docs(prompts): text-only turns no longer end an autonomous run

One tool was doing three jobs (wait on the user, wait on other agents, and
- wrongly - wait for a long-running command), so the driver had to guess which
one an agent meant and used parent_id as the proxy: the root waits for a human,
everyone else waits for agents. That proxy is wrong, since the user can message
any agent from the TUI's agent tree.

Tool identity now carries the intent, and the coordinator records it as a
wait_kind that survives snapshot/restore:

  respond_to_user  -> wait_kind="user",   never auto-resumed (root or not)
  wait_for_agents  -> wait_kind="agents", auto-resumed on a 300s timer
  recovery exhaust -> wait_kind="stalled"

respond_to_user fuses the message and the yield into one call, so there is no
way to answer and then forget to stop - the two-step that gpt-4o-mini skipped
2/2 in live testing. Plain text still renders as before.

Auto-resume is also bounded now: an agent that re-parks after every timeout
burned a model turn every 300s for the rest of the scan (and, since parked
children notify their parent, spammed the parent's inbox on the same cycle).
After _MAX_IDLE_AUTO_RESUMES it stays parked until a real message arrives.

.github/workflows/build-release.yml

@@ -6,6 +6,9 @@ on:
       - 'v*'
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   build:
     strategy:
@@ -18,19 +21,23 @@ jobs:
             target: macos-x86_64
           - os: ubuntu-22.04
             target: linux-x86_64
+          - os: ubuntu-22.04-arm
+            target: linux-arm64
           - os: windows-latest
             target: windows-x86_64
 
     runs-on: ${{ matrix.os }}
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262  # v4.4.0
+        with:
+          persist-credentials: false
 
-      - uses: actions/setup-python@v5
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065  # v5.6.0
         with:
           python-version: '3.12'
 
-      - uses: astral-sh/setup-uv@v5
+      - uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86  # v5.4.2
 
       - name: Build
         shell: bash
@@ -38,6 +45,20 @@ jobs:
           uv sync --frozen
           uv run pyinstaller strix.spec --noconfirm
 
+          if [[ "${{ runner.os }}" == "Windows" ]]; then
+            dist/strix.exe --version
+          else
+            dist/strix --version
+          fi
+
+          if [[ "${{ matrix.target }}" == "linux-arm64" ]]; then
+            file dist/strix
+            file dist/strix | grep -q "ARM aarch64" || {
+              echo "::error::linux-arm64 artifact is not an ARM aarch64 binary"
+              exit 1
+            }
+          fi
+
           VERSION=$(grep '^version' pyproject.toml | head -1 | sed 's/.*"\(.*\)"/\1/')
           mkdir -p dist/release
 
@@ -50,7 +71,7 @@ jobs:
             tar -C dist/release -czvf "dist/release/strix-${VERSION}-${{ matrix.target }}.tar.gz" "strix-${VERSION}-${{ matrix.target }}"
           fi
 
-      - uses: actions/upload-artifact@v4
+      - uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02  # v4.6.2
         with:
           name: strix-${{ matrix.target }}
           path: |
@@ -65,13 +86,13 @@ jobs:
       contents: write
 
     steps:
-      - uses: actions/download-artifact@v4
+      - uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093  # v4.3.0
         with:
           path: release
           merge-multiple: true
 
       - name: Create Release
-        uses: softprops/action-gh-release@v2
+        uses: softprops/action-gh-release@3bb12739c298aeb8a4eeaf626c5b8d85266b0e65  # v2.6.2
         with:
           prerelease: ${{ !startsWith(github.ref, 'refs/tags/') }}
           generate_release_notes: true

.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!"

README.md

@@ -267,6 +267,20 @@ export STRIX_REASONING_EFFORT="high"  # control thinking effort (default: high,
 > [!NOTE]
 > Strix automatically saves your configuration to `~/.strix/cli-config.json`, so you don't have to re-enter it on every run.
 
+#### Sign in with a ChatGPT subscription
+
+Instead of a metered API key, you can run Strix on your ChatGPT Plus/Pro subscription:
+
+```bash
+strix auth login chatgpt      # sign in with your ChatGPT account
+
+export STRIX_LLM="chatgpt/gpt-5.4"   # chatgpt/<model> runs on the subscription
+strix --target ./app-directory
+
+strix auth status             # show the active sign-in
+strix auth logout             # forget the sign-in
+```
+
 **Recommended models for best results:**
 
 - [OpenAI GPT-5.4](https://openai.com/api/) - `openai/gpt-5.4`

docs/advanced/configuration.mdx

@@ -19,6 +19,14 @@ Configure Strix using environment variables or a config file.
   Custom API base URL. Also accepts `OPENAI_API_BASE`, `LITELLM_BASE_URL`, or `OLLAMA_API_BASE`.
 </ParamField>
 
+<ParamField path="LLM_EXTRA_HEADERS" type="string">
+  Extra HTTP headers sent on every LLM request, as a JSON object (e.g.
+  `{"X-Feature-Key":"value","X-Tenant":"acme"}`). Useful for OpenAI-compatible
+  gateways that require attribution or routing headers in addition to the bearer
+  token. The bearer token itself still comes from `LLM_API_KEY`. Applies to both
+  the LiteLLM and native OpenAI routing paths.
+</ParamField>
+
 <ParamField path="LLM_TIMEOUT" default="300" type="integer">
   Request timeout in seconds for LLM calls.
 </ParamField>
@@ -35,6 +43,37 @@ Configure Strix using environment variables or a config file.
   Timeout in seconds for memory compression operations (context summarization).
 </ParamField>
 
+### Dedicated deduplication model
+
+Finding deduplication is a cheap, structured classification task. By default it
+runs on the main model, but you can route it to a smaller/cheaper model without
+affecting the agents that do the actual testing.
+
+<ParamField path="STRIX_DEDUPE_MODEL" type="string">
+  Model used to judge whether a candidate finding duplicates an existing report.
+  Falls back to `STRIX_LLM` when unset.
+</ParamField>
+
+<ParamField path="DEDUPE_LLM_API_KEY" type="string">
+  Optional provider key for the deduplication model.
+</ParamField>
+
+<ParamField path="DEDUPE_LLM_API_BASE" type="string">
+  Optional custom API base URL for the deduplication model. Use when the dedupe
+  model runs on a different endpoint than the main model.
+</ParamField>
+
+<ParamField path="DEDUPE_LLM_EXTRA_HEADERS" type="string">
+  Optional JSON object of extra HTTP headers sent on every deduplication-model
+  request, e.g. `{"X-Feature-Key":"value"}`. A dedicated dedupe model never
+  inherits `LLM_EXTRA_HEADERS`; set this when its endpoint needs custom headers.
+</ParamField>
+
+<ParamField path="STRIX_DEDUPE_REASONING_EFFORT" type="string">
+  Reasoning effort for the deduplication model. Defaults to the model's own
+  baseline when unset.
+</ParamField>
+
 ## Optional Features
 
 <ParamField path="PERPLEXITY_API_KEY" type="string">

docs/llm-providers/local.mdx

@@ -54,3 +54,20 @@ If you use LM Studio, vLLM, or other runners:
 export STRIX_LLM="openai/local-model"
 export LLM_API_BASE="http://localhost:1234/v1"  # Adjust port as needed
 ```
+
+### Gateways that require custom headers
+
+Some OpenAI-compatible gateways require extra HTTP headers (for attribution or
+tenant routing) alongside the bearer token. Set them with `LLM_EXTRA_HEADERS` as
+a JSON object — they are sent on every request:
+
+```bash
+export STRIX_LLM="openai/your-model"
+export LLM_API_BASE="https://your-gateway.example/v1"
+export LLM_API_KEY="your-bearer-token"          # sent as Authorization: Bearer ...
+export LLM_EXTRA_HEADERS='{"X-Feature-Key":"value","X-Tenant":"acme"}'
+```
+
+For endpoints behind a private CA, point Strix at your certificate bundle with
+the standard `SSL_CERT_FILE=/path/to/ca-bundle.pem` — never disable TLS
+verification against a real endpoint.

docs/usage/cli.mdx

@@ -61,11 +61,28 @@ strix (--target <target> | --target-list <path> | --mount <path>) [options]
   Path to a custom config file (JSON) to use instead of `~/.strix/cli-config.json`.
 </ParamField>
 
-<ParamField path="--max-budget-usd" type="number">
+<ParamField path="--max-budget" type="number">
   Maximum LLM spend in USD for the whole scan, counted cumulatively across the
   root agent and every child agent. The budget is checked after each model
-  response; once the running cost reaches the threshold, the scan stops cleanly
-  with a `stopped` status (not a failure) and the sandbox is torn down.
+  response.
+
+  In non-interactive mode (`-n`), once the running cost reaches the threshold,
+  the scan stops cleanly with a `stopped` status (not a failure) and the sandbox
+  is torn down. Sub-agents are stopped early, at 90% of the budget, reserving
+  the final slice for the root agent to wind down and produce the final report.
+
+  In interactive mode, reaching the budget pauses the scan instead of ending
+  it: every agent parks, and sending any message resumes the scan with the cap
+  extended by the original budget amount. There is no sub-agent reserve in
+  interactive mode.
+
+  As the budget is approached, graduated wrap-up warnings are surfaced to
+  **every** agent so they can finish their work and call their lifecycle tool
+  before the hard stop. The bands sit just below each role's own stop point: the
+  root is warned at **70%, 85% and 95%** (it stops at 100%), while sub-agents are
+  warned at **75%, 80% and 85%** (they stop at the 90% reserve). In interactive
+  mode every agent uses the **70%, 85% and 95%** bands. Percentages shown in the
+  warnings are the real cumulative spend against the full budget.
 
   Must be greater than `0`. Omit the flag for no limit.
 
@@ -84,6 +101,19 @@ strix (--target <target> | --target-list <path> | --mount <path>) [options]
     counts.
 </ParamField>
 
+<ParamField path="--max-turns" type="integer" default="500">
+  Maximum number of turns (one model response plus its tool round) allotted to
+  **each** agent, applied per run. When an agent reaches this limit it is
+  force-stopped.
+
+  As the limit is approached, graduated wrap-up warnings (at 70%, 85% and 95%)
+  are injected into that agent's next model turn so it can prioritise its
+  remaining work and call its lifecycle tool (`finish_scan` for the root agent,
+  `agent_finish` for sub-agents) before the hard stop.
+
+  Must be greater than `0`.
+</ParamField>
+
 ## Examples
 
 ```bash
@@ -99,6 +129,9 @@ strix --target api.example.com --instruction "Focus on IDOR and auth bypass"
 # CI/CD mode
 strix -n --target ./ --scan-mode quick
 
+# Cap cost and per-agent turns
+strix --target https://example.com --max-budget 25 --max-turns 300
+
 # Force diff-scope against a specific base ref
 strix -n --target ./ --scan-mode quick --scope-mode diff --diff-base origin/main
 
@@ -116,5 +149,6 @@ strix --mount ./huge-monorepo
 
 | Code | Meaning |
 |------|---------|
-| 0 | Scan completed, no vulnerabilities found |
+| 0 | Scan completed successfully (interactive mode always exits `0`; in headless mode, `0` means no vulnerabilities were found) |
+| 1 | A fatal error occurred before or during the scan (e.g. missing environment variables, Docker unavailable, invalid config file, diff-scope resolution failure, or an unhandled error) |
 | 2 | Vulnerabilities found (headless mode only) |

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "strix-agent"
-version = "1.3.0"
+version = "1.4.1"
 description = "Open-source AI Hackers for your apps"
 readme = "README.md"
 license = "Apache-2.0"
@@ -46,7 +46,9 @@ dependencies = [
   "caido-sdk-client>=0.2.0",
   "reportlab>=4.0",
   "pypdf>=5.0",
-  "cryptography>=42",
+  # Cap <49: 49.x drops the universal2 macOS wheel (arm64-only), which breaks
+  # the Intel macOS (macos-x86_64) release build's `uv sync --frozen`.
+  "cryptography>=48.0.1,<49",
 ]
 
 [project.optional-dependencies]
@@ -77,10 +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
@@ -120,6 +122,7 @@ module = [
     "pydantic_settings.*",
     "reportlab.*",
     "pypdf.*",
+    "pygments.*",
 ]
 ignore_missing_imports = true
 disable_error_code = ["import-untyped"]
@@ -213,12 +216,17 @@ ignore = [
 # Test doubles use fixture tokens/passwords and match a callee signature whose
 # args they intentionally ignore.
 "tests/test_viewer_auth.py" = ["S105", "S106", "ARG001"]
+"tests/test_codex_auth.py" = ["S105", "S106", "SLF001"]
+# Stdlib HTTP handler overrides (do_GET/do_POST).
+"strix/interface/auth_cli.py" = ["N802"]
+"tests/test_codex_streaming.py" = ["N802"]
+"tests/test_disable_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"]
@@ -251,9 +259,16 @@ ignore = [
 "strix/core/runner.py" = ["TC003", "PLR0912", "PLR0915", "PLC0415"]
 # ReportState carries scan artifact/report fields and
 # a runtime ``Callable`` annotation on ``vulnerability_found_callback``.
-"strix/report/state.py" = ["TC003", "PLR0912", "PLR0915", "E501", "PERF401"]
+"strix/report/state.py" = ["TC003", "PLR0912", "PLR0915", "E501", "PERF401", "PLC0415"]
 "strix/report/usage.py" = ["PLC0415"]
+# Lazy import of strix.config.models avoids a circular dependency between the
+# report pipeline and the config layer.
+"strix/report/dedupe.py" = ["PLC0415"]
+"strix/telemetry/logging.py" = ["PLC0415"]
 "strix/config/models.py" = ["PLC0415"]
+# Heavy inference deps (httpx, openai) imported lazily so auth-status checks
+# don't pull them in.
+"strix/config/codex.py" = ["PLC0415"]
 # Interface utility branches per scope-mode / target-type combination;
 # splitting would obscure the decision tree without simplifying it.
 "strix/interface/utils.py" = ["PLR0912", "BLE001", "PLC0415"]

scripts/install.sh

@@ -4,7 +4,7 @@ set -euo pipefail
 
 APP=strix
 REPO="usestrix/strix"
-STRIX_IMAGE="ghcr.io/usestrix/strix-sandbox:1.0.0"
+STRIX_IMAGE="ghcr.io/usestrix/strix-sandbox:1.1.0"
 
 MUTED='\033[0;2m'
 RED='\033[0;31m'
@@ -41,7 +41,7 @@ fi
 
 combo="$os-$arch"
 case "$combo" in
-  linux-x86_64|macos-x86_64|macos-arm64|windows-x86_64)
+  linux-x86_64|linux-arm64|macos-x86_64|macos-arm64|windows-x86_64)
     ;;
   *)
     echo -e "${RED}Unsupported OS/Arch: $os/$arch${NC}"

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

@@ -16,13 +16,14 @@ 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,
     send_message_to_agent,
     stop_agent,
     view_agent_graph,
-    wait_for_message,
+    wait_for_agents,
 )
 from strix.tools.finish.tool import finish_scan
 from strix.tools.load_skill.tool import load_skill
@@ -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 +43,13 @@ 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.respond.tool import respond_to_user
 from strix.tools.thinking.tool import think
 from strix.tools.todo.tools import (
     create_todo,
@@ -103,8 +111,36 @@ 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 +148,7 @@ 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 +163,7 @@ 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 +195,35 @@ 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 +264,16 @@ 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 +282,10 @@ 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 +311,10 @@ 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)
@@ -273,6 +346,10 @@ def _make_shell_configurator(*, chat_completions: bool) -> Any:
     return configure
 
 
+# Tools that hand control away by parking the agent rather than ending the scan.
+_PARKING_TOOLS: frozenset[str] = frozenset({"respond_to_user", "wait_for_agents"})
+
+
 def _lifecycle_tool_completed(tool_name: str, output: Any) -> bool:
     if tool_name == "agent_finish":
         completion_key = "agent_completed"
@@ -291,7 +368,7 @@ def _lifecycle_tool_completed(tool_name: str, output: Any) -> bool:
 
 
 def _wait_tool_parked(tool_name: str, output: Any) -> bool:
-    if tool_name != "wait_for_message" or not isinstance(output, str):
+    if tool_name not in _PARKING_TOOLS or not isinstance(output, str):
         return False
     try:
         parsed = json.loads(output)
@@ -343,6 +420,8 @@ _BASE_TOOLS: tuple[Tool, ...] = (
     web_search,
     create_vulnerability_report,
     create_dependency_report,
+    list_reports,
+    get_report,
     list_requests,
     view_request,
     repeat_request,
@@ -351,7 +430,7 @@ _BASE_TOOLS: tuple[Tool, ...] = (
     scope_rules,
     view_agent_graph,
     send_message_to_agent,
-    wait_for_message,
+    wait_for_agents,
     create_agent,
     stop_agent,
 )
@@ -435,11 +514,17 @@ def build_strix_agent(
         )
 
     agent_tools = [*_EXTRA_TOOLS, *(extra_tools or [])]
+    if interactive:
+        # Yielding to the user is only meaningful when one is attached.
+        agent_tools.append(respond_to_user)
     if is_root:
         tools: list[Tool] = [*_BASE_TOOLS, *agent_tools, finish_scan]
     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 +544,8 @@ 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

@@ -31,28 +31,26 @@ INTER-AGENT MESSAGES:
 
 {% if interactive %}
 INTERACTIVE BEHAVIOR:
-- You are in an interactive conversation with a user
-- CRITICAL: A message WITHOUT a tool call IMMEDIATELY STOPS your entire execution and waits for user input. This is a HARD SYSTEM CONSTRAINT, not a suggestion.
-  - Statements like "Planning the assessment..." or "I'll now scan..." or "Starting with..." WITHOUT a tool call will HALT YOUR WORK COMPLETELY. The system interprets no-tool-call as "I'm done, waiting for the user."
-  - If you want to plan, call the think tool. If you want to act, call the appropriate tool. There is NO valid reason to output text without a tool call while working on a task.
-  - The ONLY time you may send a message without a tool call is when you are genuinely DONE and presenting final results, or when you NEED the user to answer a question before continuing.
-- EVERY message while working MUST contain exactly one tool call — this is what keeps execution moving. No tool call = execution stops.
-- You may include brief explanatory text BEFORE the tool call
-- Respond naturally when the user asks questions or gives instructions
-- For simple conversation, acknowledgements, or direct questions that you can answer from current context, reply in plain text and stop. Do NOT call think just to prepare wording.
-- If you use a tool to answer a user question (for example list_todos, view_agent_graph, or a file read), then after the tool result arrives, provide the answer in plain text and stop unless the user explicitly asked you to continue working.
-- Never loop through think or other tools just to prepare, polish, confirm, or announce a final answer. Once you know the answer, say it.
-- NEVER send empty messages — if you have nothing to do or say, call the wait_for_message tool
-- If you catch yourself about to describe multiple steps without a tool call, STOP and call the think tool instead
+- You are in an interactive conversation with a user.
+- HOW EXECUTION ENDS: your turn ends ONLY when you make an explicit lifecycle tool call. Plain text NEVER ends your turn and NEVER hands control to the user — text is shown to the user, and then execution continues.
+  - To answer the user and hand control back, call respond_to_user. It delivers your message AND parks you for their reply in one call, so there is no way to answer and then forget to stop. This is the ONLY way to yield to the user.
+  - To wait on another AGENT (a child's report, a peer's reply), call wait_for_agents. That is not a way to reach the user.
+  - To end the whole engagement, call the lifecycle tool: finish_scan (root) or agent_finish (subagent).
+  - A turn that ends with plain text and no tool call does NOT stop you: the system nudges you to continue and will re-run you. Do not rely on going silent to pause — it will not pause you.
+- Answering a user question: put the answer in respond_to_user's message. Do not write the answer as plain text and then fall silent — that does not reach a stopping point, it just triggers a continuation nudge.
+- You may include brief explanatory text before a tool call, and you can narrate while you work — plain text is shown to the user as you go. Narrating is free; respond_to_user is specifically the act of WAITING for the user, so do not call it just to give a status update.
+- Respond naturally when the user asks questions or gives instructions.
+- While actively working on a task, every turn should carry exactly one tool call — use think to plan, the appropriate tool to act, and respond_to_user only when you genuinely need the user.
+- Never loop through think or other tools just to prepare, polish, confirm, or announce an answer. Once you know the answer, send it with respond_to_user.
 {% else %}
 AUTONOMOUS BEHAVIOR:
 - Work autonomously by default
 - You should NOT ask for user input or confirmation - you should always proceed with your task autonomously.
 - Minimize user messaging: avoid redundancy and repetition; consolidate updates into a single concise message
-- NEVER send an empty or blank message. If you have no content to output or need to wait (for user input, subagent results, or any other reason), you MUST call the wait_for_message tool (or another appropriate tool) instead of emitting an empty response.
-- If there is nothing to execute and no user query to answer any more: do NOT send filler/repetitive text — either call wait_for_message or finish your work (subagents: agent_finish; root: finish_scan)
-- While the agent loop is running, almost every output MUST be a tool call. Do NOT send plain text messages; act via tools. If idle, use wait_for_message; when done, use agent_finish (subagents) or finish_scan (root)
-- A text-only turn — even one — IMMEDIATELY ends the scan/run with no report written. The lifecycle tools (``finish_scan`` for root, ``agent_finish`` for subagents) are the ONLY valid way to terminate. If you find yourself wanting to say "Done!" or "Scan complete" without a tool call, call the lifecycle tool instead — the report and termination signal both flow through it.
+- NEVER send an empty or blank message. If you have no content to output or need to wait for subagent results, you MUST call the wait_for_agents tool (or another appropriate tool) instead of emitting an empty response.
+- There is no user attached to this run, so there is nobody to ask and nothing to yield to. If there is nothing left to execute: do NOT send filler/repetitive text — either call wait_for_agents (only if you are genuinely expecting another agent to message you) or finish your work (subagents: agent_finish; root: finish_scan)
+- While the agent loop is running, almost every output MUST be a tool call. Do NOT send plain text messages; act via tools. If waiting on another agent, use wait_for_agents; when done, use agent_finish (subagents) or finish_scan (root)
+- A text-only turn does nothing: it neither ends the run nor yields — it just wastes a turn and forces a retry. The lifecycle tools (``finish_scan`` for root, ``agent_finish`` for subagents) are the ONLY way to terminate, and the report flows through them. If you find yourself wanting to say "Done!" or "Scan complete" without a tool call, call the lifecycle tool instead.
 {% endif %}
 </communication_rules>
 
