feat(llm): Claude prompt caching (system + tools + rolling tail), Bedrock-safe

Enables Anthropic/Bedrock prompt caching for Claude routes via LiteLLM
cache_control_injection_points: system prompt + latest message everywhere,
plus tool_config on Bedrock Converse only (the sole route whose transform
consumes it; elsewhere it leaks as an unknown top-level field that native
Anthropic 400-rejects). Unmapped Bedrock Claude models run uncached instead
of crashing. Adds STRIX_PROMPT_CACHE opt-out (default on).

.gitignore

@@ -1,8 +1,8 @@
 # Node / local-viewer SPA source (the built bundle in
-# strix/viewer/static/ is committed and shipped; do not ignore it)
+# strix/interface/viewer/static/ is committed and shipped; do not ignore it)
 node_modules/
-strix/viewer/frontend/node_modules/
-strix/viewer/frontend/.vite/
+strix/interface/viewer/frontend/node_modules/
+strix/interface/viewer/frontend/.vite/
 
 # Python
 __pycache__/

CONTRIBUTING.md

@@ -102,16 +102,16 @@ We welcome feature ideas! Please:
 ## 🖥️ Local viewer SPA
 
 `strix view` serves a prebuilt web UI whose source lives in
-`strix/viewer/frontend/` (a Vite + React project) and whose built output is
-committed to `strix/viewer/static/` and shipped in the package. End users never
-run a JS build. If you change anything under `strix/viewer/frontend/`, rebuild
+`strix/interface/viewer/frontend/` (a Vite + React project) and whose built output is
+committed to `strix/interface/viewer/static/` and shipped in the package. End users never
+run a JS build. If you change anything under `strix/interface/viewer/frontend/`, rebuild
 and commit the output:
 
 ```bash
-make viewer   # or: cd strix/viewer/frontend && npm ci && npm run build
+make viewer   # or: cd strix/interface/viewer/frontend && npm ci && npm run build
 ```
 
-Commit both the source change and the regenerated `strix/viewer/static/`.
+Commit both the source change and the regenerated `strix/interface/viewer/static/`.
 
 ## 🤝 Community
 

Makefile

@@ -69,8 +69,8 @@ clean:
 
 viewer:
 	@echo "🖥️  Building the local-viewer SPA..."
-	cd strix/viewer/frontend && npm ci && npm run build
-	@echo "✅ Viewer built to strix/viewer/static/ (commit the changes)."
+	cd strix/interface/viewer/frontend && npm ci && npm run build
+	@echo "✅ Viewer built to strix/interface/viewer/static/ (commit the changes)."
 
 dev: format lint type-check
 	@echo "✅ Development cycle complete!"

pyproject.toml

@@ -79,10 +79,10 @@ build-backend = "hatchling.build"
 
 [tool.hatch.build.targets.wheel]
 packages = ["strix"]
-# The prebuilt viewer bundle under strix/viewer/static/ ships automatically
+# The prebuilt viewer bundle under strix/interface/viewer/static/ ships automatically
 # (hatchling includes non-.py files under the package). The Vite SOURCE lives
-# under the package dir too (strix/viewer/frontend/) but must never ship in the wheel.
-exclude = ["strix/viewer/frontend", "strix/viewer/frontend/**"]
+# under the package dir too (strix/interface/viewer/frontend/) but must never ship in the wheel.
+exclude = ["strix/interface/viewer/frontend", "strix/interface/viewer/frontend/**"]
 
 # ============================================================================
 # Type Checking Configuration
@@ -222,10 +222,10 @@ ignore = [
 "tests/test_codex_streaming.py" = ["N802"]
 "tests/test_report_pdf.py" = ["S105", "S106"]
 # Stdlib HTTP handler overrides (do_GET/do_POST) and lazy imports that avoid a
-# circular dependency with strix.telemetry / strix.viewer.report_pdf.
-"strix/viewer/server.py" = ["N802", "PLC0415"]
+# circular dependency with strix.telemetry / strix.interface.viewer.report_pdf.
+"strix/interface/viewer/server.py" = ["N802", "PLC0415"]
 # Lazy telemetry import to avoid importing PostHog before the viewer starts.
-"strix/viewer/cli.py" = ["PLC0415"]
+"strix/interface/viewer/cli.py" = ["PLC0415"]
 # Lazy imports inside functions to avoid circular dependency with
 # strix.telemetry / strix.report.dedupe / cvss.
 "strix/tools/notes/tools.py" = ["PLC0415", "TC002"]

strix.spec

@@ -26,7 +26,7 @@ for tcss_file in strix_root.rglob('*.tcss'):
     datas.append((str(tcss_file), str(rel_path.parent)))
 
 # Prebuilt local-viewer SPA (served by `strix view`).
-viewer_static = strix_root / 'viewer' / 'static'
+viewer_static = strix_root / 'interface' / 'viewer' / 'static'
 for asset in viewer_static.rglob('*'):
     if asset.is_file():
         rel_path = asset.relative_to(project_root)
@@ -158,12 +158,12 @@ hiddenimports = [
     'strix.report.dedupe',
     'strix.report.state',
     'strix.report.writer',
-    'strix.viewer',
-    'strix.viewer.auth',
-    'strix.viewer.cli',
-    'strix.viewer.report_pdf',
-    'strix.viewer.server',
-    'strix.viewer.transcript',
+    'strix.interface.viewer',
+    'strix.interface.viewer.auth',
+    'strix.interface.viewer.cli',
+    'strix.interface.viewer.report_pdf',
+    'strix.interface.viewer.server',
+    'strix.interface.viewer.transcript',
 
     # PDF report generation + encryption
     'reportlab',

strix/agents/factory.py

@@ -16,6 +16,7 @@ from agents.tool import CustomTool, FunctionTool, Tool
 from pydantic import ValidationError
 
 from strix.agents.prompt import render_system_prompt
+from strix.config import load_settings
 from strix.tools.agents_graph.tools import (
     agent_finish,
     create_agent,
@@ -33,6 +34,7 @@ from strix.tools.notes.tools import (
     list_notes,
     update_note,
 )
+from strix.tools.output_store import bound_and_store, bound_text
 from strix.tools.proxy.tools import (
     list_requests,
     list_sitemap,
@@ -41,7 +43,12 @@ from strix.tools.proxy.tools import (
     view_request,
     view_sitemap_entry,
 )
-from strix.tools.reporting.tool import create_dependency_report, create_vulnerability_report
+from strix.tools.reporting.tool import (
+    create_dependency_report,
+    create_vulnerability_report,
+    get_report,
+    list_reports,
+)
 from strix.tools.thinking.tool import think
 from strix.tools.todo.tools import (
     create_todo,
@@ -103,8 +110,36 @@ def _extract_custom_input(tool: CustomTool, raw_input: str | dict[str, Any]) ->
     return value if isinstance(value, str) else ""
 
 
+def _tool_output_limits() -> tuple[int, int]:
+    context = load_settings().context
+    return context.tool_output_max_lines, context.tool_output_max_bytes
+
+
+async def _bound_result(result: Any) -> Any:
+    if not isinstance(result, str):
+        return result
+    max_lines, max_bytes = _tool_output_limits()
+    return await bound_and_store(result, max_lines=max_lines, max_bytes=max_bytes)
+
+
 def _format_tool_error(exc: Exception) -> str:
-    return str(exc) or exc.__class__.__name__
+    message = str(exc) or exc.__class__.__name__
+    max_lines, max_bytes = _tool_output_limits()
+    return bound_text(message, max_lines=max_lines, max_bytes=max_bytes)
+
+
+def _with_bounded_result(tool: FunctionTool) -> FunctionTool:
+    """Cap a tool's result size before it enters history (idempotent)."""
+    if getattr(tool, "_strix_bounded", False):
+        return tool
+    invoke_tool = tool.on_invoke_tool
+
+    async def invoke(ctx: Any, raw_input: str) -> Any:
+        return await _bound_result(await invoke_tool(ctx, raw_input))
+
+    tool.on_invoke_tool = invoke
+    tool._strix_bounded = True  # type: ignore[attr-defined]
+    return tool
 
 
 def _function_tool_with_error_result(tool: FunctionTool) -> FunctionTool:
@@ -112,7 +147,7 @@ def _function_tool_with_error_result(tool: FunctionTool) -> FunctionTool:
 
     async def invoke(ctx: Any, raw_input: str) -> Any:
         try:
-            return await invoke_tool(ctx, raw_input)
+            return await _bound_result(await invoke_tool(ctx, raw_input))
         except Exception as exc:  # noqa: BLE001 - tool errors should be model-visible results.
             logger.debug("Tool %s failed; returning error as result", tool.name, exc_info=True)
             return _format_tool_error(exc)
@@ -127,7 +162,7 @@ def _custom_tool_as_function_tool(tool: CustomTool) -> FunctionTool:
         if not custom_input:
             return f"`{_custom_tool_input_field(tool)}` must be a non-empty string."
         try:
-            return await tool.on_invoke_tool(ctx, custom_input)
+            return await _bound_result(await tool.on_invoke_tool(ctx, custom_input))
         except Exception as exc:  # noqa: BLE001 - matches SDK CustomTool error-as-result behavior.
             logger.debug("Tool %s failed; returning error as result", tool.name, exc_info=True)
             return _format_tool_error(exc)
@@ -159,12 +194,35 @@ def _custom_tool_as_function_tool(tool: CustomTool) -> FunctionTool:
     )
 
