboard default | `tracing/` |
+| **`call_model_input_filter`** | **Mutate input list before every model call** | `run_config.py:61`, `run_internal/turn_preparation.py:55-80` |
+| `SandboxAgent` (v0.14.0) | Pre-configured agent with sandbox session | `sandbox/`, `extensions/sandbox/` |
+| `Manifest` + capabilities + entries | Sandbox config (env, mounts, capabilities) | `sandbox/manifest.py`, `sandbox/capabilities/` |
+| `MultiProvider`, `LitellmModel`, `AnyLLMModel` | Non-OpenAI provider routing | `models/multi_provider.py`, `extensions/models/` |
+| MCP support | 4 transports (HostedMCPTool, StreamableHttp, Sse, Stdio) | `mcp/` |
+
+Sandbox backends shipped: **UnixLocal, Docker, E2B, Daytona, Modal, Runloop, Vercel, Blaxel, Cloudflare**.
+
+---
+
+## 3. Per-subsystem mapping (revised)
+
+### 3.1 Agent loop & multi-agent (Strix §5)
+
+| Strix capability | SDK equivalent | Match | Notes |
+|---|---|---|---|
+| Single-agent loop with `max_iterations=300` | `Runner.run(max_turns=...)` | Partial | Default is 10; raise via `RunConfig(max_turns=300)`. |
+| 85% / N-3 turn warnings | `RunHooks.on_llm_start` checks `len(input_items)` and pushes a warning user-message | Bridgeable | ~20 LOC. |
+| Streaming early-truncate at `</function>` | `result.cancel(mode="after_turn")` (turn-level only) | Partial | Lose token savings on over-generating models. ~50–100 LOC custom Model wrapper if we want it back. |
+| `AgentState` (parent_id, sandbox_id, audit) | `RunState` (per-run) + `RunContextWrapper.context` (per-agent dict) | Partial | Audit trail moves into hooks/tracer; identity into context dict. |
+| **Concurrent multi-agent graph** | **`asyncio.create_task(Runner.run(...))` + shared `SandboxRunConfig.session` + `MessageBus` + `call_model_input_filter`** | **1:1 (bridge built in §4)** | True parallel children, peer-to-peer messaging, wait/timeout, agent graph view. |
+| `view_agent_graph` text rendering | Bus traversal helper | 1:1 | Ours, ~30 LOC. |
+| Subagent identity injection (`<agent_delegation>` XML) | Set `agent_id`/`parent_id`/`agent_name` in `RunContextWrapper.context`; child instructions are a callable that pulls from context | 1:1 | Same effect, no XML. |
+| Cancellation (`cancel_current_execution`) | `task.cancel()` on the `asyncio.Task` we own (one per agent in the bus) | 1:1 | Identical primitive. |
+| Interactive "waiting state" with timeout | `wait_for_message` tool polls bus inbox via `asyncio.sleep` | 1:1 | Same semantics, ~20 LOC. |
+| Subagent stat aggregation | `RunHooks.on_llm_end` pushes usage to bus; `on_agent_end` finalizes | 1:1 | Cleaner than today's `_completed_agent_llm_totals` lock-protected dict. |
+| Lifecycle hooks (implicit today) | `RunHooks` + `AgentHooks` (7 hooks) | **Gain** | Use these to wire tracer + stats. |
+| Memory compression (90K, last-15 floor, LLM summary) | Custom `Session` subclass with `compact()` hook | Bridgeable | ~150 LOC. Ports our existing `MemoryCompressor` strategy. |
+
+### 3.2 LLM layer (Strix §6)
+
+| Strix capability | SDK equivalent | Match |
+|---|---|---|
+| `litellm.acompletion` multi-provider | Native OpenAI + `LitellmModel` (extras: `litellm`) + `AnyLLMModel` (extras: `any-llm`) | 1:1 — pick `LitellmModel` for parity. |
+| `MultiProvider` prefix routing (`openai/`, `litellm/anthropic/`) | `MultiProvider` + `MultiProviderMap` | 1:1 — direct equivalent. |
+| Strix model aliasing (`strix/claude-sonnet-4.6` → `anthropic/claude-sonnet-4-6` + custom `api_base`) | Custom `ModelProvider` subclass reading our alias map | Bridgeable | ~50 LOC. |
+| Anthropic prompt caching auto-injection | `ModelSettings(extra_body={"cache_control": {"type": "ephemeral"}})` per Anthropic agent | Partial | Per-agent manual or via a small `make_anthropic_settings()` helper. ~30 LOC. |
+| Reasoning effort (env > config > scan-mode default) | `ModelSettings(reasoning=Reasoning(effort=...))` | 1:1. |
+| Streaming early-exit at `</function>` | None native | Partial — lose token savings; custom Model subclass to restore. |
+| Per-chunk streaming timeout (Bedrock fix) | None native | Partial — wrap streaming in custom Model subclass if Bedrock matters. |
+| Retries (`min(90, 2*2^n)`, max 5, custom `_should_retry`) | `ModelSettings(retry=ModelRetrySettings(...))` + `retry_policies.*` | **Gain** — composable. |
+| Memory compression with pentest-tuned summary prompt | Custom `Session` subclass | Bridgeable | ~150 LOC. |
+| `_strip_images()` for vision-less models | None automatic | Wrap as Model subclass or pre-filter. ~40 LOC. |
+| Per-call `RequestStats` w/ cost via `litellm.completion_cost` | `Usage` (tokens only) | Partial — wire `litellm.completion_cost` in `on_llm_end` hook. ~20 LOC. |
+| Vulnerability dedup (separate LLM call) | Function tool that calls a nested `Runner.run` or direct LiteLLM | 1:1 — port as-is. |
+| Custom Jinja system prompt | `Agent.instructions: str | Callable[..., str]` | 1:1 — pre-render Jinja before agent creation, or pass an async callable. |
+
+### 3.3 Tool system (Strix §7)
+
+All 13 Strix tools port. Multi-agent-graph tools are now in §4.
+
+| Strix tool | SDK primitive | Effort |
+|---|---|---|
+| `@register_tool` w/ env-conditional registration | `@function_tool` + per-agent `tools=[...]` list assembled via env checks at agent build time | Low |
+| Local-vs-sandbox dispatch | All tools are `@function_tool` async. Sandbox tools are wrappers that POST to our existing FastAPI tool server. **Network isolation + Bearer auth survive at the transport layer.** | Medium |
+| Result XML wrap + 10KB head/tail truncation + screenshot extraction | `ToolOutputText` / `ToolOutputImage` / `ToolOutputFileContent`; truncation logic in our wrapper | Low–Medium |
+| Sequential tool execution | **SDK runs tool calls in parallel within a turn** (see §5). Net gain. Verify our stateful tools are reentrant-safe (browser singleton already is via its `threading.Lock`). | n/a |
+| Argument validation → error string | Pydantic from signature; default `failure_error_function` returns error string | 1:1 |
+| Browser (Playwright, 24 actions) | `ComputerTool(computer=AsyncComputer subclass)` — keeps our Playwright code as the implementation | ~200 LOC |
+| Terminal (libtmux, custom PS1 exit-code regex) | `ShellTool(executor=...)` w/ libtmux, or `@function_tool`. **Wrap libtmux calls in `asyncio.to_thread` ourselves** | ~300 LOC |
+| Python (IPython, stateful) | `@function_tool` + module-level kernel dict keyed by `agent_id` from context | ~200 LOC |
+| Caido proxy (7 GraphQL tools) | 7× `@function_tool` | ~150 LOC |
+| Notes (in-memory + JSONL + wiki MD) | 5× `@function_tool` | ~100 LOC |
+| Todos (in-memory) | 6× `@function_tool` | ~80 LOC |
+| Reporting (CVSS, dedup) | `@function_tool` + Pydantic + cvss lib + nested Runner for dedup | ~150 LOC |
+| Web search (Perplexity) | `@function_tool` (we keep Perplexity, ignore SDK's OpenAI-only `WebSearchTool`) | ~50 LOC |
+| File edit (openhands-aci + ripgrep) | `@function_tool` wrappers | ~60 LOC |
+| Finish scan (root-only guard) | `@function_tool` + context-introspection guard (`parent_id is None`) | ~50 LOC |
+| Thinking | Trivial `@function_tool` | ~10 LOC |
+| **Multi-agent graph (6 tools)** | **§4** — `function_tool` over `MessageBus` | ~400 LOC |
+| **`load_skill`** | **`function_tool` returning skill content as tool output (accepted tradeoff)** | ~60 LOC |
+| `current_agent_id` ContextVar propagation | `RunContextWrapper.context["agent_id"]` + `get_agent_id(ctx)` helper | Low |
+| Tool guardrails (manual arg validation today) | `ToolInputGuardrail` / `ToolOutputGuardrail` | **Gain** |
+
+### 3.4 Sandbox / runtime (Strix §8)
+
+| Strix capability | SDK equivalent | Match |
+|---|---|---|
+| Custom Kali image | `DockerSandboxClientOptions(image="ghcr.io/.../strix-sandbox:0.1.13")` | 1:1 |
+| `cap_add=NET_ADMIN,NET_RAW` + `extra_hosts=host.docker.internal` | **Subclass `DockerSandboxClient`, inject into `containers.create()` kwargs** | Bridgeable | ~80 LOC |
+| Caido HTTPS proxy + CA cert + system-wide proxy env | Image-baked (Dockerfile + entrypoint stay as-is); `Manifest.environment` for runtime overrides; custom `CaidoCapability` for the 7 Caido tools + system-prompt instruction block | Bridgeable | ~200 LOC capability |
+| FastAPI tool server + Bearer auth | **Stays in the image.** Function tools wrap HTTP calls to it. Network isolation + Bearer auth preserved at transport layer. SDK's "in-process tools" model becomes "function tool that POSTs to localhost:48081 inside our shared session." | 1:1 in effect |
+| One container per scan, shared by all agents | `SandboxRunConfig(session=shared_session)` passed into every `Runner.run` call | 1:1 |
+| Random host port allocation | We pre-allocate via `socket.bind(0)` and pass to `DockerSandboxClientOptions(exposed_ports=...)` | 1:1 |
+| Healthcheck polling | External loop after `client.create()`, polling `session.exec("curl -fs localhost:48081/health")` | Bridgeable | ~30 LOC |
+| Container reuse keyed by scan_id | We track our own session map | 1:1 |
+| Local source tar-pipe to `/workspace` | `Manifest.entries={"sources": LocalDir(src=Path)}` | 1:1+ — SDK is a strict superset (LocalDir, LocalFile, GitRepo, S3Mount, …) |
+| Multi-agent silo via `agent_id` ContextVar | `RunContextWrapper.context["agent_id"]` extracted in stateful tools | 1:1 (explicit instead of implicit) |
+| Cleanup via async `docker rm -f` | `await client.delete(session)` wrapped in `try/finally` | 1:1 |
+
+### 3.5 Interface, prompts, skills, config, telemetry (Strix §9–§13)
+
+| Strix capability | SDK equivalent | Match |
+|---|---|---|
+| Textual TUI | Re-point at `Runner.run_streamed().stream_events()` | Bridgeable — our existing TUI code, new event source |
+| Headless / `-n` flag / exit code 2 | `Runner.run()` + app-layer exit codes | 1:1 |
+| CLI args | App layer; SDK has no CLI | 1:1 — keep our argparse |
+| Run directory layout | Custom trace processor + result-persistence layer | Bridgeable | ~100 LOC |
+| Built-in tracing | `tracing/` w/ custom processors; default exports to OpenAI dashboard — disable for local-only | Partial | ~40 LOC custom JSONL processor |
+| OTel / Traceloop export | Custom processor wrapping OTLP | ~30 LOC |
+| Scrubadub PII redaction | Custom trace processor — keeps our scrubadub + regex stack | ~60 LOC |
+| Live streaming content updates 2 Hz | `RunResultStreaming.stream_events()` (event-driven, not polled) | **Gain** |
+| PostHog anonymous telemetry | Keep our own implementation | 1:1 |
+| Sessions / persistence | 8+ backends (SQLite, Redis, SQLAlchemy, Mongo, Dapr, Encrypted, OpenAIResponsesCompaction, …) | **Gain** — we have nothing today |
+| Input/output/tool guardrails | Three-layer guardrail system | **Gain** |
+| Lifecycle hooks | `RunHooks` / `AgentHooks` | **Gain** |
+| Jinja system prompt rendering (32 KB) | `Agent.instructions: Callable[..., str]` runs at run start | 1:1 — pre-render Jinja in callable |
+| Tool-call requirement enforcement | `ModelSettings(tool_choice="required")` + `Agent.reset_tool_choice=True` | **Gain** — native enforcement |
+| Skills as Markdown playbooks | App-layer string management (read MD, render to instructions or tool output) | 1:1 |
+| Dynamic skill injection mid-run | **`load_skill` returns skill content as tool output (accepted tradeoff)** | Lossy but acceptable |
+| Vulnerability prompts (NoSQLi etc.) | App-layer string management | 1:1 |
+| Config file `~/.strix/cli-config.json` w/ `--config` override | Keep our `Config` class; sets env vars before SDK init | 1:1 |
+| `RunConfig` per-run knobs | `RunConfig` dataclass — strict superset | **Gain** |
+| Agent graph visualization | `agents.extensions.visualization.draw_graph()` (static Graphviz) + our `view_agent_graph` tool (live) | 1:1 |
+| Logging | `openai.agents` + `openai.agents.tracing` loggers | 1:1 |
+
+---
+
+## 4. Multi-agent design — concrete bridge
+
+This was the contested section in the previous evaluation. **It's bridgeable, the bridge is small, and the architecture is identical to today's Strix in shape — just lives in our code on top of SDK primitives.**
+
+### 4.1 The key SDK hook
+
+`run_config.py:61` defines:
+
+```python
+CallModelInputFilter = Callable[[CallModelData[Any]], MaybeAwaitable[ModelInputData]]
+```
+
+This filter runs **before every model call** (`run_internal/turn_preparation.py:55-80`). It receives the input list + instructions and returns a (possibly mutated) `ModelInputData(input=[...], instructions=...)`. **This is the exact injection point Strix uses today** in `_check_agent_messages` at the top of every iteration. It's the missing piece.
+
+### 4.2 Architecture
+
+```
+ ┌──────────────────────────────────────────────┐
+ │ AgentMessageBus (we own; ~150 LOC) │
+ │ inboxes: {agent_id -> list[msg]} │
+ │ tasks: {agent_id -> asyncio.Task} │
+ │ statuses: {agent_id -> running|...} │
+ │ parent_of: {agent_id -> parent_id|None} │
+ │ stats_live, stats_completed (under lock) │
+ └─┬──────────────────────────────────┬─────────┘
+ │ │
+ │ create_agent (function_tool) │ on_llm_end / on_agent_end
+ │ asyncio.create_task( │ (RunHooks)
+ │ Runner.run(child, ..., │
+ │ run_config=RunConfig( │ ──► record_usage,
+ │ sandbox=SandboxRunConfig( │ finalize_stats
+ │ session=SHARED), │
+ │ call_model_input_filter= │
+ │ inject_messages_filter, │
+ │ ), │
+ │ context={"bus": bus, │
+ │ "agent_id": child, │
+ │ "parent_id": me, │
+ │ "session": ...}) │
+ │ ) │
+ ▼ ▼
+ Child Runner runs in parallel Parent's next LLM call:
+ (asyncio task, true call_model_input_filter
+ I/O concurrency). drains inbox, appends msgs
+ as user-role items.