@@ -188,7 +186,7 @@ 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 +213,7 @@ 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 +17,8 @@ from strix.config.loader import (
     persist_current,
 )
 from strix.config.settings import (
+    ContextSettings,
+    DedupeSettings,
     IntegrationSettings,
     LlmSettings,
     RuntimeSettings,
@@ -26,6 +28,8 @@ from strix.config.settings import (
 
 
 __all__ = [
+    "ContextSettings",
+    "DedupeSettings",
     "IntegrationSettings",
     "LlmSettings",
     "RuntimeSettings",

strix/config/codex.py

@@ -0,0 +1,404 @@
+"""ChatGPT (Codex) subscription auth: OAuth login, token refresh, and the OpenAI
+client that routes inference through the ChatGPT backend.
+
+Mirrors OpenAI's Codex CLI: OAuth 2.0 + PKCE against ``auth.openai.com``, with the
+access token sent as a ``Bearer`` token to ``chatgpt.com/backend-api/codex``. Using
+a ChatGPT subscription outside OpenAI's own products is not officially supported by
+OpenAI; the user chooses this path knowingly. The OAuth constants are OpenAI's own
+Codex CLI values (the backend only accepts that client).
+"""
+
+from __future__ import annotations
+
+import base64
+import contextlib
+import hashlib
+import json
+import logging
+import secrets
+import threading
+import time
+import urllib.parse
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+import requests
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+
+    from openai import AsyncOpenAI
+
+
+logger = logging.getLogger(__name__)
+
+
+PROVIDER = "codex"
+
+CLIENT_ID = "app_EMoamEEZ73f0CkXaXp7hrann"
+AUTHORIZE_URL = "https://auth.openai.com/oauth/authorize"
+TOKEN_URL = "https://auth.openai.com/oauth/token"  # noqa: S105  # nosec B105 - URL, not a secret
+CALLBACK_HOST = "localhost"
+CALLBACK_PORT = 1455
+CALLBACK_PATH = "/auth/callback"
+REDIRECT_URI = f"http://{CALLBACK_HOST}:{CALLBACK_PORT}{CALLBACK_PATH}"
+SCOPE = "openid profile email offline_access"
+
+CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex"
+ORIGINATOR = "codex_cli_rs"
+_ACCOUNT_CLAIM = "https://api.openai.com/auth"
+
+_TOKEN_TIMEOUT = 30
+_EXPIRY_SKEW_S = 300
+
+_refresh_lock = threading.Lock()
+
+# Kept separate from cli-config.json so OAuth tokens never land in the env-var config.
+AUTH_PATH = Path.home() / ".strix" / "subscription-auth.json"
+
+
+def _read_store() -> dict[str, Any]:
+    try:
+        data = json.loads(AUTH_PATH.read_text(encoding="utf-8"))
+    except (OSError, json.JSONDecodeError):
+        return {}
+    return data if isinstance(data, dict) else {}
+
+
+def _write_store(data: dict[str, Any]) -> None:
+    AUTH_PATH.parent.mkdir(parents=True, exist_ok=True)
+    tmp = AUTH_PATH.with_suffix(".json.tmp")
+    tmp.write_text(json.dumps(data, indent=2), encoding="utf-8")
+    with contextlib.suppress(OSError):
+        tmp.chmod(0o600)
+    tmp.replace(AUTH_PATH)
+    with contextlib.suppress(OSError):
+        AUTH_PATH.chmod(0o600)
+
+
+def read_record() -> dict[str, Any] | None:
+    record = _read_store().get(PROVIDER)
+    if not isinstance(record, dict) or record.get("type") != "oauth":
+        return None
+    if not (record.get("access") and record.get("refresh") and record.get("account_id")):
+        return None
+    return record
+
+
+def is_authenticated() -> bool:
+    return read_record() is not None
+
+
+def save_record(record: dict[str, Any]) -> None:
+    data = _read_store()
+    data[PROVIDER] = record
+    _write_store(data)
+
+
+def logout() -> None:
+    data = _read_store()
+    if PROVIDER not in data:
+        return
+    del data[PROVIDER]
+    if data:
+        _write_store(data)
+        return
+    with contextlib.suppress(OSError):
+        AUTH_PATH.unlink()
+
+
+@contextlib.contextmanager
+def _refresh_guard() -> Iterator[None]:
+    """Serialize token refresh within (lock) and across (flock) Strix processes,
+    so concurrent runs can't both spend the single-use refresh token."""
+    with _refresh_lock:
+        try:
+            import fcntl
+
+            lock_path = AUTH_PATH.with_suffix(".lock")
+            lock_path.parent.mkdir(parents=True, exist_ok=True)
+            handle = lock_path.open("w")
+        except (ImportError, OSError):
+            yield
+            return
+        try:
+            with contextlib.suppress(OSError):
+                fcntl.flock(handle.fileno(), fcntl.LOCK_EX)
+            yield
+        finally:
+            with contextlib.suppress(OSError):
+                fcntl.flock(handle.fileno(), fcntl.LOCK_UN)
+            handle.close()
+
+
+class CodexAuthError(Exception):
+    def __init__(self, code: str, message: str | None = None) -> None:
+        self.code = code
+        super().__init__(message or code)
+
+
+class CodexContentGuardrailError(Exception):
+    """The ChatGPT backend refused a request via its content guardrail.
+    Terminal — retrying identical content never clears the block."""
+
+    def __init__(self, model: str, original: BaseException | None = None) -> None:
+        self.model = model
+        self.original = original
+        super().__init__(
+            f"'{model}' was blocked by ChatGPT's content guardrails "
+            f"(flagged as a possible cybersecurity risk). "
+            f"Set STRIX_LLM to a model that isn't blocked and re-run."
+        )
+
+
+_GUARDRAIL_MARKERS = (
+    "flagged for possible cybersecurity risk",
+    "trusted access for cyber",
+)
+
+
+def is_content_guardrail_error(exc: BaseException) -> bool:
+    if isinstance(exc, CodexContentGuardrailError):
+        return True
+    text = str(exc).lower()
+    return any(marker in text for marker in _GUARDRAIL_MARKERS)
+
+
+def _b64url(raw: bytes) -> str:
+    return base64.urlsafe_b64encode(raw).rstrip(b"=").decode("ascii")
+
+
+def generate_pkce() -> tuple[str, str]:
+    verifier = _b64url(secrets.token_bytes(64))
+    challenge = _b64url(hashlib.sha256(verifier.encode("ascii")).digest())
+    return verifier, challenge
+
+
+def create_state() -> str:
+    return secrets.token_hex(16)
+
+
+def build_authorize_url(challenge: str, state: str) -> str:
+    params = {
+        "response_type": "code",
+        "client_id": CLIENT_ID,
+        "redirect_uri": REDIRECT_URI,
+        "scope": SCOPE,
+        "code_challenge": challenge,
+        "code_challenge_method": "S256",
+        "state": state,
+        "id_token_add_organizations": "true",
+        "codex_cli_simplified_flow": "true",
+        "originator": ORIGINATOR,
+    }
+    return f"{AUTHORIZE_URL}?{urllib.parse.urlencode(params)}"
+
+
+def parse_redirect_input(value: str) -> tuple[str | None, str | None]:
+    """Extract ``(code, state)`` from a pasted redirect URL, ``code#state``,
+    query string, or bare code."""
+    value = (value or "").strip()
+    if not value:
+        return None, None
+    with contextlib.suppress(ValueError):
+        parsed = urllib.parse.urlparse(value)
+        if parsed.scheme and parsed.query:
+            query = urllib.parse.parse_qs(parsed.query)
+            return _first(query, "code"), _first(query, "state")
+    if "#" in value:
+        code, _, state = value.partition("#")
+        return code or None, state or None
+    if "code=" in value:
+        query = urllib.parse.parse_qs(value)
+        return _first(query, "code"), _first(query, "state")
+    return value, None
+
+
+def _first(query: dict[str, list[str]], key: str) -> str | None:
+    values = query.get(key)
+    return values[0] if values else None
+
+
+def _post_form(payload: dict[str, str]) -> dict[str, Any]:
+    try:
+        response = requests.post(
+            TOKEN_URL,
+            data=payload,
+            headers={"Accept": "application/json"},
+            timeout=_TOKEN_TIMEOUT,
+        )
+    except requests.RequestException as exc:
+        raise CodexAuthError("unavailable", str(exc)) from exc
+    if response.status_code >= 400:
+        detail = response.text[:300]
+        raise CodexAuthError("token_http_error", f"HTTP {response.status_code}: {detail}")
+    data = json.loads(response.content or b"{}")
+    if not isinstance(data, dict):
+        raise CodexAuthError("bad_response", "token endpoint returned non-object")
+    return data
+
+
+def _record_from_token_response(
+    data: dict[str, Any], refresh_fallback: str | None = None
+) -> dict[str, Any]:
+    access = data.get("access_token")
+    # A refresh response may omit refresh_token when it isn't rotated; keep the old one.
+    refresh = data.get("refresh_token") or refresh_fallback
+    expires_in = data.get("expires_in")
+    if not isinstance(access, str) or not access:
+        raise CodexAuthError("bad_response", "token response missing access_token")
+    if not isinstance(refresh, str) or not refresh:
+        raise CodexAuthError("bad_response", "token response missing refresh_token")
+    account_id = _account_id_from_jwt(access) or _account_id_from_jwt(
+        data.get("id_token") if isinstance(data.get("id_token"), str) else ""
+    )
+    if not account_id:
+        raise CodexAuthError("no_account_id", "could not read chatgpt_account_id from token")
+    ttl = expires_in if isinstance(expires_in, int | float) else 3600
+    return {
+        "type": "oauth",
+        "provider": PROVIDER,
+        "access": access,
+        "refresh": refresh,
+        "account_id": account_id,
+        "expires_at": time.time() + ttl,
+    }
+
+
+def exchange_code(code: str, verifier: str) -> dict[str, Any]:
+    data = _post_form(
+        {
+            "grant_type": "authorization_code",
+            "client_id": CLIENT_ID,
+            "code": code,
+            "code_verifier": verifier,
+            "redirect_uri": REDIRECT_URI,
+        }
+    )
+    return _record_from_token_response(data)
+
+
+def refresh_tokens(refresh_token: str) -> dict[str, Any]:
+    data = _post_form(
+        {
+            "grant_type": "refresh_token",
+            "client_id": CLIENT_ID,
+            "refresh_token": refresh_token,
+        }
+    )
+    return _record_from_token_response(data, refresh_fallback=refresh_token)
+
+
+def _account_id_from_jwt(token: str | None) -> str | None:
+    """Read the account id claim without verifying the JWT (the server enforces
+    authenticity on use); it feeds the ``chatgpt-account-id`` header."""
+    if not token or token.count(".") != 2:
+        return None
+    payload_b64 = token.split(".")[1]
+    padding = "=" * (-len(payload_b64) % 4)
+    try:
+        payload = json.loads(base64.urlsafe_b64decode(payload_b64 + padding))
+    except (ValueError, json.JSONDecodeError):
+        return None
+    if not isinstance(payload, dict):
+        return None
+    auth = payload.get(_ACCOUNT_CLAIM)
+    if isinstance(auth, dict):
+        account_id = auth.get("chatgpt_account_id")
+        if isinstance(account_id, str) and account_id:
+            return account_id
+    organizations = payload.get("organizations")
+    if isinstance(organizations, list) and organizations and isinstance(organizations[0], dict):
+        org_id = organizations[0].get("id")
+        if isinstance(org_id, str) and org_id:
+            return org_id
+    return None
+
+
+def _near_expiry(record: dict[str, Any]) -> bool:
+    expires_at = record.get("expires_at")
+    if not isinstance(expires_at, int | float):
+        return True
+    return expires_at - _EXPIRY_SKEW_S <= time.time()
+
+
+def get_valid_token() -> tuple[str, str]:
+    """Return ``(access_token, account_id)``, refreshing under the cross-process
+    guard if near expiry."""
+    record = read_record()
+    if record is None:
+        raise CodexAuthError("not_authenticated", "not signed in; run: strix auth login")
+    if not _near_expiry(record):
+        return record["access"], record["account_id"]
+    with _refresh_guard():
+        record = read_record()
+        if record is None:
+            raise CodexAuthError("not_authenticated", "not signed in; run: strix auth login")
+        if not _near_expiry(record):
+            return record["access"], record["account_id"]
+        try:
+            refreshed = refresh_tokens(record["refresh"])
+        except CodexAuthError:
+            # A peer process may have already spent this single-use refresh token.
+            latest = read_record()
+            if latest and latest["refresh"] != record["refresh"] and not _near_expiry(latest):
+                return latest["access"], latest["account_id"]
+            raise
+        save_record(refreshed)
+        return refreshed["access"], refreshed["account_id"]
+
+
+def build_openai_client() -> AsyncOpenAI:
+    """An ``AsyncOpenAI`` for the ChatGPT backend. A per-request hook re-stamps a
+    fresh bearer token so long scans survive token expiry."""
+    import asyncio
+
+    import httpx
+    from openai import AsyncOpenAI
+
+    get_valid_token()  # fail fast at configure time if the sign-in is dead
+
+    async def _auth_hook(request: httpx.Request) -> None:
+        access, account_id = await asyncio.to_thread(get_valid_token)
+        request.headers["Authorization"] = f"Bearer {access}"
+        request.headers["chatgpt-account-id"] = account_id
+
+    http_client = httpx.AsyncClient(
+        timeout=httpx.Timeout(600.0, connect=30.0),
+        event_hooks={"request": [_auth_hook]},
+    )
+    return AsyncOpenAI(
+        api_key="strix-codex-oauth",  # placeholder; the hook overwrites Authorization
+        base_url=CODEX_BASE_URL,
+        http_client=http_client,
+        default_headers={
+            "OpenAI-Beta": "responses=experimental",
+            "originator": ORIGINATOR,
+        },
+    )
+
+
+_subscription_client: AsyncOpenAI | None = None
+
+
+def get_subscription_client() -> AsyncOpenAI:
+    global _subscription_client  # noqa: PLW0603
+    if _subscription_client is None:
+        _subscription_client = build_openai_client()
+    return _subscription_client
+
+
+SUBSCRIPTION_PREFIX = "chatgpt/"
+
+
+def subscription_model(model_name: str | None) -> str | None:
+    """The model slug behind a ``chatgpt/<model>`` STRIX_LLM, or None."""
+    name = (model_name or "").strip()
+    if not name.lower().startswith(SUBSCRIPTION_PREFIX):
+        return None
+    return name[len(SUBSCRIPTION_PREFIX) :] or None
+
+
+def auth_mode(model_name: str | None) -> str:
+    return "subscription" if subscription_model(model_name) else "api_key"

strix/config/models.py

@@ -2,23 +2,50 @@
 
 from __future__ import annotations
 
+import contextlib
+import inspect
 import os
-from typing import TYPE_CHECKING
+import time
+from typing import TYPE_CHECKING, Any
 
-from agents import set_default_openai_api, set_default_openai_key, set_tracing_disabled
+from agents import (
+    set_default_openai_api,
+    set_default_openai_key,
+    set_tracing_disabled,
+)
+from agents.model_settings import ModelSettings
+from agents.models.fake_id import FAKE_RESPONSES_ID
+from agents.models.interface import Model
 from agents.models.multi_provider import MultiProvider
+from agents.models.openai_responses import OpenAIResponsesModel
 from agents.retry import (
     ModelRetryBackoffSettings,
     ModelRetrySettings,
     RetryPolicyContext,
     retry_policies,
 )
+from openai.types.responses import Response, ResponseCompletedEvent
+from openai.types.responses.response_usage import ResponseUsage
+from openai.types.shared import Reasoning
+
+from strix.config import codex
+from strix.config.loader import load_settings
 
 
 if TYPE_CHECKING:
-    from agents.models.interface import ModelProvider
+    from collections.abc import AsyncIterator
 
-    from strix.config.settings import Settings
+    from agents.agent_output import AgentOutputSchemaBase
+    from agents.handoffs import Handoff
+    from agents.items import ModelResponse, TResponseInputItem, TResponseStreamEvent
+    from agents.models.interface import ModelProvider, ModelTracing
+    from agents.retry import ModelRetryAdvice, ModelRetryAdviceRequest
+    from agents.tool import Tool
+    from agents.usage import Usage
+    from openai import AsyncOpenAI
+    from openai.types.responses.response_prompt_param import ResponsePromptParam
+
+    from strix.config.settings import LlmSettings, ReasoningEffort, Settings
 
 
 def request_timeout_extra_args(timeout_s: float | None) -> dict[str, float] | None:
@@ -33,9 +60,211 @@ def _retry_statusless_provider_errors(context: RetryPolicyContext) -> bool:
     normalized = context.normalized
     if normalized.is_abort:
         return False
+    if codex.is_content_guardrail_error(context.error):
+        return False
     return normalized.status_code is None
 