 
-def _configure_chat_completions_filesystem_tools(toolset: Any) -> None:
+def _bound_custom_tool(tool: CustomTool) -> CustomTool:
+    """Bound a native ``CustomTool`` result in place (Responses path)."""
+    invoke_tool = tool.on_invoke_tool
+
+    async def invoke(ctx: Any, raw_input: str) -> Any:
+        return await _bound_result(await invoke_tool(ctx, raw_input))
+
+    tool.on_invoke_tool = invoke
+    return tool
+
+
+def _configure_filesystem_tools(toolset: Any, *, chat_completions: bool) -> None:
     for name, tool in vars(toolset).items():
-        if isinstance(tool, CustomTool):
-            setattr(toolset, name, _custom_tool_as_function_tool(tool))
+        if chat_completions:
+            if isinstance(tool, CustomTool):
+                setattr(toolset, name, _custom_tool_as_function_tool(tool))
+            elif isinstance(tool, FunctionTool):
+                setattr(toolset, name, _function_tool_with_error_result(tool))
+        elif isinstance(tool, CustomTool):
+            setattr(toolset, name, _bound_custom_tool(tool))
         elif isinstance(tool, FunctionTool):
-            setattr(toolset, name, _function_tool_with_error_result(tool))
+            setattr(toolset, name, _with_bounded_result(tool))
+
+
+def _make_filesystem_configurator(*, chat_completions: bool) -> Any:
+    def configure(toolset: Any) -> None:
+        _configure_filesystem_tools(toolset, chat_completions=chat_completions)
+
+    return configure
 
 
 _CHARS_ESCAPE_RE = re.compile(r"\\(?:u[0-9a-fA-F]{4}|x[0-9a-fA-F]{2}|[0abtnvfr\\])")
@@ -205,6 +263,16 @@ def _format_validation_error(tool_name: str, exc: ValidationError) -> str:
     return f"{tool_name}: invalid arguments — " + "; ".join(parts)
 
 
+def _apply_shell_output_cap(parsed: dict[str, Any]) -> None:
+    """Clamp the SDK shell tools' ``max_output_tokens`` to the configured
+    ceiling; a smaller explicit value is respected."""
+    ceiling = load_settings().context.tool_output_max_tokens
+    requested = parsed.get("max_output_tokens")
+    parsed["max_output_tokens"] = (
+        ceiling if not isinstance(requested, int) or requested > ceiling else requested
+    )
+
+
 def _wrap_exec_command(tool: FunctionTool) -> FunctionTool:
     invoke_tool = tool.on_invoke_tool
 
@@ -213,8 +281,10 @@ def _wrap_exec_command(tool: FunctionTool) -> FunctionTool:
             parsed = json.loads(raw_input)
         except (json.JSONDecodeError, TypeError):
             parsed = None
-        if isinstance(parsed, dict) and "shell" not in parsed:
-            parsed["shell"] = "bash"
+        if isinstance(parsed, dict):
+            if "shell" not in parsed:
+                parsed["shell"] = "bash"
+            _apply_shell_output_cap(parsed)
             raw_input = json.dumps(parsed)
         try:
             return await invoke_tool(ctx, raw_input)
@@ -240,8 +310,10 @@ def _wrap_write_stdin(tool: FunctionTool) -> FunctionTool:
             parsed = json.loads(raw_input)
         except json.JSONDecodeError:
             parsed = None
-        if isinstance(parsed, dict) and isinstance(parsed.get("chars"), str):
-            parsed["chars"] = _decode_chars_escape(parsed["chars"])
+        if isinstance(parsed, dict):
+            if isinstance(parsed.get("chars"), str):
+                parsed["chars"] = _decode_chars_escape(parsed["chars"])
+            _apply_shell_output_cap(parsed)
             raw_input = json.dumps(parsed)
         try:
             return await invoke_tool(ctx, raw_input)
@@ -343,6 +415,8 @@ _BASE_TOOLS: tuple[Tool, ...] = (
     web_search,
     create_vulnerability_report,
     create_dependency_report,
+    list_reports,
+    get_report,
     list_requests,
     view_request,
     repeat_request,
@@ -440,6 +514,9 @@ def build_strix_agent(
     else:
         tools = [*_BASE_TOOLS, *agent_tools, agent_finish]
     _ensure_unique_tool_names(tools)
+    tools = [
+        _with_bounded_result(tool) if isinstance(tool, FunctionTool) else tool for tool in tools
+    ]
 
     logger.info(
         "Built %s agent '%s' (skills=%d, tools=%d, scan_mode=%s, whitebox=%s)",
@@ -459,8 +536,8 @@ def build_strix_agent(
         model=None,
         capabilities=[
             Filesystem(
-                configure_tools=(
-                    _configure_chat_completions_filesystem_tools if chat_completions_tools else None
+                configure_tools=_make_filesystem_configurator(
+                    chat_completions=chat_completions_tools,
                 ),
             ),
             Shell(

strix/agents/prompts/system_prompt.jinja

@@ -188,7 +188,7 @@ EFFICIENCY TACTICS:
   script fail with `ModuleNotFoundError`.
 - `exec_command` runs each command in a fresh non-interactive shell (plain
   pipes, no TTY). To drive an interactive or long-running process with
-  `write_stdin` — REPLs, `ssh`/`nc`/`ftp`, `msfconsole`, or to send Ctrl-C —
+  `write_stdin` — REPLs, `ssh`/`nc`/`ftp`, `sqlmap`, or to send Ctrl-C —
   you MUST start it with `exec_command(cmd="...", tty=true)` and then
   `write_stdin(session_id=<id>, chars="...")`. Calling `write_stdin` on a
   default (non-TTY) command or on a process that has already exited fails with
@@ -215,6 +215,7 @@ VALIDATION REQUIREMENTS:
 - A vulnerability is ONLY considered reported when a reporting agent uses create_vulnerability_report (or create_dependency_report for known-CVE dependency/supply-chain findings) with full details. Mentions in agent_finish, finish_scan, or generic messages are NOT sufficient
 - Reporting and fixing are ONE step, not two: when source is available, the reporting agent derives the concrete fix and files it INLINE via create_vulnerability_report (`code_locations` with `fix_before`/`fix_after` + `fix_pr_body`) — the report is not complete without it. Do NOT report first and then spawn a separate downstream agent to re-derive and re-apply the same patch; that just re-does the analysis and wastes tokens. (Do not silently patch a finding WITHOUT filing a report — the report, with its embedded fix, is the deliverable.)
 - DEDUPLICATION: The create_vulnerability_report tool uses LLM-based deduplication. If it rejects your report as a duplicate, DO NOT attempt to re-submit the same vulnerability. Accept the rejection and move on to testing other areas. The vulnerability has already been reported by another agent
+- REVIEWING FILED FINDINGS (orchestrator/root agent): use list_reports to see every vulnerability filed so far in this scan (by any agent, root or child) — metadata-first with per-severity counts — and get_report to read one finding in full by its id. These are read-only orchestration tools: the root agent uses them to track coverage, avoid dispatching work on already-covered ground, assemble the finish_scan executive summary, and reason about attack-chaining across confirmed findings. Leaf/specialist agents should NOT call them — just do your assigned testing and file findings. Each entry shows which agent filed it (agent_name), and your own entries are flagged by_you. list_notes/get_note do the same for notes.
 </execution_guidelines>
 
 <vulnerability_focus>

strix/config/__init__.py

@@ -17,6 +17,7 @@ from strix.config.loader import (
     persist_current,
 )
 from strix.config.settings import (
+    ContextSettings,
     DedupeSettings,
     IntegrationSettings,
     LlmSettings,
@@ -27,6 +28,7 @@ from strix.config.settings import (
 
 
 __all__ = [
+    "ContextSettings",
     "DedupeSettings",
     "IntegrationSettings",
     "LlmSettings",

strix/config/models.py

@@ -429,3 +429,67 @@ def is_known_openai_bare_model(model_name: str) -> bool:
         return False
     entry = litellm.model_cost.get(name)
     return bool(entry and entry.get("litellm_provider") == "openai")
+
+
+def is_claude_model(model_name: str) -> bool:
+    return "claude" in (model_name or "").strip().lower()
+
+
+def is_bedrock_route(model_name: str) -> bool:
+    """Whether ``model_name`` resolves to an AWS Bedrock route.
+
+    Matches the ``bedrock/...`` LiteLLM route prefix and bare Bedrock model ids
+    (``[region.]anthropic.claude-...``).
+    """
+    name = (model_name or "").strip().lower()
+    return name.startswith("bedrock/") or "anthropic." in name
+
+
+def _prompt_cache_name_candidates(model_name: str) -> list[str]:
+    """Candidate LiteLLM model-map keys for ``model_name``, most→least specific.
+
+    LiteLLM keys the same model under several names (``bedrock/global.anthropic.
+    claude-opus-4-1``, ``anthropic.claude-opus-4-1``, ``claude-opus-4-1``) and not
+    every provider/region-prefixed variant is present for every model. Strip the
+    LiteLLM route prefix, then leading dotted segments (region, then provider) so
+    a prefixed name still resolves to a bare key.
+    """
+    name = (model_name or "").strip().lower()
+    for prefix in ("litellm/", "bedrock/"):
+        if name.startswith(prefix):
+            name = name[len(prefix) :]
+            break
+    candidates = [name]
+    rest = name
+    while "." in rest:
+        rest = rest.split(".", 1)[1]
+        candidates.append(rest)
+    return candidates
+
+
+def bedrock_route_supports_prompt_caching(model_name: str) -> bool:
+    """Whether LiteLLM can confirm this Bedrock model supports prompt caching.
+
+    Bedrock's Converse API rejects unknown request fields outright
+    (``ValidationException: cache_control_injection_points: Extra inputs are
+    not permitted``), and LiteLLM only consumes the cache marker for models its
+    (statically bundled) model map recognises as cache-capable. For a Bedrock
+    model missing from that map — a just-released model, or any model when the
+    remote model-map refresh fails and a stale local copy is used — the marker
+    would pass straight through and fail every call, so callers must withhold
+    it unless support is confirmed here.
+    """
+    import litellm
+
+    checker = getattr(getattr(litellm, "utils", None), "supports_prompt_caching", None)
+    for cand in _prompt_cache_name_candidates(model_name):
+        if checker is not None:
+            # supports_prompt_caching raises for models missing from the map;
+            # keep checking the remaining name candidates.
+            with contextlib.suppress(Exception):
+                if checker(cand):
+                    return True
+        entry = litellm.model_cost.get(cand)
+        if entry and entry.get("supports_prompt_caching"):
+            return True
+    return False

strix/config/settings.py

@@ -40,6 +40,10 @@ class LlmSettings(BaseSettings):
         default=False,
         alias="STRIX_FORCE_REQUIRED_TOOL_CHOICE",
     )
+    prompt_cache: bool = Field(
+        default=True,
+        alias="STRIX_PROMPT_CACHE",
+    )
     timeout: int = Field(default=300, alias="LLM_TIMEOUT")
 