+```
+
+### 4.3 The bus (~150 LOC)
+
+```python
+# strix/orchestration/bus.py
+import asyncio
+from dataclasses import dataclass, field
+
+@dataclass
+class AgentMessageBus:
+ inboxes: dict[str, list[dict]] = field(default_factory=dict)
+ tasks: dict[str, asyncio.Task] = field(default_factory=dict)
+ statuses: dict[str, str] = field(default_factory=dict)
+ parent_of: dict[str, str | None] = field(default_factory=dict)
+ names: dict[str, str] = field(default_factory=dict)
+ stats_live: dict[str, dict] = field(default_factory=dict)
+ stats_completed: dict[str, dict] = field(default_factory=dict)
+ _lock: asyncio.Lock = field(default_factory=asyncio.Lock)
+
+ async def register(self, agent_id, name, parent_id):
+ async with self._lock:
+ self.inboxes[agent_id] = []
+ self.statuses[agent_id] = "running"
+ self.parent_of[agent_id] = parent_id
+ self.names[agent_id] = name
+
+ async def send(self, target, msg):
+ async with self._lock:
+ self.inboxes.setdefault(target, []).append(msg)
+
+ async def drain(self, agent_id):
+ async with self._lock:
+ msgs = self.inboxes.get(agent_id, [])
+ self.inboxes[agent_id] = []
+ return msgs
+
+ async def record_usage(self, agent_id, usage):
+ async with self._lock:
+ stats = self.stats_live.setdefault(agent_id, {"in": 0, "out": 0, "cached": 0, "cost": 0})
+ stats["in"] += usage.input_tokens
+ stats["out"] += usage.output_tokens
+ stats["cached"] += usage.input_tokens_details.cached_tokens or 0
+
+ async def finalize(self, agent_id, status):
+ async with self._lock:
+ self.statuses[agent_id] = status
+ self.stats_completed[agent_id] = self.stats_live.pop(agent_id, {})
+
+ async def total_stats(self):
+ async with self._lock:
+ agg = {"in": 0, "out": 0, "cached": 0, "cost": 0}
+ for s in (*self.stats_live.values(), *self.stats_completed.values()):
+ for k, v in s.items():
+ agg[k] = agg.get(k, 0) + v
+ return agg
+```
+
+### 4.4 The injector (~30 LOC)
+
+```python
+# strix/orchestration/filter.py
+from agents.run_config import CallModelData, ModelInputData
+
+async def inject_messages_filter(data: CallModelData) -> ModelInputData:
+ bus = data.context["bus"]
+ agent_id = data.context["agent_id"]
+ pending = await bus.drain(agent_id)
+ if not pending:
+ return data.model_data
+ new_input = list(data.model_data.input)
+ for msg in pending:
+ sender = msg.get("from", "unknown")
+ if sender == "user":
+ new_input.append({"role": "user", "content": msg["content"]})
+ else:
+ new_input.append({
+ "role": "user",
+ "content": (
+ f"<inter_agent_message from='{sender}' "
+ f"type='{msg.get('type', 'info')}' "
+ f"priority='{msg.get('priority', 'normal')}'>"
+ f"{msg['content']}"
+ f"</inter_agent_message>"
+ ),
+ })
+ return ModelInputData(input=new_input, instructions=data.model_data.instructions)
+```
+
+### 4.5 The hooks (~50 LOC)
+
+```python
+# strix/orchestration/hooks.py
+from agents import RunHooks
+
+class StrixOrchestrationHooks(RunHooks):
+ async def on_llm_end(self, ctx, agent, response):
+ bus = ctx.context["bus"]
+ await bus.record_usage(ctx.context["agent_id"], response.usage)
+
+ async def on_agent_end(self, ctx, agent, output):
+ bus = ctx.context["bus"]
+ await bus.finalize(ctx.context["agent_id"], "completed")
+
+ async def on_tool_start(self, ctx, agent, tool):
+ # Bridge to our existing Tracer
+ ctx.context["tracer"].log_tool_start(ctx.context["agent_id"], tool.name)
+
+ async def on_tool_end(self, ctx, agent, tool, result):
+ ctx.context["tracer"].log_tool_end(ctx.context["agent_id"], tool.name, result)
+```
+
+### 4.6 The six multi-agent tools (~250 LOC, replacing 839 LOC of `agents_graph_actions.py`)
+
+```python
+# strix/tools/agents_graph.py
+import asyncio, uuid
+from agents import function_tool, RunContextWrapper, Runner
+from agents.run import RunConfig
+from agents.sandbox import SandboxRunConfig
+
+@function_tool
+async def create_agent(
+ ctx: RunContextWrapper,
+ name: str,
+ task: str,
+ inherit_context: bool = True,
+ skills: list[str] | None = None,
+) -> str:
+ bus = ctx.context["bus"]
+ parent_id = ctx.context["agent_id"]
+ child_id = uuid.uuid4().hex[:8]
+ await bus.register(child_id, name, parent_id)
+
+ child_agent = build_strix_agent(name=name, skills=skills or [])
+ history = (
+ await ctx.context["session"].get_items() if inherit_context else []
+ )
+
+ bus.tasks[child_id] = asyncio.create_task(
+ Runner.run(
+ child_agent,
+ input=history + [{"role": "user", "content": task}],
+ run_config=RunConfig(
+ sandbox=SandboxRunConfig(session=ctx.context["sandbox_session"]),
+ call_model_input_filter=inject_messages_filter,
+ model_settings=ctx.context["model_settings"],
+ max_turns=300,
+ ),
+ context={
+ "bus": bus,
+ "agent_id": child_id,
+ "parent_id": parent_id,
+ "agent_name": name,
+ "session": ctx.context["session"],
+ "sandbox_session": ctx.context["sandbox_session"],
+ "tracer": ctx.context["tracer"],
+ "model_settings": ctx.context["model_settings"],
+ },
+ hooks=StrixOrchestrationHooks(),
+ )
+ )
+ return f"Spawned agent {child_id} ({name}) running in parallel."
+
+@function_tool
+async def send_message_to_agent(
+ ctx: RunContextWrapper,
+ target_agent_id: str,
+ message: str,
+ message_type: str = "info",
+ priority: str = "normal",
+) -> str:
+ await ctx.context["bus"].send(target_agent_id, {
+ "from": ctx.context["agent_id"],
+ "content": message,
+ "type": message_type,
+ "priority": priority,
+ })
+ return f"Message queued for {target_agent_id}."
+
+@function_tool
+async def wait_for_message(
+ ctx: RunContextWrapper, reason: str, timeout_seconds: int = 600
+) -> str:
+ bus = ctx.context["bus"]
+ me = ctx.context["agent_id"]
+ bus.statuses[me] = "waiting"
+ deadline = asyncio.get_event_loop().time() + timeout_seconds
+ while asyncio.get_event_loop().time() < deadline:
+ if bus.inboxes.get(me):
+ bus.statuses[me] = "running"
+ return "Message arrived. Continue your task."
+ await asyncio.sleep(1)
+ bus.statuses[me] = "running"
+ return f"Timed out after {timeout_seconds}s. Continue or call agent_finish."
+
+@function_tool
+async def agent_status(ctx: RunContextWrapper, agent_id: str) -> str:
+ bus = ctx.context["bus"]
+ if agent_id not in bus.statuses:
+ return f"Unknown agent {agent_id}."
+ return (
+ f"agent={bus.names.get(agent_id)} status={bus.statuses[agent_id]} "
+ f"parent={bus.parent_of.get(agent_id)} "
+ f"pending_msgs={len(bus.inboxes.get(agent_id, []))}"
+ )
+
+@function_tool
+async def view_agent_graph(ctx: RunContextWrapper) -> str:
+ bus = ctx.context["bus"]
+ lines = []
+ roots = [aid for aid, p in bus.parent_of.items() if p is None]
+ def render(aid, depth):
+ lines.append(" " * depth + f"- {bus.names.get(aid, '?')} ({aid}) [{bus.statuses.get(aid)}]")
+ for child, p in bus.parent_of.items():
+ if p == aid:
+ render(child, depth + 1)
+ for root in roots:
+ render(root, 0)
+ return "\n".join(lines) or "No agents."
+
+@function_tool
+async def agent_finish(
+ ctx: RunContextWrapper,
+ result_summary: str,
+ findings: list[dict] | None = None,
+ success: bool = True,
+ report_to_parent: bool = True,
+ final_recommendations: list[str] | None = None,
+) -> str:
+ bus = ctx.context["bus"]
+ me = ctx.context["agent_id"]
+ parent = bus.parent_of.get(me)
+ if parent is None:
+ return "Error: agent_finish is for subagents. Root agent must call finish_scan."
+ if report_to_parent:
+ report_xml = (
+ f"<agent_completion_report from='{bus.names.get(me)}' agent_id='{me}' "
+ f"success='{success}'>\n"
+ f" <summary>{result_summary}</summary>\n"
+ f" <findings>{findings or []}</findings>\n"
+ f" <recommendations>{final_recommendations or []}</recommendations>\n"
+ f"</agent_completion_report>"
+ )
+ await bus.send(parent, {"from": me, "content": report_xml, "type": "completion"})
+ await bus.finalize(me, "completed" if success else "failed")
+ return "Reported to parent. This agent will exit."
+```
+
+### 4.7 Capability-by-capability mapping
+
+| Strix today | SDK bridge | Identical? |
+|---|---|---|
+| Daemon-thread subagent (`threading.Thread`) | `asyncio.create_task(Runner.run(...))` | **Yes in effect.** LLM calls are I/O-bound; both designs get the same effective concurrency. We were never CPU-bound at the agent level. |
+| Shared `/workspace` Kali sandbox | Shared `SandboxRunConfig(session=...)` passed to every child's `RunConfig` | Yes |
+| `_agent_messages` inbox | `AgentMessageBus.inboxes` | Yes (renamed) |
+| Per-iteration message check (`_check_agent_messages` at top of `agent_loop`) | `call_model_input_filter` runs before every LLM call (SDK guarantees this in `turn_preparation.py:55-80`) | Yes |
+| `<inter_agent_message>` XML wrap | Filter formats as user-role items with same XML envelope | Yes |
+| `_completed_agent_llm_totals` aggregation | `RunHooks.on_agent_end` snapshots into `bus.stats_completed`, locked | Yes (cleaner) |
+| `wait_for_message` tool with timeout | Tool that polls `bus.inboxes[me]` in `asyncio.sleep` loop | Yes |
+| `view_agent_graph` text output | Bus traversal helper | Yes |
+| Identity injection via `<agent_delegation>` XML | Set identity in `RunContextWrapper.context`; agent's instructions are a callable that pulls from context | Equivalent (no XML wrapping; identity still flows) |
+| Cancellation cascade | `bus.tasks[child_id].cancel()` | Yes — same `asyncio.Task.cancel()` primitive |
+| Stop on parent | Walk descendants via `parent_of`, cancel each task | Yes (same as Strix today) |
+
+### 4.8 What this design does NOT lose
+
+- **True concurrency** at the LLM-I/O boundary. (Python threading was never giving us CPU parallelism for our workload anyway.)
+- **Shared sandbox** semantics — same Kali container, same `/workspace`, same Caido capture, same proxy state.
+- **Cross-sibling messaging** — fully bridged via the bus + filter.
+- **Stat aggregation** — cleaner via hooks.
+- **Per-agent state silo** for stateful tools (browser, terminal, python) — `RunContextWrapper.context["agent_id"]` is the explicit equivalent of the implicit `current_agent_id` ContextVar.
+
+### 4.9 What this design does lose (small)
+
+- **Per-agent task slot serialization** (Strix's tool server cancels a previous in-flight tool when a new one for the same agent arrives). Not actually needed under the SDK because each agent's run loop only emits a new tool call after the previous resolves.
+- **Implicit ContextVar magic** — became explicit `ctx.context["agent_id"]` extraction. ~3 LOC helper makes it ergonomic.
+
+---
+
+## 5. Tool execution semantics — what changes
+
+This is the operational gotcha most likely to surprise during migration. Source-verified from `tool_execution.py` and `tool_server.py` (Strix), `run_internal/tool_execution.py` and `tool.py` (SDK).
+
+### 5.1 Side-by-side
+
+| Dimension | Strix | OpenAI SDK |
+|---|---|---|
+| **Tool calls within one model turn** | **Sequential** (`for inv in invocations` at `executor.py:324`) | **Parallel** (`asyncio.create_task` per call, drained via `asyncio.wait FIRST_COMPLETED` at `tool_execution.py:1412-1430`) |
+| **Default per-tool timeout** | 120s (`STRIX_SANDBOX_EXECUTION_TIMEOUT`) + 30s host buffer = 150s outer | **None.** Must opt in via `@function_tool(timeout=N)` |
+| **Local/host-side tool timeout** | None — runs in main loop | None unless `timeout_seconds` is set |
+| **Sandbox/remote tool timeout** | 120s `asyncio.wait_for` server-side + 150s httpx outer client-side | N/A — SDK has no remote tool concept; we wrap HTTP in a function tool and set timeout ourselves |
+| **Connect timeout** | 10s for httpx → sandbox | None built-in — pass `httpx.Timeout(connect=10)` in our tool body |
+| **Sync function offload** | Tool server: `asyncio.to_thread(tool_func, ...)` always (`tool_server.py:83`) | **No auto-offload.** Sync code blocks the loop unless we wrap with `asyncio.to_thread` ourselves |
+| **Per-agent serialization** | Yes — `agent_tasks[agent_id]`; new request cancels previous (`tool_server.py:94-97`) | No — concurrent calls allowed; not needed anyway since SDK only emits next tool after current resolves |
+| **One-failure-cancels-siblings** | N/A (sequential) | `isolate_parallel_failures=True` by default for multi-call turns (`tool_execution.py:1370`) |
+| **Cancellation primitive** | `task.cancel()` on host; SIGTERM cancels all server tasks | `asyncio.shield(invoke_task)` (`tool_execution.py:1766`) + outer cancellation; `result.cancel()` for whole-run |
+| **Timeout error format** | Returns `"Tool timed out after 120s"` string to the LLM | `default_tool_timeout_error_message(...)` string (or `ToolTimeoutError` if `timeout_behavior="raise_exception"`) |
+| **Stateful-tool threading (browser)** | Dedicated daemon thread + own event loop, lock-serialized (`browser_instance.py:34-48`) | Whatever we build inside the tool function (we keep our existing approach) |
+
+### 5.2 Migration implications
+
+1. **Parallel tool calls become a feature.** Strix is sequential; SDK runs them concurrently. For the model emitting `terminal_execute("nmap ...")` + `web_search("CVE-X")` in one turn, this is faster. We verify reentrancy on:
+ - Browser singleton (already lock-serialized — fine).
+ - Terminal: per-`(agent_id, terminal_id)` tmux session (fine).
+ - Python: per-`(agent_id, session_id)` IPython kernel (fine).
+ - Notes/Todos: thread-safe via existing RLocks (fine).
+
+2. **We re-impose default timeouts via a small factory.**
+
+ ```python
+ # strix/tools/_decorator.py
+ from agents import function_tool
+
+ def strix_tool(*, timeout: float = 120, **kwargs):
+ """Strix-flavored function_tool with our defaults."""
+ return function_tool(
+ timeout=timeout,
+ timeout_behavior="error_as_result",
+ **kwargs,
+ )
+ ```
+
+ Used everywhere we'd write `@function_tool` today.
+
+3. **Sync code wraps in `asyncio.to_thread`.** Our existing libtmux / IPython / Caido sync code goes inside an `async def` tool body:
+
+ ```python
+ @strix_tool(timeout=30)
+ async def terminal_execute(ctx, command: str, ...) -> str:
+ def _run():
+ # libtmux sync code here
+ return session.send_keys(...)
+ return await asyncio.to_thread(_run)
+ ```
+
+ We lose the tool server's auto-offload-everything trick, but we gain explicit control.
+
+4. **Connect timeout becomes our responsibility** for sandbox-bound function tools:
+
+ ```python
+ _SANDBOX_TIMEOUT = httpx.Timeout(timeout=150, connect=10)
+
+ @strix_tool(timeout=160) # outer SDK timeout > inner httpx
+ async def _post_to_sandbox(tool_name, kwargs, ctx):
+ async with httpx.AsyncClient() as client:
+ r = await client.post(..., timeout=_SANDBOX_TIMEOUT)
+ return r.json()
+ ```
+
+5. **Tool error formatting** — set a default `failure_error_function` on `RunConfig` to keep our existing `<tool_result><error>...</error></tool_result>` shape if we want it; otherwise the SDK's default error string is acceptable.
+
+---
+
+## 6. Sandbox bridge
+
+### 6.1 Custom DockerSandboxClient (~80 LOC)
+
+The SDK's `DockerSandboxClient.create()` doesn't expose `cap_add` or `extra_hosts`. Subclass and inject:
+
+```python
+# strix/runtime/strix_docker_client.py
+from agents.sandbox.sandboxes.docker import DockerSandboxClient
+
+class StrixDockerSandboxClient(DockerSandboxClient):
+ """Adds NET_ADMIN, NET_RAW capabilities and host.docker.internal mapping
+ needed for raw-socket pentest tools and host-served-app testing."""
+
+ async def _create_container_kwargs(self, *args, **kwargs):
+ create_kwargs = await super()._create_container_kwargs(*args, **kwargs)
+ create_kwargs.setdefault("cap_add", []).extend(["NET_ADMIN", "NET_RAW"])
+ create_kwargs.setdefault("extra_hosts", {})["host.docker.internal"] = "host-gateway"
+ return create_kwargs
+```
+
+(Exact override point depends on SDK internals — may need to wrap `containers.create` directly. ~80 LOC including verification + tests.)
+
+### 6.2 Caido as a Capability (~200 LOC)
+
+Caido stays in the image (Dockerfile + entrypoint don't change). On the SDK side, it becomes a custom `Capability`:
+
+```python
+# strix/runtime/caido_capability.py
+from agents.sandbox.capabilities import Capability
+
+class CaidoCapability(Capability):
+ async def process_manifest(self, manifest):
+ manifest.environment.update({
+ "http_proxy": "http://127.0.0.1:48080",
+ "https_proxy": "http://127.0.0.1:48080",
+ "ALL_PROXY": "http://127.0.0.1:48080",
+ })
+
+ def tools(self):
+ return [list_requests, view_request, send_request,
+ repeat_request, scope_rules, list_sitemap, view_sitemap_entry]
+
+ async def instructions(self, manifest):
+ return "<caido_proxy>All HTTP/HTTPS traffic in this sandbox is captured by Caido. ...</caido_proxy>"
+```
+
+### 6.3 Tool server stays put
+
+The FastAPI tool server keeps running on `:48081` inside the container. Each Strix tool becomes an `@strix_tool` that POSTs to it with our existing Bearer token. **Network isolation, Bearer auth, and the entire image build pipeline are unchanged.** What changes is only the host-side dispatcher: instead of `tools/executor.py`, it's `function_tool` bodies that call the same endpoint.
+
+---
+
+## 7. What we still lose control over
+
+Smaller list than before. All accepted as tradeoffs.
+
+1. **Streaming early-truncate at `</function>`.** Token waste on over-generating models. Custom Model wrapper if Bedrock economics matter; otherwise live with it.
+2. **Per-chunk streaming timeout** (the Bedrock `60abc09` fix). Same answer — wrap if Bedrock matters.
+3. **Mid-run system prompt mutation (`load_skill`).** Skills become tool outputs (model sees them in conversation history). Slightly more memory-compressor-eviction-prone. Acceptable.
+4. **Anthropic prompt cache auto-injection.** Becomes per-agent manual `extra_body` setting via a small `make_anthropic_settings()` helper.
+5. **Cost tracking.** SDK tracks tokens, not cost. Wire `litellm.completion_cost` in `on_llm_end` hook (~20 LOC).
+6. **Vision-less model image stripping.** No automatic fallback. Wrap as Model subclass if non-vision providers matter.
+7. **Identity injection in delegation** moves from XML to context dict. Equivalent — no real loss.
+
+---
+
+## 8. What we gain
+
+Net upgrades from the SDK. Things we don't have today:
+
+| Gain | Detail |
+|---|---|
+| **Sessions / persistence** | 8+ backends (`SQLiteSession`, `RedisSession`, `SQLAlchemySession`, `MongoDBSession`, `DaprSession`, `EncryptedSession`, `OpenAIConversationsSession`, `OpenAIResponsesCompactionSession`). `RunState.to_json()` resumable runs. We currently have nothing. |
+| **Three-layer guardrails** | `@input_guardrail` / `@output_guardrail` / `@tool_input_guardrail` / `@tool_output_guardrail` with `allow / reject_content / raise_exception` semantics. Our existing manual arg validation becomes a tool guardrail. |
+| **Formal lifecycle hooks** | 7 explicit hooks (`on_llm_start/end`, `on_agent_start/end`, `on_handoff`, `on_tool_start/end`). Replaces our implicit tracer integration. |
+| **Composable retry policies** | `retry_policies.any/provider_suggested/network_error/http_status(...)`. Cleaner than our hard-coded `min(90, 2*2^n)` loop. |
+| **HITL approvals** | `@function_tool(needs_approval=True)`, `state.approve()/reject()` resume flow. We don't have this. |
+| **Parallel tool calls within a turn** | Free speedup for multi-tool model turns. |
+| **Native `tool_choice="required"` enforcement** | The hardened tool-call requirement (commit `4f90a56`) becomes a model setting. |
+| **MCP support** | 4 transports — useful if we ever want to expose Strix tools to other agents (Claude.com, etc.). |
+| **Built-in tracing dashboard** | When we send to the OpenAI backend (off by default for us). |
+| **Active maintenance** | Backed by OpenAI; Strix's harness layer becomes mostly glue. |
+
+---
+
+## 9. Effort estimate (revised)
+
+Wrapper / extension approach (no SDK forking).