 
+class _CodexResponsesModel(OpenAIResponsesModel):
+    """Responses model for the ChatGPT subscription backend (always streamed, stateless)."""
+
+    def __init__(
+        self,
+        model: str,
+        openai_client: AsyncOpenAI,
+        *,
+        reasoning_effort: ReasoningEffort | None = None,
+    ) -> None:
+        super().__init__(model, openai_client)
+        self._reasoning_effort = reasoning_effort
+
+    def _codex_settings(self, model_settings: ModelSettings) -> ModelSettings:
+        overrides = ModelSettings(store=False, response_include=["reasoning.encrypted_content"])
+        effort = self._reasoning_effort
+        if effort and effort != "none":
+            # Clamp to efforts the backend accepts.
+            if effort == "minimal":
+                effort = "low"
+            elif effort == "xhigh":
+                effort = "high"
+            overrides = overrides.resolve(ModelSettings(reasoning=Reasoning(effort=effort)))
+        return model_settings.resolve(overrides)
+
+    async def _fetch_response(self, *args: Any, stream: bool = False, **kwargs: Any) -> Any:
+        if len(args) >= 3:  # model_settings is positional arg 2
+            args = (*args[:2], self._codex_settings(args[2]), *args[3:])
+        try:
+            events = await super()._fetch_response(*args, stream=True, **kwargs)  # type: ignore[call-overload]
+        except Exception as exc:
+            guardrail = self._as_guardrail(exc)
+            if guardrail is not None:
+                raise guardrail from exc
+            raise
+        guarded = self._guarded(events)
+        if stream:
+            return guarded
+        final_response = None
+        async for event in guarded:
+            if getattr(event, "type", None) == "response.completed":
+                final_response = event.response
+        if final_response is None:
+            msg = "ChatGPT backend stream ended without a completed response"
+            raise RuntimeError(msg)
+        return final_response
+
+    def _as_guardrail(self, exc: BaseException) -> codex.CodexContentGuardrailError | None:
+        if isinstance(exc, codex.CodexContentGuardrailError):
+            return exc
+        if codex.is_content_guardrail_error(exc):
+            return codex.CodexContentGuardrailError(self.model, exc)
+        return None
+
+    async def _guarded(self, events: Any) -> AsyncIterator[Any]:
+        """Convert mid-stream guardrail rejections and close the stream on exit."""
+        try:
+            async for event in events:
+                yield event
+        except Exception as exc:
+            guardrail = self._as_guardrail(exc)
+            if guardrail is not None:
+                raise guardrail from exc
+            raise
+        finally:
+            await self._aclose(events)
+
+    @staticmethod
+    async def _aclose(events: Any) -> None:
+        aclose = getattr(events, "aclose", None)
+        if callable(aclose):
+            with contextlib.suppress(Exception):
+                await aclose()
+            return
+        close = getattr(events, "close", None)
+        if callable(close):
+            with contextlib.suppress(Exception):
+                result = close()
+                if inspect.isawaitable(result):
+                    await result
+
+
+class _NonStreamingModel(Model):
+    """Serve the SDK's streamed run loop from a single non-streaming request.
+
+    Some OpenAI-compatible gateways do not support Server-Sent Events, or
+    deliver them unreliably (dropping structured tool-call deltas, or stalling
+    mid-stream so the whole turn waits out the read timeout). The SDK run loop
+    Strix uses only issues streamed requests, so such a gateway fails every
+    turn. Opt in with ``LLM_DISABLE_STREAMING=true`` to wrap the resolved model
+    so each turn makes one non-streaming ``get_response`` (``stream:false`` on
+    the wire) and the completed result is replayed as a single terminal stream
+    event. The run loop then executes tools and emits run items from that final
+    response exactly as it would for a real stream, so nothing else changes.
+    """
+
+    def __init__(self, inner: Model) -> None:
+        self._inner = inner
+
+    async def close(self) -> None:
+        await self._inner.close()
+
+    def get_retry_advice(self, request: ModelRetryAdviceRequest) -> ModelRetryAdvice | None:
+        return self._inner.get_retry_advice(request)
+
+    async def get_response(
+        self,
+        system_instructions: str | None,
+        input: str | list[TResponseInputItem],  # noqa: A002
+        model_settings: ModelSettings,
+        tools: list[Tool],
+        output_schema: AgentOutputSchemaBase | None,
+        handoffs: list[Handoff],
+        tracing: ModelTracing,
+        *,
+        previous_response_id: str | None,
+        conversation_id: str | None,
+        prompt: ResponsePromptParam | None,
+    ) -> ModelResponse:
+        return await self._inner.get_response(
+            system_instructions,
+            input,
+            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: str | None,
+        input: str | list[TResponseInputItem],  # noqa: A002
+        model_settings: ModelSettings,
+        tools: list[Tool],
+        output_schema: AgentOutputSchemaBase | None,
+        handoffs: list[Handoff],
+        tracing: ModelTracing,
+        *,
+        previous_response_id: str | None,
+        conversation_id: str | None,
+        prompt: ResponsePromptParam | None,
+    ) -> AsyncIterator[TResponseStreamEvent]:
+        response = await self._inner.get_response(
+            system_instructions,
+            input,
+            model_settings,
+            tools,
+            output_schema,
+            handoffs,
+            tracing,
+            previous_response_id=previous_response_id,
+            conversation_id=conversation_id,
+            prompt=prompt,
+        )
+        yield _completed_stream_event(response, getattr(self._inner, "model", None))
+
+
+def _completed_stream_event(
+    model_response: ModelResponse, model_name: object | None
+) -> TResponseStreamEvent:
+    """Wrap a non-streamed ``ModelResponse`` as the terminal event of a stream.
+
+    The run loop builds its authoritative per-turn response solely from the
+    ``response.completed`` event, so a single event carrying the full output
+    and usage is all it needs.
+    """
+    response = Response(
+        id=model_response.response_id or FAKE_RESPONSES_ID,
+        created_at=time.time(),
+        model=str(model_name) if model_name else "",
+        object="response",
+        output=list(model_response.output),
+        tool_choice="auto",
+        tools=[],
+        parallel_tool_calls=False,
+        usage=_response_usage(model_response.usage),
+    )
+    return ResponseCompletedEvent(
+        response=response,
+        sequence_number=0,
+        type="response.completed",
+    )
+
+
+def _response_usage(usage: Usage | None) -> ResponseUsage | None:
+    if usage is None:
+        return None
+    return ResponseUsage(
+        input_tokens=usage.input_tokens,
+        output_tokens=usage.output_tokens,
+        total_tokens=usage.total_tokens,
+        input_tokens_details=usage.input_tokens_details,
+        output_tokens_details=usage.output_tokens_details,
+    )
+
+
 class StrixProvider(MultiProvider):
     """Route any non-OpenAI prefix through LiteLLM with the prefix preserved,
     so users type ``deepseek/deepseek-chat`` rather than
@@ -59,6 +288,23 @@ class StrixProvider(MultiProvider):
             return self._get_fallback_provider("litellm"), f"ollama_chat/{stripped_model_name}"
         return self._get_fallback_provider("litellm"), original_model_name
 
+    def get_model(self, model_name: str | None) -> Model:
+        llm = load_settings().llm
+        slug = codex.subscription_model(model_name)
+        if slug:
+            # The ChatGPT subscription backend is always streamed; it has no
+            # non-streaming mode to fall back to, so LLM_DISABLE_STREAMING
+            # does not apply here.
+            return _CodexResponsesModel(
+                slug,
+                codex.get_subscription_client(),
+                reasoning_effort=llm.reasoning_effort,
+            )
+        model = super().get_model(model_name)
+        if llm.disable_streaming:
+            return _NonStreamingModel(model)
+        return model
+
 
 DEFAULT_MODEL_RETRY = ModelRetrySettings(
     max_retries=5,
@@ -77,39 +323,42 @@ DEFAULT_MODEL_RETRY = ModelRetrySettings(
 )
 
 RECOMMENDED_MODEL_NAMES = (
-    "openai/gpt-5.6",
     "openai/gpt-5.6-sol",
     "openai/gpt-5.6-terra",
-    "openai/gpt-5.5",
+    "openai/gpt-5.6-luna",
+    "openai/gpt-5.6",
     "openai/gpt-5.5-pro",
+    "openai/gpt-5.5",
     "openai/gpt-5.4",
     "openai/gpt-5.3-codex",
     "anthropic/claude-fable-5",
+    "anthropic/claude-opus-5",
     "anthropic/claude-opus-4-8",
-    "anthropic/claude-opus-4-7",
     "anthropic/claude-sonnet-5",
     "anthropic/claude-sonnet-4-6",
     "vertex_ai/gemini-3.1-pro-preview",
     "gemini/gemini-3.1-pro-preview",
+    "gemini/gemini-3.6-flash",
     "deepseek/deepseek-v4-pro",
     "deepseek/deepseek-v4-flash",
+    "dashscope/qwen3.8-max",
     "dashscope/qwen3.7-max-2026-06-08",
+    "moonshot/kimi-k3",
     "moonshot/kimi-k2.7-code",
-    "moonshot/kimi-k2.6",
 )
 
 _RECOMMENDED_MODEL_NAME_SET = frozenset(name.lower() for name in RECOMMENDED_MODEL_NAMES)
 
 FRONTIER_MODEL_FAMILIES = (
-    (("azure", "azure_ai", "bedrock_mantle", "openai"), ("gpt-5",)),
+    (("azure", "azure_ai", "bedrock_mantle", "chatgpt", "openai"), ("gpt-5",)),
     (
         ("anthropic", "azure_ai", "bedrock", "claude", "databricks", "snowflake", "vertex_ai"),
-        ("claude-fable-5", "claude-opus-4", "claude-sonnet-5", "claude-sonnet-4"),
+        ("claude-fable-5", "claude-opus-5", "claude-opus-4", "claude-sonnet-5", "claude-sonnet-4"),
     ),
     (("google", "gemini", "vertex_ai"), ("gemini-3",)),
     (("deepseek",), ("deepseek-v4", "deepseek-r1", "deepseek-reasoner")),
-    (("alibaba", "dashscope", "qwen"), ("qwen3.7", "qwen3.5", "qwen3-max")),
-    (("moonshot", "moonshotai", "kimi"), ("kimi-k2.7", "kimi-k2.6", "kimi-k2.5")),
+    (("alibaba", "dashscope", "qwen"), ("qwen3.8", "qwen3.7", "qwen3-max")),
+    (("moonshot", "moonshotai", "kimi"), ("kimi-k3", "kimi-k2.7", "kimi-k2.6")),
 )
 
 
@@ -117,6 +366,8 @@ def configure_sdk_model_defaults(settings: Settings) -> None:
     """Apply Strix config to SDK-native defaults."""
     llm = settings.llm
     set_tracing_disabled(True)
+    if codex.subscription_model(llm.model):
+        return
     _configure_litellm_compatibility()
     _configure_openrouter_attribution(llm.model)
     if llm.api_key:
@@ -129,6 +380,7 @@ def configure_sdk_model_defaults(settings: Settings) -> None:
         set_default_openai_api("chat_completions")
     else:
         set_default_openai_api("responses")
+    _configure_extra_headers(llm)
 
 
 def _mirror_api_key_to_provider_env(model_name: str | None, api_key: str) -> None:
@@ -163,6 +415,51 @@ def _configure_litellm_compatibility() -> None:
     litellm.suppress_debug_info = True
 
     _register_litellm_cost_callback()
+    _install_openrouter_stream_cost_capture()
+
+
+def _install_openrouter_stream_cost_capture() -> None:
+    """Preserve OpenRouter's per-stream cost, which LiteLLM drops when streaming.
+
+    OpenRouter reports the real charge in ``usage.cost`` of the final stream
+    chunk, but LiteLLM rebuilds streamed responses from token-only fields and
+    discards it (its non-streamed path stashes the cost in hidden params; the
+    streaming path does not). Every scan streams, so without this the cost is
+    lost and Strix falls back to a cost-map estimate that is missing entirely
+    for new models (e.g. kimi-k3), reporting $0. Subclass the OpenRouter
+    streaming handler to record the cost keyed by response id so the cost
+    callback can recover the exact charge for the matching rebuilt response.
+    """
+    import litellm
+    from litellm.llms.openrouter.chat.transformation import (
+        OpenRouterChatCompletionStreamingHandler,
+        OpenrouterConfig,
+    )
+
+    from strix.report.state import streamed_openrouter_costs
+
+    class _StrixOpenRouterStreamingHandler(OpenRouterChatCompletionStreamingHandler):
+        def chunk_parser(self, chunk: dict[str, Any]) -> Any:
+            stream = super().chunk_parser(chunk)
+            streamed_openrouter_costs.remember(
+                chunk.get("id") or getattr(stream, "id", None), chunk.get("usage")
+            )
+            return stream
+
+    class _StrixOpenrouterConfig(OpenrouterConfig):
+        def get_model_response_iterator(
+            self, streaming_response: Any, sync_stream: bool, json_mode: bool | None = False
+        ) -> Any:
+            return _StrixOpenRouterStreamingHandler(
+                streaming_response=streaming_response,
+                sync_stream=sync_stream,
+                json_mode=json_mode,
+            )
+
+    # LiteLLM's provider-config factory reads litellm.OpenrouterConfig at call
+    # time, so overriding the attribute is enough for the subclass to take
+    # effect. (type: ignore — mypy rejects reassigning a class attribute.)
+    litellm.OpenrouterConfig = _StrixOpenrouterConfig  # type: ignore[misc]
 
 
 _OPENROUTER_ATTRIBUTION_HEADERS = {
@@ -188,6 +485,43 @@ def _configure_openrouter_attribution(model_name: str | None) -> None:
     litellm.headers = {**existing, **_OPENROUTER_ATTRIBUTION_HEADERS}  # type: ignore[assignment]
 
 
+def _configure_extra_headers(llm: LlmSettings) -> None:
+    """Send user-provided default headers on every LLM request.
+
+    Some OpenAI-compatible endpoints require extra HTTP headers (e.g. request
+    attribution or tenant routing) alongside the bearer token. Users supply
+    them via ``LLM_EXTRA_HEADERS``; they are applied to both routing paths:
+    the LiteLLM route (``litellm.headers``) and the SDK-native OpenAI route
+    (a default client carrying ``default_headers``), so they take effect
+    regardless of the ``STRIX_LLM`` prefix.
+    """
+    headers = llm.extra_headers
+    if not headers:
+        return
+    _merge_litellm_headers(headers)
+    _register_openai_client_with_headers(llm, headers)
+
+
+def _merge_litellm_headers(headers: dict[str, str]) -> None:
+    import litellm
+
+    current: object = litellm.headers
+    existing: dict[str, str] = current if isinstance(current, dict) else {}
+    litellm.headers = {**existing, **headers}  # type: ignore[assignment]
+
+
+def _register_openai_client_with_headers(llm: LlmSettings, headers: dict[str, str]) -> None:
+    from agents import set_default_openai_client
+    from openai import AsyncOpenAI
+
+    client = AsyncOpenAI(
+        api_key=llm.api_key or "not-needed",
+        base_url=llm.api_base,
+        default_headers=dict(headers),
+    )
+    set_default_openai_client(client, use_for_tracing=False)
+
+
 def _register_litellm_cost_callback() -> None:
     import litellm
 
@@ -211,6 +545,8 @@ def _configure_litellm_default(name: str, value: str) -> None:
 
 def uses_chat_completions_tool_schema(model_name: str, settings: Settings) -> bool:
     """Return whether the resolved SDK route can only receive JSON function tools."""
+    if codex.subscription_model(model_name):
+        return False
     model = model_name.strip().lower()
     if "/" in model and not model.startswith("openai/"):
         return True
@@ -313,3 +649,45 @@ 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:
+    name = (model_name or "").strip().lower()
+    return name.startswith("bedrock/") or "anthropic." in name
+
+
+def _prompt_cache_name_candidates(model_name: str) -> list[str]:
+    # LiteLLM's model map keys the same model under several names; strip the
+    # route prefix, then leading dotted segments (region, provider).
+    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:
+    # Bedrock rejects the cache marker for models LiteLLM's map doesn't
+    # recognise as cache-capable, so callers withhold it unless 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:
+            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

@@ -35,19 +35,67 @@ class LlmSettings(BaseSettings):
             "OLLAMA_API_BASE",
         ),
     )
+    extra_headers: dict[str, str] | None = Field(
+        default=None,
+        alias="LLM_EXTRA_HEADERS",
+    )
     reasoning_effort: ReasoningEffort = Field(default="high", alias="STRIX_REASONING_EFFORT")
     force_required_tool_choice: bool = Field(
         default=False,
         alias="STRIX_FORCE_REQUIRED_TOOL_CHOICE",
     )
+    prompt_cache: bool = Field(
+        default=True,
+        alias="STRIX_PROMPT_CACHE",
+    )
+    disable_streaming: bool = Field(
+        default=False,
+        alias="LLM_DISABLE_STREAMING",
+    )
     timeout: int = Field(default=300, alias="LLM_TIMEOUT")
 
 
+class DedupeSettings(BaseSettings):
+    model_config = _BASE_CONFIG
+
+    model: str | None = Field(default=None, alias="STRIX_DEDUPE_MODEL")
+    reasoning_effort: ReasoningEffort | None = Field(
+        default=None,
+        alias="STRIX_DEDUPE_REASONING_EFFORT",
+    )
+    api_key: str | None = Field(default=None, alias="DEDUPE_LLM_API_KEY")
+    api_base: str | None = Field(default=None, alias="DEDUPE_LLM_API_BASE")
+    extra_headers: dict[str, str] | None = Field(
+        default=None,
+        alias="DEDUPE_LLM_EXTRA_HEADERS",
+    )
+
+
+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
 
     image: str = Field(
-        default="ghcr.io/usestrix/strix-sandbox:1.0.0",
+        default="ghcr.io/usestrix/strix-sandbox:1.1.0",
         alias="STRIX_IMAGE",
     )
     backend: str = Field(default="docker", alias="STRIX_RUNTIME_BACKEND")
@@ -85,7 +133,9 @@ class Settings(BaseSettings):
     model_config = _BASE_CONFIG
 
     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/agents.py

@@ -14,13 +14,20 @@ from strix.core.sessions import session_write_lock
 
 
 if TYPE_CHECKING:
+    from collections.abc import Callable
+
     from agents.items import TResponseInputItem
     from agents.memory import Session
 
 
 logger = logging.getLogger(__name__)
 
-Status = Literal["running", "waiting", "completed", "stopped", "crashed", "failed"]
+Status = Literal["running", "waiting", "completed", "stopped", "crashed", "failed", "budget_paused"]
+
+# Why an agent parked. The user can message any agent, so this - not the agent's
+# position in the tree - decides whether waiting is bounded: only an agent waiting
+# on other agents is re-checked on a timer.
+WaitKind = Literal["user", "agents", "stalled"]
 
 
 @dataclass(slots=True)
@@ -30,6 +37,8 @@ class AgentRuntime:
     stream: Any | None = None
     interrupt_on_message: bool = False
     wake: asyncio.Event = field(default_factory=asyncio.Event)
+    mailbox: list[dict[str, Any]] = field(default_factory=list)
+    user_wake_required: bool = False
 
 
 class AgentCoordinator:
@@ -41,11 +50,18 @@ class AgentCoordinator:
         self.names: dict[str, str] = {}
         self.metadata: dict[str, dict[str, Any]] = {}
         self.pending_counts: dict[str, int] = {}
+        self.errors: dict[str, str] = {}
+        self.recovery_counts: dict[str, int] = {}
+        self.idle_resume_counts: dict[str, int] = {}
+        self.wait_kinds: dict[str, WaitKind] = {}
         self.runtimes: dict[str, AgentRuntime] = {}
         self._lock = asyncio.Lock()
         self._snapshot_path: Path | None = None
         self.is_shutting_down = False
         self._budget_stopped = False
+        self._reserve_stopped = False
+        self._budget_paused = False
+        self._extend_budget: Callable[[], None] | None = None
 
     def set_snapshot_path(self, path: Path) -> None:
         self._snapshot_path = path
@@ -64,6 +80,71 @@ class AgentCoordinator:
             for runtime in self.runtimes.values():
                 runtime.wake.set()
 
+    @property
+    def reserve_stopped(self) -> bool:
+        return self._reserve_stopped
+
+    @property
+    def budget_paused(self) -> bool:
+        return self._budget_paused
+
+    def set_budget_extender(self, extend: Callable[[], None]) -> None:
+        self._extend_budget = extend
+
+    async def pause_for_budget(self, agent_id: str) -> None:
+        async with self._lock:
+            self._budget_paused = True
+        await self.set_status(agent_id, "budget_paused")
+
+    async def resume_from_budget_pause(self, *, exclude: str | None = None) -> None:
+        async with self._lock:
+            if not self._budget_paused:
+                return
+            self._budget_paused = False
+            paused = [aid for aid, status in self.statuses.items() if status == "budget_paused"]
+        if self._extend_budget is not None:
+            self._extend_budget()
+        for aid in paused:
+            await self.set_status(aid, "waiting")
+            if aid != exclude:
+                await self.send(
+                    aid,
+                    {
+                        "from": "system",
+                        "type": "budget_extended",
+                        "content": (
+                            "[Budget] The user extended the scan budget \u2014 continue your "
+                            "current task."
+                        ),
+                    },
+                )
+
+    async def reset_budget_stops(
+        self,
+        *,
+        budget_stopped: bool,
+        reserve_stopped: bool,
+        budget_paused: bool = False,
+    ) -> None:
+        async with self._lock:
+            self._budget_stopped = budget_stopped
+            self._reserve_stopped = reserve_stopped
+            if not budget_paused:
+                self._budget_paused = False
+                for aid, status in self.statuses.items():
+                    if status == "budget_paused":
+                        self.statuses[aid] = "waiting"
+        await self._maybe_snapshot()
+
+    async def claim_reserve_notification(self) -> str | None:
+        async with self._lock:
+            if self._reserve_stopped:
+                return None
+            self._reserve_stopped = True
+            for runtime in self.runtimes.values():
+                runtime.wake.set()
+            return next((aid for aid, parent in self.parent_of.items() if parent is None), None)
+
     async def register(
         self,
         agent_id: str,
@@ -107,62 +188,120 @@ class AgentCoordinator:
         async with self._lock:
             if agent_id in self.statuses:
                 self.statuses[agent_id] = "running"
+                self.errors.pop(agent_id, None)
+                self.wait_kinds.pop(agent_id, None)
+                self.runtimes.setdefault(agent_id, AgentRuntime()).user_wake_required = False
         await self._maybe_snapshot()
 
-    async def park_waiting(self, agent_id: str) -> None:
+    async def park_waiting(self, agent_id: str, *, wait_kind: WaitKind) -> None:
+        """Park an agent, recording what it is waiting on so the driver can time it."""
+        async with self._lock:
+            if agent_id in self.statuses:
+                self.wait_kinds[agent_id] = wait_kind
         await self.set_status(agent_id, "waiting")
 
-    async def set_status(self, agent_id: str, status: Status | str) -> None:
+    async def wait_kind_of(self, agent_id: str) -> WaitKind | None:
+        async with self._lock:
+            return self.wait_kinds.get(agent_id)
+
+    async def record_recovery(self, agent_id: str) -> int:
+        """Count a turn that ended without a lifecycle tool call; return the new total.
+
+        Persisted so a resumed agent cannot earn a fresh nudge budget on every
+        auto-resume and loop forever.
+        """
+        async with self._lock:
+            count = self.recovery_counts.get(agent_id, 0) + 1
+            self.recovery_counts[agent_id] = count
+        await self._maybe_snapshot()
+        return count
+
+    async def reset_recovery(self, agent_id: str) -> None:
+        """Clear the nudge budget after real progress (new message or a lifecycle tool)."""
+        async with self._lock:
+            if self.recovery_counts.pop(agent_id, None) is None:
+                return
+        await self._maybe_snapshot()
+
+    async def record_idle_resume(self, agent_id: str) -> int:
+        """Count an auto-resume that no message triggered; return the new total.
+
+        An agent that parks again after every auto-resume would otherwise burn a
+        model turn per timeout for the rest of the scan.
+        """
+        async with self._lock:
+            count = self.idle_resume_counts.get(agent_id, 0) + 1
+            self.idle_resume_counts[agent_id] = count
+        await self._maybe_snapshot()
+        return count
+
+    async def reset_idle_resumes(self, agent_id: str) -> None:
+        async with self._lock:
+            if self.idle_resume_counts.pop(agent_id, None) is None:
+                return
+        await self._maybe_snapshot()
+
+    async def set_status(
+        self, agent_id: str, status: Status | str, *, error: str | None = None
+    ) -> None:
         async with self._lock:
             if agent_id not in self.statuses:
                 return
             self.statuses[agent_id] = status  # type: ignore[assignment]
+            if error is not None:
+                self.errors[agent_id] = error
+            elif status == "running":
+                self.errors.pop(agent_id, None)
             runtime = self.runtimes.setdefault(agent_id, AgentRuntime())
+            runtime.user_wake_required = status in {"failed", "crashed"}
             runtime.wake.set()
         logger.info("agent.status %s=%s", agent_id, status)
         await self._maybe_snapshot()
 