 
@@ -55,6 +59,26 @@ class DedupeSettings(BaseSettings):
     api_base: str | None = Field(default=None, alias="DEDUPE_LLM_API_BASE")
 
 
+class ContextSettings(BaseSettings):
+    """Context-window management: per-tool-output caps and history compaction."""
+
+    model_config = _BASE_CONFIG
+
+    auto_compact: bool = Field(default=True, alias="STRIX_CONTEXT_AUTO_COMPACT")
+    compact_buffer_tokens: int = Field(default=20_000, gt=0, alias="STRIX_CONTEXT_BUFFER_TOKENS")
+    keep_tokens: int = Field(default=8_000, gt=0, alias="STRIX_CONTEXT_KEEP_TOKENS")
+    fallback_context_tokens: int = Field(
+        default=200_000, gt=0, alias="STRIX_CONTEXT_FALLBACK_TOKENS"
+    )
+    summary_max_tokens: int = Field(default=4_096, gt=0, alias="STRIX_CONTEXT_SUMMARY_TOKENS")
+    tool_output_max_tokens: int = Field(default=8_000, gt=0, alias="STRIX_TOOL_OUTPUT_MAX_TOKENS")
+    tool_output_max_lines: int = Field(default=2_000, gt=0, alias="STRIX_TOOL_OUTPUT_MAX_LINES")
+    # Floor above the truncation-notice size so a preview always fits.
+    tool_output_max_bytes: int = Field(
+        default=50 * 1024, ge=1024, alias="STRIX_TOOL_OUTPUT_MAX_BYTES"
+    )
+
+
 class RuntimeSettings(BaseSettings):
     model_config = _BASE_CONFIG
 
@@ -99,6 +123,7 @@ class Settings(BaseSettings):
     llm: LlmSettings = Field(default_factory=LlmSettings)
     dedupe: DedupeSettings = Field(default_factory=DedupeSettings)
     runtime: RuntimeSettings = Field(default_factory=RuntimeSettings)
+    context: ContextSettings = Field(default_factory=ContextSettings)
     telemetry: TelemetrySettings = Field(default_factory=TelemetrySettings)
     integrations: IntegrationSettings = Field(default_factory=IntegrationSettings)
     viewer: ViewerSettings = Field(default_factory=ViewerSettings)

strix/core/execution.py

@@ -22,6 +22,7 @@ from strix.core.sessions import (
     open_agent_session,
     strip_all_images_from_session,
 )
+from strix.llm.compaction import is_context_overflow, maybe_compact
 
 
 if TYPE_CHECKING:
@@ -40,6 +41,41 @@ logger = logging.getLogger(__name__)
 StreamEventSink = Callable[[str, Any], None]
 
 _INPUT_REJECTION_CODES = frozenset({400, 404, 422})
+_MAX_COMPACTIONS_PER_CYCLE = 2
+
+
+def _run_config_model(run_config: RunConfig) -> str | None:
+    return run_config.model if isinstance(run_config.model, str) else None
+
+
+def _agent_instructions(agent: Any) -> str:
+    instructions = getattr(agent, "instructions", None)
+    return instructions if isinstance(instructions, str) else ""
+
+
+def _agent_tools_text(agent: Any) -> str:
+    parts: list[str] = []
+    for tool in getattr(agent, "tools", []) or []:
+        name = getattr(tool, "name", "")
+        description = getattr(tool, "description", "") or ""
+        schema = getattr(tool, "params_json_schema", "") or ""
+        parts.append(f"{name} {description} {schema}")
+    return "\n".join(parts)
+
+
+async def _compact_session(
+    agent: Any, session: Session, run_config: RunConfig, *, force: bool
+) -> bool:
+    model = _run_config_model(run_config)
+    if session is None or model is None:
+        return False
+    return await maybe_compact(
+        session,
+        model=model,
+        instructions=_agent_instructions(agent),
+        tools_text=_agent_tools_text(agent),
+        force=force,
+    )
 
 
 async def run_agent_loop(
@@ -350,6 +386,7 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
     hooks: RunHooks[dict[str, Any]] | None,
 ) -> RunResultBase | None:
     image_strips = 0
+    compactions = 0
     while True:
         try:
             await coordinator.mark_running(agent_id)
@@ -360,6 +397,10 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                         await enforce_image_budget(session, max_images)
                     except Exception:
                         logger.exception("image-budget enforcement failed for %s", agent_id)
+                try:
+                    await _compact_session(agent, session, run_config, force=False)
+                except Exception:
+                    logger.exception("proactive compaction failed for %s", agent_id)
             stream = Runner.run_streamed(
                 agent,
                 input=input_data,
@@ -428,6 +469,25 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                     )
                     input_data = []
                     continue
+            if (
+                compactions < _MAX_COMPACTIONS_PER_CYCLE
+                and session is not None
+                and is_context_overflow(exc)
+            ):
+                try:
+                    compacted = await _compact_session(agent, session, run_config, force=True)
+                except Exception:
+                    logger.exception("overflow compaction recovery failed for %s", agent_id)
+                    compacted = False
+                if compacted:
+                    compactions += 1
+                    logger.info(
+                        "Compacted %s session after context overflow; retrying (%d)",
+                        agent_id,
+                        compactions,
+                    )
+                    input_data = []
+                    continue
             if not interactive:
                 raise
             if isinstance(exc, MaxTurnsExceeded):

strix/core/inputs.py

@@ -10,6 +10,9 @@ from openai.types.shared import Reasoning
 
 from strix.config.models import (
     DEFAULT_MODEL_RETRY,
+    bedrock_route_supports_prompt_caching,
+    is_bedrock_route,
+    is_claude_model,
     is_known_openai_bare_model,
     model_supports_reasoning,
     request_timeout_extra_args,
@@ -128,6 +131,7 @@ def make_model_settings(
     model_name: str,
     force_required_tool_choice: bool = False,
     request_timeout: float | None = None,
+    prompt_cache: bool = True,
 ) -> ModelSettings:
     model_settings = ModelSettings(
         parallel_tool_calls=False,
@@ -145,9 +149,95 @@ def make_model_settings(
         )
     if force_required_tool_choice and _accepts_required_tool_choice(model_name):
         model_settings = model_settings.resolve(ModelSettings(tool_choice="required"))
+
+    cache_extra_args = _prompt_cache_extra_args(model_name) if prompt_cache else None
+    if cache_extra_args:
+        # Merge into any existing extra_args (e.g. the request timeout) rather
+        # than relying on resolve()'s dict-merge semantics, so it is obvious at
+        # the call site that unrelated LiteLLM options are preserved.
+        model_settings = model_settings.resolve(
+            ModelSettings(
+                extra_args={**(model_settings.extra_args or {}), **cache_extra_args},
+            ),
+        )
     return model_settings
 
 
+def _prompt_cache_extra_args(model_name: str) -> dict[str, Any] | None:
+    """LiteLLM ``extra_args`` that enable Anthropic/Bedrock prompt caching.
+
+    A Strix scan is a long, multi-turn agentic loop that re-sends a large,
+    STABLE prefix every turn — the system prompt plus the tool schemas — AND an
+    append-only conversation transcript that only grows. Without caching
+    breakpoints the whole request is re-tokenised and billed at the full input
+    rate on every turn; on Claude that is the single biggest lever on scan cost
+    (measured: ``cache-read 0% -> ~66%`` on a real scan once these points are set).
+
+    This mirrors the caching policy of production agent harnesses (e.g.
+    anomalyco/opencode's ``cache-policy``): cache the tool schemas, the system
+    prompt, and the latest conversation message, capped at Anthropic's 4
+    breakpoints. We express it through LiteLLM's ``cache_control_injection_points``
+    — the ``AnthropicCacheControlHook`` fires on that kwarg and emits the
+    provider-appropriate breakpoint (Anthropic ``cache_control``; Bedrock
+    Converse ``cachePoint``). ``LitellmModel`` forwards ``ModelSettings.extra_args``
+    straight into ``litellm.acompletion()``, so passing the points there is all
+    that is required; this is the LiteLLM-config-layer approach the Agents SDK
+    maintainer prescribed over a native ``ModelSettings`` caching flag
+    (openai/openai-agents-python#3008 / #3009).
+
+    Returns ``None`` (a strict no-op — the hook never fires) for every route
+    that would not benefit or could break:
+      - non-Claude models, and
+      - Bedrock Claude routes LiteLLM can't confirm as cache-capable. Bedrock's
+        Converse API rejects unknown request fields outright
+        (``ValidationException: cache_control_injection_points: Extra inputs are
+        not permitted``) and LiteLLM only consumes the marker for models its
+        model map recognises; an unmapped Bedrock model would pass the marker
+        straight through and crash the first call. Only Bedrock hard-rejects, so
+        only Bedrock is guarded — gating Anthropic-native/Vertex/OpenRouter on
+        confirmed support would needlessly disable caching for capable models.
+    """
+    if not is_claude_model(model_name):
+        return None
+    if is_bedrock_route(model_name) and not bedrock_route_supports_prompt_caching(model_name):
+        return None
+
+    points = _prompt_cache_injection_points(model_name)
+    return {"cache_control_injection_points": points}
+
+
+def _prompt_cache_injection_points(model_name: str) -> list[dict[str, Any]]:
+    """Cache breakpoints for a Claude route (system + tools + latest message).
+
+    At most 3 of Anthropic's 4 allowed breakpoints, leaving headroom:
+      - system prompt (``role: system``) — the largest repeated span.
+      - tool schemas (``tool_config``) — Bedrock Converse ONLY. LiteLLM's
+        ``tool_config`` location is implemented solely by the Bedrock Converse
+        transform (which appends a ``cachePoint`` to the tool list); on any other
+        route it is not consumed and would leak onto the wire as an unknown
+        top-level ``cache_control_injection_points`` field. It is also redundant
+        elsewhere: Anthropic orders tools BEFORE the system prompt, so the system
+        breakpoint already caches the tool schemas in the shared prefix.
+      - latest message (``index: -1``) — a ROLLING breakpoint on the last
+        message. A scan transcript is append-only (prior turns are immutable, each
+        turn just appends new assistant/tool messages), so without it the growing
+        body is re-sent at full input price every turn and cache-read decays as a
+        denominator effect even though the prefix keeps hitting. Re-caching the
+        whole prefix-so-far each turn keeps cache-read high on long scans. (This
+        is the Strix analogue of opencode's ``latest-user-message``; ``index: -1``
+        tracks the true tail because Strix appends tool-role, not user-role,
+        messages each turn.)
+
+    Unrecognised locations degrade gracefully on older LiteLLM — they are simply
+    not injected (no error).
+    """
+    points: list[dict[str, Any]] = [{"location": "message", "role": "system"}]
+    if is_bedrock_route(model_name):
+        points.append({"location": "tool_config"})
+    points.append({"location": "message", "index": -1})
+    return points
+
+
 def child_initial_input(
     *,
     name: str,