+
+| Area | LOC | Days |
+|---|---:|---:|
+| `MultiProvider` config + Strix model alias `ModelProvider` | 60 | 0.5 |
+| Anthropic cache-control helper (`make_anthropic_settings`) | 30 | 0.25 |
+| Streaming early-truncate Model wrapper *(optional, if Bedrock matters)* | 100 | 1.5 |
+| Per-chunk timeout Model wrapper *(optional)* | 100 | 1.5 |
+| Vision-less `_strip_images` Model wrapper *(optional)* | 50 | 0.5 |
+| Cost tracking via `on_llm_end` hook | 30 | 0.5 |
+| Custom `Session` w/ memory compression + pentest summary prompt | 150 | 2 |
+| `strix_tool` decorator (re-imposes our defaults) | 30 | 0.25 |
+| Sandbox tool wrapper (httpx → tool server, Bearer auth, connect timeout) | 80 | 0.5 |
+| Tool ports: browser (AsyncComputer), terminal (libtmux executor + asyncio.to_thread), python (IPython), proxy (7×), notes (5×), todos (6×), reporting (CVSS+dedup), web_search (Perplexity), file_edit (openhands-aci+rg), finish, think | 1500 | 7 |
+| **Multi-agent: `MessageBus` + `inject_messages_filter` + 6 graph tools + `StrixOrchestrationHooks`** | **400** | **4** |
+| `StrixDockerSandboxClient` subclass (`cap_add` + `extra_hosts`) | 80 | 0.5 |
+| `CaidoCapability` (env vars + 7 Caido tools + instructions block) | 200 | 1 |
+| Healthcheck polling layer | 30 | 0.25 |
+| Per-agent state silo helper + ports of stateful tools to use it | 100 | 1 |
+| Custom JSONL trace processor + OTel + scrubadub | 150 | 1.5 |
+| Run-directory persistence (vulns/, notes/, wiki/, penetration_test_report.md) | 100 | 1 |
+| Jinja-rendered `Agent.instructions` callable builder | 60 | 0.5 |
+| Skill-loading workaround (skill content as tool output) | 60 | 0.5 |
+| Config file → env-var bridge | 30 | 0.25 |
+| TUI re-pointing at `Runner.run_streamed().stream_events()` | 200 | 2 |
+| End-to-end tests against migrated harness (smoke + multi-agent + sandbox) | 400 | 4 |
+| **Core (single-provider, no streaming optimizations)** | **~3,800** | **~25 days** |
+| **Full parity (multi-provider + streaming optimizations)** | **~4,000–4,500** | **~30–35 days** |
+
+---
+
+## 10. Migration plan (step-by-step)
+
+Branch: `harness-migration` (already cut). Spike-first; mainline-last.
+
+### Phase 1 — Foundation (~5 days)
+
+1. **Provider layer.** Wire `MultiProvider` + `LitellmModel` for Anthropic. Custom `ModelProvider` for Strix model aliases. Verify our existing models all resolve.
+2. **`strix_tool` decorator.** Re-imposes our 120s default timeout + `error_as_result` behavior + structured error formatting.
+3. **`StrixDockerSandboxClient`.** Subclass injecting `cap_add` + `extra_hosts`. Verify `nmap` works inside a session.
+4. **Custom `Session`.** Port `MemoryCompressor` strategy. Validate against current production transcripts.
+5. **Trace processor.** Custom JSONL exporter + scrubadub PII filter. Wire into `set_default_trace_processors()`.
+
+### Phase 2 — Tool ports (~8 days)
+
+Port one tool category at a time, with end-to-end tests after each:
+
+1. **Sandbox dispatcher** — single function tool that POSTs to FastAPI server. All sandbox-resident tools share this transport.
+2. **Browser** as `ComputerTool` + `AsyncComputer` subclass that wraps existing Playwright code.
+3. **Terminal** — `@strix_tool` wrapping libtmux behind `asyncio.to_thread`.
+4. **Python** — `@strix_tool` wrapping IPython.
+5. **Caido proxy** — 7 GraphQL tools.
+6. **Notes / Todos / Reporting / Web search / File edit / Finish / Thinking** — straightforward `@strix_tool` ports.
+
+### Phase 3 — Multi-agent orchestration (~4 days)
+
+1. **`AgentMessageBus`** + tests.
+2. **`inject_messages_filter`** + tests against synthetic message streams.
+3. **`StrixOrchestrationHooks`** for stat aggregation + tracer wiring.
+4. **6 graph tools** (`create_agent`, `send_message_to_agent`, `wait_for_message`, `agent_status`, `view_agent_graph`, `agent_finish`).
+5. **End-to-end test**: root spawns 2 children in parallel, children exchange messages, both finish, root aggregates stats. Compare to today's baseline.
+
+### Phase 4 — Sandbox + Caido capability (~2 days)
+
+1. **`CaidoCapability`** wires env vars + 7 Caido tools + system-prompt instruction block.
+2. **Healthcheck polling** loop after `client.create()`.
+3. **Container reuse** keyed by scan_id (we own this map; SDK just gives us the session primitive).
+
+### Phase 5 — Interface + persistence (~3 days)
+
+1. **TUI** re-pointed at `Runner.run_streamed().stream_events()`.
+2. **Run-directory layout** rebuilt as a custom processor + result-persistence layer.
+3. **CLI flags** unchanged (we keep our argparse).
+4. **Config file → env-var bridge** unchanged (we keep our `Config` class).
+
+### Phase 6 — Validation (~4 days)
+
+1. Smoke: every tool runs in a sandbox.
+2. Multi-agent: parallel children + messaging + cancel.
+3. Bedrock + Anthropic + OpenAI parity test.
+4. Memory compression at 90K tokens.
+5. PII redaction in traces.
+6. Run an existing pentest end-to-end and diff outputs against the Strix baseline.
+
+---
+
+## 11. Risks & open questions
+
+1. **`CallModelInputFilter` re-runs on every model call.** If we drain the inbox in the filter and the model call retries (e.g. retryable HTTP error), do we lose messages? Need to verify SDK retry behavior — does `call_model_input_filter` re-run on retry, or does the filtered input get cached for the retry? **Action: read `run_internal/turn_preparation.py:55-80` + retry path before Phase 3.** If messages would be lost, the fix is to drain into a per-call buffer that only commits on successful response.
+
+2. **`Session` semantics under parallel children.** When children share the same `Session` for sandbox state, do their LLM histories cross-contaminate? Children should use distinct logical sessions for history (per-agent) but share the sandbox session. **Action: verify `Session` and `SandboxRunConfig.session` are independent — they are, but write a test.**
+
+3. **`isolate_parallel_failures=True` default.** When the model emits multiple tool calls in one turn and one fails, all siblings get cancelled. We may want `False` for our use case (a failed `nmap` shouldn't kill an in-flight `web_search`). **Action: configure per `RunConfig` once we see real behavior.**
+
+4. **Sandbox tool concurrency under parallel calls.** Today's tool server has per-agent task slot serialization (one tool in flight per agent). Under SDK's parallel-tool-calls model, we'd issue multiple POSTs concurrently for the same agent. Tool server's current behavior is to **cancel the previous task** (`tool_server.py:94-97`), which would break us. **Action: relax tool server to allow concurrent same-agent tool calls, OR set `parallel_tool_calls=False` on `ModelSettings` and stay sequential.** The latter is the safer migration default; revisit later.
+
+5. **Bedrock per-chunk-timeout regression.** Without the custom Model wrapper, Bedrock TCP-stalls return as a class of failure. **Action: decide whether Bedrock matters enough to invest the 1.5 days. If it does, build the wrapper in Phase 1.**
+
+6. **Streaming early-exit at `</function>` cost.** Wasted tokens on every multi-turn for over-generating models. Quantify against a representative scan; if cost delta is small, skip the wrapper.
+
+7. **Memory-compressor eviction risk for tool-output skills.** When `load_skill` returns content as tool output, the compressor may summarize the skill content into oblivion after 15+ messages. **Action: tag skill-load tool outputs in the conversation and configure the compressor to preserve them.**
+
+---
+
+*Revision 2 — incorporates: (a) `call_model_input_filter`-based multi-agent bridge; (b) accepted tradeoffs on XML / skills / sandbox subclass; (c) tool execution semantic deltas (parallel by default, no default timeouts, no auto-offload).*
diff --git a/PLAYBOOK.md b/PLAYBOOK.md
new file mode 100644
index 00000000..436cd974
--- /dev/null
+++ b/PLAYBOOK.md
@@ -0,0 +1,1401 @@
+# Strix → OpenAI Agents SDK Migration Playbook
+
+> The day-1 engineering reference. Distillation of two audit rounds and four implementation deep-dives into actionable specs. All SDK references verified against `openai-agents` v0.14.6 at `/tmp/openai-agents`. Strix references at `9fb1012`.
+
+---
+
+## Reading order
+
+1. `HARNESS_WIKI.md` — what we have today.
+2. `MIGRATION_EVALUATION.md` rev 2 — the architectural plan.
+3. `AUDIT.md` — first audit; five plan corrections (C1–C5).
+4. `AUDIT_R2.md` — Round 1 audit; seven additional corrections (C6–C12).
+5. **This file** — file-by-file specs, per-tool contracts, test plans, standards.
+
+---
+
+## Table of contents
+
+1. [Consolidated corrections register](#1-consolidated-corrections-register)
+2. [Foundation files (concrete skeletons)](#2-foundation-files-concrete-skeletons)
+3. [Sandbox + Caido capability](#3-sandbox--caido-capability)
+4. [Per-tool migration contracts](#4-per-tool-migration-contracts)
+5. [Standards: logging, error formatting, observability](#5-standards-logging-error-formatting-observability)
+6. [Test plans per phase](#6-test-plans-per-phase)
+7. [Rollback, CI, cross-cutting](#7-rollback-ci-cross-cutting)
+8. [Phase ordering and day-1 commit list](#8-phase-ordering-and-day-1-commit-list)
+
+---
+
+## 1. Consolidated corrections register
+
+Twelve corrections to apply, ordered by phase. Every fix is bounded and source-verified.
+
+| # | Severity | Defect | Fix shape | Apply in |
+|---|---|---|---|---|
+| **C1** | Blocker | SDK runs tools in parallel within a turn (`run_internal/tool_execution.py:1414, 1424`); Strix tool server cancels previous task for same agent (`tool_server.py:94-97`). | Phase 1 safe default: `ModelSettings(parallel_tool_calls=False)` + `RunConfig(isolate_parallel_failures=False)`. Phase 6 relaxation: remove the cancel logic, allow concurrent same-agent tool calls. | Phase 1 / 6 |
+| **C2** | Blocker | Plan §3.2 said `extra_body["cache_control"]` injects Anthropic prompt cache. Verified — that lands at request level, not on system message. Caches nothing. | `AnthropicCachingLitellmModel` subclass — override `get_response`/`stream_response` to inject `cache_control` on the system message before super delegation. | Phase 0 |
+| **C3** | Blocker | `DockerSandboxClient._create_container()` (`sandbox/sandboxes/docker.py:1434-1477`) has no kwarg-injection hook. | Subclass + duplicate parent body verbatim, append `create_kwargs.setdefault("cap_add", []).extend(["NET_ADMIN","NET_RAW"])` and `create_kwargs.setdefault("extra_hosts", {})["host.docker.internal"] = "host-gateway"`. Pin SDK version. | Phase 0 |
+| **C4** | Blocker | Subagent `agent_finish` returns from a tool; SDK loop continues to `max_turns` unless told to stop. | Every child Agent: `tool_use_behavior={"stop_at_tool_names": ["agent_finish"]}`. Root Agent: `tool_use_behavior={"stop_at_tool_names": ["finish_scan"]}`. | Phase 3 |
+| **C5** | High | TUI today polls `tracer.streaming_content` per-chunk; SDK's `Runner.run_streamed().stream_events()` is event-driven. | `StrixStreamAccumulator` consumes `RawResponsesStreamEvent` + `RunItemStreamEvent` and synthesizes legacy tracer API. Hooks bridge for non-streamed children. | Phase 5 |
+| **C6** | Critical | `notes/notes_actions.py:_append_note_event` writes JSONL without lock. Two concurrent agents corrupt file. | Wrap `f.write(...)` in `with _notes_lock:`. Same for `_persist_wiki_note()`. | Phase 2 |
+| **C7** | Critical | `telemetry/tracer.py:_append_event_record` calls `append_jsonl_record` without acquiring `_get_events_write_lock()`. | Wrap append in the existing lock. Apply identically in our custom `TracingProcessor`. | Phase 1 |
+| **C8** | High | Subagent crash in daemon thread is silent — parent's `wait_for_message` polls forever. Same shape post-migration if a child `Runner.run` task raises. | `StrixOrchestrationHooks.on_agent_end` detects crash (output is None or `agent_finish` flag absent in context); pushes synthetic `<agent_crash>` message to parent inbox so filter surfaces it. | Phase 3 |
+| **C9** | High | Root cancellation does NOT cascade to children spawned via `asyncio.create_task` in `create_agent` tool. | `cancel_run_with_descendants(bus, root_id)` walks `bus.parent_of` tree leaf-first and `task.cancel()`s each. Wired to CLI signal handler + TUI stop. | Phase 3 |
+| **C10** | Medium | Memory compressor LLM call exception bubbles to agent loop, killing the whole iteration. Defeats the purpose. | Custom `Session` wraps compressor invocation in try/except; on failure, returns uncompressed history. Downstream context-window error is itself retryable. | Phase 1 |
+| **C11** | Medium | Strix today fails fast on 401/403/400. SDK retry policy default may include them. | Explicit `ModelRetrySettings.policy = retry_policies.any(network_error(), http_status([429,500,502,503,504]))` — note 401/403/400 NOT included. Bake into `make_run_config()` factory. | Phase 1 |
+| **C12** | Medium | `_completed_agent_llm_totals` read by tracer without lock. | Bus `total_stats()` reads under `asyncio.Lock`; tracer goes through bus. | Phase 3 (in design) |
+
+---
+
+## 2. Foundation files (concrete skeletons)
+
+Seven load-bearing modules, plus three supporting ones. Every skeleton is non-stub — copy-edit, fill, ship.
+
+### 2.1 `strix/llm/anthropic_cache_wrapper.py`
+
+```python
+from typing import Any
+from agents.extensions.models.litellm_model import LitellmModel
+from agents.items import ModelResponse, TResponseInputItem
+from agents.models.interface import ModelTracing
+from agents.model_settings import ModelSettings
+from agents.tool import Tool
+from agents.handoffs import Handoff
+from agents.agent_output import AgentOutputSchemaBase
+from openai.types.responses.response_prompt_param import ResponsePromptParam
+
+
+class AnthropicCachingLitellmModel(LitellmModel):
+ """Inject cache_control on the system message for Anthropic models.
+
+ Why: SDK ModelSettings.extra_body lands cache_control at request level,
+ not on the system message — Anthropic only caches when cache_control is
+ attached to the message itself.
+ """
+
+ def _is_anthropic(self) -> bool:
+ m = self.model.lower()
+ return "anthropic/" in m or "claude" in m
+
+ def _patch(self, items: list[TResponseInputItem]) -> list[TResponseInputItem]:
+ if not self._is_anthropic():
+ return items
+ out: list[TResponseInputItem] = []
+ for item in items:
+ if isinstance(item, dict) and item.get("role") == "system":
+ content = item["content"]
+ if isinstance(content, str):
+ content = [{
+ "type": "text", "text": content,
+ "cache_control": {"type": "ephemeral"},
+ }]
+ out.append({**item, "content": content})
+ else:
+ out.append(item)
+ return out
+
+ # F1 (AUDIT_R3): SDK Model.get_response declares the first 7 params positional-first.
+ async def get_response(
+ self,
+ system_instructions,
+ input,
+ model_settings,
+ tools,
+ output_schema,
+ handoffs,
+ tracing,
+ *,
+ previous_response_id=None,
+ conversation_id=None,
+ prompt=None,
+ ) -> ModelResponse:
+ patched = self._patch(input if isinstance(input, list) else [input])
+ return await super().get_response(
+ system_instructions,
+ patched,
+ model_settings,
+ tools,
+ output_schema,
+ handoffs,
+ tracing,
+ previous_response_id=previous_response_id,
+ conversation_id=conversation_id,
+ prompt=prompt,
+ )
+
+ async def stream_response(
+ self,
+ system_instructions,
+ input,
+ model_settings,
+ tools,
+ output_schema,
+ handoffs,
+ tracing,
+ *,
+ previous_response_id=None,
+ conversation_id=None,
+ prompt=None,
+ ):
+ patched = self._patch(input if isinstance(input, list) else [input])
+ async for ev in super().stream_response(
+ system_instructions,
+ patched,
+ model_settings,
+ tools,
+ output_schema,
+ handoffs,
+ tracing,
+ previous_response_id=previous_response_id,
+ conversation_id=conversation_id,
+ prompt=prompt,
+ ):
+ yield ev
+```
+
+**Tests:** assert system message acquires `cache_control` for Anthropic; assert non-Anthropic passes through; `cache_read_input_tokens > 0` on call 2 of identical-prompt sequence.
+
+---
+
+### 2.2 `strix/runtime/strix_docker_client.py`
+
+```python
+import uuid
+from typing import Any
+from docker.models.containers import Container
+from agents.sandbox.sandboxes.docker import (
+ DockerSandboxClient,
+ _build_docker_volume_mounts,
+ _manifest_requires_fuse,
+ _manifest_requires_sys_admin,
+ _docker_port_key,
+ parse_repository_tag,
+)
+from agents.sandbox.manifest import Manifest
+
+
+class StrixDockerSandboxClient(DockerSandboxClient):
+ """Adds NET_ADMIN, NET_RAW capabilities and host.docker.internal mapping.