-    async def send(self, target_agent_id: str, message: dict[str, Any]) -> bool:
-        """Deliver a user/peer message by appending it to the target SDK session."""
+    async def send(
+        self, target_agent_id: str, message: dict[str, Any], *, interrupt: bool = True
+    ) -> bool:
+        """Queue a user/peer message in the target's mailbox and wake it."""
+        from_user = message.get("from") == "user"
+        if from_user and self._budget_paused:
+            await self.resume_from_budget_pause(exclude=target_agent_id)
         async with self._lock:
             if target_agent_id not in self.statuses:
                 logger.debug("agent.send dropped unknown target=%s", target_agent_id)
                 return False
             runtime = self.runtimes.setdefault(target_agent_id, AgentRuntime())
-            session = runtime.session
-            stream = runtime.stream
-            interrupt = runtime.interrupt_on_message
-        if session is None:
-            logger.warning(
-                "agent.send dropped target=%s because its SDK session is not attached",
-                target_agent_id,
-            )
-            return False
-        try:
-            async with session_write_lock(session):
-                await session.add_items([self._message_to_session_item(message)])
-        except Exception:
-            logger.exception(
-                "agent.send failed to append to SDK session target=%s",
-                target_agent_id,
-            )
-            return False
-        async with self._lock:
+            runtime.mailbox.append(dict(message))
             self.pending_counts[target_agent_id] = self.pending_counts.get(target_agent_id, 0) + 1
-            self.runtimes.setdefault(target_agent_id, AgentRuntime()).wake.set()
-        if stream is not None and interrupt:
+            if from_user:
+                runtime.user_wake_required = False
+            runtime.wake.set()
+            stream = runtime.stream
+            interrupt_on_message = runtime.interrupt_on_message
+        if stream is not None and interrupt and interrupt_on_message:
             stream.cancel(mode="immediate")
         await self._maybe_snapshot()
         return True
 
-    async def wait_for_message(self, agent_id: str) -> None:
+    async def wait_for_message(self, agent_id: str, *, timeout: float | None = None) -> bool:
+        """Wait until a message is ready for ``agent_id``; False on ``timeout``."""
         while True:
             async with self._lock:
-                if self._budget_stopped or self.pending_counts.get(agent_id, 0) > 0:
-                    return
-                wake = self.runtimes.setdefault(agent_id, AgentRuntime()).wake
+                runtime = self.runtimes.setdefault(agent_id, AgentRuntime())
+                reserve_exit = self._reserve_stopped and self.parent_of.get(agent_id) is not None
+                pending_ready = (
+                    self.pending_counts.get(agent_id, 0) > 0 and not runtime.user_wake_required
+                )
+                if self._budget_stopped or reserve_exit or pending_ready:
+                    return True
+                wake = runtime.wake
                 wake.clear()
-            await wake.wait()
+            if timeout is None:
+                await wake.wait()
+            else:
+                try:
+                    await asyncio.wait_for(wake.wait(), timeout)
+                except TimeoutError:
+                    return False
 
     async def consume_pending(
         self,
@@ -170,17 +309,38 @@ class AgentCoordinator:
         *,
         include_items: bool = False,
     ) -> tuple[int, list[Any]]:
+        """Drain the agent's mailbox into its own SDK session."""
         async with self._lock:
-            count = self.pending_counts.get(agent_id, 0)
+            runtime = self.runtimes.setdefault(agent_id, AgentRuntime())
+            queued = list(runtime.mailbox)
+            runtime.mailbox.clear()
+            count = max(self.pending_counts.get(agent_id, 0), len(queued))
             self.pending_counts[agent_id] = 0
-            session = self.runtimes.get(agent_id, AgentRuntime()).session
+            session = runtime.session
         if count <= 0:
             return 0, []
+        items = [self._message_to_session_item(m) for m in queued]
+        if items:
+            if session is None:
+                logger.warning(
+                    "agent %s has no SDK session attached; %d queued messages were not persisted",
+                    agent_id,
+                    len(items),
+                )
+            else:
+                try:
+                    async with session_write_lock(session):
+                        await session.add_items(items)
+                except Exception:
+                    logger.exception(
+                        "failed to append %d queued messages to the session of %s",
+                        len(items),
+                        agent_id,
+                    )
         await self._maybe_snapshot()
-        if not include_items or session is None:
+        if not include_items:
             return count, []
-        items = await session.get_items()
-        return count, list(items[-count:])
+        return count, items
 
     async def request_stop(self, agent_id: str) -> None:
         async with self._lock:
@@ -246,9 +406,14 @@ class AgentCoordinator:
 
     async def graph_snapshot(
         self,
-    ) -> tuple[dict[str, str | None], dict[str, Status], dict[str, str]]:
+    ) -> tuple[dict[str, str | None], dict[str, Status], dict[str, str], dict[str, str]]:
         async with self._lock:
-            return dict(self.parent_of), dict(self.statuses), dict(self.names)
+            return (
+                dict(self.parent_of),
+                dict(self.statuses),
+                dict(self.names),
+                dict(self.errors),
+            )
 
     def _message_to_session_item(self, message: dict[str, Any]) -> TResponseInputItem:
         sender = str(message.get("from", "unknown"))
@@ -286,6 +451,18 @@ class AgentCoordinator:
                 "names": dict(self.names),
                 "metadata": {aid: dict(md) for aid, md in self.metadata.items()},
                 "pending_counts": dict(self.pending_counts),
+                "recovery_counts": dict(self.recovery_counts),
+                "idle_resume_counts": dict(self.idle_resume_counts),
+                "wait_kinds": dict(self.wait_kinds),
+                "mailboxes": {
+                    aid: [dict(m) for m in runtime.mailbox]
+                    for aid, runtime in self.runtimes.items()
+                    if runtime.mailbox
+                },
+                "errors": dict(self.errors),
+                "budget_stopped": self._budget_stopped,
+                "reserve_stopped": self._reserve_stopped,
+                "budget_paused": self._budget_paused,
             }
 
     async def restore(self, snap: dict[str, Any]) -> None:
@@ -295,6 +472,19 @@ class AgentCoordinator:
             self.names = dict(snap.get("names", {}))
             self.metadata = {aid: dict(md) for aid, md in snap.get("metadata", {}).items()}
             self.pending_counts = dict(snap.get("pending_counts", {}))
+            self.errors = dict(snap.get("errors", {}))
+            self.recovery_counts = dict(snap.get("recovery_counts", {}))
+            self.idle_resume_counts = dict(snap.get("idle_resume_counts", {}))
+            self.wait_kinds = dict(snap.get("wait_kinds", {}))
+            mailboxes = snap.get("mailboxes", {})
+            if isinstance(mailboxes, dict):
+                for aid, msgs in mailboxes.items():
+                    if isinstance(msgs, list):
+                        runtime = self.runtimes.setdefault(aid, AgentRuntime())
+                        runtime.mailbox = [dict(m) for m in msgs if isinstance(m, dict)]
+            self._budget_stopped = bool(snap.get("budget_stopped", False))
+            self._reserve_stopped = bool(snap.get("reserve_stopped", False))
+            self._budget_paused = bool(snap.get("budget_paused", False))
             for aid in self.statuses:
                 self.runtimes.setdefault(aid, AgentRuntime())
 

strix/core/execution.py

@@ -9,19 +9,32 @@ import uuid
 from collections.abc import Callable
 from typing import TYPE_CHECKING, Any, cast
 
+import litellm
 from agents import RunConfig, Runner
 from agents.exceptions import AgentsException, MaxTurnsExceeded, UserError
 from agents.sandbox.errors import ExecTransportError
 from docker import errors as docker_errors  # type: ignore[import-untyped, unused-ignore]
-from openai import APIError
+from openai import (
+    APIConnectionError,
+    APIError,
+    APITimeoutError,
+)
 
-from strix.core.hooks import BudgetExceededError
+from strix.config import codex
+from strix.core.hooks import (
+    BudgetExceededError,
+    BudgetPausedError,
+    SubagentBudgetReservedError,
+)
 from strix.core.inputs import child_initial_input
 from strix.core.sessions import (
     enforce_image_budget,
     open_agent_session,
+    replace_session_items,
+    seed_initial_input,
     strip_all_images_from_session,
 )
+from strix.llm.compaction import is_context_overflow, maybe_compact
 
 
 if TYPE_CHECKING:
@@ -40,6 +53,120 @@ logger = logging.getLogger(__name__)
 StreamEventSink = Callable[[str, Any], None]
 
 _INPUT_REJECTION_CODES = frozenset({400, 404, 422})
+_MAX_COMPACTIONS_PER_CYCLE = 2
+
+
+class ProviderRefusalError(AgentsException):
+    """Raised when a provider returns a structured refusal instead of an exception."""
+
+
+def _structured_provider_refusal(result: Any) -> str | None:
+    for item in getattr(result, "new_items", ()) or ():
+        raw_item = getattr(item, "raw_item", None)
+        for content in getattr(raw_item, "content", ()) or ():
+            if getattr(content, "type", None) != "refusal":
+                continue
+            refusal = getattr(content, "refusal", None)
+            if isinstance(refusal, str) and refusal.strip():
+                return refusal.strip()
+            return "The model provider refused this request."
+    return None
+
+
+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,
+    )
+
+
+_MAX_TRANSIENT_MODEL_RETRIES = 5
+_TRANSIENT_MODEL_RETRY_BASE_DELAY_S = 2.0
+_TRANSIENT_MODEL_RETRY_MAX_DELAY_S = 90.0
+
+
+def _model_error_status_code(exc: BaseException) -> int | None:
+    code = getattr(exc, "status_code", None)
+    return code if isinstance(code, int) else None
+
+
+def _is_transient_model_error(exc: BaseException) -> bool:
+    if codex.is_content_guardrail_error(exc):
+        return False
+    if isinstance(
+        exc, APITimeoutError | APIConnectionError | TimeoutError | ConnectionError | OSError
+    ):
+        return True
+    code = _model_error_status_code(exc)
+    if code is not None:
+        return bool(litellm._should_retry(code))
+    return isinstance(exc, APIError)
+
+
+def _transient_model_retry_delay(attempt: int) -> float:
+    delay = _TRANSIENT_MODEL_RETRY_BASE_DELAY_S * float(2 ** (attempt - 1))
+    return min(delay, _TRANSIENT_MODEL_RETRY_MAX_DELAY_S)
+
+
+async def _salvage_stream_to_session(
+    session: Session,
+    pre_run_items: list[Any],
+    stream: Any,
+    agent_id: str,
+) -> None:
+    """Persist a crashed run's full history so a revived agent loses no context."""
+    if stream is None:
+        return
+    try:
+        replay = list(stream.to_input_list())
+    except Exception:
+        logger.exception("could not build salvage history for %s", agent_id)
+        return
+    desired = list(pre_run_items) + replay
+    if len(desired) <= len(pre_run_items):
+        return
+    try:
+        await replace_session_items(session, desired)
+    except Exception:
+        logger.exception("salvaging crashed run history failed for %s", agent_id)
+
+
+async def _seed_and_prepare_first_input(
+    session: Session | None, initial_input: Any, *, start_parked: bool
+) -> Any:
+    """Persist the opening input up front so it survives a first-turn crash."""
+    if initial_input and session is not None and not start_parked:
+        with contextlib.suppress(Exception):
+            if await seed_initial_input(session, initial_input):
+                return []
+    return initial_input
 
 
 async def run_agent_loop(
@@ -64,13 +191,29 @@ async def run_agent_loop(
     )
     result: RunResultBase | None = None
 
+    first_cycle_input = await _seed_and_prepare_first_input(
+        session, initial_input, start_parked=start_parked
+    )
+
+    budget_stopped = coordinator.budget_stopped
+    reserve_stopped = coordinator.reserve_stopped
+    if budget_stopped:
+        await coordinator.set_status(agent_id, "stopped")
+        raise BudgetExceededError("scan budget reached")
+    if reserve_stopped and context.get("parent_id") is not None:
+        await coordinator.set_status(agent_id, "stopped")
+        raise SubagentBudgetReservedError("scan reached the sub-agent budget reserve")
+
+    if reserve_stopped and start_parked and interactive and context.get("parent_id") is None:
+        await coordinator.send(agent_id, _reserve_notice())
+
     if not (start_parked and interactive):
-        if interactive:
-            result = await _run_cycle(
+        with contextlib.suppress(BudgetPausedError):
+            result = await _run_until_lifecycle(
                 agent,
                 coordinator,
                 agent_id,
-                input_data=initial_input,
+                initial_input=first_cycle_input,
                 run_config=run_config,
                 context=context,
                 max_turns=max_turns,
@@ -79,26 +222,14 @@ async def run_agent_loop(
                 event_sink=event_sink,
                 hooks=hooks,
             )
-        else:
-            result = await _run_noninteractive_until_lifecycle(
-                agent,
-                coordinator,
-                agent_id,
-                initial_input=initial_input,
-                run_config=run_config,
-                context=context,
-                max_turns=max_turns,
-                session=session,
-                event_sink=event_sink,
-                hooks=hooks,
-            )
 
     if not interactive:
         return result
 
     while True:
+        timeout = await _plain_waiting_timeout(coordinator, agent_id)
         try:
-            await coordinator.wait_for_message(agent_id)
+            woke = await coordinator.wait_for_message(agent_id, timeout=timeout)
         except asyncio.CancelledError:
             return result
 
@@ -106,20 +237,53 @@ async def run_agent_loop(
             await coordinator.set_status(agent_id, "stopped")
             raise BudgetExceededError("scan budget reached")
 
+        if coordinator.reserve_stopped and context.get("parent_id") is not None:
+            await coordinator.set_status(agent_id, "stopped")
+            raise SubagentBudgetReservedError("scan reached the sub-agent budget reserve")
+
+        if woke:
+            # Real input is real progress, so the nudge budget starts over. A bare
+            # auto-resume is not: it must not hand a wedged agent a fresh budget.
+            await coordinator.reset_recovery(agent_id)
+            await coordinator.reset_idle_resumes(agent_id)
+        else:
+            idle_resumes = await coordinator.record_idle_resume(agent_id)
+            if idle_resumes >= _MAX_IDLE_AUTO_RESUMES:
+                logger.warning(
+                    "agent %s auto-resumed %d times without hearing from anyone; "
+                    "leaving it parked until a real message arrives",
+                    agent_id,
+                    idle_resumes,
+                )
+                await coordinator.park_waiting(agent_id, wait_kind="stalled")
+                await _notify_parent_on_stall(coordinator, agent_id)
+                continue
+            logger.info("agent %s reached its waiting timeout; auto-resuming", agent_id)
+            await coordinator.send(
+                agent_id,
+                {
+                    "from": "system",
+                    "type": "auto_resume",
+                    "content": "Waiting timeout reached. Resuming execution.",
+                },
+                interrupt=False,
+            )
+
         await coordinator.consume_pending(agent_id)
-        result = await _run_cycle(
-            agent,
-            coordinator,
-            agent_id,
-            input_data=[],
-            run_config=run_config,
-            context=context,
-            max_turns=max_turns,
-            session=session,
-            interactive=interactive,
-            event_sink=event_sink,
-            hooks=hooks,
-        )
+        with contextlib.suppress(BudgetPausedError):
+            result = await _run_until_lifecycle(
+                agent,
+                coordinator,
+                agent_id,
+                initial_input=[],
+                run_config=run_config,
+                context=context,
+                max_turns=max_turns,
+                session=session,
+                interactive=True,
+                event_sink=event_sink,
+                hooks=hooks,
+            )
 
 
 async def spawn_child_agent(
@@ -267,7 +431,10 @@ async def respawn_subagents(
                 await coordinator.set_status(child_id, "crashed")
 
 
-async def _run_noninteractive_until_lifecycle(
+_INTERACTIVE_TOOL_RECOVERY_LIMIT = 3
+
+
+async def _run_until_lifecycle(
     agent: Any,
     coordinator: AgentCoordinator,
     agent_id: str,
@@ -277,21 +444,167 @@ async def _run_noninteractive_until_lifecycle(
     context: dict[str, Any],
     max_turns: int,
     session: Session | None,
+    interactive: bool,
     event_sink: StreamEventSink | None,
     hooks: RunHooks[dict[str, Any]] | None,
 ) -> RunResultBase | None:
-    """Non-chat mode keeps running until finish_scan / agent_finish settles status."""
+    """Drive an agent until an explicit lifecycle tool settles its status.
+
+    A turn that ends without ``finish_scan``, ``agent_finish``,
+    ``respond_to_user``, or ``wait_for_agents`` leaves the agent ``running``:
+    plain text never terminates a run and never yields to the user. Such a turn
+    is nudged back into a tool call, bounded by a recovery limit.
+    """
     result: RunResultBase | None = None
     input_data: Any = initial_input
-    invalid_final_outputs = 0
-    invalid_final_output_limit = max(1, max_turns)
+    recovery_limit = _INTERACTIVE_TOOL_RECOVERY_LIMIT if interactive else max(1, max_turns)
 
     while True:
         if coordinator.budget_stopped:
             await coordinator.set_status(agent_id, "stopped")
             raise BudgetExceededError("scan budget reached")
 
-        result = await _run_cycle(
+        if coordinator.reserve_stopped and context.get("parent_id") is not None:
+            await coordinator.set_status(agent_id, "stopped")
+            raise SubagentBudgetReservedError("scan reached the sub-agent budget reserve")
+
+        if interactive:
+            result = await _run_cycle_parked(
+                agent,
+                coordinator,
+                agent_id,
+                input_data=input_data,
+                run_config=run_config,
+                context=context,
+                max_turns=max_turns,
+                session=session,
+                event_sink=event_sink,
+                hooks=hooks,
+            )
+        else:
+            result = await _run_cycle(
+                agent,
+                coordinator,
+                agent_id,
+                input_data=input_data,
+                run_config=run_config,
+                context=context,
+                max_turns=max_turns,
+                session=session,
+                interactive=False,
+                event_sink=event_sink,
+                hooks=hooks,
+            )
+
+        status = await _agent_status(coordinator, agent_id)
+        if status != "running":
+            await coordinator.reset_recovery(agent_id)
+            return result
+
+        recoveries = await coordinator.record_recovery(agent_id)
+        logger.warning(
+            "agent %s ended a turn without a lifecycle tool call (interactive=%s); "
+            "forcing tool continuation (%d/%d): %s",
+            agent_id,
+            interactive,
+            recoveries,
+            recovery_limit,
+            _final_output_preview(result),
+        )
+
+        if recoveries >= recovery_limit:
+            return await _exhausted_recovery(coordinator, agent_id, result, interactive=interactive)
+
+        input_data = await _append_tool_required_message(
+            session=session,
+            context=context,
+            attempt=recoveries,
+            limit=recovery_limit,
+            interactive=interactive,
+        )
+
+
+async def _exhausted_recovery(
+    coordinator: AgentCoordinator,
+    agent_id: str,
+    result: RunResultBase | None,
+    *,
+    interactive: bool,
+) -> RunResultBase | None:
+    """Settle an agent that never recovered into a tool call.
+
+    Interactive runs park instead of dying: a human is attached and can message
+    any agent, so the scan stays resumable. Autonomous runs have nobody to
+    resume them, so they fail loudly.
+    """
+    if not interactive:
+        await coordinator.set_status(agent_id, "crashed")
+        await _notify_parent_on_terminal(coordinator, agent_id, "crashed")
+        raise MaxTurnsExceeded(
+            "Agent exhausted recovery attempts without calling finish_scan or agent_finish."
+        )
+
+    logger.warning(
+        "agent %s exhausted tool-call recovery attempts; parking until a message arrives",
+        agent_id,
+    )
+    await coordinator.park_waiting(agent_id, wait_kind="stalled")
+    # A parked child owes its parent a completion report it can no longer send. The
+    # parent is an agent, not a watching human, so nothing else tells it to stop
+    # waiting and it burns its full timeout on a message that is never coming.
+    await _notify_parent_on_stall(coordinator, agent_id)
+    return result
+
+
+_WAITING_AUTO_RESUME_TIMEOUT_S = 300.0
+
+# An agent that parks again after every auto-resume makes no progress, so stop
+# spending a model turn per timeout and leave it parked for a real message.
+_MAX_IDLE_AUTO_RESUMES = 3
+
+
+async def _plain_waiting_timeout(
+    coordinator: AgentCoordinator,
+    agent_id: str,
+) -> float | None:
+    """Auto-resume timeout for a parked agent; None waits until a message arrives.
+
+    Driven by what the agent is waiting on, not by where it sits in the graph:
+    the user can message any agent, so an agent awaiting a human parks
+    indefinitely whether or not it is the root. Only an agent awaiting other
+    agents is re-checked on a timer, and only until it has spent its idle
+    budget re-parking without hearing anything.
+    """
+    async with coordinator._lock:
+        status = coordinator.statuses.get(agent_id)
+        has_error = agent_id in coordinator.errors
+        runtime = coordinator.runtimes.get(agent_id)
+        gated = runtime.user_wake_required if runtime is not None else False
+        wait_kind = coordinator.wait_kinds.get(agent_id)
+        idle_resumes = coordinator.idle_resume_counts.get(agent_id, 0)
+    if status != "waiting" or has_error or gated:
+        return None
+    if wait_kind != "agents" or idle_resumes >= _MAX_IDLE_AUTO_RESUMES:
+        return None
+    return _WAITING_AUTO_RESUME_TIMEOUT_S
+
+
+async def _run_cycle_parked(
+    agent: Any,
+    coordinator: AgentCoordinator,
+    agent_id: str,
+    *,
+    input_data: Any,
+    run_config: RunConfig,
+    context: dict[str, Any],
+    max_turns: int,
+    session: Session | None,
+    event_sink: StreamEventSink | None,
+    hooks: RunHooks[dict[str, Any]] | None,
+) -> RunResultBase | None:
+    """Interactive run cycle that parks on any error instead of killing the runner."""
+    try:
+        return await _run_cycle(
             agent,
             coordinator,
             agent_id,
@@ -300,39 +613,17 @@ async def _run_noninteractive_until_lifecycle(
             context=context,
             max_turns=max_turns,
             session=session,
-            interactive=False,
+            interactive=True,
             event_sink=event_sink,
             hooks=hooks,
         )
-
-        status = await _agent_status(coordinator, agent_id)
-        if status != "running":
-            return result
-
-        invalid_final_outputs += 1
-        logger.warning(
-            "agent %s produced non-lifecycle final output in non-interactive mode; "
-            "forcing tool continuation (%d/%d): %s",
-            agent_id,
-            invalid_final_outputs,
-            invalid_final_output_limit,
-            _final_output_preview(result),
-        )
-
-        if invalid_final_outputs >= invalid_final_output_limit:
-            await coordinator.set_status(agent_id, "crashed")
-            await _notify_parent_on_crash(coordinator, agent_id, "crashed")
-            raise MaxTurnsExceeded(
-                "Agent exhausted non-interactive recovery attempts without calling "
-                "finish_scan or agent_finish."
-            )
-
-        input_data = await _append_noninteractive_tool_required_message(
-            session=session,
-            context=context,
-            attempt=invalid_final_outputs,
-            limit=invalid_final_output_limit,
-        )
+    except (BudgetExceededError, BudgetPausedError, SubagentBudgetReservedError):
+        raise
+    except Exception as exc:
+        logger.exception("error escaped the run cycle for %s; parking as failed", agent_id)
+        await coordinator.set_status(agent_id, "failed", error=str(exc) or type(exc).__name__)
+        await _notify_parent_on_terminal(coordinator, agent_id, "failed")
+        return None
 