strix/core/runner.py

@@ -3,10 +3,12 @@
 from __future__ import annotations
 
 import contextlib
+import io
 import json
 import logging
 import uuid
 from collections.abc import Callable
+from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 from agents import RunConfig
@@ -40,6 +42,10 @@ from strix.core.paths import run_dir_for, runtime_state_dir
 from strix.core.sessions import open_agent_session
 from strix.runtime import session_manager
 from strix.telemetry.logging import set_scan_id, setup_scan_logging
+from strix.tools.output_store import (
+    WORKSPACE_SPILL_DIR,
+    configure_spill_writer,
+)
 
 
 if TYPE_CHECKING:
@@ -203,6 +209,20 @@ async def run_strix_scan(
     )
     logger.info("Sandbox ready for scan %s", scan_id)
 
+    sandbox_session = bundle["session"]
+
+    async def _spill_to_workspace(output_id: str, text: str) -> str | None:
+        """Write an oversized tool result into the sandbox; return its path or None."""
+        path = f"{WORKSPACE_SPILL_DIR}/{output_id}.txt"
+        try:
+            await sandbox_session.write(Path(path), io.BytesIO(text.encode("utf-8")))
+        except Exception:
+            logger.exception("failed to spill tool output to sandbox workspace")
+            return None
+        return path
+
+    configure_spill_writer(_spill_to_workspace)
+
     sessions_to_close: list[SQLiteSession] = []
 
     try:
@@ -216,6 +236,7 @@ async def run_strix_scan(
             model_name=resolved_model,
             force_required_tool_choice=settings.llm.force_required_tool_choice,
             request_timeout=settings.llm.timeout,
+            prompt_cache=settings.llm.prompt_cache,
         )
         run_config = RunConfig(
             model=resolved_model,
@@ -399,6 +420,7 @@ async def run_strix_scan(
                 await coordinator.set_status(root_id, "failed")
         raise
     finally:
+        configure_spill_writer(None)
         for s in sessions_to_close:
             with contextlib.suppress(Exception):
                 s.close()

strix/core/sessions.py

@@ -92,6 +92,39 @@ async def _rewrite_session(
         return True
 
 
+async def replace_session_items(
+    session: Session,
+    new_items: list[Any],
+    *,
+    expected_len: int | None = None,
+) -> bool:
+    """Overwrite the session's items, restoring the originals on failure.
+
+    When ``expected_len`` is given, the rewrite is skipped if the session no
+    longer has that many items (a concurrent writer changed it), so a slow
+    compaction summary can't clobber newer turns.
+    """
+    async with session_write_lock(session):
+        original = list(await session.get_items())
+        if expected_len is not None and len(original) != expected_len:
+            logger.warning(
+                "skipping session rewrite: expected %d items, found %d",
+                expected_len,
+                len(original),
+            )
+            return False
+        rebuilt = cast("list[TResponseInputItem]", new_items)
+        await session.clear_session()
+        try:
+            await session.add_items(rebuilt)
+        except Exception:
+            logger.exception("session rewrite failed; restoring original items")
+            await session.clear_session()
+            await session.add_items(original)
+            raise
+        return True
+
+
 async def strip_all_images_from_session(session: Session) -> bool:
     """Replace every image tool output with a text placeholder (rejection recovery)."""
 

strix/interface/main.py

@@ -952,7 +952,7 @@ def main() -> None:
     # `strix view [<run>]` is a viewer-only subcommand, dispatched before the
     # scan argument parser (which requires a target) and before any scan setup.
     if len(sys.argv) > 1 and sys.argv[1] == "view":
-        from strix.viewer.cli import run_view
+        from strix.interface.viewer.cli import run_view
 
         run_view(sys.argv[2:])
         return

strix/interface/tui/app.py

@@ -1862,7 +1862,7 @@ class StrixTUIApp(App):  # type: ignore[misc]
                 webbrowser.open(self._viewer_url)
             return
         try:
-            from strix.viewer.server import authorized_url, bundle_is_built, serve
+            from strix.interface.viewer.server import authorized_url, bundle_is_built, serve
 
             if not bundle_is_built():
                 self._set_viewer_cta("[#eab308]Viewer UI not built[/]")

strix/interface/tui/renderers/notes_renderer.py

@@ -7,6 +7,13 @@ from .base_renderer import BaseToolRenderer
 from .registry import register_tool_renderer
 
 
+def _author_label(note: dict[str, Any]) -> str:
+    if note.get("by_you"):
+        return "you"
+    agent_name = note.get("agent_name")
+    return str(agent_name).strip() if agent_name else ""
+
+
 @register_tool_renderer
 class CreateNoteRenderer(BaseToolRenderer):
     tool_name: ClassVar[str] = "create_note"
@@ -123,6 +130,9 @@ class ListNotesRenderer(BaseToolRenderer):
                     text.append("\n  - ")
                     text.append(title)
                     text.append(f" ({category})", style="dim")
+                    author = _author_label(note)
+                    if author:
+                        text.append(f" by {author}", style="dim")
 
                     if note_content:
                         text.append("\n    ")
@@ -156,6 +166,9 @@ class GetNoteRenderer(BaseToolRenderer):
             text.append("\n  ")
             text.append(title)
             text.append(f" ({category})", style="dim")
+            author = _author_label(note)
+            if author:
+                text.append(f" by {author}", style="dim")
             if content:
                 text.append("\n  ")
                 text.append(content, style="dim")

strix/interface/tui/renderers/reporting_renderer.py

@@ -431,3 +431,117 @@ class CreateDependencyReportRenderer(BaseToolRenderer):
 
         css_classes = cls.get_css_classes("completed")
         return Static(padded, classes=css_classes)
+
+
+_LIST_SEVERITY_COLORS = {
+    "critical": "#dc2626",
+    "high": "#ea580c",
+    "medium": "#d97706",
+    "low": "#65a30d",
+    "info": "#0284c7",
+    "none": "#6b7280",
+}
+
+
+def _severity_style(severity: Any) -> str:
+    return _LIST_SEVERITY_COLORS.get(str(severity or "").lower(), "#d97706")
+
+
+def _author_label(report: dict[str, Any]) -> str:
+    if report.get("by_you"):
+        return "you"
+    agent_name = report.get("agent_name")
+    return str(agent_name).strip() if agent_name else ""
+
+
+@register_tool_renderer
+class ListReportsRenderer(BaseToolRenderer):
+    tool_name: ClassVar[str] = "list_reports"
+    css_classes: ClassVar[list[str]] = ["tool-call", "reporting-tool"]
+
+    @classmethod
+    def render(cls, tool_data: dict[str, Any]) -> Static:
+        result = _coerce_dict(tool_data.get("result"))
+
+        text = Text()
+        text.append("◆ ", style="#ef4444")
+        text.append("reports", style="dim")
+
+        if isinstance(tool_data.get("result"), str) and str(tool_data["result"]).strip():
+            text.append("\n  ")
+            text.append(str(tool_data["result"]).strip(), style="dim")
+        elif result.get("success"):
+            total = result.get("total_count", 0)
+            reports = _coerce_list_of_dicts(result.get("reports"))
+            counts = _coerce_dict(result.get("severity_counts"))
+
+            text.append(f" ({total})", style="dim")
+            for sev, count in counts.items():
+                text.append("  ")
+                text.append(f"{sev} {count}", style=_severity_style(sev))
+
+            if not reports:
+                text.append("\n  ")
+                text.append("No reports filed yet", style="dim")
+            else:
+                for report in reports:
+                    rid = str(report.get("id", "")).strip()
+                    title = str(report.get("title", "")).strip() or "(untitled)"
+                    severity = str(report.get("severity", "")).strip()
+                    text.append("\n  - ")
+                    if severity:
+                        text.append(severity.upper(), style=f"bold {_severity_style(severity)}")
+                        text.append(" ")
+                    if rid:
+                        text.append(f"{rid} ", style="dim")
+                    text.append(title)
+                    author = _author_label(report)
+                    if author:
+                        text.append(f" ({author})", style="dim")
+        else:
+            text.append("\n  ")
+            text.append("Loading...", style="dim")
+
+        css_classes = cls.get_css_classes("completed")
+        return Static(text, classes=css_classes)
+
+
+@register_tool_renderer
+class GetReportRenderer(BaseToolRenderer):
+    tool_name: ClassVar[str] = "get_report"
+    css_classes: ClassVar[list[str]] = ["tool-call", "reporting-tool"]
+
+    @classmethod
+    def render(cls, tool_data: dict[str, Any]) -> Static:
+        result = _coerce_dict(tool_data.get("result"))
+
+        text = Text()
+        text.append("◆ ", style="#ef4444")
+        text.append("report read", style="dim")
+
+        report = _coerce_dict(result.get("report")) if result.get("success") else {}
+        if report:
+            rid = str(report.get("id", "")).strip()
+            title = str(report.get("title", "")).strip() or "(untitled)"
+            severity = str(report.get("severity", "")).strip()
+            text.append("\n  ")
+            if severity:
+                text.append(severity.upper(), style=f"bold {_severity_style(severity)}")
+                text.append(" ")
+            if rid:
+                text.append(f"{rid} ", style="dim")
+            text.append(title)
+            author = _author_label(report)
+            if author:
+                text.append(f" ({author})", style="dim")
+            target = str(report.get("target", "")).strip()
+            if target:
+                text.append("\n  ")
+                text.append(target, style="dim")
+        else:
+            text.append("\n  ")
+            detail = result.get("error") if result.get("success") is False else None
+            text.append(str(detail) if detail else "Loading...", style="dim")
+
+        css_classes = cls.get_css_classes("completed")
+        return Static(text, classes=css_classes)