+
+ Body is a verbatim copy of DockerSandboxClient._create_container
+ (sandbox/sandboxes/docker.py:1434-1477) with two appends before the
+ final containers.create() call. SDK has no kwarg-injection hook;
+ pin openai-agents version, re-merge on bump.
+ """
+
+ async def _create_container(
+ self,
+ image: str,
+ *,
+ manifest: Manifest | None = None,
+ exposed_ports: tuple[int, ...] = (),
+ session_id: uuid.UUID | None = None,
+ ) -> Container:
+ if not self.image_exists(image):
+ repo, tag = parse_repository_tag(image)
+ self.docker_client.images.pull(repo, tag=tag or None, all_tags=False)
+
+ environment: dict[str, str] | None = None
+ if manifest:
+ environment = await manifest.environment.resolve()
+
+ create_kwargs: dict[str, Any] = {
+ "entrypoint": ["tail"],
+ "image": image,
+ "detach": True,
+ "command": ["-f", "/dev/null"],
+ "environment": environment,
+ }
+
+ if manifest is not None:
+ mounts = _build_docker_volume_mounts(manifest, session_id=session_id)
+ if mounts:
+ create_kwargs["mounts"] = mounts
+ if _manifest_requires_fuse(manifest):
+ create_kwargs.update(
+ devices=["/dev/fuse"],
+ cap_add=["SYS_ADMIN"],
+ security_opt=["apparmor:unconfined"],
+ )
+ elif _manifest_requires_sys_admin(manifest):
+ create_kwargs.update(
+ cap_add=["SYS_ADMIN"],
+ security_opt=["apparmor:unconfined"],
+ )
+
+ if exposed_ports:
+ create_kwargs["ports"] = {
+ _docker_port_key(p): ("127.0.0.1", None) for p in exposed_ports
+ }
+
+ # ---- STRIX additions ----
+ create_kwargs.setdefault("cap_add", []).extend(["NET_ADMIN", "NET_RAW"])
+ create_kwargs.setdefault("extra_hosts", {})["host.docker.internal"] = "host-gateway"
+ # -------------------------
+
+ return self.docker_client.containers.create(**create_kwargs)
+```
+
+**Tests:** mock `docker_client.containers.create`; assert kwargs contain `cap_add ⊇ {NET_ADMIN, NET_RAW}` and `extra_hosts["host.docker.internal"] == "host-gateway"`. Live test: spin up real container, run `nmap -sS scanme.nmap.org` via `session.exec`, assert exit 0.
+
+---
+
+### 2.3 `strix/orchestration/bus.py`
+
+```python
+import asyncio
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class AgentMessageBus:
+ inboxes: dict[str, list[dict]] = field(default_factory=dict)
+ tasks: dict[str, asyncio.Task] = field(default_factory=dict)
+ statuses: dict[str, str] = field(default_factory=dict) # running|waiting|completed|crashed|stopped
+ parent_of: dict[str, str | None] = field(default_factory=dict)
+ names: dict[str, str] = field(default_factory=dict)
+ stats_live: dict[str, dict[str, Any]] = field(default_factory=dict)
+ stats_completed: dict[str, dict[str, Any]] = field(default_factory=dict)
+ _lock: asyncio.Lock = field(default_factory=asyncio.Lock)
+
+ async def register(self, agent_id: str, name: str, parent_id: str | None) -> None:
+ async with self._lock:
+ self.inboxes[agent_id] = []
+ self.statuses[agent_id] = "running"
+ self.parent_of[agent_id] = parent_id
+ self.names[agent_id] = name
+ self.stats_live[agent_id] = {"in": 0, "out": 0, "cached": 0, "cost": 0.0, "calls": 0}
+
+ async def send(self, target: str, msg: dict) -> None:
+ async with self._lock:
+ self.inboxes.setdefault(target, []).append(msg)
+
+ async def drain(self, agent_id: str) -> list[dict]:
+ async with self._lock:
+ msgs = self.inboxes.get(agent_id, [])
+ self.inboxes[agent_id] = []
+ return msgs
+
+ async def record_usage(self, agent_id: str, usage) -> None:
+ if usage is None:
+ return
+ async with self._lock:
+ s = self.stats_live.setdefault(agent_id, {"in": 0, "out": 0, "cached": 0, "cost": 0.0, "calls": 0})
+ s["in"] += getattr(usage, "input_tokens", 0) or 0
+ s["out"] += getattr(usage, "output_tokens", 0) or 0
+ details = getattr(usage, "input_tokens_details", None)
+ s["cached"] += getattr(details, "cached_tokens", 0) or 0 if details else 0
+ s["calls"] += 1
+
+ async def finalize(self, agent_id: str, status: str) -> None:
+ # C13 (AUDIT_R3): clear inbox/parent/name to avoid orphaned-message memory leak
+ # when sibling agents try to send to a finished agent.
+ async with self._lock:
+ self.statuses[agent_id] = status
+ self.stats_completed[agent_id] = self.stats_live.pop(agent_id, {})
+ self.inboxes.pop(agent_id, None)
+ self.parent_of.pop(agent_id, None)
+ self.names.pop(agent_id, None)
+
+ async def total_stats(self) -> dict[str, Any]:
+ async with self._lock:
+ agg = {"in": 0, "out": 0, "cached": 0, "cost": 0.0, "calls": 0}
+ for s in (*self.stats_live.values(), *self.stats_completed.values()):
+ for k, v in s.items():
+ agg[k] = agg.get(k, 0) + v
+ return agg
+
+ async def cancel_descendants(self, root_agent_id: str) -> None:
+ """Walk parent_of tree leaf-first; cancel each task. (C9)"""
+ async with self._lock:
+ queue = [root_agent_id]
+ order: list[str] = []
+ while queue:
+ aid = queue.pop()
+ order.append(aid)
+ queue.extend(c for c, p in self.parent_of.items() if p == aid)
+ tasks = [self.tasks[a] for a in reversed(order) if a in self.tasks]
+ for t in tasks:
+ if not t.done():
+ t.cancel()
+ await asyncio.gather(*[t for t in tasks if not t.done()], return_exceptions=True)
+```
+
+**Tests:** stress concurrent `send`/`drain`; FIFO ordering; `cancel_descendants` cancels children before parent; `total_stats` snapshot is consistent.
+
+---
+
+### 2.4 `strix/orchestration/filter.py`
+
+```python
+import logging
+from agents.run_config import CallModelData, ModelInputData
+
+logger = logging.getLogger(__name__)
+
+
+async def inject_messages_filter(data: CallModelData) -> ModelInputData:
+ """Drain bus inbox; append as user-role items wrapped in <inter_agent_message>.
+
+ Filter runs once per turn; output captured for retries (verified) — safe to drain.
+ C14 (AUDIT_R3): wrap whole body in try/except so a filter bug never tears down
+ the run. On any exception, return unmodified data.model_data.
+ """
+ try:
+ if not isinstance(data.context, dict):
+ return data.model_data
+ bus = data.context.get("bus")
+ agent_id = data.context.get("agent_id")
+ if bus is None or agent_id is None:
+ return data.model_data
+ pending = await bus.drain(agent_id)
+ if not pending:
+ return data.model_data
+
+ new_input = list(data.model_data.input)
+ for msg in pending:
+ sender = msg.get("from", "unknown")
+ if sender == "user":
+ new_input.append({"role": "user", "content": msg["content"]})
+ else:
+ new_input.append({
+ "role": "user",
+ "content": (
+ f"<inter_agent_message from='{sender}' "
+ f"type='{msg.get('type', 'info')}' "
+ f"priority='{msg.get('priority', 'normal')}'>"
+ f"{msg['content']}"
+ f"</inter_agent_message>"
+ ),
+ })
+ return ModelInputData(input=new_input, instructions=data.model_data.instructions)
+ except Exception:
+ logger.exception("inject_messages_filter failed; proceeding without injection")
+ return data.model_data
+```
+
+**Tests:** N pending messages → N items appended in FIFO; user-from-user skips XML wrap; empty inbox passes through; messages preserved across forced LLM retry.
+
+---
+
+### 2.5 `strix/orchestration/hooks.py`
+
+```python
+import logging
+from agents.lifecycle import RunHooks
+# F2 (AUDIT_R3): on_agent_start/on_agent_end receive AgentHookContext, NOT RunContextWrapper.
+from agents.lifecycle import AgentHookContext # type: ignore[attr-defined]
+from agents.run_context import RunContextWrapper
+
+logger = logging.getLogger(__name__)
+
+
+class StrixOrchestrationHooks(RunHooks):
+ # C15: every hook body wrapped in try/except so a hook bug never tears down the run.
+
+ async def on_llm_start(
+ self,
+ context: RunContextWrapper,
+ agent,
+ system_prompt: str | None,
+ input_items: list,
+ ) -> None:
+ try:
+ if not isinstance(input_items, list):
+ return
+ max_turns = context.context.get("max_turns", 300)
+ cur = context.context.get("turn_count", 0)
+ if cur == int(max_turns * 0.85):
+ input_items.append({
+ "role": "user",
+ "content": "<system_warning>You are at 85% of your iteration "
+ "budget. Begin consolidating findings.</system_warning>",
+ })
+ elif cur == max_turns - 3:
+ input_items.append({
+ "role": "user",
+ "content": "<system_warning>You have 3 iterations left. Your next "
+ "tool call MUST be the finish tool.</system_warning>",
+ })
+ except Exception:
+ logger.exception("on_llm_start failed")
+
+ async def on_llm_end(self, context: RunContextWrapper, agent, response) -> None:
+ try:
+ bus = context.context.get("bus")
+ if bus and (aid := context.context.get("agent_id")):
+ await bus.record_usage(aid, getattr(response, "usage", None))
+ context.context["turn_count"] = context.context.get("turn_count", 0) + 1
+ except Exception:
+ logger.exception("on_llm_end failed")
+
+ async def on_agent_start(self, context: AgentHookContext, agent) -> None:
+ # F2: context is AgentHookContext, not RunContextWrapper.
+ try:
+ cap = next(
+ (c for c in (getattr(agent, "capabilities", None) or [])
+ if hasattr(c, "_healthcheck_task")),
+ None,
+ )
+ if cap and getattr(cap, "_healthcheck_task", None) is not None:
+ await cap._healthcheck_task
+ except Exception:
+ logger.exception("on_agent_start failed")
+
+ async def on_agent_end(self, context: AgentHookContext, agent, output) -> None:
+ # F2: context is AgentHookContext.
+ try:
+ bus = context.context.get("bus")
+ if bus is None or (me := context.context.get("agent_id")) is None:
+ return
+ crashed = (output is None) or not context.context.get("agent_finish_called", False)
+ parent = bus.parent_of.get(me)
+ if crashed and parent is not None:
+ await bus.send(parent, {
+ "from": me,
+ "content": (
+ f"<agent_crash agent_id='{me}' name='{bus.names.get(me, me)}'>"
+ "Agent terminated without calling agent_finish. "
+ "Stop waiting on this child."
+ "</agent_crash>"
+ ),
+ "type": "crash",
+ })
+ await bus.finalize(me, "crashed" if crashed else "completed")
+ except Exception:
+ logger.exception("on_agent_end failed")
+
+ async def on_tool_start(self, context: RunContextWrapper, agent, tool) -> None:
+ try:
+ if tracer := context.context.get("tracer"):
+ tracer.log_tool_start(context.context.get("agent_id", "?"), tool.name)
+ except Exception:
+ logger.exception("on_tool_start failed")
+
+ # F2: on_tool_end's `result` param is typed `str` in the SDK.
+ async def on_tool_end(
+ self, context: RunContextWrapper, agent, tool, result: str,
+ ) -> None:
+ try:
+ if tool.name in ("agent_finish", "finish_scan"):
+ context.context["agent_finish_called"] = True
+ if tracer := context.context.get("tracer"):
+ tracer.log_tool_end(context.context.get("agent_id", "?"), tool.name, result)
+ except Exception:
+ logger.exception("on_tool_end failed")
+
+ async def on_handoff(self, context: RunContextWrapper, from_agent, to_agent) -> None:
+ # We don't use SDK handoffs in Strix; multi-agent goes through bus.
+ pass
+```
+
+**Tests:** crash detection fires when agent_finish not called; warnings injected at thresholds; usage recording aggregates.
+
+---
+
+### 2.6 `strix/tools/_decorator.py`
+
+```python
+from agents import function_tool
+
+
+def strix_tool(*, timeout: float = 120.0, timeout_behavior: str = "error_as_result", **kwargs):
+ """function_tool with Strix defaults: 120s timeout, error-as-result behavior.
+
+ SDK auto-threads sync function bodies via asyncio.to_thread (tool.py:1820-1829),
+ so writing `def foo(...)` (sync) works for libtmux/IPython/etc.
+ """
+ return function_tool(
+ timeout=timeout,
+ timeout_behavior=timeout_behavior,
+ **kwargs,
+ )
+```
+
+**Tests:** decorated tool gets timeout; timeout returns error string not exception; sync function auto-threads.
+
+---
+
+### 2.7 `strix/llm/multi_provider_setup.py`
+
+```python
+from agents.models.multi_provider import MultiProvider, MultiProviderMap
+from agents.models.interface import ModelProvider, Model
+from agents.extensions.models.litellm_model import LitellmModel
+from strix.llm.anthropic_cache_wrapper import AnthropicCachingLitellmModel
+
+# Strix model alias map
+STRIX_MODEL_MAP = {
+ "claude-sonnet-4.6": ("anthropic/claude-sonnet-4-5-20250929", "claude-sonnet-4-5"),
+ # ... port from strix/llm/utils.py:STRIX_MODEL_MAP
+}
+STRIX_API_BASE = "https://models.strix.ai/api/v1"
+
+
+class StrixModelProvider(ModelProvider):
+ """Resolve `strix/<short>` aliases via STRIX_MODEL_MAP, route to proxy."""
+
+ def get_model(self, model_name: str | None) -> Model:
+ if model_name is None:
+ raise ValueError("Model name required")
+ api_model, _canonical = STRIX_MODEL_MAP.get(model_name, (model_name, model_name))
+ # Anthropic-prefixed → cache wrapper; otherwise vanilla
+ if api_model.startswith("anthropic/") or "claude" in api_model.lower():
+ return AnthropicCachingLitellmModel(model=api_model, base_url=STRIX_API_BASE)
+ return LitellmModel(model=api_model, base_url=STRIX_API_BASE)
+
+
+class LitellmAnthropicProvider(ModelProvider):
+ """Resolve `litellm/anthropic/...` directly to AnthropicCachingLitellmModel."""
+
+ def get_model(self, model_name: str | None) -> Model:
+ # model_name arrives post-prefix-strip, e.g. "anthropic/claude-3-5-sonnet"
+ return AnthropicCachingLitellmModel(model=model_name)
+
+
+def build_multi_provider() -> MultiProvider:
+ pmap = MultiProviderMap()
+ pmap.add_provider("strix", StrixModelProvider())
+ pmap.add_provider("litellm/anthropic", LitellmAnthropicProvider())
+ # default openai/* and others fall through to MultiProvider's built-in routing
+ return MultiProvider(provider_map=pmap)
+```
+
+**Tests:** alias resolution; route `strix/claude-sonnet-4.6` → `AnthropicCachingLitellmModel("anthropic/claude-sonnet-4-5-20250929", base_url=STRIX_API_BASE)`; unknown prefix falls through.
+
+---
+
+### 2.8 `strix/llm/strix_session.py`
+
+```python
+import logging
+from typing import Any
+from agents.memory.session import SessionABC
+
+logger = logging.getLogger(__name__)
+
+
+class StrixSession(SessionABC):
+ """Wraps an underlying Session; injects memory compression at 90K tokens.
+
+ On compressor failure: returns uncompressed history (C10) and logs warning.
+ Downstream context-window error is itself retryable.
+ """
+
+ def __init__(self, underlying: SessionABC, compressor):
+ self._underlying = underlying
+ self._compressor = compressor # Strix's existing MemoryCompressor
+
+ async def get_items(self, limit: int | None = None) -> list[Any]:
+ items = await self._underlying.get_items(limit=limit)
+ try:
+ return await self._compressor.compress_history(items)
+ except Exception as e:
+ logger.warning("Memory compression failed (%s) — returning uncompressed", e)
+ return items
+
+ async def add_items(self, items: list[Any]) -> None:
+ await self._underlying.add_items(items)
+
+ async def pop_item(self) -> Any | None:
+ return await self._underlying.pop_item()
+
+ async def clear_session(self) -> None:
+ await self._underlying.clear_session()
+```
+
+**Tests:** compression triggers > 90K tokens; failure returns uncompressed; underlying contract satisfied.
+
+---
+
+### 2.9 `strix/telemetry/strix_processor.py`
+
+```python
+import json
+import logging
+import threading
+from pathlib import Path
+from typing import Any
+from agents.tracing.processor_interface import TracingProcessor
+from agents.tracing.spans import Span
+from agents.tracing.traces import Trace
+from strix.telemetry.utils import TelemetrySanitizer
+
+logger = logging.getLogger(__name__)
+
+_FILE_LOCKS: dict[Path, threading.Lock] = {}
+_GUARD = threading.Lock()
+
+
+def _lock_for(path: Path) -> threading.Lock:
+ with _GUARD:
+ return _FILE_LOCKS.setdefault(path, threading.Lock())
+
+
+class StrixTracingProcessor(TracingProcessor):
+ """Custom processor: writes events.jsonl in our schema, scrubs PII, supports OTel.
+
+ C7 — JSONL writes locked via per-path threading.Lock.
+ C16 — every write wrapped in try/except so disk-full doesn't tear down the run.
+ F3 — every hook method is SYNC (def, not async def) per SDK contract.