 
 async def _run_cycle(  # noqa: PLR0912, PLR0915
@@ -350,7 +641,11 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
     hooks: RunHooks[dict[str, Any]] | None,
 ) -> RunResultBase | None:
     image_strips = 0
+    compactions = 0
+    model_retries = 0
     while True:
+        stream: Any = None
+        pre_run_items: list[Any] = []
         try:
             await coordinator.mark_running(agent_id)
             if session is not None:
@@ -360,6 +655,12 @@ 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)
+                with contextlib.suppress(Exception):
+                    pre_run_items = list(await session.get_items())
             stream = Runner.run_streamed(
                 agent,
                 input=input_data,
@@ -380,9 +681,9 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                                 logger.exception("stream event sink failed for %s", agent_id)
                     if stream.run_loop_exception is not None:
                         raise stream.run_loop_exception
-                except BudgetExceededError:
-                    # A RuntimeError subclass: re-raise explicitly so it is never
-                    # mistaken for the LiteLLM "after shutdown" race below.
+                    if refusal := _structured_provider_refusal(stream):
+                        raise ProviderRefusalError(refusal)
+                except (BudgetExceededError, BudgetPausedError, SubagentBudgetReservedError):
                     raise
                 except RuntimeError as stream_exc:
                     if "after shutdown" not in str(stream_exc):
@@ -401,6 +702,15 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                     )
             finally:
                 await coordinator.detach_stream(agent_id, stream)
+        except BudgetPausedError as exc:
+            logger.info("agent %s paused at the scan budget limit: %s", agent_id, exc)
+            await coordinator.pause_for_budget(agent_id)
+            raise
+        except SubagentBudgetReservedError as exc:
+            logger.info("sub-agent %s stopped at the budget reserve: %s", agent_id, exc)
+            await coordinator.set_status(agent_id, "stopped")
+            await _notify_root_on_budget_reserve(coordinator)
+            raise
         except BudgetExceededError as exc:
             logger.info(
                 "agent %s reached the scan budget limit; stopping the scan: %s", agent_id, exc
@@ -428,6 +738,48 @@ 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 model_retries < _MAX_TRANSIENT_MODEL_RETRIES and _is_transient_model_error(exc):
+                model_retries += 1
+                delay = _transient_model_retry_delay(model_retries)
+                logger.warning(
+                    "transient model/provider error for %s; replaying turn "
+                    "(attempt %d/%d, backoff %.1fs): %r",
+                    agent_id,
+                    model_retries,
+                    _MAX_TRANSIENT_MODEL_RETRIES,
+                    delay,
+                    exc,
+                )
+                await asyncio.sleep(delay)
+                if session is not None:
+                    input_data = []
+                continue
+            if session is not None:
+                await _salvage_stream_to_session(session, pre_run_items, stream, agent_id)
+            if isinstance(exc, ProviderRefusalError):
+                logger.warning("agent %s refused by the model provider: %s", agent_id, exc)
+                await coordinator.set_status(agent_id, "failed", error=str(exc))
+                await _notify_parent_on_terminal(coordinator, agent_id, "failed")
+                return None
             if not interactive:
                 raise
             if isinstance(exc, MaxTurnsExceeded):
@@ -437,31 +789,11 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
             else:
                 status = "crashed"
             logger.exception("agent run failed for %s; parking as %s", agent_id, status)
-            await coordinator.set_status(agent_id, status)
-            await _notify_parent_on_crash(coordinator, agent_id, status)
-            if context.get("parent_id") is None and status in {"failed", "crashed"}:
-                raise
+            await coordinator.set_status(agent_id, status, error=str(exc) or type(exc).__name__)
+            await _notify_parent_on_terminal(coordinator, agent_id, status)
             return None
         else:
-            await _settle_run_result(coordinator, agent_id, interactive)
-            return stream
-
-
-async def _settle_run_result(
-    coordinator: AgentCoordinator,
-    agent_id: str,
-    interactive: bool,
-) -> None:
-    async with coordinator._lock:
-        current_status = coordinator.statuses.get(agent_id)
-
-    if current_status != "running":
-        return
-
-    if not interactive:
-        return
-
-    await coordinator.set_status(agent_id, "waiting")
+            return cast("RunResultBase | None", stream)
 
 
 async def _agent_status(coordinator: AgentCoordinator, agent_id: str) -> Status | None:
@@ -479,23 +811,37 @@ def _final_output_preview(result: RunResultBase | None) -> str:
     return text[:300]
 
 
-async def _append_noninteractive_tool_required_message(
+async def _append_tool_required_message(
     *,
     session: Session | None,
     context: dict[str, Any],
     attempt: int,
     limit: int,
+    interactive: bool,
 ) -> list[dict[str, str]]:
     finish_tool = "finish_scan" if context.get("parent_id") is None else "agent_finish"
-    message = (
-        "Your previous response ended the autonomous Strix run without a lifecycle tool call. "
-        "That is invalid in non-interactive mode; plain text final answers are ignored. "
-        "Continue immediately and call exactly one tool. "
-        f"If your work is complete, call {finish_tool}. "
-        "If you are blocked waiting for another agent, call wait_for_message. "
-        "Otherwise use the appropriate execution or planning tool. "
-        f"This is recovery attempt {attempt}/{limit}."
-    )
+    if interactive:
+        message = (
+            "Your previous message ended a turn without a tool call. Plain text never ends "
+            "execution and never hands control to the user: it is shown to the user, and the "
+            "run continues. Continue immediately and call exactly one tool. "
+            "If you have something to tell the user and nothing to do until they reply, "
+            "call respond_to_user. "
+            "If you are blocked waiting for another agent, call wait_for_agents. "
+            f"If the whole engagement is complete, call {finish_tool}. "
+            "Otherwise use the appropriate execution or planning tool. "
+            f"This is recovery attempt {attempt}/{limit}."
+        )
+    else:
+        message = (
+            "Your previous response ended the autonomous Strix run without a lifecycle tool "
+            "call. That is invalid in non-interactive mode; plain text final answers are "
+            "ignored. Continue immediately and call exactly one tool. "
+            f"If your work is complete, call {finish_tool}. "
+            "If you are blocked waiting for another agent, call wait_for_agents. "
+            "Otherwise use the appropriate execution or planning tool. "
+            f"This is recovery attempt {attempt}/{limit}."
+        )
     item = {"role": "user", "content": message}
     if session is None:
         return [item]
@@ -504,12 +850,61 @@ async def _append_noninteractive_tool_required_message(
     return []
 
 
-async def _notify_parent_on_crash(
+_TERMINAL_NOTICE = {
+    "crashed": (
+        "[Agent crash] {name} ({agent_id}) terminated unexpectedly. "
+        "Stop waiting on this child unless you want to message it again."
+    ),
+    "failed": (
+        "[Agent failed] {name} ({agent_id}) stopped with an error and will not "
+        "send a completion report. Stop waiting on this child unless you want to "
+        "message it again."
+    ),
+    "stopped": (
+        "[Agent capped] {name} ({agent_id}) hit its turn limit and was stopped "
+        "before finishing. It will not send a completion report, so stop waiting "
+        "on this child; account for its capped subtask and continue."
+    ),
+}
+
+
+_STALL_NOTICE = (
+    "[Agent stalled] {name} ({agent_id}) kept ending turns without a tool call and is "
+    "parked until it receives a message. It will not send a completion report on its "
+    "own: either message it with a concrete next step to unblock it, or stop waiting on "
+    "it and account for its unfinished subtask."
+)
+
+
+async def _notify_parent_on_stall(
+    coordinator: AgentCoordinator,
+    agent_id: str,
+) -> None:
+    """Tell the parent that a child parked mid-task, so it stops waiting blindly."""
+    async with coordinator._lock:
+        parent = coordinator.parent_of.get(agent_id)
+        name = coordinator.names.get(agent_id, agent_id)
+    if parent is None:
+        return
+    await coordinator.send(
+        parent,
+        {
+            "from": agent_id,
+            "type": "stalled",
+            "priority": "high",
+            "content": _STALL_NOTICE.format(name=name, agent_id=agent_id),
+        },
+        interrupt=False,
+    )
+
+
+async def _notify_parent_on_terminal(
     coordinator: AgentCoordinator,
     agent_id: str,
     status: str,
 ) -> None:
-    if status != "crashed":
+    template = _TERMINAL_NOTICE.get(status)
+    if template is None:
         return
     async with coordinator._lock:
         parent = coordinator.parent_of.get(agent_id)
@@ -520,16 +915,36 @@ async def _notify_parent_on_crash(
         parent,
         {
             "from": agent_id,
-            "type": "crash",
+            "type": status,
             "priority": "high",
-            "content": (
-                f"[Agent crash] {name} ({agent_id}) terminated unexpectedly. "
-                "Stop waiting on this child unless you want to message it again."
-            ),
+            "content": template.format(name=name, agent_id=agent_id),
         },
+        interrupt=False,
     )
 
 
+def _reserve_notice() -> dict[str, Any]:
+    return {
+        "from": "system",
+        "type": "budget_reserve_stop",
+        "priority": "high",
+        "content": (
+            "[Budget reserve] The scan has reached the sub-agent budget reserve: every "
+            "sub-agent is being force-stopped as soon as its in-flight turn completes, and "
+            "none will send a completion report. Their confirmed vulnerabilities are "
+            "already filed as they were found. Do not wait on any sub-agents and do not "
+            "spawn new ones — wrap up now and call finish_scan."
+        ),
+    }
+
+
+async def _notify_root_on_budget_reserve(coordinator: AgentCoordinator) -> None:
+    root = await coordinator.claim_reserve_notification()
+    if root is None:
+        return
+    await coordinator.send(root, _reserve_notice())
+
+
 async def _start_child_runner(
     *,
     parent_ctx: dict[str, Any],
@@ -581,6 +996,8 @@ async def _start_child_runner(
             )
         except BudgetExceededError:
             logger.info("child %s stopped after reaching the scan budget limit", child_id)
+        except SubagentBudgetReservedError:
+            logger.info("child %s stopped at the sub-agent budget reserve", child_id)
 
     task_handle = asyncio.create_task(_child_loop(), name=f"agent-{name}-{child_id}")
     await coordinator.attach_runtime(child_id, task=task_handle)

strix/core/hooks.py

@@ -14,26 +14,210 @@ from strix.report.state import get_global_report_state
 if TYPE_CHECKING:
     from agents import RunContextWrapper
     from agents.agent import Agent
-    from agents.items import ModelResponse
+    from agents.items import ModelResponse, TResponseInputItem
 
 
 logger = logging.getLogger(__name__)
 
 
+_STAGE_LABELS: tuple[str, ...] = ("NOTICE", "URGENT", "CRITICAL")
+_TURN_WARN_BANDS: tuple[float, ...] = (0.70, 0.85, 0.95)
+_ROOT_BUDGET_WARN_BANDS: tuple[float, ...] = (0.70, 0.85, 0.95)
+_SUBAGENT_BUDGET_WARN_BANDS: tuple[float, ...] = (0.75, 0.80, 0.85)
+_SUBAGENT_BUDGET_RESERVE = 0.90
+
+
 class BudgetExceededError(RuntimeError):
     """Raised when the accumulated LLM cost reaches the configured budget."""
 
 
-class ReportUsageHooks(RunHooks[dict[str, Any]]):
-    """Persist SDK-native usage after every model response."""
+class SubagentBudgetReservedError(RuntimeError):
+    """Raised to stop a single sub-agent once the reserve threshold is crossed."""
 
-    def __init__(self, *, model: str, max_budget_usd: float | None = None) -> None:
+
+class BudgetPausedError(RuntimeError):
+    """Raised to park one agent when an interactive scan reaches its budget."""
+
+
+def recomputed_budget_flags(
+    cost: float,
+    max_budget_usd: float | None,
+    *,
+    interactive: bool,
+) -> tuple[bool, bool]:
+    """Return the (budget_stopped, reserve_stopped) flags a resumed scan should carry."""
+    if max_budget_usd is None:
+        return False, False
+    if interactive:
+        return False, False
+    budget_stopped = cost >= max_budget_usd
+    reserve_stopped = cost >= max_budget_usd * _SUBAGENT_BUDGET_RESERVE
+    return budget_stopped, reserve_stopped
+
+
+def _crossed_stage(fraction: float, bands: tuple[float, ...]) -> int | None:
+    crossed: int | None = None
+    for index, band in enumerate(bands):
+        if fraction >= band:
+            crossed = index
+    return crossed
+
+
+_ROOT_DIRECTIVES: tuple[str, ...] = (
+    (
+        "As the root agent, begin planning your wind-down of the whole scan: avoid "
+        "starting large new lines of investigation, and keep your required objectives on "
+        "track so you can call finish_scan comfortably before the limit."
+    ),
+    (
+        "As the root agent, prioritize wrapping up the whole scan now: stop opening new "
+        "lines of investigation, close out only what is essential, and move toward calling "
+        "finish_scan to compile and deliver the final report."
+    ),
+    (
+        "As the root agent, STOP all other work on the whole scan and finish immediately: "
+        "secure your findings and call finish_scan now — anything left unfinished when the "
+        "limit is hit is discarded."
+    ),
+)
+_SUBAGENT_DIRECTIVES: tuple[str, ...] = (
+    (
+        "As a sub-agent, begin planning your wind-down: avoid starting large new subtasks, "
+        "and if you are close to a confirmed, validated vulnerability, drive it to a result "
+        "you can report."
+    ),
+    (
+        "As a sub-agent, prioritize wrapping up your task now: report any confirmed, "
+        "validated vulnerability, finish work that is nearly done rather than starting "
+        "anything new, and prepare to call agent_finish."
+    ),
+    (
+        "As a sub-agent, STOP all other work and finish immediately: report any confirmed "
+        "vulnerability right now and call agent_finish to hand your results back to your "
+        "parent before you are cut off."
+    ),
+)
+
+
+def _wrapup_directive(context: RunContextWrapper[dict[str, Any]], stage: int) -> str:
+    is_root = context.context.get("parent_id") is None
+    directives = _ROOT_DIRECTIVES if is_root else _SUBAGENT_DIRECTIVES
+    return directives[stage]
+
+
+def _urgency(stage: int) -> str:
+    return _STAGE_LABELS[stage]
+
+
+class ReportUsageHooks(RunHooks[dict[str, Any]]):
+    """Persist SDK-native usage and warn/stop as turn and cost budgets are consumed."""
+
+    def __init__(
+        self,
+        *,
+        model: str,
+        max_budget_usd: float | None = None,
+        max_turns: int | None = None,
+        interactive: bool = False,
+    ) -> None:
         if max_budget_usd is not None and (
             not math.isfinite(max_budget_usd) or max_budget_usd <= 0
         ):
             raise ValueError("max_budget_usd must be a finite number greater than 0")
+        if max_turns is not None and max_turns <= 0:
+            raise ValueError("max_turns must be a positive integer")
         self._model = model
         self._max_budget_usd = max_budget_usd
+        self._budget_increment = max_budget_usd
+        self._max_turns = max_turns
+        self._interactive = interactive
+
+    def extend_budget(self) -> None:
+        if self._max_budget_usd is None or self._budget_increment is None:
+            return
+        self._max_budget_usd += self._budget_increment
+
+    async def on_llm_start(
+        self,
+        context: RunContextWrapper[dict[str, Any]],
+        agent: Agent[dict[str, Any]],  # noqa: ARG002
+        system_prompt: str | None,  # noqa: ARG002
+        input_items: list[TResponseInputItem],
+    ) -> None:
+        try:
+            self._maybe_warn_turns(context, input_items)
+            self._maybe_warn_budget(context, input_items)
+        except Exception:
+            logger.exception("budget/turn warning injection failed")
+
+    def _maybe_warn_turns(
+        self,
+        context: RunContextWrapper[dict[str, Any]],
+        input_items: list[TResponseInputItem],
+    ) -> None:
+        if not self._max_turns:
+            return
+        usage = getattr(context, "usage", None)
+        requests = getattr(usage, "requests", None)
+        if not isinstance(requests, int):
+            return
+        turns_used = requests + 1
+        stage = _crossed_stage(turns_used / self._max_turns, _TURN_WARN_BANDS)
+        if stage is None:
+            return
+        remaining = max(self._max_turns - turns_used, 0)
+        pct = round(100 * turns_used / self._max_turns)
+        content = (
+            f"[{_urgency(stage)}] Turn budget: {turns_used}/{self._max_turns} used ({pct}%). "
+            f"About {remaining} turn(s) remain before this agent is force-stopped and any "
+            f"in-progress work is discarded. {_wrapup_directive(context, stage)}"
+        )
+        input_items.append({"role": "user", "content": content})
+
+    def _maybe_warn_budget(
+        self,
+        context: RunContextWrapper[dict[str, Any]],
+        input_items: list[TResponseInputItem],
+    ) -> None:
+        if self._max_budget_usd is None:
+            return
+        report_state = get_global_report_state()
+        if report_state is None:
+            return
+        cost = report_state.get_total_llm_cost()
+        is_root = context.context.get("parent_id") is None
+        if self._interactive:
+            bands = _ROOT_BUDGET_WARN_BANDS
+        else:
+            bands = _ROOT_BUDGET_WARN_BANDS if is_root else _SUBAGENT_BUDGET_WARN_BANDS
+        stage = _crossed_stage(cost / self._max_budget_usd, bands)
+        if stage is None:
+            return
+        pct = round(100 * cost / self._max_budget_usd)
+        reserve_pct = round(_SUBAGENT_BUDGET_RESERVE * 100)
+        if self._interactive:
+            content = (
+                f"[{_urgency(stage)}] Scan cost budget: ${cost:.2f}/${self._max_budget_usd:.2f} "
+                f"spent ({pct}%). This budget is shared across every agent in the scan; when it "
+                "is reached all agents are paused until the user chooses to continue. "
+                f"{_wrapup_directive(context, stage)}"
+            )
+        elif is_root:
+            content = (
+                f"[{_urgency(stage)}] Scan cost budget: ${cost:.2f}/${self._max_budget_usd:.2f} "
+                f"spent ({pct}%). This budget is shared across every agent in the scan; when it "
+                "is reached the whole scan is stopped immediately, and sub-agents are stopped at "
+                f"{reserve_pct}% to reserve the remainder for your final report. "
+                f"{_wrapup_directive(context, stage)}"
+            )
+        else:
+            content = (
+                f"[{_urgency(stage)}] Scan cost budget: ${cost:.2f}/${self._max_budget_usd:.2f} "
+                f"spent ({pct}%). This budget is shared across every agent in the scan; "
+                f"sub-agents are stopped at {reserve_pct}% to leave the remainder for the root "
+                f"agent's final report. {_wrapup_directive(context, stage)}"
+            )
+        input_items.append({"role": "user", "content": content})
 
     async def on_llm_end(
         self,
@@ -66,6 +250,21 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
         if self._max_budget_usd is not None:
             cost = report_state.get_total_llm_cost()
             if cost >= self._max_budget_usd:
+                if self._interactive:
+                    raise BudgetPausedError(
+                        f"Scan budget of ${self._max_budget_usd:.2f} reached "
+                        f"(spent ${cost:.4f}); pausing until the user continues"
+                    )
                 raise BudgetExceededError(
                     f"Token budget of ${self._max_budget_usd:.2f} exceeded (spent ${cost:.4f})"
                 )
+            is_root = ctx.get("parent_id") is None
+            if not self._interactive and not is_root:
+                reserve_limit = self._max_budget_usd * _SUBAGENT_BUDGET_RESERVE
+                if cost >= reserve_limit:
+                    raise SubagentBudgetReservedError(
+                        f"Sub-agent budget reserve reached: spent ${cost:.4f} of "
+                        f"${self._max_budget_usd:.2f} "
+                        f"(>= {round(_SUBAGENT_BUDGET_RESERVE * 100)}% reserve); stopping this "
+                        "sub-agent so the root agent can finish the scan."
+                    )