strix/interface/viewer/__init__.py

@@ -6,7 +6,7 @@ directly from the run's on-disk files. No cloud dependency, no file picker.
 
 from __future__ import annotations
 
-from strix.viewer.server import serve
+from strix.interface.viewer.server import serve
 
 
 __all__ = ["serve"]

strix/interface/viewer/auth.py

@@ -155,7 +155,7 @@ def _post_json(path: str, payload: dict[str, Any], *, timeout: int) -> tuple[int
         method="POST",
     )
     try:
-        with urllib.request.urlopen(request, timeout=timeout) as response:  # noqa: S310
+        with urllib.request.urlopen(request, timeout=timeout) as response:  # noqa: S310  # nosec B310
             return response.status, _parse_body(response.read())
     except urllib.error.HTTPError as exc:
         return exc.code, _parse_body(exc.read())

strix/interface/viewer/cli.py

@@ -16,8 +16,8 @@ from strix.core.paths import (
     run_record_path,
     runs_base_dir,
 )
-from strix.viewer.server import authorized_url, bundle_is_built, serve
-from strix.viewer.transcript import read_run_summary
+from strix.interface.viewer.server import authorized_url, bundle_is_built, serve
+from strix.interface.viewer.transcript import read_run_summary
 
 
 if TYPE_CHECKING:
@@ -58,7 +58,7 @@ def run_view(argv: list[str]) -> None:
     if not bundle_is_built():
         console.print(
             "[bold red]Viewer UI is not built.[/]\n"
-            "Build it with: [cyan]cd strix/viewer/frontend && npm ci && npm run build[/]"
+            "Build it with: [cyan]cd strix/interface/viewer/frontend && npm ci && npm run build[/]"
         )
         raise SystemExit(1)
 

strix/interface/viewer/frontend/src/components/live/tool-renderers/NotesRenderer.tsx

@@ -49,6 +49,9 @@ export default function NotesRenderer({ toolName, args, result }: ToolRendererPr
             <div className="mt-1.5 text-[#999] text-[13px]">
               {note.title ?? "(untitled)"}
               <span className="text-[#555] ml-1">({note.category ?? "general"})</span>
+              {(note.by_you || note.agent_name) && (
+                <span className="text-[#666] ml-1 text-xs">by {note.by_you ? "you" : note.agent_name}</span>
+              )}
             </div>
             {note.content && <div className="mt-1"><Markdown text={note.content} /></div>}
           </>
@@ -74,6 +77,9 @@ export default function NotesRenderer({ toolName, args, result }: ToolRendererPr
                 <span className="text-[#555] mr-1">-</span>
                 <span className="text-[#999]">{n.title ?? "(untitled)"}</span>
                 <span className="text-[#555] ml-1">({n.category ?? "general"})</span>
+                {(n.by_you || n.agent_name) && (
+                  <span className="text-[#666] ml-1 text-xs">by {n.by_you ? "you" : n.agent_name}</span>
+                )}
                 {n.content && <div className="ml-3"><Markdown text={n.content} /></div>}
               </div>
             ))}

strix/interface/viewer/frontend/src/components/live/tool-renderers/ReportListRenderer.tsx

@@ -0,0 +1,121 @@
+"use client";
+
+import type { ToolRendererProps } from "@/types/events";
+import { TruncatedText } from "./ToolCard";
+import Markdown from "./Markdown";
+
+const SEVERITY_COLORS: Record<string, string> = {
+  critical: "text-red-400", high: "text-orange-400", medium: "text-yellow-400",
+  low: "text-blue-400", info: "text-cyan-400", none: "text-[#888]",
+};
+
+interface ReportEntry {
+  id?: string;
+  title?: string;
+  severity?: string;
+  cvss?: number;
+  cve?: string;
+  cwe?: string;
+  target?: string;
+  endpoint?: string;
+  method?: string;
+  description_preview?: string;
+  description?: string;
+  agent_name?: string;
+  by_you?: boolean;
+}
+
+function authorTag(r: ReportEntry) {
+  if (!r.agent_name && !r.by_you) return null;
+  const label = r.by_you ? "you" : r.agent_name;
+  return <span className="text-[#666] text-xs ml-1.5">({label})</span>;
+}
+
+function sevBadge(severity: string | undefined) {
+  const sev = String(severity ?? "").toLowerCase();
+  const color = SEVERITY_COLORS[sev] ?? "text-yellow-400";
+  return <span className={`font-semibold text-[13px] ${color}`}>{sev.toUpperCase() || "—"}</span>;
+}
+
+export default function ReportListRenderer({ toolName, result }: ToolRendererProps) {
+  const res = result as Record<string, unknown> | null;
+  const ok = res != null && typeof res === "object" && res.success === true;
+
+  if (toolName === "get_report") {
+    const report = ok ? (res.report as ReportEntry | undefined) : undefined;
+    return (
+      <div>
+        <span className="text-red-400/80 font-semibold text-sm">report</span>
+        {report ? (
+          <div className="mt-1.5 space-y-2">
+            <div className="flex items-center gap-2 flex-wrap">
+              {sevBadge(report.severity)}
+              {report.cvss != null && <span className="text-[#888] text-[13px]">CVSS {report.cvss}</span>}
+              {report.id && <span className="text-[#555] font-mono text-[13px]">{report.id}</span>}
+              {report.cve && <span className="text-[#888] font-mono text-[13px]">{report.cve}</span>}
+              {report.cwe && <span className="text-[#888] font-mono text-[13px]">{report.cwe}</span>}
+              {(report.agent_name || report.by_you) && (
+                <span className="text-[#666] text-[13px]">{report.by_you ? "you" : report.agent_name}</span>
+              )}
+            </div>
+            {report.title && <div className="text-[15px] text-white/80 font-semibold">{report.title}</div>}
+            {(report.target || report.endpoint) && (
+              <div className="text-[13px] text-[#888] font-mono">
+                {report.target}{report.endpoint ? ` ${report.method ?? ""} ${report.endpoint}` : ""}
+              </div>
+            )}
+            {report.description && <TruncatedText text={report.description} maxLines={20} />}
+          </div>
+        ) : (
+          <div className="mt-1 text-[#555] text-xs">
+            {(res && typeof res === "object" && (res.error as string)) || "Report not found"}
+          </div>
+        )}
+      </div>
+    );
+  }
+
+  // list_reports
+  const rawReports = ok ? res.reports : null;
+  const reports: ReportEntry[] = Array.isArray(rawReports) ? (rawReports as ReportEntry[]) : [];
+  const total = ok && typeof res.total_count === "number" ? (res.total_count as number) : reports.length;
+  const counts = ok && res.severity_counts && typeof res.severity_counts === "object"
+    ? (res.severity_counts as Record<string, number>)
+    : {};
+  const countEntries = Object.entries(counts);
+
+  return (
+    <div>
+      <div className="flex items-center gap-2 flex-wrap">
+        <span className="text-red-400/80 font-semibold text-sm">reports</span>
+        <span className="text-[#555] text-[13px]">({total})</span>
+        {countEntries.map(([sev, n]) => (
+          <span key={sev} className="text-[13px]">
+            {sevBadge(sev)}<span className="text-[#888] ml-0.5">{n}</span>
+          </span>
+        ))}
+      </div>
+      {reports.length > 0 ? (
+        <div className="mt-1.5 space-y-1">
+          {reports.map((r, i) => (
+            <div key={r.id ?? i} className="text-[13px]">
+              <span className="text-[#555] mr-1">-</span>
+              {sevBadge(r.severity)}
+              {r.id && <span className="text-[#555] font-mono ml-1.5">{r.id}</span>}
+              <span className="text-[#999] ml-1.5">{r.title ?? "(untitled)"}</span>
+              {authorTag(r)}
+              {(r.target || r.endpoint) && (
+                <div className="ml-3 text-[#666] font-mono text-xs">
+                  {r.target}{r.endpoint ? ` ${r.method ?? ""} ${r.endpoint}` : ""}
+                </div>
+              )}
+              {r.description_preview && (
+                <div className="ml-3"><Markdown text={r.description_preview} /></div>
+              )}
+            </div>
+          ))}
+        </div>
+      ) : <div className="mt-1 text-[#555] text-xs">No reports filed yet</div>}
+    </div>
+  );
+}