+ """
+
+ def __init__(self, run_dir: Path, sanitizer: TelemetrySanitizer | None = None):
+ self.events_path = run_dir / "events.jsonl"
+ self.run_dir = run_dir
+ self.sanitizer = sanitizer or TelemetrySanitizer()
+
+ def _emit(self, event: dict[str, Any]) -> None:
+ try:
+ clean = self.sanitizer.sanitize(event)
+ with _lock_for(self.events_path):
+ with self.events_path.open("a", encoding="utf-8") as f:
+ f.write(json.dumps(clean, ensure_ascii=True) + "\n")
+ except OSError:
+ logger.exception("Failed to write event to JSONL")
+
+ def on_trace_start(self, trace: Trace) -> None: # SYNC — F3
+ self._emit({
+ "event_type": "run.started",
+ "trace_id": trace.trace_id,
+ "metadata": getattr(trace, "metadata", {}),
+ })
+
+ def on_trace_end(self, trace: Trace) -> None: # SYNC — F3
+ self._emit({"event_type": "run.completed", "trace_id": trace.trace_id})
+
+ def on_span_start(self, span: Span[Any]) -> None: # SYNC — F3
+ sd = type(span.span_data).__name__
+ if sd in ("AgentSpanData", "GenerationSpanData", "FunctionSpanData"):
+ self._emit({
+ "event_type": f"{sd.replace('SpanData', '').lower()}.started",
+ "span_id": span.span_id,
+ "trace_id": span.trace_id,
+ "data": span.span_data.export(),
+ })
+
+ def on_span_end(self, span: Span[Any]) -> None: # SYNC — F3
+ sd = type(span.span_data).__name__
+ self._emit({
+ "event_type": f"{sd.replace('SpanData', '').lower()}.completed",
+ "span_id": span.span_id,
+ "trace_id": span.trace_id,
+ "data": span.span_data.export(),
+ })
+
+ def force_flush(self) -> None: # SYNC — F3
+ # All writes are sync; nothing to flush.
+ pass
+
+ def shutdown(self) -> None: # SYNC — F3
+ pass
+```
+
+**Tests:** concurrent writes don't corrupt; PII scrubbed; OTel passthrough optional.
+
+---
+
+### 2.10 `strix/run_config_factory.py`
+
+```python
+from pathlib import Path
+from agents import RunConfig
+from agents.sandbox import SandboxRunConfig
+from agents.model_settings import ModelSettings
+from agents.retry import ModelRetrySettings, ModelRetryBackoffSettings, retry_policies
+
+from strix.runtime.strix_docker_client import StrixDockerSandboxClient
+from strix.orchestration.bus import AgentMessageBus
+from strix.orchestration.filter import inject_messages_filter
+from strix.llm.multi_provider_setup import build_multi_provider
+
+
+# Phase 1-5 default. Phase 6 flips parallel_tool_calls=True after relaxing tool server.
+_PHASE1_PARALLEL_DEFAULT = False
+
+
+def make_run_config(
+ *,
+ sandbox_session,
+ bus: AgentMessageBus,
+ model: str = "strix/claude-sonnet-4.6",
+ max_turns: int = 300,
+) -> RunConfig:
+ return RunConfig(
+ model=model,
+ model_provider=build_multi_provider(),
+ model_settings=ModelSettings(
+ parallel_tool_calls=_PHASE1_PARALLEL_DEFAULT, # C1 safe default
+ tool_choice="required",
+ retry=ModelRetrySettings(
+ max_retries=5,
+ backoff=ModelRetryBackoffSettings(
+ initial_delay=2.0, multiplier=2.0, max_delay=90.0, jitter=0.0,
+ ),
+ # C11: explicitly does NOT include 401/403/400
+ policy=retry_policies.any(
+ retry_policies.network_error(),
+ retry_policies.http_status([429, 500, 502, 503, 504]),
+ ),
+ ),
+ ),
+ sandbox=SandboxRunConfig(
+ client=StrixDockerSandboxClient(),
+ session=sandbox_session, # shared across all agents in the run
+ ),
+ call_model_input_filter=inject_messages_filter,
+ isolate_parallel_failures=False, # C1 — don't cascade-cancel siblings
+ tracing_disabled=False,
+ trace_include_sensitive_data=False,
+ max_turns=max_turns,
+ )
+
+
+def make_agent_context(
+ *,
+ bus: AgentMessageBus,
+ sandbox_session,
+ sandbox_token: str,
+ tool_server_host_port: int,
+ caido_host_port: int,
+ agent_id: str,
+ agent_name: str,
+ parent_id: str | None,
+ tracer,
+ model_settings,
+ max_turns: int = 300,
+) -> dict:
+ return {
+ "bus": bus,
+ "sandbox_session": sandbox_session,
+ "sandbox_token": sandbox_token,
+ "tool_server_host_port": tool_server_host_port,
+ "caido_host_port": caido_host_port,
+ "agent_id": agent_id,
+ "agent_name": agent_name,
+ "parent_id": parent_id,
+ "tracer": tracer,
+ "model_settings": model_settings,
+ "max_turns": max_turns,
+ "turn_count": 0,
+ "agent_finish_called": False,
+ }
+```
+
+**Tests:** RunConfig has all defaults; retry policy excludes 401; safe defaults for parallel.
+
+---
+
+## 3. Sandbox + Caido capability
+
+### 3.1 `strix/sandbox/healthcheck.py`
+
+```python
+import asyncio
+import httpx
+
+
+class SandboxNotReadyError(Exception):
+ pass
+
+
+async def wait_for_ports_ready(ports: list[int], timeout: float = 30.0, interval: float = 0.5) -> None:
+ deadline = asyncio.get_event_loop().time() + timeout
+ while asyncio.get_event_loop().time() < deadline:
+ all_ok = True
+ async with httpx.AsyncClient(timeout=5.0) as client:
+ for port in ports:
+ try:
+ r = await client.get(f"http://localhost:{port}/health")
+ if r.status_code != 200:
+ all_ok = False
+ break
+ except (httpx.RequestError, httpx.TimeoutException):
+ all_ok = False
+ break
+ if all_ok:
+ return
+ await asyncio.sleep(interval)
+ raise SandboxNotReadyError(f"Ports {ports} not ready after {timeout}s")
+```
+
+### 3.2 `strix/sandbox/caido_capability.py`
+
+```python
+import asyncio
+from typing import Literal
+from pydantic import Field
+from agents.sandbox.capabilities.capability import Capability
+from agents.sandbox.manifest import Manifest
+from agents.sandbox.session.base_sandbox_session import BaseSandboxSession
+from agents.tool import Tool
+
+# 7 Caido tools defined in strix/tools/caido/*.py and imported here
+from strix.tools.caido import (
+ list_requests, view_request, send_request, repeat_request,
+ scope_rules, list_sitemap, view_sitemap_entry,
+)
+
+
+class CaidoCapability(Capability):
+ """Caido HTTPS proxy + 7 GraphQL function tools + system prompt block."""
+
+ type: Literal["caido"] = "caido"
+ _healthcheck_task: asyncio.Task | None = Field(default=None, exclude=True)
+
+ def process_manifest(self, manifest: Manifest) -> Manifest:
+ env = dict(manifest.environment.value or {})
+ env.update({
+ "http_proxy": "http://127.0.0.1:48080",
+ "https_proxy": "http://127.0.0.1:48080",
+ "ALL_PROXY": "http://127.0.0.1:48080",
+ })
+ manifest.environment.value = env
+ return manifest
+
+ def tools(self) -> list[Tool]:
+ return [list_requests, view_request, send_request, repeat_request,
+ scope_rules, list_sitemap, view_sitemap_entry]
+
+ async def instructions(self, manifest: Manifest) -> str | None:
+ return (
+ "<caido_proxy>\n"
+ "All HTTP/HTTPS traffic in this sandbox is captured by Caido (localhost:48080).\n"
+ "Available tools: list_requests, view_request, send_request, repeat_request, "
+ "scope_rules, list_sitemap, view_sitemap_entry.\n"
+ "HTTPQL filter syntax: request.method == 'POST' && response.status >= 400.\n"
+ "</caido_proxy>"
+ )
+
+ def bind(self, session: BaseSandboxSession) -> None:
+ super().bind(session)
+ from strix.sandbox.healthcheck import wait_for_ports_ready
+ # Phase 4: caido :48080 + tool server :48081
+ self._healthcheck_task = asyncio.create_task(wait_for_ports_ready([48080, 48081]))
+```
+
+`StrixOrchestrationHooks.on_agent_start` awaits `cap._healthcheck_task` (already in 2.5 above).
+
+### 3.3 `strix/sandbox/session_manager.py`
+
+```python
+import socket
+import secrets
+from pathlib import Path
+from agents.sandbox.sandboxes.docker import DockerSandboxClientOptions
+from agents.sandbox.manifest import Manifest
+from agents.sandbox.entries import LocalDir
+from strix.runtime.strix_docker_client import StrixDockerSandboxClient
+from strix.sandbox.caido_capability import CaidoCapability
+from strix.sandbox.healthcheck import wait_for_ports_ready
+
+_session_cache = {} # scan_id -> session
+
+
+def _alloc_port() -> int:
+ s = socket.socket()
+ s.bind(("127.0.0.1", 0))
+ p = s.getsockname()[1]
+ s.close()
+ return p
+
+
+async def create_or_reuse(scan_id: str, image: str, sources_path: Path):
+ if scan_id in _session_cache:
+ return _session_cache[scan_id]
+
+ bearer = secrets.token_urlsafe(32)
+ manifest = Manifest(
+ entries={"sources": LocalDir(src=sources_path)},
+ environment={
+ "TOOL_SERVER_TOKEN": bearer,
+ "TOOL_SERVER_PORT": "48081",
+ "CAIDO_PORT": "48080",
+ "STRIX_SANDBOX_EXECUTION_TIMEOUT": "120",
+ "PYTHONUNBUFFERED": "1",
+ },
+ capabilities=[CaidoCapability()],
+ )
+
+ client = StrixDockerSandboxClient()
+ options = DockerSandboxClientOptions(image=image, exposed_ports=(48080, 48081))
+ session = await client.create(options=options, manifest=manifest)
+
+ # Resolve mapped host ports (sandbox/sandboxes/docker.py:211-260)
+ tool_server_endpoint = await session._resolve_exposed_port(48081)
+ caido_endpoint = await session._resolve_exposed_port(48080)
+ await wait_for_ports_ready([tool_server_endpoint.port, caido_endpoint.port])
+
+ bundle = {
+ "client": client,
+ "session": session,
+ "tool_server_host_port": tool_server_endpoint.port,
+ "caido_host_port": caido_endpoint.port,
+ "bearer": bearer,
+ }
+ _session_cache[scan_id] = bundle
+ return bundle
+
+
+async def cleanup(scan_id: str) -> None:
+ bundle = _session_cache.pop(scan_id, None)
+ if bundle is None:
+ return
+ try:
+ await bundle["client"].delete(bundle["session"])
+ except Exception:
+ pass # best-effort orphan reaping
+```
+
+### 3.4 `strix/tools/_sandbox_dispatch.py`
+
+```python
+import httpx
+from typing import Any
+from agents import RunContextWrapper
+
+_TIMEOUT = httpx.Timeout(connect=10.0, read=150.0, write=150.0, pool=150.0)
+
+
+async def post_to_sandbox(
+ ctx: RunContextWrapper, tool_name: str, kwargs: dict[str, Any],
+) -> dict[str, Any]:
+ """POST tool invocation to FastAPI tool server via host-mapped port.
+
+ Returns {"result": ...} or {"error": ...}; never raises.
+ """
+ port = ctx.context.get("tool_server_host_port")
+ token = ctx.context.get("sandbox_token")
+ agent_id = ctx.context.get("agent_id", "unknown")
+ if not (port and token):
+ return {"error": "Sandbox not initialized"}
+
+ url = f"http://127.0.0.1:{port}/execute"
+ body = {"agent_id": agent_id, "tool_name": tool_name, "kwargs": kwargs}
+ headers = {"Authorization": f"Bearer {token}"}
+ try:
+ async with httpx.AsyncClient(timeout=_TIMEOUT) as client:
+ r = await client.post(url, json=body, headers=headers)
+ if r.status_code == 401:
+ return {"error": "Sandbox auth failed"}
+ if r.status_code >= 400:
+ return {"error": f"Sandbox HTTP {r.status_code}: {r.text[:300]}"}
+ try:
+ return r.json()
+ except ValueError:
+ return {"error": f"Invalid sandbox response: {r.text[:200]}"}
+ except httpx.TimeoutException:
+ return {"error": f"Sandbox timeout (>{_TIMEOUT.read}s)"}
+ except httpx.RequestError as e:
+ return {"error": f"Sandbox request failed: {e!s}"[:300]}
+```
+
+---
+
+## 4. Per-tool migration contracts
+
+Compact table — one row per tool, expanded notes for non-trivial ones.
+
+### 4.1 Sandbox tools (POST to tool server)
+
+Every sandbox tool follows the same shape:
+
+```python
+from strix.tools._decorator import strix_tool
+from strix.tools._sandbox_dispatch import post_to_sandbox
+from agents import RunContextWrapper, ToolOutputText, ToolOutputImage
+
+
+@strix_tool(timeout=160) # outer SDK > inner httpx (150)
+async def <name>(ctx: RunContextWrapper, ...args) -> str | list:
+ result = await post_to_sandbox(ctx, "<server_tool_name>", {<kwargs>})
+ if "error" in result:
+ return result["error"]
+ # Tool-specific result shaping (extract screenshot, truncate, etc.)
+ return _shape_result(result)
+```
+
+| Tool | Server-side name | Args | Return shape | Notes |
+|---|---|---|---|---|
+| `browser_action` | `browser_action` | `action: str, **action_kwargs` | `[ToolOutputImage(base64), ToolOutputText(state)]` | Extract `screenshot` key as image; rest as text. 24 actions. Truncate page-source/console at server. |
+| `terminal_execute` | `terminal_execute` | `command: str, is_input=False, timeout=30, terminal_id="default", no_enter=False` | `ToolOutputText(content + status + exit_code)` | Per-`(agent_id, terminal_id)` libtmux session at server. |
+| `python_action` | `python_action` | `action: str, code: str | None, session_id="default", timeout=30` | `ToolOutputText(stdout + stderr + result_repr)` | Per-`(agent_id, session_id)` IPython kernel. |
+| `list_requests` | `list_requests` | `httpql_filter: str | None, start_page=1, end_page=1, page_size=50, sort_by="timestamp", sort_order="desc"` | `ToolOutputText(json)` | Caido. |
+| `view_request` | `view_request` | `request_id: str, search_pattern: str | None, part="response"` | `ToolOutputText(json)` | Caido. |
+| `send_request` | `send_request` | `method: str, url: str, headers={}, body=""` | `ToolOutputText(json)` | Caido. |
+| `repeat_request` | `repeat_request` | `request_id: str, modifications: dict` | `ToolOutputText(json)` | Caido. |
+| `scope_rules` | `scope_rules` | `action: Literal["list","create","update","delete"], **kwargs` | `ToolOutputText(json)` | Caido. |
+| `list_sitemap` | `list_sitemap` | `parent_id: str | None, page_size=50` | `ToolOutputText(json)` | Caido. |
+| `view_sitemap_entry` | `view_sitemap_entry` | `node_id: str` | `ToolOutputText(json)` | Caido. |
+
+### 4.2 Local tools (in-process)
+
+| Tool | Module | Signature | Notes |
+|---|---|---|---|
+| `create_note` / `list_notes` / `get_note` / `update_note` / `delete_note` | `strix/tools/notes/` | per Strix today | **Apply C6**: lock JSONL writes. Per-agent state silo via `ctx.context["agent_id"]`. |
+| `create_todo` / `list_todos` / `update_todo` / `mark_todo_done` / `mark_todo_pending` / `delete_todo` | `strix/tools/todo/` | per Strix today | In-memory only; not persisted. |
+| `create_vulnerability_report` | `strix/tools/reporting/` | All current required+optional fields | Calls existing `llm/dedupe.py` via nested `Runner.run` or direct `litellm.acompletion`. Wires `vulnerability_found_callback` via `ToolOutputGuardrail` for TUI popup. |
+| `web_search` | `strix/tools/web_search/` | `query: str` | Direct Perplexity API; do NOT use SDK's `WebSearchTool` (Responses-only, OpenAI-only). |
+| `str_replace_editor` / `list_files` / `search_files` | `strix/tools/file_edit/` | per Strix today | Wrap openhands-aci + ripgrep. SDK auto-threads sync. |
+| `finish_scan` | `strix/tools/finish/` | per Strix today | Root-only guard: `if ctx.context["parent_id"] is not None: return error`. Sets `ctx.context["agent_finish_called"] = True`. **Root agent uses `tool_use_behavior={"stop_at_tool_names": ["finish_scan"]}`.** |
+| `think` | `strix/tools/thinking/` | `thought: str` | Trivial; `return f"Recorded {len(thought)} chars"`. |
+| `load_skill` | `strix/tools/load_skill/` | `skills: list[str]` (max 5) | Returns skill content as `ToolOutputText` (accepted tradeoff). Validates against skill registry. |
+
+### 4.3 Multi-agent graph tools
+
+All in `strix/tools/agents_graph.py`. Use `bus = ctx.context["bus"]`, `me = ctx.context["agent_id"]`.
+
+```python
+@strix_tool(timeout=30)
+async def view_agent_graph(ctx) -> str:
+ bus = ctx.context["bus"]
+ lines = []
+ roots = [aid for aid, p in bus.parent_of.items() if p is None]
+
+ def render(aid, depth):
+ st = bus.statuses.get(aid, "?")
+ lines.append(" " * depth + f"- {bus.names.get(aid, aid)} ({aid}) [{st}]")
+ for c, p in bus.parent_of.items():
+ if p == aid:
+ render(c, depth + 1)
+
+ for r in roots:
+ render(r, 0)
+ return "\n".join(lines) or "(no agents)"
+
+
+@strix_tool(timeout=30)
+async def agent_status(ctx, agent_id: str) -> str:
+ bus = ctx.context["bus"]
+ if agent_id not in bus.statuses:
+ return f"Unknown agent {agent_id}"
+ return (
+ f"agent={bus.names.get(agent_id)} status={bus.statuses[agent_id]} "
+ f"parent={bus.parent_of.get(agent_id)} "
+ f"pending_msgs={len(bus.inboxes.get(agent_id, []))}"
+ )
+
+
+@strix_tool(timeout=30)
+async def send_message_to_agent(ctx, target_agent_id: str, message: str,
+ message_type: str = "info", priority: str = "normal") -> str:
+ bus = ctx.context["bus"]
+ if target_agent_id not in bus.statuses:
+ return f"Unknown agent {target_agent_id}"
+ await bus.send(target_agent_id, {
+ "from": ctx.context["agent_id"],
+ "content": message,
+ "type": message_type,
+ "priority": priority,
+ })
+ return f"Queued for {target_agent_id}"
+
+
+@strix_tool(timeout=601)
+async def wait_for_message(ctx, reason: str, timeout_seconds: int = 600) -> str:
+ import asyncio
+ bus = ctx.context["bus"]
+ me = ctx.context["agent_id"]
+ bus.statuses[me] = "waiting"
+ end = asyncio.get_event_loop().time() + timeout_seconds
+ while asyncio.get_event_loop().time() < end:
+ if bus.inboxes.get(me):
+ bus.statuses[me] = "running"
+ return "Message arrived. Continue."