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,
@@ -128,12 +131,15 @@ def make_model_settings(
     model_name: str,
     force_required_tool_choice: bool = False,
     request_timeout: float | None = None,
+    prompt_cache: bool = True,
+    extra_headers: dict[str, str] | None = None,
 ) -> ModelSettings:
     model_settings = ModelSettings(
         parallel_tool_calls=False,
         retry=DEFAULT_MODEL_RETRY,
         include_usage=True,
         extra_args=request_timeout_extra_args(request_timeout),
+        extra_headers=dict(extra_headers) if extra_headers else None,
     )
     if (
         reasoning_effort is not None
@@ -145,9 +151,38 @@ 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:
+        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 ``cache_control_injection_points`` for Claude prompt caching.
+
+    System prompt + rolling last-message breakpoint everywhere; ``tool_config``
+    only on Bedrock Converse (the only route whose LiteLLM transform consumes
+    it — elsewhere it leaks onto the wire and native Anthropic 400s). Unmapped
+    Bedrock models get no points at all: Bedrock rejects the passed-through
+    field outright.
+    """
+    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: 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 {"cache_control_injection_points": 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
@@ -29,7 +31,7 @@ from strix.core.execution import (
 from strix.core.execution import (
     spawn_child_agent as start_child_agent,
 )
-from strix.core.hooks import BudgetExceededError, ReportUsageHooks
+from strix.core.hooks import BudgetExceededError, ReportUsageHooks, recomputed_budget_flags
 from strix.core.inputs import (
     DEFAULT_MAX_TURNS,
     build_root_task,
@@ -38,8 +40,13 @@ from strix.core.inputs import (
 )
 from strix.core.paths import run_dir_for, runtime_state_dir
 from strix.core.sessions import open_agent_session
+from strix.report.state import get_global_report_state
 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:
@@ -179,6 +186,18 @@ async def run_strix_scan(
                 f"Cannot resume scan {scan_id}: missing SDK session database at {agents_db}",
             )
         await coordinator.restore(snap)
+        report_state = get_global_report_state()
+        if report_state is not None:
+            budget_stopped, reserve_stopped = recomputed_budget_flags(
+                report_state.get_total_llm_cost(),
+                max_budget_usd,
+                interactive=interactive,
+            )
+            await coordinator.reset_budget_stops(
+                budget_stopped=budget_stopped,
+                reserve_stopped=reserve_stopped,
+                budget_paused=interactive and coordinator.budget_paused,
+            )
         for aid, parent in coordinator.parent_of.items():
             if parent is None:
                 root_id = aid
@@ -203,6 +222,20 @@ 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 +249,8 @@ 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,
+            extra_headers=settings.llm.extra_headers,
         )
         run_config = RunConfig(
             model=resolved_model,
@@ -224,7 +259,14 @@ async def run_strix_scan(
             sandbox=SandboxRunConfig(client=bundle["client"], session=bundle["session"]),
             trace_include_sensitive_data=False,
         )
-        hooks = ReportUsageHooks(model=resolved_model, max_budget_usd=max_budget_usd)
+        hooks = ReportUsageHooks(
+            model=resolved_model,
+            max_budget_usd=max_budget_usd,
+            max_turns=max_turns,
+            interactive=interactive,
+        )
+        if interactive:
+            coordinator.set_budget_extender(hooks.extend_budget)
 
         scope_context = build_scope_context(scan_config)
         root_context = _merge_root_prompt_context(scope_context, extra_system_prompt_context)
@@ -238,7 +280,7 @@ async def run_strix_scan(
         )
 
         root_agent = build_strix_agent(
-            name="strix",
+            name="Strix",
             skills=skills,
             is_root=True,
             scan_mode=scan_mode,
@@ -252,7 +294,7 @@ async def run_strix_scan(
         if not is_resume:
             await coordinator.register(
                 root_id,
-                "strix",
+                "Strix",
                 parent_id=None,
                 task=root_task,
                 skills=skills,
@@ -399,6 +441,7 @@ 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

@@ -7,6 +7,7 @@ import logging
 from typing import TYPE_CHECKING, Any, cast
 from weakref import WeakKeyDictionary
 
+from agents.items import ItemHelpers
 from agents.memory import SQLiteSession
 
 
@@ -26,6 +27,18 @@ def open_agent_session(agent_id: str, path: Path) -> SQLiteSession:
     return SQLiteSession(session_id=agent_id, db_path=path)
 
 
+async def seed_initial_input(session: Session, initial_input: Any) -> bool:
+    """Commit an agent's opening identity/task input before its first run cycle."""
+    items = ItemHelpers.input_to_new_input_list(initial_input)
+    if not items:
+        return False
+    async with session_write_lock(session):
+        if await session.get_items():
+            return False
+        await session.add_items(items)
+    return True
+
+
 _IMAGE_REJECTED_TEXT = "[image rejected by the model]"
 _IMAGE_ELIDED_TEXT = "[older screenshot elided to bound context memory]"
 _INHERITED_IMAGE_TEXT = "[screenshot omitted from inherited context]"
@@ -92,6 +105,39 @@ 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/auth_cli.py

@@ -0,0 +1,419 @@
+"""`strix auth` — ChatGPT subscription sign-in (login / status / logout).
+
+Signing in only stores OAuth tokens (``~/.strix/subscription-auth.json``); model
+selection stays with ``STRIX_LLM``. A ``chatgpt/<model>`` STRIX_LLM runs on the
+subscription.
+"""
+
+from __future__ import annotations
+
+import argparse
+import base64
+import logging
+import threading
+import webbrowser
+from http.server import BaseHTTPRequestHandler, HTTPServer
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+from urllib.parse import parse_qs, urlparse
+
+from rich.console import Console
+from rich.panel import Panel
+from rich.text import Text
+
+from strix.config import codex, load_settings
+
+
+if TYPE_CHECKING:
+    from collections.abc import Callable
+
+
+logger = logging.getLogger(__name__)
+
+_CALLBACK_TIMEOUT_S = 300
+
+# CLI-facing name for the login provider. Internally this is the Codex OAuth
+# flow (``codex.PROVIDER``), but users know it as ChatGPT, so that's what the
+# command and messaging say. ``codex`` is accepted as an alias.
+LOGIN_PROVIDER = "chatgpt"
+_ACCEPTED_PROVIDERS = frozenset({LOGIN_PROVIDER, codex.PROVIDER})
+
+_USAGE = "Usage:\n  strix auth login chatgpt [--manual]\n  strix auth status\n  strix auth logout"
+
+
+def run_auth(argv: list[str]) -> int:
+    """Entry point for ``strix auth …``. Returns a process exit code."""
+    console = Console()
+    # Bare `strix auth` (no subcommand) defaults to login.
+    subcommand = argv[0] if argv else "login"
+    rest = argv[1:]
+
+    if subcommand in ("-h", "--help", "help"):
+        console.print(_USAGE)
+        return 0
+
+    handlers: dict[str, Callable[[], int]] = {
+        "login": lambda: _login(console, rest),
+        "status": lambda: _status(console),
+        "logout": lambda: _logout(console),
+    }
+    handler = handlers.get(subcommand)
+    if handler is not None:
+        return handler()
+
+    console.print(f"[red]Unknown auth command:[/] {subcommand}\n")
+    console.print(_USAGE)
+    return 2
+
+
+def _login(console: Console, argv: list[str]) -> int:
+    parser = argparse.ArgumentParser(prog="strix auth login", add_help=True)
+    parser.add_argument(
+        "provider",
+        nargs="?",
+        default=LOGIN_PROVIDER,
+        help="Model provider to sign in with (default: chatgpt).",
+    )
+    parser.add_argument(
+        "--manual",
+        action="store_true",
+        help="Skip the local callback server and paste the redirect URL by hand.",
+    )
+    try:
+        args = parser.parse_args(argv)
+    except SystemExit as exc:  # argparse already printed the message
+        return int(exc.code or 2)
+
+    if args.provider.lower() not in _ACCEPTED_PROVIDERS:
+        console.print(
+            f"[red]Unsupported provider:[/] {args.provider}. "
+            f"Only '{LOGIN_PROVIDER}' (ChatGPT subscription) is supported."
+        )
+        return 2
+
+    verifier, challenge = codex.generate_pkce()
+    state = codex.create_state()
+    authorize_url = codex.build_authorize_url(challenge, state)
+
+    console.print()
+    console.print("[bold]Signing in with ChatGPT[/] [dim](provider: chatgpt)[/]")
+    console.print(
+        "[dim]This uses your ChatGPT Plus/Pro plan for inference instead of a metered API key.[/]"
+    )
+    console.print()
+
+    try:
+        record = _run_oauth_flow(console, authorize_url, verifier, state, manual=args.manual)
+    except codex.CodexAuthError as exc:
+        return _fail(console, exc)
+    except KeyboardInterrupt:
+        console.print("\n[yellow]Sign-in cancelled.[/]")
+        return 130
+
+    codex.save_record(record)
+    _print_success(console)
+    return 0
+
+
+def _run_oauth_flow(
+    console: Console,
+    authorize_url: str,
+    verifier: str,
+    state: str,
+    *,
+    manual: bool,
+) -> dict[str, Any]:
+    """Drive the browser (or manual) OAuth flow and return a token record."""
+    server = None if manual else _try_start_callback_server()
+
+    console.print("Open this URL in your browser to authorize:")
+    console.print(f"[cyan]{authorize_url}[/]")
+    console.print()
+    if not manual:
+        try:
+            webbrowser.open(authorize_url)
+        except Exception:  # noqa: BLE001 - opening a browser is best-effort
+            logger.debug("could not open browser", exc_info=True)
+
+    if server is not None:
+        console.print("[dim]Waiting for you to finish signing in…[/]")
+        result = server.wait(_CALLBACK_TIMEOUT_S)
+        server.shutdown()
+        if result is not None:
+            code, returned_state, error = result
+            if error:
+                raise codex.CodexAuthError("oauth_error", error)
+            return _finish(code, returned_state, verifier, state, require_state=True)
+        console.print("[yellow]Timed out waiting for the browser. Falling back to manual paste.[/]")
+
+    # Manual fallback: the user completes sign-in and pastes the redirect URL
+    # (the browser lands on a localhost page that won't load if no server is up;
+    # the address bar still holds the code+state).
+    console.print()
+    try:
+        pasted = console.input("Paste the full redirect URL (or code#state): ").strip()
+    except EOFError as exc:
+        raise codex.CodexAuthError("no_input", "no redirect URL provided") from exc
+    code, returned_state = codex.parse_redirect_input(pasted)
+    return _finish(code, returned_state, verifier, state, require_state=False)
+
+
+def _finish(
+    code: str | None,
+    returned_state: str | None,
+    verifier: str,
+    expected_state: str,
+    *,
+    require_state: bool,
+) -> dict[str, Any]:
+    if not code:
+        raise codex.CodexAuthError("no_code", "no authorization code found in the redirect")
+    # The loopback callback from OpenAI always carries state, so a missing or
+    # mismatched value there is forged (CSRF) and must be rejected. Manual paste
+    # is user-initiated (the user copies their own redirect), so state is only
+    # validated when the pasted value includes it.
+    if require_state and returned_state is None:
+        raise codex.CodexAuthError("state_mismatch", "missing state in callback; possible CSRF")
+    if returned_state is not None and returned_state != expected_state:
+        raise codex.CodexAuthError("state_mismatch", "state did not match; possible CSRF")
+    return codex.exchange_code(code, verifier)
+
+
+class _CallbackServer:
+    """A one-shot local HTTP server that catches the OAuth redirect."""
+
+    def __init__(self, httpd: HTTPServer, event: threading.Event, holder: dict[str, Any]) -> None:
+        self._httpd = httpd
+        self._event = event
+        self._holder = holder
+        self._thread = threading.Thread(target=httpd.serve_forever, daemon=True)
+        self._thread.start()
+
+    def wait(self, timeout: float) -> tuple[str | None, str | None, str | None] | None:
+        if not self._event.wait(timeout):
+            return None
+        return (
+            self._holder.get("code"),
+            self._holder.get("state"),
+            self._holder.get("error"),
+        )
+
+    def shutdown(self) -> None:
+        self._httpd.shutdown()
+        self._httpd.server_close()
+
+
+def _try_start_callback_server() -> _CallbackServer | None:
+    event = threading.Event()
+    holder: dict[str, Any] = {}
+
+    class Handler(BaseHTTPRequestHandler):
+        def log_message(self, *args: Any) -> None:  # silence default stderr logging
+            pass
+
+        def do_GET(self) -> None:
+            parsed = urlparse(self.path)
+            if parsed.path != codex.CALLBACK_PATH:
+                self.send_response(404)
+                self.end_headers()
+                return
+            query = parse_qs(parsed.query)
+            holder["code"] = _first(query, "code")
+            holder["state"] = _first(query, "state")
+            holder["error"] = _first(query, "error_description") or _first(query, "error")
+            body = _render_callback_html().encode("utf-8")
+            self.send_response(200)
+            self.send_header("Content-Type", "text/html; charset=utf-8")
+            self.send_header("Content-Length", str(len(body)))
+            self.end_headers()
+            self.wfile.write(body)
+            event.set()
+
+    try:
+        httpd = HTTPServer(("127.0.0.1", codex.CALLBACK_PORT), Handler)
+    except OSError:
+        logger.debug("could not bind callback port %d", codex.CALLBACK_PORT, exc_info=True)
+        return None
+    return _CallbackServer(httpd, event, holder)
+
+
+def _first(query: dict[str, list[str]], key: str) -> str | None:
+    values = query.get(key)
+    return values[0] if values else None
+
+
+def _status(console: Console) -> int:
+    record = codex.read_record()
+    if record is None:
+        console.print("[yellow]Not signed in.[/] Run [cyan]strix auth login chatgpt[/] to sign in.")
+        return 1
+    settings = load_settings()
+    console.print("[green]Signed in[/] with a ChatGPT subscription.")
+    console.print(f"  Account: [bold]{record.get('account_id')}[/]")
+    if codex.subscription_model(settings.llm.model):
+        console.print(f"  Runs use the subscription (STRIX_LLM=[bold]{settings.llm.model}[/]).")
+    else:
+        console.print(
+            "  [yellow]Note:[/] set [cyan]STRIX_LLM[/] to e.g. [cyan]chatgpt/gpt-5.4[/] "
+            "to run on the subscription."
+        )
+    return 0
+
+
+def _logout(console: Console) -> int:
+    codex.logout()
+    console.print("[green]Signed out.[/] Stored subscription credentials removed.")
+    return 0
+
+
+def _fail(console: Console, exc: codex.CodexAuthError) -> int:
+    error_text = Text()
+    error_text.append("SIGN-IN FAILED", style="bold red")
+    error_text.append("\n\n", style="white")
+    error_text.append(f"{exc}", style="white")
+    console.print()
+    console.print(
+        Panel(
+            error_text,
+            title="[bold white]STRIX",
+            title_align="left",
+            border_style="red",
+            padding=(1, 2),
+        )
+    )
+    return 1
+
+
+def _print_success(console: Console) -> None:
+    text = Text()
+    text.append("Signed in with your ChatGPT subscription", style="bold #22c55e")
+    text.append("\n\n", style="white")
+    text.append("Set ", style="white")
+    text.append("STRIX_LLM", style="bold white")
+    text.append(" to a ", style="white")
+    text.append("chatgpt/", style="bold cyan")
+    text.append(" model (e.g. ", style="white")
+    text.append("chatgpt/gpt-5.4", style="bold cyan")
+    text.append(") — runs are billed to your ChatGPT plan.", style="white")
+    text.append("\n\n", style="white")
+    text.append("Run a scan as usual, e.g. ", style="white")
+    text.append("strix --target https://example.com", style="bold cyan")
+    console.print()
+    console.print(
+        Panel(
+            text,
+            title="[bold white]STRIX",
+            title_align="left",
+            border_style="#22c55e",
+            padding=(1, 2),
+        )
+    )
+    console.print()
+
+
+_LOGO_PATH = Path(__file__).resolve().parent.parent / "viewer" / "static" / "logo.png"
+
+
+def _logo_img_tag() -> str:
+    """Return an ``<img>`` for the Strix logo as an inline data URI, or "".
+
+    The callback page is served offline by the local OAuth server, so the logo
+    is embedded rather than linked. Missing/unreadable file degrades to just the
+    "Strix" wordmark.
+    """
+    try:
+        data = _LOGO_PATH.read_bytes()
+    except OSError:
+        return ""
+    encoded = base64.b64encode(data).decode("ascii")
+    return f'<img class="logo" src="data:image/png;base64,{encoded}" alt="" />'
+
+
+def _render_callback_html() -> str:
+    return _CALLBACK_HTML.replace("<!--LOGO-->", _logo_img_tag())
+
+
+_CALLBACK_HTML = """<!doctype html>
+<html lang="en"><head><meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>Strix — signed in</title>
+<style>
+  :root { color-scheme: dark; }
+  * { box-sizing: border-box; }
+  body {
+    margin: 0; min-height: 100vh; padding: 24px;
+    font-family: 'Geist', 'Geist Sans', ui-sans-serif, system-ui, -apple-system,
+      "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
+    -webkit-font-smoothing: antialiased; -moz-osx-font-smoothing: grayscale;
+    background: #000; color: #ededed;
+    display: flex; flex-direction: column; align-items: center; justify-content: center;
+  }
+  .topbar {
+    position: absolute; top: 20px; left: 22px;
+    display: flex; align-items: center; gap: 6px; text-decoration: none;
+  }
+  .topbar .logo { width: 40px; height: 40px; display: block; }
+  .topbar span {
+    font-size: 1.1rem; font-weight: 600; letter-spacing: -.01em; color: #fff;
+    transition: color .15s ease;
+  }
+  .topbar:hover span { color: #c9c9c9; }
+  .brand {
+    font-size: 2.1rem; font-weight: 700; letter-spacing: -.02em; color: #fff;
+    text-align: center; margin: 0 0 10px;
+  }
+  h1 {
+    font-size: 1.35rem; font-weight: 600; letter-spacing: -.01em; color: #f5f5f5;
+    text-align: center; margin: 0 0 28px;
+  }
+  .card {
+    width: 100%; max-width: 430px; text-align: center;
+    background: #171717; border: 1px solid rgba(255, 255, 255, .06);
+    border-radius: 24px; padding: 40px 40px 34px;
+  }
+  .badge {
+    margin: 0 auto 22px; width: 52px; height: 52px; border-radius: 50%;
+    display: flex; align-items: center; justify-content: center; font-size: 23px; color: #fff;
+    background: rgba(255, 255, 255, .05); border: 1px solid rgba(255, 255, 255, .14);
+  }
+  .msg { margin: 0 auto; max-width: 34ch; color: #b5b5b5; line-height: 1.6; font-size: .98rem; }
+  .rule { height: 1px; background: rgba(255, 255, 255, .07); margin: 26px 0 0; }
+  .tagline { margin: 22px 0 0; color: #7c7c7c; font-size: .9rem; line-height: 1.55; }
+  .tagline b { color: #ededed; font-weight: 500; }
+  .links {
+    margin-top: 18px; display: flex; gap: 8px; justify-content: center;
+    align-items: center; flex-wrap: wrap; font-size: .84rem;
+  }
+  .links a { color: #a3a3a3; text-decoration: none; transition: color .15s ease; }
+  .links a:hover { color: #fff; }
+  .links .dot { color: #3a3a3a; }
+  .close { margin: 24px 0 0; color: #5a5a5a; font-size: .78rem; text-align: center; }
+</style></head>
+<body>
+  <a class="topbar" href="https://strix.ai" target="_blank" rel="noopener"
+     aria-label="Strix — strix.ai">
+    <!--LOGO-->
+    <span>Strix</span>
+  </a>
+  <div class="brand">Strix</div>
+  <h1>You're signed in</h1>
+  <main class="card">
+    <div class="badge">✓</div>
+    <p class="msg">Strix is connected to your ChatGPT subscription. Head back to your
+      terminal — your security test runs there.</p>
+    <div class="rule"></div>
+    <p class="tagline">Autonomous AI hackers that <b>find and fix</b> your app's
+      vulnerabilities.</p>
+    <nav class="links">
+      <a href="https://strix.ai" target="_blank" rel="noopener">strix.ai</a>
+      <span class="dot">·</span>
+      <a href="https://docs.strix.ai" target="_blank" rel="noopener">docs</a>
+      <span class="dot">·</span>
+      <a href="https://discord.gg/strix-ai" target="_blank" rel="noopener">community</a>
+    </nav>
+  </main>
+  <p class="close">You can close this tab.</p>
+</body></html>"""
+
+
+__all__ = ["run_auth"]

strix/interface/cli.py

@@ -13,6 +13,7 @@ from rich.panel import Panel
 from rich.text import Text
 
 from strix.config import load_settings
+from strix.core.inputs import DEFAULT_MAX_TURNS
 from strix.core.runner import run_strix_scan
 from strix.report.state import ReportState, set_global_report_state
 from strix.runtime import session_manager
@@ -184,6 +185,7 @@ async def run_cli(args: Any) -> None:  # noqa: PLR0915
                     local_sources=getattr(args, "local_sources", None) or [],
                     interactive=bool(getattr(args, "interactive", False)),
                     max_budget_usd=getattr(args, "max_budget_usd", None),
+                    max_turns=getattr(args, "max_turns", DEFAULT_MAX_TURNS),
                 )
             finally:
                 stop_updates.set()

strix/interface/main.py

@@ -5,9 +5,9 @@ Strix Agent Interface
 
 import argparse
 import asyncio
+import os
 import shutil
 import sys
-import time
 from datetime import UTC, datetime
 from pathlib import Path
 