strix/interface/viewer/frontend/src/components/live/tool-renderers/index.ts

@@ -12,6 +12,7 @@ import FileEditRenderer from "./FileEditRenderer";
 import ApplyPatchRenderer from "./ApplyPatchRenderer";
 import ViewImageRenderer from "./ViewImageRenderer";
 import VulnReportRenderer from "./VulnReportRenderer";
+import ReportListRenderer from "./ReportListRenderer";
 import ProxyRenderer from "./ProxyRenderer";
 import ThinkRenderer from "./ThinkRenderer";
 import AgentCommsRenderer from "./AgentCommsRenderer";
@@ -101,7 +102,7 @@ const CATEGORY_TOOLS: Record<ToolCategory, readonly string[]> = {
   filesystem: ["apply_patch", "view_image", "str_replace_editor", "list_files", "search_files"],
   // Caido proxy tools (legacy: send_request)
   proxy: ["list_requests", "view_request", "repeat_request", "list_sitemap", "view_sitemap_entry", "scope_rules", "send_request"],
-  reporting: ["create_vulnerability_report"],
+  reporting: ["create_vulnerability_report", "list_reports", "get_report"],
   thinking: ["think"],
   agents: ["create_agent", "agent_finish", "send_message_to_agent", "wait_for_message", "view_agent_graph", "stop_agent"],
   search: ["web_search"],
@@ -128,6 +129,8 @@ const RENDERER_OVERRIDES: Partial<Record<string, ComponentType<ToolRendererProps
   finish_scan: FinishRenderer,
   apply_patch: ApplyPatchRenderer,
   view_image: ViewImageRenderer,
+  list_reports: ReportListRenderer,
+  get_report: ReportListRenderer,
 };
 
 /**

strix/interface/viewer/frontend/vite.config.ts

@@ -5,7 +5,7 @@ import { fileURLToPath, URL } from "node:url";
 
 // The viewer is served as static files by a stdlib Python server on an
 // arbitrary ephemeral port, so all asset URLs must be relative (base: "./").
-// The build output is committed at strix/viewer/static and shipped.
+// The build output is committed at strix/interface/viewer/static and shipped.
 export default defineConfig({
   base: "./",
   plugins: [react(), tailwindcss()],

strix/interface/viewer/report_pdf.py

@@ -38,7 +38,7 @@ from reportlab.platypus import (
     TableStyle,
 )
 
-from strix.viewer.transcript import (
+from strix.interface.viewer.transcript import (
     primary_target,
     read_run_summary,
     read_vulnerabilities,

strix/interface/viewer/server.py

@@ -27,8 +27,8 @@ from typing import TYPE_CHECKING, Any
 from urllib.parse import parse_qs, unquote, urlencode, urlsplit
 
 from strix.core.paths import run_record_path
-from strix.viewer import auth
-from strix.viewer.transcript import (
+from strix.interface.viewer import auth
+from strix.interface.viewer.transcript import (
     build_run_state,
     primary_target,
     read_report_markdown,
@@ -367,7 +367,7 @@ def _make_handler(state: _ViewerState) -> type[BaseHTTPRequestHandler]:
                 self._send_json(HTTPStatus.CONFLICT, {"error": "run_not_finished"})
                 return
 
-            from strix.viewer.report_pdf import build_encrypted_report
+            from strix.interface.viewer.report_pdf import build_encrypted_report
 
             pdf_bytes, password, filename = build_encrypted_report(run_dir)
             run_name = str(summary.get("run_name") or run_dir.name)

strix/interface/viewer/static/index.html

@@ -6,8 +6,8 @@
     <meta name="viewport" content="width=device-width, initial-scale=1.0" />
     <meta name="color-scheme" content="dark" />
     <title>Strix Results</title>
-    <script type="module" crossorigin src="./assets/index-Dd1cyttN.js"></script>
-    <link rel="stylesheet" crossorigin href="./assets/index-vV8wxCG6.css">
+    <script type="module" crossorigin src="./assets/index-DzvI_0HX.js"></script>
+    <link rel="stylesheet" crossorigin href="./assets/index-C3kQ5kk8.css">
   </head>
   <body>
     <div id="root"></div>

strix/interface/viewer/transcript.py

@@ -44,8 +44,8 @@ def build_run_state(run_dir: Path) -> dict[str, Any]:
     Reuses the Textual-free ``TuiLiveView`` projection so the viewer and the TUI
     share one parser for ``agents.json`` + ``agents.db`` and never drift.
     """
-    # Imported lazily so importing strix.viewer does not eagerly pull the TUI.
-    from strix.interface.tui.live_view import TuiLiveView  # noqa: PLC0415
+    # Imported lazily so importing strix.interface.viewer does not eagerly pull the TUI.
+    from strix.interface.tui.live_view import TuiLiveView
 
     view = TuiLiveView()
     view.hydrate_from_run_dir(run_dir)

strix/llm/__init__.py

@@ -0,0 +1 @@
+"""LLM-facing context management: model-aware budgets and history compaction."""