+ await asyncio.sleep(1.0)
+ bus.statuses[me] = "running"
+ return f"Timed out after {timeout_seconds}s. Continue or call agent_finish."
+
+
+@strix_tool(timeout=120)
+async def create_agent(
+ ctx, name: str, task: str,
+ inherit_context: bool = True, skills: list[str] | None = None,
+) -> str:
+ import asyncio, uuid
+ from agents import Runner
+ from strix.run_config_factory import make_run_config, make_agent_context
+ from strix.orchestration.hooks import StrixOrchestrationHooks
+ from strix.agents.factory import build_strix_agent # build child Agent w/ tool_use_behavior
+
+ bus = ctx.context["bus"]
+ parent = ctx.context["agent_id"]
+ cid = uuid.uuid4().hex[:8]
+ await bus.register(cid, name, parent)
+
+ child_agent = build_strix_agent(name=name, skills=skills or [], is_root=False)
+
+ # Inherited context: parent's session items
+ parent_session = ctx.context.get("session")
+ history = await parent_session.get_items() if (inherit_context and parent_session) else []
+ initial_input = history + [
+ # Identity injection (replaces Strix's <agent_delegation> XML)
+ {"role": "user", "content": f"<agent_delegation>You are agent {name} ({cid}). "
+ f"Parent is {parent}. Do not echo this block.</agent_delegation>"},
+ {"role": "user", "content": task},
+ ]
+
+ child_ctx = make_agent_context(
+ bus=bus,
+ sandbox_session=ctx.context["sandbox_session"],
+ sandbox_token=ctx.context["sandbox_token"],
+ tool_server_host_port=ctx.context["tool_server_host_port"],
+ caido_host_port=ctx.context["caido_host_port"],
+ agent_id=cid,
+ agent_name=name,
+ parent_id=parent,
+ tracer=ctx.context["tracer"],
+ model_settings=ctx.context["model_settings"],
+ max_turns=ctx.context["max_turns"],
+ )
+
+ bus.tasks[cid] = asyncio.create_task(
+ Runner.run(
+ child_agent,
+ input=initial_input,
+ run_config=make_run_config(
+ sandbox_session=ctx.context["sandbox_session"],
+ bus=bus,
+ model=ctx.context.get("model", "strix/claude-sonnet-4.6"),
+ max_turns=ctx.context["max_turns"],
+ ),
+ context=child_ctx,
+ hooks=StrixOrchestrationHooks(),
+ )
+ )
+ return f"Spawned {name} ({cid}) running in parallel"
+
+
+@strix_tool(timeout=30)
+async def agent_finish(
+ ctx, result_summary: str, findings: list[dict] | None = None,
+ success: bool = True, report_to_parent: bool = True,
+ final_recommendations: list[str] | None = None,
+) -> str:
+ bus = ctx.context["bus"]
+ me = ctx.context["agent_id"]
+ parent = bus.parent_of.get(me)
+ if parent is None:
+ return "Error: agent_finish is for subagents. Root must call finish_scan."
+ ctx.context["agent_finish_called"] = True
+
+ if report_to_parent:
+ report_xml = (
+ f"<agent_completion_report from='{bus.names.get(me)}' agent_id='{me}' "
+ f"success='{success}'>\n"
+ f" <summary>{result_summary}</summary>\n"
+ f" <findings>{findings or []}</findings>\n"
+ f" <recommendations>{final_recommendations or []}</recommendations>\n"
+ f"</agent_completion_report>"
+ )
+ await bus.send(parent, {"from": me, "content": report_xml, "type": "completion"})
+
+ # Whitebox wiki update (M10 — preserved)
+ if ctx.context.get("is_whitebox") and findings:
+ # Append delta to shared wiki note via existing notes API
+ from strix.tools.notes.notes_actions import append_note_content
+ append_note_content("wiki", f"\n## Update from {bus.names.get(me)}\n{result_summary}\n")
+
+ return "Reported to parent. This agent will exit."
+```
+
+**Every child agent built by `build_strix_agent`:** `tool_use_behavior={"stop_at_tool_names": ["agent_finish"]}` (C4).
+
+---
+
+## 5. Standards: logging, error formatting, observability
+
+### 5.1 Logging
+
+| Logger | Level | When |
+|---|---|---|
+| `strix.orchestration.bus` | DEBUG | every send/drain/finalize |
+| `strix.orchestration.hooks` | INFO | agent start/end; WARNING on crash detection |
+| `strix.tools.<name>` | DEBUG | tool entry/exit; WARNING on user-recoverable error |
+| `strix.runtime.strix_docker_client` | INFO | container created; ERROR on creation failure |
+| `strix.sandbox.caido_capability` | INFO | bind/healthcheck-task lifecycle |
+| `strix.sandbox.session_manager` | INFO | session create/reuse/cleanup |
+| `strix.llm.anthropic_cache_wrapper` | DEBUG | cache_control injection events |
+| `strix.telemetry.strix_processor` | WARNING | sanitization mismatches |
+
+Verbose mode flag: `--verbose` → `agents.set_default_logging("DEBUG")` + `enable_verbose_stdout_logging()`. Production keeps `OPENAI_AGENTS_DONT_LOG_MODEL_DATA=1` (default) so model I/O isn't logged.
+
+Never logged anywhere: `LLM_API_KEY`, `TOOL_SERVER_TOKEN`, anything the existing `TelemetrySanitizer` patterns match.
+
+### 5.2 Error formatting
+
+| Layer | Recipient | Format |
+|---|---|---|
+| Tool body | LLM | Plain `str` returned from tool. Convention: `"Error: <reason>"` for failures. |
+| Tool body | User | Visible only via TUI when popup-worthy (vulnerability popup hooks via `ToolOutputGuardrail`). |
+| Sandbox dispatch | LLM | `{"error": "..."}` → returned as `Error: ...` string. |
+| Hook | Logs only | `logger.exception(...)` for unhandled in hooks; never propagated (would tear down the run). |
+| Agent crash | Parent | `<agent_crash agent_id='X' name='Y'>...</agent_crash>` user message via bus. |
+| Run-level exception | CLI | Exit code 1 (general failure), 2 (vulns found in headless), 130 (Ctrl+C). |
+| Run-level exception | Trace | `error.kind`, `error.message`, `error.traceback` span attributes (post-scrub). |
+
+### 5.3 Observability during migration
+
+While both old (Strix) and new (SDK-based) live in parallel:
+
+- **Comparison metrics**: per-run, log `total_input_tokens`, `total_output_tokens`, `cost`, `findings_count`, `wall_clock_seconds`. Diff old vs new with same target set.
+- **Drift detection**: identical `--target` should produce equivalent (not byte-identical) findings. Track findings-by-CWE histogram.
+- **Canary mode**: feature flag `STRIX_USE_SDK_HARNESS=1` or separate CLI command (`strix-sdk`) until cut-over. Two CLIs let users opt in.
+- **Event compatibility**: `events.jsonl` schema must remain compatible — the custom `TracingProcessor` emits the same `event_type` values as today's tracer.
+
+---
+
+## 6. Test plans per phase
+
+### Phase 0 — Spike & corrections
+
+**Smoke tests** (must all pass before Phase 1 starts):
+- `test_anthropic_cache_wrapper_injection` — system message has `cache_control`; non-Anthropic passes through.
+- `test_anthropic_cache_hit_rate` — three identical-prompt calls; call 2+ shows `cache_read_input_tokens > 0`.
+- `test_strix_docker_client_caps` — mock `containers.create`; assert `cap_add ⊇ {NET_ADMIN, NET_RAW}` and `extra_hosts["host.docker.internal"]=="host-gateway"`.
+- `test_strix_docker_client_live` — real Kali image; `session.exec(["nmap", "-sS", "scanme.nmap.org"])` returns exit 0.
+- `test_message_bus_two_children_exchange` — minimal `MessageBus` with two synthetic agents; verify FIFO message delivery via filter injection.
+- `test_tool_use_behavior_stop_on_finish` — toy agent with `agent_finish`-equivalent + `stop_at_tool_names`; assert SDK terminates exactly when tool is called.
+- `test_tool_server_concurrent_same_agent_calls` — issue two simultaneous POSTs to local tool server with same `agent_id`; current code cancels first; observe behavior; choose Phase 1 default (`parallel_tool_calls=False`).
+
+### Phase 1 — Foundation
+
+**Unit tests:**
+- `MultiProvider` resolution: `strix/claude-sonnet-4.6` → `AnthropicCachingLitellmModel` with correct base URL.
+- `strix_tool` decorator: timeout applied; sync auto-threaded.
+- `StrixSession`: 100K-token history triggers compression; compressor failure returns uncompressed.
+- `StrixTracingProcessor`: 100 concurrent span writes, JSONL is valid (no interleaved bytes); PII patterns scrubbed.
+- `RunConfig` factory: retry policy excludes 401 explicitly; `parallel_tool_calls=False` default.
+
+**Integration tests:**
+- Full single-agent run with `LitellmModel` against an Anthropic mock; verify cache_control reaches the model.
+- Cost tracking via `RunHooks.on_llm_end` matches `litellm.cost_per_token` calculation within 1%.
+
+### Phase 2 — Tool ports
+
+**Per-tool smoke tests** (one per `@strix_tool`):
+- Browser: `launch` → `goto("https://example.com")` → screenshot returned as `ToolOutputImage`; viewport 1280x720; per-agent tab keying.
+- Terminal: `terminal_execute("echo hello")` → status=completed, exit_code=0. Special key: `terminal_execute("C-c", is_input=True)` no error.
+- Python: `python_action(action="execute", code="x=1; x+1")` → result `"2"`.
+- Caido: `list_requests(httpql_filter="request.method == 'GET'")` returns JSON with `requests` array.
+- Notes: concurrent `create_note` from two agents; assert JSONL parses line-by-line (C6 fix).
+- Reporting: full `create_vulnerability_report` payload; CVSS computed; persisted to `vulnerabilities/vuln_*.json`; dedup roundtrip works.
+- File edit: `str_replace_editor("create", "/workspace/foo.txt", file_text="x")` → file exists.
+
+**Reentrancy tests:** if/when Phase 6 enables parallel: assert two parallel `terminal_execute` calls on same agent with different `terminal_id` don't collide (different libtmux sessions).
+
+### Phase 3 — Multi-agent
+
+**Unit tests:**
+- `AgentMessageBus`: 1000 concurrent `send`/`drain` ops; FIFO maintained; `total_stats` snapshot consistent.
+- `inject_messages_filter`: pending messages appended in order; user-from-user skips XML wrap; idempotent under retry simulation.
+- `StrixOrchestrationHooks.on_agent_end`: crash detection fires when `agent_finish_called` False; sends `<agent_crash>` to parent.
+- `bus.cancel_descendants`: tree of N=10 agents; all cancelled leaf-first.
+
+**Integration:**
+- Root spawns 2 children in parallel; children exchange messages; both finish via `agent_finish`; root reads completion reports; `bus.total_stats` aggregates correctly.
+- Root + 1 child; user Ctrl+C mid-run; both tasks cancelled; processor flushes all pending events.
+- Crashed child scenario: child `Runner.run` raises `RuntimeError`; parent receives `<agent_crash>` within next turn.
+
+### Phase 4 — Sandbox + Caido
+
+**Integration:**
+- `create_or_reuse(scan_id="x", ...)` twice → same session; `session.exec(["echo", "hi"])` works.
+- Caido capability: container has `http_proxy=http://127.0.0.1:48080` env; `curl https://example.com` is captured by Caido (verify via `list_requests` shows the request).
+- Healthcheck: stop tool server inside container; `wait_for_ports_ready` raises after 30s.
+
+### Phase 5 — Interface + persistence
+
+**Tests:**
+- `StrixStreamAccumulator`: feed mock `RunResultStreaming` events; tracer's `streaming_content[agent_id]` updates.
+- TUI: under simulated agent run, refresh latency < 500ms after each LLM chunk.
+- Run-directory: post-run, `events.jsonl` valid; `vulnerabilities/` populated; `penetration_test_report.md` non-empty.
+
+### Phase 6 — Validation + tool server relaxation
+
+**End-to-end:**
+- Full pentest against a stable target (a vulnerable webapp pinned in tests). Compare findings (count + CVSS distribution + CWE histogram) vs Strix baseline within agreed tolerance.
+- Stress: 3 child agents, parallel tool calls within turn, no JSONL corruption, no message loss.
+- After tool server relaxation: `parallel_tool_calls=True`; multi-tool turn fires correctly; sibling failure isolation works.
+
+---
+
+## 7. Rollback, CI, cross-cutting
+
+### 7.1 Branching & cutover
+
+- Branch: `harness-migration` (already created).
+- Cutover trigger: Phase 6 acceptance criteria met (golden-output diff within tolerance).
+- Pre-cutover: SDK-based path lives behind `STRIX_USE_SDK_HARNESS=1`. Old path unchanged.
+- Cutover: flip default to SDK-based; legacy path stays for one release as fallback.
+- Removal of legacy: after one release with no rollback signals.
+
+### 7.2 Rollback procedure
+
+If post-cut regression discovered:
+1. Set `STRIX_USE_SDK_HARNESS=0` in deployed config (re-uses legacy harness).
+2. Investigate via traces/events.jsonl from the failed run.
+3. Fix, re-run end-to-end suite, re-cut.
+
+### 7.3 Data compatibility
+
+- `events.jsonl` schema preserved (custom processor emits same `event_type`s).
+- `vulnerabilities/vuln_*.json` schema preserved (same Pydantic shape).
+- `penetration_test_report.md` schema preserved.
+- `notes/notes.jsonl` schema preserved.
+- Old runs are still analyzable by new tooling (no migration script needed).
+
+### 7.4 SDK version pinning
+
+- `pyproject.toml`: `openai-agents==0.14.6` (exact pin while we duplicate `_create_container`).
+- CI nightly: re-run our Phase 0 spike against SDK `latest`. If green, bump pin and re-test.
+- Track upstream PR to add `additional_create_kwargs` hook that would let us drop the body duplication.
+
+### 7.5 CI changes
+
+- New test buckets: `tests/orchestration/`, `tests/sandbox/`, `tests/llm_wrappers/`, `tests/integration/`.
+- Snapshot tests for golden outputs (one or two stable targets).
+- Mypy strict on new SDK types; install `openai-agents` package types.
+- Lint rules for tool decorators (`@strix_tool` not bare `@function_tool` for new tools).
+
+### 7.6 Drop / add deps
+
+- Drop direct: `litellm[proxy]>=1.81.1,<1.82.0` (becomes transitive via `openai-agents[litellm]`).
+- Add: `openai-agents[litellm]==0.14.6`.
+- Keep: `tenacity`, `pydantic`, `rich`, `docker`, `textual`, `cvss`, `traceloop-sdk`, `opentelemetry-exporter-otlp-proto-http`, `scrubadub`, `defusedxml`.
+
+---
+
+## 8. Phase ordering and day-1 commit list
+
+### Phase ordering (compressed)
+
+- **Phase 0** — Foundations + spikes. Land the 3 wrapper classes, run all Phase 0 smoke tests.
+- **Phase 1** — Foundation files (rest), provider layer, `strix_tool`, custom `Session`, custom `TracingProcessor`, `RunConfig` factory.
+- **Phase 2** — Tool ports. Sandbox dispatcher first, then all 30+ tools incrementally.
+- **Phase 3** — Multi-agent: bus, filter, hooks, six graph tools. End-to-end parallel-children test.
+- **Phase 4** — `CaidoCapability` + healthcheck + session cache by scan_id.
+- **Phase 5** — TUI / streaming accumulator / run-directory persistence / turn warnings.
+- **Phase 6** — Validation + tool server relaxation + parallel tool calls flip.
+
+### Day-1 commits (Phase 0)
+
+In order; each is independent until linked at Phase 0 acceptance.
+
+1. `strix/llm/anthropic_cache_wrapper.py` — `AnthropicCachingLitellmModel`. Tests: cache injection + non-Anthropic passthrough.
+2. `strix/runtime/strix_docker_client.py` — `StrixDockerSandboxClient`. Tests: kwarg injection mock + live nmap.
+3. `strix/orchestration/bus.py` — `AgentMessageBus`. Tests: concurrency stress + cancel_descendants.
+4. `strix/orchestration/filter.py` — `inject_messages_filter`. Tests: FIFO + retry-stable.
+5. `strix/orchestration/hooks.py` — `StrixOrchestrationHooks`. Tests: crash detection + warning injection.
+6. `strix/tools/_decorator.py` — `strix_tool`. Tests: timeout + sync auto-threading.
+7. `strix/llm/multi_provider_setup.py` — `MultiProviderMap` + `StrixModelProvider`. Tests: alias resolution.
+
+### Phase 0 acceptance gate
+
+All seven module unit tests green. The seven Phase 0 smoke tests in §6 green. Tool server concurrent-call test executed; result documented. Only then proceed to Phase 1.
+
+---
+
+*This playbook merges every audit finding into a single engineering reference. When implementation discovers something the playbook missed, treat it as a bug in the playbook — update this file, then write the code.*
diff --git a/TESTING_STRATEGY.md b/TESTING_STRATEGY.md
new file mode 100644
index 00000000..501f9d97
--- /dev/null
+++ b/TESTING_STRATEGY.md
@@ -0,0 +1,597 @@
+# Migration Testing Strategy
+
+> Seven-layer safety net to prove the SDK-migrated harness produces behavior identical to the legacy harness, with no silent feature loss. Centered on **behavioral parity diffing** (the only test that doesn't require knowing in advance what could break) and a **feature inventory matrix** (every feature has a row, every row has a test, no row → no proof).