@@ -20,6 +20,7 @@ from rich.text import Text
 
 from strix.config import (
     apply_config_override,
+    codex,
     load_settings,
     persist_current,
 )
@@ -30,9 +31,17 @@ from strix.config.models import (
     is_known_openai_bare_model,
     is_recommended_or_frontier_model,
 )
+from strix.core.inputs import DEFAULT_MAX_TURNS, make_model_settings
 from strix.core.paths import run_dir_for, runtime_state_dir
 from strix.interface.cli import run_cli
 from strix.interface.tui import run_tui
+from strix.interface.update_check import (
+    is_binary_install,
+    notify_update,
+    prompt_update_if_available,
+    self_update,
+    start_background_check,
+)
 from strix.interface.utils import (
     assign_workspace_subdirs,
     build_final_stats_text,
@@ -85,6 +94,16 @@ def validate_environment() -> None:
 
     settings = load_settings()
 
+    if codex.subscription_model(settings.llm.model):
+        if not codex.is_authenticated():
+            console.print(
+                f"[red]STRIX_LLM={settings.llm.model} uses your ChatGPT subscription, "
+                "but you're not signed in.[/] Run [cyan]strix auth login chatgpt[/] first."
+            )
+            sys.exit(1)
+        logger.info("Environment OK (ChatGPT subscription)")
+        return
+
     if not settings.llm.model:
         missing_required_vars.append("STRIX_LLM")
 
@@ -192,7 +211,7 @@ def validate_environment() -> None:
             padding=(1, 2),
         )
 
-        logger.error("Missing required env vars: %s", missing_required_vars)
+        logger.debug("Missing required env vars: %s", missing_required_vars)
         console.print("\n")
         console.print(panel)
         console.print()
@@ -205,7 +224,7 @@ def validate_environment() -> None:
 
 def check_docker_installed() -> None:
     if shutil.which("docker") is None:
-        logger.error("Docker CLI not found in PATH")
+        logger.debug("Docker CLI not found in PATH")
         console = Console()
         error_text = Text()
         error_text.append("DOCKER NOT INSTALLED", style="bold red")
@@ -267,6 +286,29 @@ def _provider_import_hint(exc: BaseException, model: str) -> str | None:
     return None
 
 
+def _subscription_error_hint(exc: BaseException) -> str | None:
+    """Return an actionable hint for a known ChatGPT-subscription error, or None."""
+    if not codex.subscription_model(load_settings().llm.model):
+        return None
+    joined = " ".join(_exception_messages(exc)).lower()
+    if "not supported when using codex with a chatgpt account" in joined:
+        return (
+            "This model isn't available on your ChatGPT subscription. "
+            "Set STRIX_LLM to a model your plan includes (e.g. chatgpt/gpt-5.4)."
+        )
+    if (
+        "error code: 401" in joined
+        or "http 401" in joined
+        or "unauthorized" in joined
+        or "invalid_grant" in joined
+    ):
+        return (
+            "Your ChatGPT sign-in has expired or was revoked. Sign in again:\n"
+            "  strix auth login chatgpt"
+        )
+    return None
+
+
 async def warm_up_llm(show_model_warning: bool = True) -> None:
     console = Console()
     logger.info("Warming up LLM connection")
@@ -276,8 +318,8 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
         settings = load_settings()
         configure_sdk_model_defaults(settings)
         llm = settings.llm
-
         raw_model = (llm.model or "").strip()
+
         if (
             raw_model
             and "/" not in raw_model
@@ -340,7 +382,13 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
             model.get_response(
                 system_instructions="You are a helpful assistant.",
                 input="Reply with just 'OK'.",
-                model_settings=ModelSettings(),
+                model_settings=make_model_settings(
+                    None,
+                    model_name=raw_model,
+                    request_timeout=llm.timeout,
+                    prompt_cache=False,
+                    extra_headers=llm.extra_headers,
+                ),
                 tools=[],
                 output_schema=None,
                 handoffs=[],
@@ -353,23 +401,75 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
         )
         logger.info("LLM warm-up succeeded for model %s", (llm.model or "").strip())
 
+        if settings.dedupe.model:
+            from strix.report.dedupe import _dedupe_extra_args
+
+            dedupe_model = settings.dedupe.model.strip()
+            raw_model = dedupe_model
+            deduper = StrixProvider().get_model(dedupe_model)
+            # Match the runtime path: send the dedupe key/endpoint per call so a
+            # separate-provider dedupe model authenticates during warm-up too.
+            deduper_extra = _dedupe_extra_args(settings.dedupe)
+            # A dedicated dedupe model may route to another provider, which must
+            # never receive the main endpoint's headers; it has its own
+            # DEDUPE_LLM_EXTRA_HEADERS.
+            deduper_settings = make_model_settings(
+                None,
+                model_name=dedupe_model,
+                request_timeout=llm.timeout,
+                prompt_cache=False,
+                extra_headers=settings.dedupe.extra_headers,
+            )
+            if deduper_extra:
+                merged = {**(deduper_settings.extra_args or {}), **deduper_extra}
+                deduper_settings = deduper_settings.resolve(ModelSettings(extra_args=merged))
+            await asyncio.wait_for(
+                deduper.get_response(
+                    system_instructions="You are a helpful assistant.",
+                    input="Reply with just 'OK'.",
+                    model_settings=deduper_settings,
+                    tools=[],
+                    output_schema=None,
+                    handoffs=[],
+                    tracing=ModelTracing.DISABLED,
+                    previous_response_id=None,
+                    conversation_id=None,
+                    prompt=None,
+                ),
+                timeout=llm.timeout,
+            )
+            logger.info("LLM warm-up succeeded for dedupe model %s", dedupe_model)
+
     except Exception as e:
-        logger.exception("LLM warm-up failed")
+        logger.debug("LLM warm-up failed", exc_info=True)
         error_text = Text()
-        error_text.append("LLM CONNECTION FAILED", style="bold red")
-        error_text.append("\n\n", style="white")
-        error_text.append("Could not establish connection to the language model.\n", style="white")
-        error_text.append("Please check your configuration and try again.\n", style="white")
-        hint = _provider_import_hint(e, raw_model)
-        if hint is not None:
-            error_text.append(f"\n{hint}\n", style="bold yellow")
-        error_text.append(f"\nError: {e}", style="dim white")
+        sub_hint = _subscription_error_hint(e)
+        if sub_hint is not None:
+            # The model/backend answered with a clear, actionable rejection —
+            # show that instead of a generic "connection failed".
+            border_style = "yellow"
+            error_text.append("MODEL NOT AVAILABLE ON SUBSCRIPTION", style="bold yellow")
+            error_text.append("\n\n", style="white")
+            error_text.append(f"{sub_hint}\n", style="white")
+            error_text.append(f"\nDetails: {e}", style="dim white")
+        else:
+            border_style = "red"
+            error_text.append("LLM CONNECTION FAILED", style="bold red")
+            error_text.append("\n\n", style="white")
+            error_text.append(
+                "Could not establish connection to the language model.\n", style="white"
+            )
+            error_text.append("Please check your configuration and try again.\n", style="white")
+            hint = _provider_import_hint(e, raw_model)
+            if hint is not None:
+                error_text.append(f"\n{hint}\n", style="bold yellow")
+            error_text.append(f"\nError: {e}", style="dim white")
 
         panel = Panel(
             error_text,
             title="[bold white]STRIX",
             title_align="left",
-            border_style="red",
+            border_style=border_style,
             padding=(1, 2),
         )
 
@@ -400,6 +500,16 @@ def _positive_budget(value: str) -> float:
     return budget
 
 
+def _positive_int(value: str) -> int:
+    try:
+        parsed = int(value)
+    except ValueError as exc:
+        raise argparse.ArgumentTypeError(f"invalid int value: {value!r}") from exc
+    if parsed <= 0:
+        raise argparse.ArgumentTypeError("must be an integer greater than 0")
+    return parsed
+
+
 def parse_arguments() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
         description="Strix Multi-Agent Cybersecurity Penetration Testing Tool",
@@ -448,6 +558,14 @@ Examples:
         version=f"strix {get_version()}",
     )
 
+    parser.add_argument(
+        "--update",
+        action="store_true",
+        help="Update strix to the latest version and exit. Self-updates the "
+        "standalone binary install; for pip/pipx/uv installs, prints the "
+        "matching upgrade command instead.",
+    )
+
     parser.add_argument(
         "-t",
         "--target",
@@ -547,10 +665,27 @@ Examples:
     )
 
     parser.add_argument(
-        "--max-budget-usd",
+        "--max-budget",
+        dest="max_budget_usd",
+        metavar="USD",
         type=_positive_budget,
         default=None,
-        help="Maximum LLM cost in USD (> 0). The scan stops cleanly when this limit is reached.",
+        help=(
+            "Maximum LLM cost in USD (> 0). The scan stops cleanly when this limit is reached. "
+            "Graduated wrap-up warnings are sent to all agents as it is approached."
+        ),
+    )
+
+    parser.add_argument(
+        "--max-turns",
+        dest="max_turns",
+        metavar="N",
+        type=_positive_int,
+        default=DEFAULT_MAX_TURNS,
+        help=(
+            "Maximum turns per agent (> 0, default %(default)s). Each agent is force-stopped "
+            "when it reaches this limit, with graduated wrap-up warnings as it is approached."
+        ),
     )
 
     parser.add_argument(
@@ -566,6 +701,9 @@ Examples:
 
     args = parser.parse_args()
 
+    if args.update:
+        sys.exit(0 if self_update() else 1)
+
     if args.instruction and args.instruction_file:
         parser.error(
             "Cannot specify both --instruction and --instruction-file. Use one or the other."
@@ -664,6 +802,7 @@ def _persist_run_record(args: argparse.Namespace) -> None:
         "status": "running",
         "start_time": datetime.now(UTC).isoformat(),
         "end_time": None,
+        "auth_mode": codex.auth_mode(load_settings().llm.model),
         "targets_info": args.targets_info,
         "scan_mode": args.scan_mode,
         "instruction": args.instruction,
@@ -721,9 +860,7 @@ def _load_resume_state(args: argparse.Namespace, parser: argparse.ArgumentParser
         args.scan_mode = persisted_scan_mode
 
 
-def display_completion_message(
-    args: argparse.Namespace, results_path: Path, web_url: str | None = None
-) -> None:
+def display_completion_message(args: argparse.Namespace, results_path: Path) -> None:
     console = Console()
     report_state = get_global_report_state()
 
@@ -762,28 +899,12 @@ def display_completion_message(
     results_text.append(str(results_path), style="#60a5fa")
     panel_parts.extend(["\n", results_text])
 
-    if web_url:
-        web_text = Text()
-        web_text.append("\n")
-        web_text.append("View in web", style="dim")
-        web_text.append("  ")
-        # OSC-8 hyperlink: clickable in modern terminals, falls back to the URL.
-        web_text.append(web_url, style=f"#60a5fa link {web_url}")
-        panel_parts.extend(["\n", web_text])
-
-        reopen_text = Text()
-        reopen_text.append("\n")
-        reopen_text.append("Reopen", style="dim")
-        reopen_text.append("       ")
-        reopen_text.append(f"strix view {args.run_name}", style="#22c55e")
-        panel_parts.extend(["\n", reopen_text])
-    else:
-        view_text = Text()
-        view_text.append("\n")
-        view_text.append("View", style="dim")
-        view_text.append("         ")
-        view_text.append(f"strix view {args.run_name}", style="#22c55e")
-        panel_parts.extend(["\n", view_text])
+    view_text = Text()
+    view_text.append("\n")
+    view_text.append("View", style="dim")
+    view_text.append("    ")
+    view_text.append(f"strix view {args.run_name}", style="#22c55e")
+    panel_parts.extend(["\n", view_text])
 
     if not scan_completed:
         resume_text = Text()
@@ -814,6 +935,8 @@ def display_completion_message(
         "[#60a5fa]discord.gg/strix-ai[/]"
     )
     console.print()
+    if not args.non_interactive:
+        notify_update(console)
 
 
 def pull_docker_image() -> None:
@@ -841,7 +964,7 @@ def pull_docker_image() -> None:
                 last_update = process_pull_line(line, layers_info, status, last_update)
 
         except DockerException as e:
-            logger.exception("Failed to pull docker image %s", image)
+            logger.debug("Failed to pull docker image %s", image, exc_info=True)
             console.print()
             error_text = Text()
             error_text.append("FAILED TO PULL IMAGE", style="bold red")
@@ -875,16 +998,29 @@ 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 auth …` manages model-subscription sign-in and exits; it needs no
+    # target, Docker, or scan setup.
+    if len(sys.argv) > 1 and sys.argv[1] == "auth":
+        from strix.interface.auth_cli import run_auth
+
+        sys.exit(run_auth(sys.argv[2:]))
+
     args = parse_arguments()
 
     if args.config:
         apply_config_override(validate_config_file(args.config))
 
+    start_background_check()
+    if not args.non_interactive and prompt_update_if_available(Console()):
+        if is_binary_install() and sys.platform != "win32":
+            os.execv(sys.executable, sys.argv)  # noqa: S606  # nosec B606
+        sys.exit(0)
+
     check_docker_installed()
     pull_docker_image()
 
@@ -941,6 +1077,7 @@ def main() -> None:
 
     _telemetry_start_kwargs = {
         "model": load_settings().llm.model,
+        "auth_mode": codex.auth_mode(load_settings().llm.model),
         "scan_mode": args.scan_mode,
         "is_whitebox": is_whitebox_scan(args.targets_info),
         "interactive": not args.non_interactive,
@@ -975,39 +1112,7 @@ def main() -> None:
 
     results_path = run_dir_for(args.run_name)
 
-    # For an interactive run, host the local viewer so the completion panel can
-    # show a clickable "View in web" link. Skipped in non-interactive/CI runs
-    # (no TTY to serve and it would block the process).
-    viewer_httpd = None
-    web_url = None
-    if not args.non_interactive and sys.stdout.isatty():
-        from strix.viewer.server import authorized_url, bundle_is_built, serve
-
-        if bundle_is_built():
-            try:
-                viewer_httpd, base_url, token = serve(results_path, open_browser=False)
-                # The completion panel's "View in web" link must authorize the
-                # browser, so hand it the tokened URL rather than the bare host.
-                web_url = authorized_url(base_url, token)
-                posthog.viewer_opened(source="post_scan", live=False)
-            except Exception:
-                logger.debug("could not start local viewer", exc_info=True)
-                viewer_httpd, web_url = None, None
-
-    display_completion_message(args, results_path, web_url=web_url)
-
-    if viewer_httpd is not None:
-        console = Console()
-        console.print("[dim]Hosting the local viewer. Press Ctrl-C to stop.[/]")
-        console.print()
-        try:
-            while True:
-                time.sleep(1)
-        except KeyboardInterrupt:
-            console.print("\n[dim]Viewer stopped.[/]")
-        finally:
-            viewer_httpd.shutdown()
-            viewer_httpd.server_close()
+    display_completion_message(args, results_path)
 
     if args.non_interactive:
         report_state = get_global_report_state()

strix/interface/tui/app.py

@@ -15,6 +15,7 @@ from typing import TYPE_CHECKING, Any, ClassVar
 
 
 if TYPE_CHECKING:
+    from pygments.token import _TokenType
     from textual.timer import Timer
 
 from rich.align import Align
@@ -34,6 +35,7 @@ from textual.widgets.tree import TreeNode
 from strix.config import load_settings
 from strix.config.models import is_recommended_or_frontier_model
 from strix.core.hooks import BudgetExceededError
+from strix.core.inputs import DEFAULT_MAX_TURNS
 from strix.core.runner import run_strix_scan
 from strix.interface.tui.live_view import TuiLiveView
 from strix.interface.tui.messages import send_user_message_to_agent
@@ -42,6 +44,12 @@ from strix.interface.tui.renderers.agent_message_renderer import AgentMessageRen
 from strix.interface.tui.renderers.user_message_renderer import UserMessageRenderer
 from strix.interface.utils import build_tui_stats_text
 from strix.report.state import ReportState, set_global_report_state
+from strix.report.writer import (
+    guess_language_name,
+    parse_fenced_code,
+    resolve_lexer,
+    safe_fence,
+)
 from strix.runtime import session_manager
 
 
@@ -330,12 +338,11 @@ class VulnerabilityDetailScreen(ModalScreen):  # type: ignore[misc]
             return "#65a30d"
         return "#6b7280"
 
-    def _highlight_python(self, code: str) -> Text:
+    def _highlight_python(self, code: str, language: str | None = None) -> Text:
         try:
-            from pygments.lexers import PythonLexer
             from pygments.styles import get_style_by_name
 
-            lexer = PythonLexer()
+            lexer = resolve_lexer(language, code)
             style = get_style_by_name("native")
             colors = {
                 token: f"#{style_def['color']}" for token, style_def in style if style_def["color"]
@@ -346,7 +353,7 @@ class VulnerabilityDetailScreen(ModalScreen):  # type: ignore[misc]
                 if not token_value:
                     continue
                 color = None
-                tt = token_type
+                tt: _TokenType | None = token_type
                 while tt:
                     if tt in colors:
                         color = colors[tt]
@@ -501,10 +508,11 @@ class VulnerabilityDetailScreen(ModalScreen):  # type: ignore[misc]
 
         poc_script_code = vuln.get("poc_script_code", "")
         if poc_script_code:
+            poc_language, poc_code = parse_fenced_code(poc_script_code)
             text.append("\n\n")
             text.append("PoC Code", style=self.FIELD_STYLE)
             text.append("\n")
-            text.append_text(self._highlight_python(poc_script_code))
+            text.append_text(self._highlight_python(poc_code, poc_language))
 
         remediation_steps = vuln.get("remediation_steps", "")
         if remediation_steps:
@@ -601,9 +609,12 @@ class VulnerabilityDetailScreen(ModalScreen):  # type: ignore[misc]
                 lines.append(vuln["poc_description"])
                 lines.append("")
             if vuln.get("poc_script_code"):
-                lines.append("```python")
-                lines.append(vuln["poc_script_code"])
-                lines.append("```")
+                poc_language, poc_code = parse_fenced_code(vuln["poc_script_code"])
+                fence_lang = poc_language or guess_language_name(poc_code)
+                fence = safe_fence(poc_code)
+                lines.append(f"{fence}{fence_lang}")
+                lines.append(poc_code)
+                lines.append(fence)
 
         if vuln.get("code_locations"):
             lines.extend(["", "## Code Analysis", ""])
@@ -619,7 +630,9 @@ class VulnerabilityDetailScreen(ModalScreen):  # type: ignore[misc]
                 if loc.get("label"):
                     lines.append(f"  {loc['label']}")
                 if loc.get("snippet"):
-                    lines.append(f"```\n{loc['snippet']}\n```")
+                    snippet = str(loc["snippet"])
+                    snippet_fence = safe_fence(snippet)
+                    lines.append(f"{snippet_fence}\n{snippet}\n{snippet_fence}")
                 if loc.get("fix_before") or loc.get("fix_after"):
                     lines.append("**Suggested Fix:**")
                     lines.append("```diff")
@@ -802,6 +815,8 @@ class StrixTUIApp(App):  # type: ignore[misc]
         self._scan_stop_event = threading.Event()
         self._scan_completed = threading.Event()
         self._scan_error: BaseException | None = None
+        self._error_noted_agents: set[str] = set()
+        self._budget_pause_notified = False
 
         self._spinner_frame_index: int = 0
         self._sweep_num_squares: int = 6
@@ -1015,26 +1030,50 @@ class StrixTUIApp(App):  # type: ignore[misc]
             else:
                 self._agent_graph_sync_future = None
                 try:
-                    parent_of, statuses, names = future.result()
+                    parent_of, statuses, names, errors = future.result()
                 except Exception:
                     logger.exception("TUI agent graph sync failed")
                 else:
                     for agent_id, status in statuses.items():
+                        error = errors.get(agent_id)
                         self.live_view.upsert_agent(
                             agent_id,
                             name=names.get(agent_id, agent_id),
                             parent_id=parent_of.get(agent_id),
                             status=status,
+                            error_message=error or "",
                         )
+                        if error:
+                            if agent_id not in self._error_noted_agents:
+                                self._error_noted_agents.add(agent_id)
+                                self.live_view.record_agent_error(agent_id, error)
+                        else:
+                            self._error_noted_agents.discard(agent_id)
+                    self._notify_budget_pause(statuses)
 
         if self._scan_loop is None or self._scan_loop.is_closed():
             return
 
-        async def collect() -> tuple[dict[str, str | None], dict[str, Any], dict[str, str]]:
+        async def collect() -> tuple[
+            dict[str, str | None], dict[str, Any], dict[str, str], dict[str, str]
+        ]:
             return await self.coordinator.graph_snapshot()
 
         self._agent_graph_sync_future = asyncio.run_coroutine_threadsafe(collect(), self._scan_loop)
 