strix/llm/compaction.py

@@ -0,0 +1,354 @@
+"""Provider-agnostic conversation compaction.
+
+When an agent's session grows past the model's usable context window, older
+turns are summarised into a single checkpoint while the most recent turns are
+kept verbatim. This runs for every LiteLLM provider (not just OpenAI), keeps a
+security-focused structured summary, and preserves tool-call/tool-result
+pairing so the trimmed history is still valid provider input.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, Any
+
+import litellm
+from litellm.exceptions import BadRequestError, ContextWindowExceededError
+
+from strix.config import load_settings
+from strix.core.sessions import replace_session_items, session_write_lock
+from strix.llm.context_budget import context_window, count_tokens, output_limit
+
+
+if TYPE_CHECKING:
+    from agents.memory import Session
+
+
+logger = logging.getLogger(__name__)
+
+_CHECKPOINT_TAG = "<conversation-checkpoint>"
+_TOOL_OUTPUT_MAX_CHARS = 2_000
+_MIN_ITEMS_TO_COMPACT = 6
+_HEAD_TRUNCATED_MARKER = "\n\n[... older conversation omitted to fit the summary request ...]\n\n"
+
+
+# Providers that don't type overflow errors (OpenRouter maps every 400 to a
+# plain BadRequestError) leave only the message to go on, so we match it the way
+# LiteLLM's own checker does — but with rate-limit exclusions first, so a
+# throttling 429 is never mistaken for an overflow and sent into compaction.
+_OVERFLOW_EXCLUSIONS = (
+    "rate limit",
+    "too many requests",
+    "throttling",
+    "service unavailable",
+    "quota",
+)
+_OVERFLOW_MARKERS = (
+    "context length",
+    "context window",
+    "context_length_exceeded",
+    "prompt is too long",
+    "input is too long",
+    "input length",
+    "maximum prompt length",
+    "reduce the length of the messages",
+    "too many tokens",
+    "token limit exceeded",
+    "request entity too large",
+)
+
+
+def is_context_overflow(exc: BaseException) -> bool:
+    """Whether ``exc`` is a model context-window-overflow error.
+
+    LiteLLM types most providers' overflow as ContextWindowExceededError, but its
+    OpenRouter branch raises a plain BadRequestError, so for that we fall back to
+    matching the provider message.
+    """
+    if isinstance(exc, ContextWindowExceededError):
+        return True
+    if isinstance(exc, BadRequestError):
+        msg = str(exc).lower()
+        if any(x in msg for x in _OVERFLOW_EXCLUSIONS):
+            return False
+        return any(x in msg for x in _OVERFLOW_MARKERS)
+    return False
+
+
+_SUMMARY_INSTRUCTIONS = """\
+You are compacting the earlier part of an autonomous security-testing agent's \
+conversation so it fits the model context window. Produce a dense, factual \
+record that lets the agent continue with no loss of important state.
+
+This is a security engagement: dropped findings mean lost vulnerabilities. Be \
+EXHAUSTIVE, not concise. Enumerate every distinct item as its own bullet — \
+never merge, deduplicate, generalise, or omit distinct findings, credentials, \
+or dead ends, even if they seem minor or repetitive. If the source mentions \
+five vulnerabilities, list five. Copy exact values verbatim: URLs, endpoints, \
+file paths, parameters, payloads, credentials, tokens, keys, hashes, cracked \
+passwords, software versions, and error messages — never paraphrase or \
+placeholder them. Do not invent anything and do not describe this compaction \
+process.
+
+Return Markdown with exactly these sections:
+
+## Objective
+The overall goal and target scope.
+
+## Vulnerabilities & Findings
+One bullet per DISTINCT vulnerability or finding (SQLi, XSS, SSRF, auth bypass, \
+misconfig, etc.). For each: type, exact location (URL/endpoint/param/file), the \
+verbatim payload or proof, confirmation status, and impact. List them all.
+
+## Credentials & Secrets
+One bullet per credential, secret, API key, token, hash, or cracked password, \
+copied verbatim with where it applies. Write "(none)" only if truly none.
+
+## System & Recon Details
+Architecture, tech stack, versions, discovered endpoints/paths/params, and \
+other weak points worth keeping.
+
+## Work State
+- Completed: what has been verified or finished.
+- Active: what is in progress right now.
+- Blocked: anything stuck and why.
+
+## Failed Attempts & Dead Ends
+One bullet per approach already tried that did not work (including WAF blocks, \
+filtered inputs, non-exploitable leads) so they are not repeated. Write \
+"(none)" only if truly none.
+
+## Next Move
+The concrete next step(s) the agent intended to take.
+
+## Relevant Files
+Files/notes/reports created or modified and their purpose."""
+
+
+def _content_text(content: Any) -> str:
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts: list[str] = []
+        for block in content:
+            if not isinstance(block, dict):
+                continue
+            text = block.get("text")
+            if isinstance(text, str):
+                parts.append(text)
+            elif block.get("type") in {"input_image", "image_url", "output_image"}:
+                parts.append("[image]")
+        return "\n".join(parts)
+    return ""
+
+
+def _truncate(text: str, limit: int) -> str:
+    return text if len(text) <= limit else f"{text[:limit]}\n[truncated]"
+
+
+def _serialize_item(item: Any) -> str:
+    if not isinstance(item, dict):
+        return str(item)
+    item_type = item.get("type")
+    role = item.get("role")
+    if item_type == "function_call":
+        args = _truncate(str(item.get("arguments", "")), _TOOL_OUTPUT_MAX_CHARS)
+        return f"[tool_call {item.get('name', '?')}] {args}"
+    if item_type == "function_call_output":
+        output = item.get("output")
+        text = output if isinstance(output, str) else _content_text(output)
+        return f"[tool_result] {_truncate(text, _TOOL_OUTPUT_MAX_CHARS)}"
+    if item_type == "reasoning":
+        return ""
+    if role or item_type == "message":
+        return f"[{role or 'assistant'}] {_content_text(item.get('content'))}".strip()
+    return ""
+
+
+def _serialize_items(items: list[Any]) -> str:
+    return "\n".join(s for s in (_serialize_item(item) for item in items) if s)
+
+
+def _is_tool_call(item: Any) -> bool:
+    return isinstance(item, dict) and item.get("type") == "function_call"
+
+
+def _is_tool_output(item: Any) -> bool:
+    return isinstance(item, dict) and item.get("type") == "function_call_output"
+
+
+def _open_calls_at(items: list[Any]) -> list[int]:
+    """Prefix count of tool calls still awaiting their result at each index;
+    a split is only safe where this is zero."""
+    balance = [0] * (len(items) + 1)
+    for i, item in enumerate(items):
+        delta = 1 if _is_tool_call(item) else -1 if _is_tool_output(item) else 0
+        balance[i + 1] = max(0, balance[i] + delta)
+    return balance
+
+
+def _select_split(model: str, items: list[Any], keep_tokens: int) -> int:
+    """Index where the kept-verbatim recent tail begins: walk newest→oldest to
+    ``keep_tokens``, then snap to a point with no tool call left open."""
+    total = 0
+    split = len(items)
+    for i in range(len(items) - 1, -1, -1):
+        total += count_tokens(model, _serialize_item(items[i]))
+        if total > keep_tokens:
+            break
+        split = i
+    open_calls = _open_calls_at(items)
+    while split > 0 and open_calls[split] != 0:
+        split -= 1
+    return split
+
+
+def _previous_summary(head: list[Any]) -> str | None:
+    for item in head:
+        if isinstance(item, dict) and item.get("role") == "user":
+            text = _content_text(item.get("content"))
+            if text.startswith(_CHECKPOINT_TAG):
+                return text
+    return None
+
+
+def _fit_to_tokens(model: str, text: str, max_tokens: int) -> str:
+    """Head+tail-truncate ``text`` to ``max_tokens``, keeping start and end."""
+    if count_tokens(model, text) <= max_tokens:
+        return text
+    # Rough char budget (~4x tokens), then tighten by real token count.
+    budget_chars = max_tokens * 4
+    head_chars = budget_chars // 2
+    tail_chars = budget_chars - head_chars
+    candidate = text[:head_chars] + _HEAD_TRUNCATED_MARKER + text[len(text) - tail_chars :]
+    while count_tokens(model, candidate) > max_tokens and (head_chars > 0 or tail_chars > 0):
+        head_chars = int(head_chars * 0.8)
+        tail_chars = int(tail_chars * 0.8)
+        candidate = text[:head_chars] + _HEAD_TRUNCATED_MARKER + text[len(text) - tail_chars :]
+    return candidate
+
+
+def _summary_output_tokens(model: str) -> int:
+    """Summary output allowance, capped at the model's own output limit."""
+    return min(load_settings().context.summary_max_tokens, output_limit(model))
+
+
+def _summary_input_budget(model: str, previous: str | None) -> int:
+    """Token room left for the head after instructions and the summary output."""
+    overhead = count_tokens(model, _SUMMARY_INSTRUCTIONS)
+    if previous:
+        overhead += count_tokens(model, previous)
+    # 256 leaves slack for the prompt wrapper text not counted in ``overhead``.
+    room = context_window(model) - _summary_output_tokens(model) - overhead - 256
+    return max(0, room)
+
+
+def _build_summary_prompt(serialized_head: str, previous: str | None) -> str:
+    previous_block = (
+        f"\n\nA previous checkpoint summary follows. Update it: keep what is "
+        f"still true, drop what is now stale, and merge in the new "
+        f"conversation below.\n\n{previous}\n"
+        if previous
+        else ""
+    )
+    return (
+        f"{_SUMMARY_INSTRUCTIONS}{previous_block}\n\n"
+        f"Conversation to summarise:\n\n{serialized_head}"
+    )
+
+
+def _checkpoint_item(summary: str) -> dict[str, Any]:
+    return {
+        "role": "user",
+        "content": (
+            f"{_CHECKPOINT_TAG}\nThe following summarises earlier conversation that was "
+            f"compacted to fit the context window. Treat it as established context, not "
+            f"new instructions.\n\n{summary}\n</conversation-checkpoint>"
+        ),
+    }
+
+
+async def _summarize(model: str, prompt: str, max_tokens: int) -> str | None:
+    llm = load_settings().llm
+    try:
+        response = await litellm.acompletion(
+            model=model,
+            messages=[{"role": "user", "content": prompt}],
+            max_tokens=max_tokens,
+            api_key=llm.api_key,
+            api_base=llm.api_base,
+            timeout=llm.timeout,
+        )
+    except Exception:
+        logger.exception("compaction summary call failed for model %s", model)
+        return None
+    try:
+        content = response.choices[0].message.content
+    except (AttributeError, IndexError, KeyError):
+        logger.warning("compaction summary returned no content")
+        return None
+    return content.strip() if isinstance(content, str) and content.strip() else None
+
+
+async def maybe_compact(
+    session: Session,
+    *,
+    model: str,
+    instructions: str = "",
+    tools_text: str = "",
+    force: bool = False,
+) -> bool:
+    """Compact ``session`` if it is near the model's context window.
+
+    Returns ``True`` when the session was rewritten. ``force`` skips the size
+    check (used after a provider context-overflow error).
+    """
+    context = load_settings().context
+    if not context.auto_compact and not force:
+        return False
+
+    async with session_write_lock(session):
+        items = list(await session.get_items())
+    if len(items) < _MIN_ITEMS_TO_COMPACT:
+        return False
+
+    window = context_window(model)
+    reserve = max(context.compact_buffer_tokens, output_limit(model))
+    budget = max(context.keep_tokens, window - reserve)
+    used = count_tokens(model, "\n".join((instructions, tools_text, _serialize_items(items))))
+    if not force and used <= budget:
+        return False
+
+    split = _select_split(model, items, context.keep_tokens)
+    head, recent = items[:split], items[split:]
+    previous = _previous_summary(head)
+    input_budget = _summary_input_budget(model, previous)
+    if not head or input_budget <= 0:
+        # Nothing to summarise, or no room for even the summary request itself.
+        if head:
+            logger.warning(
+                "skipping compaction for %s: no room to summarise within its context window", model
+            )
+        return False
+
+    serialized_head = _fit_to_tokens(model, _serialize_items(head), input_budget)
+    summary = await _summarize(
+        model,
+        _build_summary_prompt(serialized_head, previous),
+        _summary_output_tokens(model),
+    )
+    if summary is None:
+        return False
+
+    new_items = [_checkpoint_item(summary), *recent]
+    rewritten = await replace_session_items(session, new_items, expected_len=len(items))
+    if rewritten:
+        logger.info(
+            "compacted %s: %d items (~%d tok) -> %d items (summary + %d recent)",
+            model,
+            len(items),
+            used,
+            len(new_items),
+            len(recent),
+        )
+    return rewritten

strix/llm/context_budget.py

@@ -0,0 +1,76 @@
+"""Model-aware token budgets, resolved from LiteLLM model metadata with a
+large configurable fallback for models LiteLLM doesn't map.
+"""
+
+from __future__ import annotations
+
+import logging
+from functools import lru_cache
+from typing import Any
+
+import litellm
+
+from strix.config import load_settings
+
+
+logger = logging.getLogger(__name__)
+
+# LiteLLM keys models without the routing prefix users type (``openai/``,
+# ``litellm/``, ``ollama/`` ...). Strip a leading provider segment on lookup.
+_STRIPPABLE_PREFIXES = ("openai/", "litellm/", "any-llm/", "ollama/", "ollama_chat/")
+
+_DEFAULT_OUTPUT_TOKENS = 8_192
+
+
+def _lookup_key(model: str) -> str:
+    for prefix in _STRIPPABLE_PREFIXES:
+        if model.startswith(prefix):
+            return model[len(prefix) :]
+    return model
+
+
+def _safe_get_model_info(model: str) -> dict[str, Any] | None:
+    try:
+        return dict(litellm.get_model_info(model))
+    except Exception:  # noqa: BLE001 - unmapped models raise; caller falls back.
+        return None
+
+
+@lru_cache(maxsize=128)
+def _model_info(model: str) -> dict[str, int]:
+    for candidate in (model, _lookup_key(model)):
+        info = _safe_get_model_info(candidate)
+        if info is not None:
+            return {
+                "max_input_tokens": int(
+                    info.get("max_input_tokens") or info.get("max_tokens") or 0
+                ),
+                "max_output_tokens": int(info.get("max_output_tokens") or 0),
+            }
+    logger.debug("No LiteLLM model info for %r; using configured fallbacks", model)
+    return {"max_input_tokens": 0, "max_output_tokens": 0}
+
+
+def context_window(model: str) -> int:
+    """Input token capacity for ``model`` (configured fallback when unmapped)."""
+    resolved = _model_info(model)["max_input_tokens"]
+    return resolved or load_settings().context.fallback_context_tokens
+
+
+def output_limit(model: str) -> int:
+    """Max output tokens for ``model`` (a conservative default when unmapped)."""
+    return _model_info(model)["max_output_tokens"] or _DEFAULT_OUTPUT_TOKENS
+
+
+def count_tokens(model: str, text: str) -> int:
+    """Token count for ``text`` under ``model``.
+
+    Falls back to UTF-8 byte length (a guaranteed upper bound) when LiteLLM
+    can't count, so budget checks stay conservative.
+    """
+    if not text:
+        return 0
+    try:
+        return int(litellm.token_counter(model=_lookup_key(model), text=text))
+    except Exception:  # noqa: BLE001 - tokenizer may be unavailable for some models.
+        return len(text.encode("utf-8"))