+
+---
+
+## Table of contents
+
+1. [Threat model — what we're afraid of](#1-threat-model)
+2. [Testing layers (cheap → expensive)](#2-testing-layers)
+3. [Feature inventory matrix](#3-feature-inventory-matrix)
+4. [Behavioral parity diffing — how it actually works](#4-behavioral-parity-diffing)
+5. [Replay infrastructure (deterministic LLM + sandbox)](#5-replay-infrastructure)
+6. [Live shadow mode (production-grade canary)](#6-live-shadow-mode)
+7. [Per-correction test mapping](#7-per-correction-test-mapping)
+8. [CI gating and cutover criteria](#8-ci-gating-and-cutover-criteria)
+9. [Manual smoke checklist](#9-manual-smoke-checklist)
+10. [Post-cutover monitoring](#10-post-cutover-monitoring)
+11. [What to do when a parity diff fails](#11-what-to-do-when-a-parity-diff-fails)
+
+---
+
+## 1. Threat model
+
+Concrete categories of regression we want to catch — every test below targets at least one row.
+
+| # | Threat | Detection difficulty | Where it hides |
+|---|---|---|---|
+| T1 | A tool stops being invocable (typo, registration miss) | Easy | Agent runs, never calls it |
+| T2 | A tool's args don't validate the same way | Medium | LLM gets different error string; no exception |
+| T3 | A tool's output format changes (XML vs structured, truncation cap, screenshot extraction) | Hard | Findings still happen, but different shape |
+| T4 | An LLM provider quirk lost (Anthropic cache, Bedrock timeout, vision strip) | Hard | Cost+latency drift; no functional break |
+| T5 | A side effect lost (wiki note auto-update, vulnerability dedup, PII scrub) | **Critical-silent** | Reports persist but a feature stops firing |
+| T6 | An event type stops being emitted to events.jsonl | Hard | Run completes, dashboards have gaps |
+| T7 | Cancellation cascade incomplete (Ctrl+C leaves orphan tasks) | Medium | Looks fine on success path; only visible on cancel |
+| T8 | Memory leak / resource leak (orphan inboxes, stale sessions) | Hard | Long runs degrade |
+| T9 | Concurrency regression (parallel tools collide, message ordering breaks) | **Critical-silent** | Pass once, fail on second concurrent run |
+| T10 | Schema drift in persisted artifacts (events.jsonl, vuln JSON, report MD) | Medium | Downstream consumers break |
+| T11 | Config / env var stops being read | Easy | Default kicks in instead of user override |
+| T12 | Exit-code change | Easy | CI integrations break |
+| T13 | TUI / CLI output format change | Easy | User experience drift |
+| T14 | A skill / prompt section silently dropped | **Critical-silent** | Agent's behavior subtly worse on specific vuln types |
+| T15 | Subagent crash silent (parent waits forever) | **Critical-silent** | Run hangs without diagnosis |
+| T16 | LLM provider routing wrong (`strix/foo` → wrong model) | Easy | API error within seconds |
+
+**Critical-silent** rows are the dangerous ones — the run *looks* fine but a feature is gone. Layer 4 (parity diffing) is specifically designed to catch these.
+
+---
+
+## 2. Testing layers
+
+Bottom-up. Each layer catches a different class of bug; together they're complementary.
+
+```
+┌────────────────────────────────────────────┐
+│ Layer 7: Manual smoke (humans, TUI, etc.) │ rare, high signal
+├────────────────────────────────────────────┤
+│ Layer 6: Live shadow / canary │ prod-grade, expensive
+├────────────────────────────────────────────┤
+│ Layer 5: Recorded replay (deterministic) │ end-to-end, reproducible
+├────────────────────────────────────────────┤
+│ Layer 4: Behavioral parity diffing │ **most powerful**
+├────────────────────────────────────────────┤
+│ Layer 3: Integration (modules together) │
+├────────────────────────────────────────────┤
+│ Layer 2: Unit (one module, mocked deps) │
+├────────────────────────────────────────────┤
+│ Layer 1: Static (mypy, ruff, signatures) │ cheapest, fastest
+└────────────────────────────────────────────┘
+```
+
+### Layer 1 — Static / pre-runtime
+
+Catches type-level mismatches without running the code.
+
+- **`mypy --strict`** against `strix/` after migration. With `openai-agents[litellm]==0.14.6` installed, mypy verifies our subclasses honor SDK's ABC contracts. **Catches: F1, F2, F3 type-fix regressions, future SDK signature drift.**
+- **`ruff` + `pyright`** as secondary type checkers; they sometimes catch what mypy misses (e.g., Pyright's stricter overload resolution).
+- **Import-surface test** (`tests/static/test_imports.py`): a test that just imports every module in `strix/` and instantiates every `RunHooks`/`Capability`/`Model`/`Session` subclass with valid kwargs. If any abstract method is unimplemented, instantiation raises. Catches T2, T11.
+- **Inventory completeness test** (`tests/static/test_inventory.py`): asserts every row in `tests/inventory/features.csv` (see §3) has a non-empty `test_id` field. Prevents adding a feature without a test.
+- **SDK version pin guard** (`tests/static/test_sdk_version.py`): asserts `agents.__version__ == "0.14.6"` exactly. We duplicate `_create_container` body; an SDK bump must be intentional. Fails CI on accidental upgrade.
+
+### Layer 2 — Unit
+
+Each module tested in isolation with mocked dependencies. One file per source file.
+
+- `tests/orchestration/test_bus.py` — `AgentMessageBus` register/send/drain/cancel_descendants/total_stats. Concurrency stress: 1000 concurrent send/drain ops, FIFO assertion.
+- `tests/orchestration/test_filter.py` — `inject_messages_filter` with synthetic `CallModelData`. Empty inbox → passthrough; 3 messages → 3 user items; user-from-user no XML wrap; bus exception → return unchanged (C14).
+- `tests/orchestration/test_hooks.py` — `StrixOrchestrationHooks`. Crash detection (output=None or `agent_finish_called=False`); turn warnings at 85% / N-3; bus errors don't propagate (C15).
+- `tests/llm/test_anthropic_cache.py` — `AnthropicCachingLitellmModel._patch`. Anthropic model → `cache_control` present; non-Anthropic → passthrough; system message wrapped correctly.
+- `tests/llm/test_multi_provider.py` — `StrixModelProvider.get_model`. Known alias → correct concrete model + base URL; unknown alias → `UserError` (C17).
+- `tests/llm/test_session.py` — `StrixSession`. Compression triggers above 90K; compressor exception → uncompressed history (C10); subsequent calls skip compression after first failure.
+- `tests/runtime/test_strix_docker_client.py` — `StrixDockerSandboxClient._create_container` with mocked `docker_client`. Assert `cap_add ⊇ {NET_ADMIN, NET_RAW}` and `extra_hosts["host.docker.internal"] == "host-gateway"`.
+- `tests/sandbox/test_caido_capability.py` — `CaidoCapability.process_manifest` injects proxy env; `tools()` returns 7 tools; `instructions()` returns non-empty string; `bind()` spawns healthcheck task.
+- `tests/sandbox/test_session_manager.py` — `create_or_reuse` cache hit returns same session; `cleanup` removes container.
+- `tests/telemetry/test_processor.py` — `StrixTracingProcessor`. Concurrent writes, JSONL is line-valid; PII patterns scrubbed; `OSError` doesn't propagate (C16).
+- `tests/tools/test_decorator.py` — `strix_tool` factory. Default 120s timeout applied; sync function auto-threaded; `error_as_result` returns string.
+- `tests/tools/test_sandbox_dispatch.py` — `post_to_sandbox`. 401 → error string; size cap enforced (C18); timeout returns error string.
+- `tests/tools/test_<each>.py` — one per ported tool. Each: smoke test (valid args → expected output shape) + one edge case (bad args, timeout, network error).
+
+### Layer 3 — Integration
+
+Multiple modules wired together; LLM and sandbox still mocked.
+
+- `tests/integration/test_single_agent_mocked.py` — Build root `Agent`, run with mocked `Model` that emits scripted tool calls; assert tracer captured expected events; assert `RunResult.final_output` populated.
+- `tests/integration/test_multi_agent_mocked.py` — Root spawns 2 children via `create_agent` tool; mocked Model scripts each child's responses; bus messaging between children; both `agent_finish`; root `finish_scan`. Assert: stat aggregation correct; messages delivered FIFO; `bus.statuses` all `completed`.
+- `tests/integration/test_cancellation_cascade.py` — Build a tree, then `bus.cancel_descendants(root)`; assert all child tasks `cancelled()`; assert leaves cancelled before parents (C9).
+- `tests/integration/test_subagent_crash.py` — Mock child raising `RuntimeError`; assert `<agent_crash>` arrives in parent inbox via filter (C8).
+- `tests/integration/test_compressor_fallback.py` — Mock compressor LLM raising; assert run continues with uncompressed history (C10).
+- `tests/integration/test_finish_scan_blocks_with_running_children.py` — Root calls `finish_scan` while child still `running`; assert error returned (C22).
+- `tests/integration/test_jsonl_concurrent_writes.py` — 50 concurrent agents writing to events.jsonl; assert every line is valid JSON (C7); same for notes (C6).
+
+### Layer 4 — Behavioral parity diffing
+
+**The most important layer.** Run the same scenario through legacy and SDK harness; diff every artifact. Detail in §4 below.
+
+### Layer 5 — Recorded replay
+
+Deterministic end-to-end tests using captured LLM and sandbox traces. Detail in §5.
+
+### Layer 6 — Live shadow / canary
+
+Run both harnesses in production for a small slice of real users; diff results in real time. Detail in §6.
+
+### Layer 7 — Manual smoke
+
+Humans operating the TUI, headless mode, multi-target scans. Detail in §9.
+
+---
+
+## 3. Feature inventory matrix
+
+The single artifact that prevents silent feature loss. **One CSV file (`tests/inventory/features.csv`) with one row per feature.** CI fails if any row is missing a test reference.
+
+### Schema
+
+```csv
+feature_id,subsystem,description,source_ref,test_id,owner,phase,status
+```
+
+- `feature_id` — stable opaque identifier (e.g., `F-AGENT-LOOP-001`).
+- `subsystem` — `agents | llm | tools | sandbox | telemetry | interface | config`.
+- `description` — one sentence.
+- `source_ref` — `path:line` to the legacy implementation OR `path:line` to the migration playbook spec.
+- `test_id` — name of the test (or test marker) that proves it survives. Empty = blocker.
+- `owner` — engineer responsible.
+- `phase` — Phase 0–6 in which the feature lands.
+- `status` — `legacy-only | both | sdk-only | parity-verified`.
+
+### Sample rows (~250 features expected)
+
+```csv
+F-AGENT-LOOP-001,agents,Agent loop with max_iterations=300,strix/agents/base_agent.py:152,test_runner_max_turns_300,allam,1,parity-verified
+F-AGENT-LOOP-002,agents,85%/N-3 turn warnings,strix/agents/base_agent.py:186,test_turn_warnings_injection,allam,3,parity-verified
+F-AGENT-LOOP-003,agents,Streaming early-truncate at </function>,strix/llm/llm.py:212,,allam,defer,legacy-only
+F-LLM-001,llm,Anthropic prompt cache control,strix/llm/llm.py:371,test_anthropic_cache_present,allam,0,parity-verified
+F-TOOL-BROWSER-001,tools,launch action,strix/tools/browser/browser_actions.py:75,test_browser_launch,allam,2,parity-verified
+F-TOOL-BROWSER-002,tools,goto action,strix/tools/browser/browser_actions.py:80,test_browser_goto,allam,2,parity-verified
+... (24 browser actions)
+F-TOOL-CAIDO-001,tools,list_requests with HTTPQL filter,strix/tools/proxy/proxy_actions.py:9,test_caido_list_requests,allam,2,parity-verified
+... (7 caido tools)
+F-MULTIAGENT-MSG-001,orchestration,send_message_to_agent FIFO,strix/tools/agents_graph/agents_graph_actions.py:495,test_bus_send_drain_fifo,allam,3,parity-verified
+F-MULTIAGENT-MSG-002,orchestration,inter_agent_message XML wrap,base_agent.py:491,test_filter_xml_wrap,allam,3,parity-verified
+F-EVENT-001,telemetry,run.started event in events.jsonl,strix/telemetry/tracer.py:87,test_event_run_started_emitted,allam,1,parity-verified
+F-EVENT-002,telemetry,tool.execution.started event,strix/telemetry/tracer.py:300,test_event_tool_execution_started,allam,1,parity-verified
+... (every event type)
+F-CLI-FLAG-target,interface,--target,--t accepts URL/repo/path,strix/interface/main.py:267,test_cli_target_inference,allam,5,parity-verified
+F-CLI-FLAG-scan-mode,interface,--scan-mode quick|standard|deep,strix/interface/main.py:295,test_cli_scan_mode,allam,5,parity-verified
+... (every CLI flag)
+F-OUTPUT-EXIT-CODE-2,interface,exit code 2 on findings in headless,strix/interface/main.py:640,test_headless_exit_code_2,allam,5,parity-verified
+F-OUTPUT-VULN-JSON,telemetry,vulnerabilities/vuln_*.json schema,strix/telemetry/tracer.py:365,test_vuln_json_schema,allam,2,parity-verified
+F-OUTPUT-REPORT-MD,interface,penetration_test_report.md template,strix/telemetry/tracer.py:400,test_report_md_template,allam,5,parity-verified
+F-PII-001,telemetry,scrubadub OpenAI key pattern,strix/telemetry/utils.py:87,test_pii_scrub_openai_key,allam,1,parity-verified
+F-PII-002,telemetry,scrubadub bearer token pattern,strix/telemetry/utils.py:91,test_pii_scrub_bearer,allam,1,parity-verified
+... (every regex pattern)
+F-SKILL-NoSQL,prompts,NoSQL injection skill,strix/prompts/vulnerabilities/nosql_injection.jinja,test_skill_nosql_loadable,allam,2,parity-verified
+F-SKILL-K8s,skills,Kubernetes security skill,strix/skills/cloud/kubernetes.md,test_skill_k8s_loadable,allam,2,parity-verified
+... (every skill file)
+```
+
+### Workflow
+
+1. **Bootstrap**: a script (`scripts/build_inventory.py`) walks the legacy code and emits a draft CSV. Engineer fills `test_id` per row.
+2. **CI gate**: `tests/static/test_inventory_completeness.py` asserts every row has `test_id` non-empty. Empty row → CI fails.
+3. **Coverage check**: a separate test asserts every `test_id` listed in the inventory actually exists as a test function (no typos, no orphans).
+4. **PR review rule**: any code change touching `strix/` must update the inventory if it adds/removes/modifies a feature.
+5. **Cutover gate**: ≥98% of inventory rows must be `parity-verified`. The remaining ≤2% are `legacy-only` (intentionally dropped) with explicit owner approval recorded in the row.
+
+This matrix is the single artifact that proves "we didn't lose anything." It's the audit trail.
+
+---
+
+## 4. Behavioral parity diffing
+
+**The strongest non-trivial test we have.** Same input, same env, both harnesses, diff every output. If the diff is empty, parity is proved.
+
+### What we diff
+
+For a fixed scenario (same target, same instruction, same model, same env):
+
+| Artifact | Diff strategy |
+|---|---|
+| `events.jsonl` | Per-line JSON normalized + sorted by `event_type` + key fields; some fields ignored (timestamps, UUIDs, span IDs); deep diff |
+| `vulnerabilities/*.json` | Group by stable identifier (target+endpoint+CVE), normalize, deep-diff each group's contents |
+| `penetration_test_report.md` | Length within ±10%; heading set identical; CWE histogram identical |
+| `notes/notes.jsonl` | Sorted by category + title; content body fuzzy-match (Levenshtein > 0.95) |
+| `wiki/*.md` | File set identical; per-file content fuzzy-match |
+| Tool call sequence (extracted from events) | Tool name multiset identical; arg signature shapes identical |
+| Token usage | Within ±5% (LLM nondeterminism + caching variance) |
+| Wall-clock duration | Within ±50% (allow for SDK overhead and parallel tool gains) |
+
+### Normalization (the careful part)
+
+LLMs are nondeterministic. We **must** normalize away noise before diffing or every diff fails.
+
+```python
+# tests/parity/normalize.py
+def normalize_events(events_jsonl_path: Path) -> list[dict]:
+ out = []
+ for line in events_jsonl_path.read_text().splitlines():
+ evt = json.loads(line)
+ # Strip noise
+ evt.pop("timestamp", None)
+ evt.pop("trace_id", None)
+ evt.pop("span_id", None)
+ evt.pop("agent_id", None) # different IDs in old vs new
+ evt.pop("scan_id", None)
+ # Normalize content fields
+ if "payload" in evt and "content" in evt["payload"]:
+ evt["payload"]["content"] = _normalize_text(evt["payload"]["content"])
+ out.append(evt)
+ # Sort by event_type, then by stable shape hash
+ out.sort(key=lambda e: (e.get("event_type", ""), _shape_hash(e)))
+ return out
+```
+
+The diff library is `deepdiff`; failures point to specific keys.
+
+### The runner
+
+```python
+# tests/parity/run_parity.py
+def run_parity(scenario_id: str) -> ParityResult:
+ """Run scenario through both harnesses with identical inputs and recorded LLM."""
+ inputs = load_scenario(scenario_id) # target, instruction, env, model
+
+ legacy_run = run_legacy(inputs, recorded_llm=RECORDED[scenario_id])
+ sdk_run = run_sdk(inputs, recorded_llm=RECORDED[scenario_id])
+
+ return ParityResult(
+ events_diff=diff_events(legacy_run.events_jsonl, sdk_run.events_jsonl),
+ vulns_diff=diff_vulns(legacy_run.vulns, sdk_run.vulns),
+ report_diff=diff_report(legacy_run.report, sdk_run.report),
+ tool_call_diff=diff_tool_calls(legacy_run.events, sdk_run.events),
+ usage_drift=compute_usage_drift(legacy_run.usage, sdk_run.usage),
+ )
+```
+
+### Scenarios to fix as parity baselines
+
+A small, hand-picked, *stable* set:
+
+| Scenario | Target | Mode | Why |
+|---|---|---|---|
+| `S01-static-blackbox` | https://juice-shop.local | quick | Standard OWASP target; many findings; broad tool exercise |
+| `S02-static-whitebox` | ./examples/vulnerable-flask-app | deep | Whitebox path; semgrep+ast paths; wiki notes |
+| `S03-multi-target` | repo + url combined | standard | Multi-target coordination |
+| `S04-diff-mode` | ./examples/vulnerable-flask-app `--scope-mode=diff` | quick | Diff scope injection |
+| `S05-cancellation` | juice-shop, kill at iter 10 | quick | Cancellation cascade |
+| `S06-multi-agent-explicit` | DVWA, instruction triggers `create_agent` | deep | Subagent flow |
+| `S07-empty-target` | nonexistent.local | quick | Failure path |
+| `S08-large-codebase` | ./examples/big-repo | standard | Compression triggered |
+
+Every scenario has a recorded LLM trace (§5). Every scenario runs in CI. Failure on any scenario blocks cutover.
+
+### What "diff is empty" means
+
+After normalization, an empty diff on every artifact means: every event the legacy harness emits, the SDK harness emits; every finding the legacy harness reports, the SDK harness reports; every tool call sequence is the same set; the report has the same structure. **It does NOT mean every byte is identical** — that's impossible with LLMs.
+
+### Bidirectional comparison
+
+We diff in both directions:
+- "What does legacy have that SDK doesn't?" — the dangerous direction (silent feature loss).
+- "What does SDK have that legacy doesn't?" — safer (new behavior), but worth review.
+
+A row like `event_type=agent.created` missing in SDK → blocker. A row like `event_type=tool.guardrail.rejected` only in SDK → review.
+
+---
+
+## 5. Replay infrastructure
+
+LLM and sandbox calls are nondeterministic (LLM) or environment-dependent (sandbox). For deterministic CI, we record once and replay forever.
+
+### LLM recording
+
+A `RecordedLLM` model that intercepts `Model.get_response`/`stream_response`, looks up the request in a fixture file, returns the recorded response.
+
+```python
+# tests/replay/recorded_llm.py
+from agents.models.interface import Model
+
+class RecordedLLM(Model):
+ def __init__(self, recording_path: Path):
+ self.recordings = json.loads(recording_path.read_text())
+ self.cursor = 0
+
+ async def get_response(self, system_instructions, input, model_settings, tools, ...):
+ key = _hash_request(system_instructions, input, model_settings)
+ if key not in self.recordings:
+ raise ValueError(f"No recording for request hash {key[:8]}; capture with --record")
+ recording = self.recordings[key]
+ return ModelResponse(
+ output=[_deserialize_item(it) for it in recording["output"]],
+ usage=Usage(**recording["usage"]),
+ response_id=recording.get("response_id"),
+ )
+
+ async def stream_response(self, ...):
+ # Same lookup; replay chunks one by one
+ ...
+```
+
+**How requests are keyed:** hash of `(system_instructions, input_items, model, tools, model_settings excerpts)`. PII-stripped before hashing. Hash collisions are vanishingly rare; if they happen, append a sequence counter.
+
+**Capture mode:** `pytest --record-llm` runs scenarios against a real LLM with a flag set; all `acompletion` calls are intercepted at a wrapper layer and serialized to `tests/replay/recordings/<scenario_id>.json`. PII scrubbed via existing `TelemetrySanitizer` before write.
+
+**Replay mode (default):** `pytest` uses `RecordedLLM` instead of real LLM; deterministic.
+
+### Sandbox recording
+
+Same pattern for sandbox HTTP calls. Wrap `post_to_sandbox`:
+
+```python
+class RecordedSandbox:
+ def __init__(self, recording_path: Path):
+ self.recordings = json.loads(recording_path.read_text())
+
+ async def post(self, agent_id, tool_name, kwargs):
+ key = _hash_call(agent_id, tool_name, kwargs)
+ if key not in self.recordings:
+ raise ValueError(f"No sandbox recording for {tool_name} hash {key[:8]}")
+ return self.recordings[key]
+```
+
+Inject via test fixture; production code path unchanged.
+
+### Recording vs replay in CI
+
+- **Replay** (default, fast): every PR runs scenarios against recorded LLM + sandbox. Catches behavioral regressions in the harness itself.
+- **Re-record** (manual, scheduled): a recurring CI job (or operator command) re-records scenarios against real LLM + real sandbox. Generates fresh recordings if the model output drifts.
+- **Drift detection**: if a re-record produces output that fails the parity diff, the migration broke (or the legacy harness changed in main; check git history).
+
+This is exactly the pattern the SDK uses internally (`inline_snapshot` library). We can adopt it directly.
+
+---
+
+## 6. Live shadow mode
+
+For the highest-confidence pre-cutover signal: run both harnesses in production, on real user runs, diff results.
+
+### Shadow runner
+
+A CLI mode `strix --target ... --shadow` that:
+
+1. Spawns the legacy harness against the target.
+2. Spawns the SDK harness against the same target (separate sandbox container).
+3. Both run to completion independently.
+4. After both finish, computes parity diff and emits a report.
+
+User sees the legacy result (no UX disruption); engineering team sees the diff.
+
+```python
+# strix/interface/shadow.py
+async def run_shadow(args):
+ legacy_task = asyncio.create_task(run_legacy(args, run_dir=Path("strix_runs/shadow_legacy")))
+ sdk_task = asyncio.create_task(run_sdk(args, run_dir=Path("strix_runs/shadow_sdk")))
+ legacy_result, sdk_result = await asyncio.gather(legacy_task, sdk_task)
+
+ diff = run_parity(legacy_result.run_dir, sdk_result.run_dir)
+ write_diff_report(diff, Path("strix_runs/shadow_diff.json"))
+ upload_diff_to_telemetry(diff)
+ return legacy_result # user gets the legacy answer; safe rollback
+```
+
+### Sampling
+
+Not every run shadows — too expensive in tokens. Configurable sampling:
+
+```bash
+STRIX_SHADOW_SAMPLE_RATE=0.1 # 10% of runs go through shadow mode
+STRIX_SHADOW_FORCE=1 # always (for engineering / staging)
+```
+
+### What we look for
+
+- **Parity rate**: % of shadow runs where diff is empty. Target: ≥99% before cutover.
+- **Drift class histograms**: which fields differ most? Track per-field over time.
+- **Resource drift**: token cost, wall-clock, sandbox memory. Plot distributions.
+
+Shadow runs **don't gate cutover** by themselves (too noisy with 1 sample), but a sustained drop in parity rate is a stop signal.
+
+---
+
+## 7. Per-correction test mapping
+
+Every one of the 25 corrections from `AUDIT.md` + `AUDIT_R2.md` + `AUDIT_R3.md` needs a test that would have caught the bug. Defensive code without proof-of-defense is theater.
+
+| # | Correction | Test | Layer |
+|---|---|---|---|
+| C1 | Tool-server slot serialization vs SDK parallel calls | `test_parallel_tool_calls_safe_default_no_collision` (Layer 3) + `test_tool_server_relax_phase6_concurrent_works` (Layer 3) | 3 |
+| C2 | Anthropic cache_control on system message | `test_anthropic_cache_control_on_system_message` (Layer 2) + `test_anthropic_cache_hit_rate` (Layer 5, recorded) | 2, 5 |
+| C3 | DockerSandboxClient subclass injects caps | `test_strix_docker_client_caps_injected` (Layer 2, mocked) + `test_strix_docker_client_nmap_works` (Layer 7, live) | 2, 7 |
+| C4 | Subagent `tool_use_behavior` | `test_subagent_exits_on_agent_finish` (Layer 3) | 3 |
+| C5 | StrixStreamAccumulator parity | `test_stream_accumulator_event_coverage` (Layer 4 vs Layer 5 baseline) | 4, 5 |
+| C6 | Notes JSONL write lock | `test_notes_jsonl_concurrent_writes_no_corruption` (Layer 3) | 3 |
+| C7 | events.jsonl write lock | `test_events_jsonl_concurrent_writes_no_corruption` (Layer 3) | 3 |
+| C8 | Subagent crash detection | `test_subagent_crash_emits_agent_crash_to_parent` (Layer 3) | 3 |
+| C9 | Cancellation cascade | `test_cancel_descendants_walks_tree_leaf_first` (Layer 3) | 3 |
+| C10 | Compressor try/except | `test_compressor_failure_returns_uncompressed` (Layer 2) | 2 |
+| C11 | Retry policy excludes 401/403/400 | `test_retry_policy_does_not_retry_401` (Layer 2) | 2 |
+| C12 | Stats snapshot under lock | `test_total_stats_consistent_under_concurrent_writes` (Layer 2) | 2 |
+| F1 | LitellmModel signature positional-first | mypy / `test_anthropic_caching_litellm_model_overrides_correctly` | 1, 2 |
+| F2 | RunHooks AgentHookContext + result:str | mypy / `test_hooks_signatures_match_sdk` | 1, 2 |
+| F3 | TracingProcessor methods sync | mypy / `test_processor_methods_are_sync` | 1, 2 |
+| C13 | Bus.finalize cleans up state | `test_finalize_clears_inbox_parent_name` (Layer 2) | 2 |
+| C14 | Filter try/except | `test_filter_exception_returns_unmodified` (Layer 2) | 2 |
+| C15 | Hooks try/except | `test_hooks_exception_does_not_propagate` (Layer 2) | 2 |
+| C16 | Processor catches OSError | `test_processor_oserror_caught_run_continues` (Layer 2) | 2 |
+| C17 | Model alias validation | `test_unknown_alias_raises_user_error` (Layer 2) | 2 |
+| C18 | Sandbox response size cap | `test_sandbox_response_too_large_returns_error` (Layer 2) | 2 |
+| C19 | Assert ≥1 enabled tool | `test_agent_build_fails_with_no_tools` (Layer 2) | 2 |
+| C20 | Per-tool timeout_behavior | `test_critical_tool_timeout_raises` + `test_idempotent_tool_timeout_returns_string` | 2 |
+| C21 | RunConfig override + context fields | `test_run_config_override_merges` + `test_context_has_is_whitebox` | 2 |
+| C22 | finish_scan checks children | `test_finish_scan_blocks_with_running_children` (Layer 3) | 3 |
+| C23 | Diff-scope injection | `test_diff_scope_in_first_user_message` (Layer 4) | 4 |
+| C24 | Run-name + Docker preflight | `test_collision_detected` + `test_docker_unavailable_clear_error` (Layer 7) | 7 |
+| C25 | Cancel mode mapping | `test_ctrl_c_immediate` + `test_tui_stop_after_turn` (Layer 7) | 7 |
+
+Every test name above is a placeholder for a real test function. The grid is the contract. CI runs all of these.
+
+---
+
+## 8. CI gating and cutover criteria
+
+### Per-PR CI (every commit)
+
+| Check | Layer | Failure means |
+|---|---|---|
+| `mypy --strict` | 1 | Type contract drift |
+| `ruff check + format` | 1 | Lint failure |
+| `pytest tests/static/` | 1 | Inventory incomplete or imports broken |
+| `pytest tests/<module>/` (unit) | 2 | Module-level regression |
+| `pytest tests/integration/` | 3 | Cross-module regression |
+| `pytest tests/parity/` against recorded scenarios | 4, 5 | Behavioral drift |
+| Inventory completeness | 1 | A feature row has no test_id |
+| Inventory test existence | 1 | A test_id is missing the test |
+| SDK version pin | 1 | Accidental SDK upgrade |
+
+### Nightly CI (scheduled, longer-running)
+
+| Check | Purpose |
+|---|---|
+| Re-record scenarios against real LLM+sandbox | Detect drift in legacy or SDK behavior over time |
+| Mutation testing on critical modules (bus, filter, hooks) | Verify tests actually catch bugs |
+| Multi-platform PyInstaller build + smoke | Catch packaging regressions on macOS arm64/x86_64, Linux, Windows |
+| Memory-pressure soak (300-turn run) | Catch leaks |
+
+### Cutover criteria
+
+To flip `STRIX_USE_SDK_HARNESS` default from `0` to `1`:
+
+- [ ] All 25 corrections (C1–C25 + F1–F3) have green tests in the grid above.
+- [ ] Inventory matrix: ≥98% of rows are `parity-verified`. Remaining ≤2% are explicitly `legacy-only` with owner sign-off.
+- [ ] Layer 4 parity diffing: 8 baseline scenarios all empty-diff (post-normalization).
+- [ ] Layer 5 replay: all recorded scenarios green.
+- [ ] Layer 7 manual smoke: TUI works on macOS + Linux; headless mode produces correct exit codes.
+- [ ] Shadow mode (Layer 6): ≥99% parity rate over a 7-run sample at minimum.
+- [ ] Mutation testing: ≥80% mutation kill rate on critical modules.
+- [ ] Memory soak: 300-turn run completes; memory growth < 1 GB; no orphan containers.
+- [ ] Engineering team signoff via PR review.
+
+### Post-cutover gates (kept for one release after flip)
+
+- Legacy harness still ships (gated by `STRIX_USE_SDK_HARNESS=0`).
+- CI continues to run parity diffing on every PR.
+- Nightly re-record runs detect any post-cutover legacy/SDK drift.
+- Production telemetry on parity rate from real users (sampled).
+
+If parity rate drops below 95% in production: emergency rollback.
+
+---
+
+## 9. Manual smoke checklist
+
+Things a human verifies before cutover. Run these on macOS and Linux at minimum.
+
+### TUI mode
+- [ ] `strix --target https://juice-shop.local` launches splash screen.
+- [ ] Agent tree renders; root expands to show subagents when spawned.
+- [ ] Streaming text appears in real time as agent generates.
+- [ ] F1 opens help screen; ESC closes.
+- [ ] Vulnerability popup appears when first finding logged.
+- [ ] Ctrl+C → confirm dialog → ESC dismisses, Y stops agent.
+- [ ] Tab cycles panels; arrow keys navigate agent tree.
+- [ ] Ctrl+Q quits cleanly; container removed; run dir intact.
+- [ ] After agent completes, prompt for follow-up message; user can type.
+
+### Headless mode
+- [ ] `strix -n --target ./examples/vulnerable-flask-app --scan-mode quick` runs.
+- [ ] Rich panels render: startup, live stats, vuln-found.
+- [ ] Final summary panel shows.
+- [ ] Exit code 2 if findings; 0 if none.
+- [ ] Ctrl+C exits 130 cleanly; container removed.
+
+### Multi-target
+- [ ] `strix -t ./repo -t https://app.local` handles both.
+- [ ] Each target gets its own `/workspace/<subdir>` mount.
+- [ ] Findings tagged by target.
+
+### Diff scope
+- [ ] `strix -n --target ./repo --scope-mode diff --diff-base origin/main` includes diff block in instruction.
+- [ ] Agent's first turn references the diff.
+
+### Config override
+- [ ] `strix --config /path/to/custom.json --target ...` overrides defaults.
+- [ ] Env vars from custom config apply; default config vars cleared.
+
+### Resilience
+- [ ] Kill the sandbox container mid-run (`docker stop`); agent surfaces error, exits gracefully.
+- [ ] Run with invalid `STRIX_LLM=strix/typo`; clear error message naming valid aliases.
+- [ ] Run without Docker daemon; preflight error message tells user to start Docker.
+
+---
+
+## 10. Post-cutover monitoring
+
+Once SDK harness is the default, watch for slow regressions.
+
+### Daily metrics
+
+- Parity rate from shadow sample (rolling 7-day).
+- Token cost per scan (rolling 7-day per scan_mode).
+- Wall-clock per scan (rolling 7-day).
+- Findings count distribution by CWE.
+- Crash rate by category (LLM, sandbox, tool, agent).
+
+### Alerts
+
+- Parity rate drops below 95% → page engineer.
+- Token cost rises >20% week-over-week (with same scan mix) → review.
+- Crash rate rises >2× baseline → page.
+- Any new error class appears in events.jsonl that wasn't present pre-cutover → review.
+
+### Telemetry (existing PostHog + custom)
+
+- Per-scan: legacy-vs-sdk path used, total cost, total findings, exit code, duration, scan_mode.
+- Per-tool: invocation count, error rate, p50/p99 latency.
+- Per-error: category, count, first/last seen.
+
+---
+
+## 11. What to do when a parity diff fails
+
+Standard incident playbook. Don't let a failed diff sit; chase root cause.
+
+1. **Examine the diff** — `tests/parity/run_parity.py` outputs a structured diff. Identify the specific field/event that drifted.
+2. **Classify the drift**:
+ - **Code bug** in our migration → fix, re-test.
+ - **Acceptable behavior change** (e.g., SDK emits a new event the legacy didn't) → update normalizer to ignore the field, document in `tests/parity/normalization_notes.md`.
+ - **Recording staleness** → re-record affected scenario; investigate why output changed.
+3. **Add a new test** if the drift wasn't caught by an existing assertion. Update inventory matrix.
+4. **Don't suppress** — never `# noqa` a parity failure. The matrix is the contract.
+
+### Common drift causes (anticipated)
+
+| Drift | Root cause | Fix |
+|---|---|---|
+| Tool call sequence differs | LLM nondeterminism on retries | Recording captures multiple valid sequences; diff accepts any |
+| Event count off by 1 | SDK emits an extra `agent.start` for handoff | Normalizer filters handoff events |
+| Token count drift > 5% | Anthropic cache hit/miss timing | Replay against recorded; if still drifts, investigate cache wrapper |
+| Vulnerability missing | Dedup decided differently | Check dedup LLM call recording; verify same prompt produces same answer |
+| Wiki note shape differs | Update format changed | Normalize whitespace; check `append_note_content` invocation |
+
+---
+
+## TL;DR
+
+Five mechanisms catch the categories of regression we're worried about:
+
+1. **Layer 4 parity diffing on 8 baseline scenarios** — catches every silent regression in tool calls, events, findings, reports. The diff itself is the signal; we don't need to enumerate failures.
+2. **Feature inventory matrix (~250 rows)** — every feature has a test; CI fails if the matrix has gaps. Prevents adding features without tests; prevents losing features without notice.
+3. **Recorded LLM + sandbox replay (Layer 5)** — deterministic CI; same input always produces same output; no token burn per PR.
+4. **Shadow mode in production (Layer 6)** — real-world parity validation; sampled at 10%; strongest signal pre-cutover.
+5. **Per-correction tests (25+3 = 28 specific tests)** — every audit finding has a test that proves the fix works.
+
+CI gates everything. Cutover criteria are explicit and measurable. Rollback is a flag flip. Post-cutover monitoring catches slow drift.
+
+The migration cannot silently lose a feature unless: (a) the feature isn't in the inventory matrix, AND (b) it isn't exercised by any of the 8 baseline scenarios, AND (c) it doesn't appear in production within the shadow sampling window. That's a narrow gap, and it gets narrower with every scenario added.