strix/report/writer.py

@@ -10,7 +10,7 @@ import re
 import tempfile
 from datetime import UTC, datetime
 from pathlib import Path
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, cast
 
 from pygments.lexers import PythonLexer, get_lexer_by_name, guess_lexer
 from pygments.lexers.special import TextLexer
@@ -74,10 +74,10 @@ def resolve_lexer(language: str | None, code: str) -> Lexer:
     try:
         lexer = guess_lexer(code)
     except ClassNotFound:
-        return PythonLexer()
+        return cast("Lexer", PythonLexer())
     # ``guess_lexer`` returns the plain-text lexer when it can't detect anything.
     if isinstance(lexer, TextLexer):
-        return PythonLexer()
+        return cast("Lexer", PythonLexer())
     return lexer
 
 

strix/skills/reconnaissance/asset_discovery.md

@@ -54,18 +54,17 @@ CT logs record nearly every publicly-trusted certificate. Query by domain (match
 
 ## Recommended Tooling
 
-Prefer the projectdiscovery suite (already available in the sandbox and pipeline-friendly with JSON output):
+These tools are available in the sandbox and are pipeline-friendly with JSON output:
 
 - **`subfinder`** — passive subdomain aggregation across many sources incl. CT: `subfinder -d example.com -all -recursive -silent -oJ -o subs.jsonl`
-- **`tlsx`** — TLS/cert data at scale; grab SANs and issuer/org to pivot: `tlsx -l hosts.txt -san -cn -tls-version -json -o tls.jsonl`
-- **`uncover`** — query Shodan/Censys/Fofa/Quake/crt.sh engines from one CLI: `uncover -q 'ssl:"Example Inc"' -e shodan,censys,fofa -json`
-- **`asnmap`** — org/domain/ASN → CIDR ranges: `asnmap -d example.com -json` / `asnmap -org "Example Inc"`
-- **`mapcidr`** — expand/aggregate CIDRs into host lists for probing: `mapcidr -cidr 192.0.2.0/24 -o hosts.txt`
-- **`dnsx`** — fast resolution, PTR, and wildcard filtering: `dnsx -l names.txt -a -aaaa -cname -ptr -resp -json -o dns.jsonl`
-- **`httpx`** — live probing + cert grab in one pass (see methodology).
+- **`httpx`** — live probing plus cert/SAN grab in one pass: `httpx -l hosts.txt -tls-grab -json` (see methodology).
 - **`naabu`** — port sweep for non-HTTP services: `naabu -list hosts.txt -top-ports 100 -verify -silent`
+- **`curl` + `jq`** — direct **crt.sh** JSON queries for CT (no key needed) and other index APIs.
+- **`openssl s_client`** — active read of a live host's cert to extract SANs/CN.
+- **`dig`** / **`nslookup`** — forward/reverse (PTR) resolution and CNAME chains.
+- **`whois`** — ASN/netblock lookups (e.g. `whois -h whois.cymru.com`).
 
-Also useful: **`amass`** (`amass intel`/`enum` for ASN, cert, and passive sources), **`cero`** (bulk SAN extraction from IPs/ranges), and direct **crt.sh** JSON queries when no keys are configured. Cross-source results — CT + passive DNS + `subfinder` together beat any single source.
+Cross-source results — CT + passive DNS + `subfinder` together beat any single source. If you need a tool that is not installed, install it into the sandbox at runtime.
 
 ## Key Techniques
 
@@ -75,7 +74,7 @@ Every new name, PTR result, CNAME target, and cert SAN becomes a fresh seed. Loo
 
 ### Cert-Fingerprint Pivoting
 
-Search Censys/Shodan (or `uncover`) by a cert's `fingerprint_sha256` to find every other host presenting the same certificate — the strongest cross-asset link for tying acquisitions and shadow infra to the target.
+Search Censys/Shodan by a cert's `fingerprint_sha256` to find every other host presenting the same certificate — the strongest cross-asset link for tying acquisitions and shadow infra to the target.
 
 ### Naming-Convention Inference
 
@@ -83,11 +82,11 @@ Wildcard SANs and observed hostnames expose the org's naming scheme; generate ta
 
 ### IP-First Discovery
 
-For ASN-owned ranges, sweep IPs directly with `naabu`/`httpx` and read served certs (`tlsx`) to find services that have no DNS name at all.
+For ASN-owned ranges, sweep IPs directly with `naabu`/`httpx` and read served certs (`httpx -tls-grab`, or `openssl s_client`) to find services that have no DNS name at all.
 
 ## Advanced Techniques
 
-- **Active SAN harvesting** across whole ranges with `tlsx`/`cero` recovers internal hostnames never logged to public CT.
+- **Active SAN harvesting** across whole ranges with `httpx -tls-grab` (or `openssl s_client`) recovers internal hostnames never logged to public CT.
 - **Favicon and response hashing** (`httpx -favicon`, hash pivots in Shodan) clusters instances of the same app across unrelated hostnames.
 - **Vhost differentials**: probe a single IP with multiple `Host:` values to unmask co-located apps behind one address.
 - **Historical CT/DNS diffing** highlights recently issued certs and newly appearing hosts — high-signal for fresh or misconfigured deployments.
@@ -108,11 +107,11 @@ For ASN-owned ranges, sweep IPs directly with `naabu`/`httpx` and read served ce
 ## Testing Methodology
 
 1. **Seed** - domains, org/legal names, known IPs, email domains, code-host org
-2. **Certificate transparency** - pull all logged certs per seed domain and org name (crt.sh, `uncover`)
-3. **SAN/CN extraction** - parse every Subject CN and SAN with `tlsx`; each new name is a new seed
-4. **Passive DNS** - resolve forward and reverse with `dnsx`; harvest historical records
-5. **ASN/IP mapping** - `asnmap` → `mapcidr` to expand owned ranges, then sweep for live hosts
-6. **Active TLS pivot** - `tlsx`/`cero` on live IPs/ports to grab SANs missing from public CT
+2. **Certificate transparency** - pull all logged certs per seed domain and org name (crt.sh, Censys/Shodan)
+3. **SAN/CN extraction** - parse every Subject CN and SAN with `httpx -tls-grab` (or `openssl s_client`); each new name is a new seed
+4. **Passive DNS** - resolve forward and reverse with `dig`; harvest historical records
+5. **ASN/IP mapping** - `whois` the netblock/ASN to expand owned ranges, then sweep for live hosts
+6. **Active TLS pivot** - `httpx -tls-grab` on live IPs/ports to grab SANs missing from public CT
 7. **Consolidate & probe** - dedupe, `httpx` probe, classify, and route to specialists
 
 ## Validation
@@ -139,13 +138,13 @@ For ASN-owned ranges, sweep IPs directly with `naabu`/`httpx` and read served ce
 ## Pro Tips
 
 1. Loop the pipeline — every SAN, PTR, and CNAME target is a new seed until the set converges.
-2. crt.sh is the cheapest high-yield source (no key); Censys/Shodan via `uncover` add cert-fingerprint and vhost pivoting when keys exist.
-3. Always cert-grab live hosts with `tlsx` — active SANs catch internal hostnames never sent to public CT.
+2. crt.sh is the cheapest high-yield source (no key); Censys/Shodan add cert-fingerprint and vhost pivoting when keys exist.
+3. Always cert-grab live hosts with `httpx -tls-grab` (or `openssl s_client`) — active SANs catch internal hostnames never sent to public CT.
 4. Internal-looking SANs (`*.internal`, `*.svc.cluster.local`, staging names) are the highest-signal leads.
 5. Wildcard SANs reveal naming conventions — seed targeted guesses instead of blind brute force.
 6. Cluster by function, not product name, so the workflow generalizes to any exposed service.
-7. Keep JSON output throughout so stages chain cleanly (`subfinder` → `dnsx` → `httpx` → `naabu`).
+7. Keep JSON output throughout so stages chain cleanly (`subfinder` → `dig` → `httpx` → `naabu`).
 
 ## Summary
 
-Broad passive discovery — CT + TLS SAN pivoting + passive DNS + ASN/IP mapping, looped until convergence — finds the assets brute force misses, especially internal-named and forgotten services leaked through certificates. Build the inventory with the projectdiscovery suite, probe and classify it generically, then route each interesting asset to the specialist skill for its class.
+Broad passive discovery — CT + TLS SAN pivoting + passive DNS + ASN/IP mapping, looped until convergence — finds the assets brute force misses, especially internal-named and forgotten services leaked through certificates. Build the inventory with `subfinder`, `httpx`, `naabu`, and CT/DNS/cert queries, probe and classify it generically, then route each interesting asset to the specialist skill for its class.