+    def _notify_budget_pause(self, statuses: dict[str, Any]) -> None:
+        paused = any(status == "budget_paused" for status in statuses.values())
+        if paused and not self._budget_pause_notified:
+            self._budget_pause_notified = True
+            self.notify(
+                "Budget limit reached \u2014 agents paused. Send a message to continue "
+                "(this extends the budget), or ctrl-q to quit.",
+                severity="warning",
+                timeout=15,
+            )
+        elif not paused:
+            self._budget_pause_notified = False
+
     def _update_agent_node(self, agent_id: str, agent_data: dict[str, Any]) -> bool:
         if agent_id not in self.agent_nodes:
             return False
@@ -1047,8 +1086,10 @@ class StrixTUIApp(App):  # type: ignore[misc]
             status_indicators = {
                 "running": "⚪",
                 "waiting": "⏸",
+                "budget_paused": "⏸",
                 "completed": "🟢",
                 "failed": "🔴",
+                "crashed": "🔴",
                 "stopped": "■",
             }
 
@@ -1234,20 +1275,30 @@ class StrixTUIApp(App):  # type: ignore[misc]
             text.append(msg)
             return (text, Text(), False)
 
-        if status == "failed":
+        if status in {"failed", "crashed"}:
             error_msg = agent_data.get("error_message", "")
             text = Text()
-            if error_msg:
-                text.append(error_msg, style="red")
-            else:
-                text.append("Scan failed", style="red")
+            text.append(error_msg or "Agent failed", style="red")
+            text.append(" · ", style="dim")
+            text.append("Send message to resume", style="dim")
             self._stop_dot_animation()
             return (text, Text(), False)
 
-        if status == "waiting":
+        if status in {"waiting", "budget_paused"}:
             text = Text()
-            text.append("Send message to resume", style="dim")
-            return (text, Text(), False)
+            keymap = Text()
+            if status == "budget_paused":
+                text.append("Budget limit reached", style="yellow")
+                text.append(" \u00b7 ", style="dim")
+                text.append("Send a message to continue", style="dim")
+                keymap = keymap_styled([("ctrl-q", "quit")])
+            else:
+                error_msg = agent_data.get("error_message") or ""
+                if error_msg:
+                    text.append(error_msg, style="red")
+                    text.append(" \u00b7 ", style="dim")
+                text.append("Send message to resume", style="dim")
+            return (text, keymap, False)
 
         if status == "running":
             if self._agent_has_real_activity(agent_id):
@@ -1472,6 +1523,7 @@ class StrixTUIApp(App):  # type: ignore[misc]
                                 coordinator=self.coordinator,
                                 interactive=True,
                                 max_budget_usd=getattr(self.args, "max_budget_usd", None),
+                                max_turns=getattr(self.args, "max_turns", DEFAULT_MAX_TURNS),
                                 event_sink=self._capture_sdk_event,
                             ),
                         )
@@ -1479,10 +1531,7 @@ class StrixTUIApp(App):  # type: ignore[misc]
                 except (KeyboardInterrupt, asyncio.CancelledError):
                     logger.info("Scan interrupted by user")
                 except BudgetExceededError:
-                    # Defensive: the runner stops the scan cleanly on budget and
-                    # returns, so this normally never propagates. Treat it as a
-                    # graceful stop, not a scan error, if it ever does.
-                    logger.info("Scan stopped: --max-budget-usd limit reached")
+                    logger.info("Scan stopped: --max-budget limit reached")
                 except (ConnectionError, TimeoutError) as e:
                     logging.exception("Network error during scan")
                     self._scan_error = e
@@ -1537,8 +1586,10 @@ class StrixTUIApp(App):  # type: ignore[misc]
         status_indicators = {
             "running": "⚪",
             "waiting": "⏸",
+            "budget_paused": "⏸",
             "completed": "🟢",
             "failed": "🔴",
+            "crashed": "🔴",
             "stopped": "■",
         }
 
@@ -1582,8 +1633,10 @@ class StrixTUIApp(App):  # type: ignore[misc]
         status_indicators = {
             "running": "⚪",
             "waiting": "⏸",
+            "budget_paused": "⏸",
             "completed": "🟢",
             "failed": "🔴",
+            "crashed": "🔴",
             "stopped": "■",
         }
 
@@ -1705,7 +1758,10 @@ class StrixTUIApp(App):  # type: ignore[misc]
             message=message,
         )
         if not submitted:
-            self.notify("Scan loop is not ready; message was not sent", severity="warning")
+            if self._scan_completed.is_set():
+                self.notify("The scan has ended; message was not sent", severity="warning")
+            else:
+                self.notify("Scan loop is not ready; message was not sent", severity="warning")
             return
 
         self._displayed_events.clear()
@@ -1838,7 +1894,7 @@ 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/live_view.py

@@ -20,7 +20,7 @@ class TuiLiveView:
         self.events: list[dict[str, Any]] = []
         self._next_event_id = 1
         self._open_assistant_event_by_agent: dict[str, dict[str, Any]] = {}
-        self._tool_event_by_call_id: dict[str, dict[str, Any]] = {}
+        self._tool_event_by_agent_and_call_id: dict[tuple[str, str], dict[str, Any]] = {}
 
     def hydrate_from_run_dir(self, run_dir: Path) -> None:
         state_dir = runtime_state_dir(run_dir)
@@ -82,10 +82,21 @@ class TuiLiveView:
             current["parent_id"] = parent_id
         if status is not None:
             current["status"] = status
-        if error_message:
+        if error_message is not None:
             current["error_message"] = error_message
         current["updated_at"] = now
 
+    def record_agent_error(self, agent_id: str, error: str) -> None:
+        self._append_event(
+            agent_id,
+            "chat",
+            {
+                "role": "assistant",
+                "content": (f"An error occurred: {error}\nI'm now waiting for new instructions."),
+                "metadata": {"source": "agent_error"},
+            },
+        )
+
     def record_user_message(self, agent_id: str, content: str) -> None:
         self._append_event(
             agent_id,
@@ -212,7 +223,8 @@ class TuiLiveView:
         timestamp: str | None = None,
     ) -> None:
         call_id = call["call_id"]
-        existing = self._tool_event_by_call_id.get(call_id)
+        event_key = (agent_id, call_id)
+        existing = self._tool_event_by_agent_and_call_id.get(event_key)
         tool_data = {
             "tool_name": call["tool_name"],
             "args": call["args"],
@@ -222,7 +234,7 @@ class TuiLiveView:
         }
         if existing is None:
             event = self._append_event(agent_id, "tool", tool_data, timestamp=timestamp)
-            self._tool_event_by_call_id[call_id] = event
+            self._tool_event_by_agent_and_call_id[event_key] = event
         else:
             existing["data"].update(tool_data)
             self._bump_event(existing, timestamp=timestamp)
@@ -238,7 +250,8 @@ class TuiLiveView:
         timestamp: str | None = None,
     ) -> None:
         call_id = output["call_id"]
-        event = self._tool_event_by_call_id.get(call_id)
+        event_key = (agent_id, call_id)
+        event = self._tool_event_by_agent_and_call_id.get(event_key)
         if event is None:
             event = self._append_event(
                 agent_id,
@@ -252,7 +265,7 @@ class TuiLiveView:
                 },
                 timestamp=timestamp,
             )
-            self._tool_event_by_call_id[call_id] = event
+            self._tool_event_by_agent_and_call_id[event_key] = event
 
         result = _parse_json_value(output["output"])
         event["data"]["result"] = result

strix/interface/tui/renderers/__init__.py

@@ -6,6 +6,7 @@ from . import (
     notes_renderer,
     proxy_renderer,
     reporting_renderer,
+    respond_renderer,
     shell_renderer,
     thinking_renderer,
     todo_renderer,
@@ -23,6 +24,7 @@ __all__ = [
     "proxy_renderer",
     "render_tool_widget",
     "reporting_renderer",
+    "respond_renderer",
     "shell_renderer",
     "thinking_renderer",
     "todo_renderer",

strix/interface/tui/renderers/agents_graph_renderer.py

@@ -117,8 +117,8 @@ class AgentFinishRenderer(BaseToolRenderer):
 
 
 @register_tool_renderer
-class WaitForMessageRenderer(BaseToolRenderer):
-    tool_name: ClassVar[str] = "wait_for_message"
+class WaitForAgentsRenderer(BaseToolRenderer):
+    tool_name: ClassVar[str] = "wait_for_agents"
     css_classes: ClassVar[list[str]] = ["tool-call", "agents-graph-tool"]
 
     @classmethod

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

@@ -1,11 +1,12 @@
 from functools import cache
 from typing import Any, ClassVar
 
-from pygments.lexers import PythonLexer
 from pygments.styles import get_style_by_name
 from rich.text import Text
 from textual.widgets import Static
 
+from strix.report.writer import parse_fenced_code, resolve_lexer
+
 from .base_renderer import BaseToolRenderer
 from .registry import register_tool_renderer
 
@@ -61,8 +62,8 @@ class CreateVulnerabilityReportRenderer(BaseToolRenderer):
         return None
 
     @classmethod
-    def _highlight_python(cls, code: str) -> Text:
-        lexer = PythonLexer()
+    def _highlight_code(cls, code: str, language: str | None) -> Text:
+        lexer = resolve_lexer(language, code)
         text = Text()
 
         for token_type, token_value in lexer.get_tokens(code):
@@ -234,10 +235,11 @@ class CreateVulnerabilityReportRenderer(BaseToolRenderer):
             text.append(poc_description)
 
         if poc_script_code:
+            poc_language, poc_code = parse_fenced_code(poc_script_code)
             text.append("\n\n")
             text.append("PoC Code", style=FIELD_STYLE)
             text.append("\n")
-            text.append_text(cls._highlight_python(poc_script_code))
+            text.append_text(cls._highlight_code(poc_code, poc_language))
 
         if remediation_steps:
             text.append("\n\n")
@@ -429,3 +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/tui/renderers/respond_renderer.py

@@ -0,0 +1,35 @@
+from typing import Any, ClassVar
+
+from rich.text import Text
+from textual.widgets import Static
+
+from .agent_message_renderer import AgentMessageRenderer
+from .base_renderer import BaseToolRenderer
+from .registry import register_tool_renderer
+
+
+@register_tool_renderer
+class RespondToUserRenderer(BaseToolRenderer):
+    """Render a reply as the agent's own prose, not as a tool call.
+
+    ``respond_to_user`` carries the message the user is meant to read, so it
+    gets the same markdown treatment as a plain assistant turn.
+    """
+
+    tool_name: ClassVar[str] = "respond_to_user"
+    css_classes: ClassVar[list[str]] = ["tool-call", "respond-tool"]
+
+    @classmethod
+    def render(cls, tool_data: dict[str, Any]) -> Static:
+        args = tool_data.get("args", {})
+        message = args.get("message", "")
+
+        text = Text()
+        if message:
+            text.append_text(AgentMessageRenderer.render_simple(message))
+            text.append("\n\n")
+        text.append("○ ", style="#6b7280")
+        text.append("waiting for your reply", style="dim")
+
+        css_classes = cls.get_css_classes(tool_data.get("status", "unknown"))
+        return Static(text, classes=css_classes)

strix/interface/update_check.py

@@ -0,0 +1,395 @@
+"""Update notifications and self-update for the strix CLI.
+
+Follows the pattern used by tools like gh, uv, and pip: a background,
+rate-limited (once per 24h) check against the release source, a cached
+result in ``~/.strix``, a non-intrusive notice with the upgrade command
+for the detected install method, and a ``strix --update`` self-update
+path for the standalone binary install.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import logging
+import os
+import platform
+import shutil
+import stat
+import subprocess
+import sys
+import tarfile
+import tempfile
+import threading
+import time
+import zipfile
+from pathlib import Path
+from typing import cast
+
+import requests
+from rich.console import Console
+from rich.prompt import Prompt
+
+from strix.telemetry._common import get_version
+
+
+logger = logging.getLogger(__name__)
+
+GITHUB_REPO = "usestrix/strix"
+PYPI_PACKAGE = "strix-agent"
+CHECK_INTERVAL_SECONDS = 24 * 60 * 60
+REQUEST_TIMEOUT_SECONDS = 5
+
+_CACHE_PATH = Path.home() / ".strix" / "update-check.json"
+
+_background_thread: threading.Thread | None = None
+
+
+def _is_disabled() -> bool:
+    return bool(os.environ.get("STRIX_NO_UPDATE_CHECK")) or any(
+        os.environ.get(key)
+        for key in ("CI", "GITHUB_ACTIONS", "GITLAB_CI", "JENKINS_URL", "BUILDKITE", "CIRCLECI")
+    )
+
+
+def is_binary_install() -> bool:
+    return bool(getattr(sys, "frozen", False))
+
+
+def get_install_method() -> str:
+    if is_binary_install():
+        return "binary"
+    prefix = str(Path(sys.prefix)).replace("\\", "/")
+    if "/pipx/" in prefix or prefix.endswith("/pipx"):
+        return "pipx"
+    if "/uv/tools/" in prefix:
+        return "uv"
+    return "pip"
+
+
+def get_upgrade_command(method: str | None = None) -> str:
+    method = method or get_install_method()
+    commands = {
+        "binary": "strix --update",
+        "pipx": "pipx upgrade strix-agent",
+        "uv": "uv tool upgrade strix-agent",
+        "pip": "pip install --upgrade strix-agent",
+    }
+    return commands[method]
+
+
+def _parse_version(value: str) -> tuple[int, ...] | None:
+    parts = value.strip().lstrip("v").split(".")
+    try:
+        return tuple(int(part) for part in parts)
+    except ValueError:
+        return None
+
+
+def _is_newer(latest: str, current: str) -> bool:
+    latest_parts = _parse_version(latest)
+    current_parts = _parse_version(current)
+    if latest_parts is None or current_parts is None:
+        return False
+    return latest_parts > current_parts
+
+
+def _fetch_latest_version() -> str | None:
+    try:
+        if is_binary_install():
+            response = requests.get(
+                f"https://api.github.com/repos/{GITHUB_REPO}/releases/latest",
+                timeout=REQUEST_TIMEOUT_SECONDS,
+            )
+            response.raise_for_status()
+            tag = response.json().get("tag_name", "")
+            return tag.lstrip("v") or None
+        response = requests.get(
+            f"https://pypi.org/pypi/{PYPI_PACKAGE}/json",
+            timeout=REQUEST_TIMEOUT_SECONDS,
+        )
+        response.raise_for_status()
+        version = response.json().get("info", {}).get("version")
+        return str(version) if version else None
+    except Exception:  # noqa: BLE001
+        logger.debug("update check failed", exc_info=True)
+        return None
+
+
+def _fetch_asset_digest(version: str, filename: str) -> str | None:
+    """Return the expected sha256 (hex) for a release asset, if the API provides one."""
+    try:
+        response = requests.get(
+            f"https://api.github.com/repos/{GITHUB_REPO}/releases/tags/v{version}",
+            timeout=REQUEST_TIMEOUT_SECONDS,
+        )
+        response.raise_for_status()
+        for asset in response.json().get("assets", []):
+            if asset.get("name") == filename:
+                digest = asset.get("digest") or ""
+                if digest.startswith("sha256:"):
+                    return digest.removeprefix("sha256:")
+    except Exception:  # noqa: BLE001
+        logger.debug("release asset digest lookup failed", exc_info=True)
+    return None
+
+
+def _sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as f:
+        for chunk in iter(lambda: f.read(1 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _read_cache() -> dict[str, object]:
+    try:
+        with _CACHE_PATH.open(encoding="utf-8") as f:
+            data = json.load(f)
+        if isinstance(data, dict):
+            return cast("dict[str, object]", data)
+    except Exception:  # noqa: BLE001, S110
+        pass  # nosec B110
+    return {}
+
+
+def _write_cache(**fields: object) -> None:
+    try:
+        cache = _read_cache()
+        cache.update(fields)
+        _CACHE_PATH.parent.mkdir(parents=True, exist_ok=True)
+        _CACHE_PATH.write_text(json.dumps(cache), encoding="utf-8")
+    except Exception:  # noqa: BLE001, S110
+        pass  # nosec B110
+
+
+def skip_version(version: str) -> None:
+    """Remember not to prompt again for this version (newer releases still notify)."""
+    _write_cache(skipped_version=version)
+
+
+def _refresh_cache() -> None:
+    latest = _fetch_latest_version()
+    if latest:
+        _write_cache(latest_version=latest, checked_at=time.time())
+
+
+def start_background_check() -> None:
+    """Refresh the cached latest-version info in a daemon thread (at most once per 24h)."""
+    global _background_thread  # noqa: PLW0603
+    if _is_disabled():
+        return
+    cache = _read_cache()
+    checked_at = cache.get("checked_at")
+    if isinstance(checked_at, int | float) and time.time() - checked_at < CHECK_INTERVAL_SECONDS:
+        return
+    _background_thread = threading.Thread(target=_refresh_cache, daemon=True)
+    _background_thread.start()
+
+
+def get_available_update(*, respect_skip: bool = True) -> str | None:
+    """Return the newer version from the cache, or None if up to date / unknown."""
+    if _is_disabled():
+        return None
+    if _background_thread is not None:
+        _background_thread.join(timeout=0.2)
+    cache = _read_cache()
+    latest = cache.get("latest_version")
+    current = get_version()
+    if not isinstance(latest, str) or current == "unknown" or not _is_newer(latest, current):
+        return None
+    if respect_skip and cache.get("skipped_version") == latest:
+        return None
+    return latest
+
+
+def notify_update(console: Console) -> None:
+    latest = get_available_update()
+    if not latest:
+        return
+    console.print(
+        f"[#eab308]A new version of strix is available:[/] "
+        f"[dim]{get_version()}[/] [dim]→[/] [bold #22c55e]{latest}[/]"
+        f"  [dim]·[/]  [#60a5fa]{get_upgrade_command()}[/]"
+    )
+    console.print()
+
+
+def run_package_upgrade(console: Console, method: str) -> bool:
+    """Upgrade a package-manager install by running its upgrade command."""
+    command = get_upgrade_command(method).split()
+    console.print(f"[dim]Running[/] [#60a5fa]{' '.join(command)}[/]")
+    try:
+        result = subprocess.run(command, check=False)  # noqa: S603
+    except OSError as e:
+        console.print(f"[bold red]Update failed:[/] {e}")
+        return False
+    if result.returncode != 0:
+        console.print(
+            f"[bold red]Update failed[/] [dim](exit code {result.returncode}).[/] "
+            f"Run it manually: [#60a5fa]{get_upgrade_command(method)}[/]"
+        )
+        return False
+    console.print("[#22c55e]✓ strix updated — restart the scan to use the new version[/]")
+    return True
+
+
+def prompt_update_if_available(console: Console) -> bool:
+    """Offer an interactive update before a scan starts.
+
+    Returns True if strix was updated (caller should re-exec / exit).
+    """
+    latest = get_available_update()
+    if not latest or not sys.stdin.isatty() or not sys.stdout.isatty():
+        return False
+    console.print()
+    console.print(
+        f"[#eab308]A new version of strix is available:[/] "
+        f"[dim]{get_version()}[/] [dim]→[/] [bold #22c55e]{latest}[/]"
+    )
+    console.print(
+        "[dim]  y — update now    n — not now (ask again next run)    s — skip this version[/]"
+    )
+    choice = Prompt.ask("Update strix?", choices=["y", "n", "s"], default="n")
+    console.print()
+    if choice == "s":
+        skip_version(latest)
+        return False
+    if choice != "y":
+        return False
+    method = get_install_method()
+    if method == "binary":
+        return self_update(console, version=latest)
+    return run_package_upgrade(console, method)
+
+
+def _release_target() -> str | None:
+    raw_os = platform.system().lower()
+    os_name = {"darwin": "macos", "linux": "linux", "windows": "windows"}.get(raw_os)
+    arch = platform.machine().lower()
+    arch = {"aarch64": "arm64", "amd64": "x86_64"}.get(arch, arch)
+    if os_name is None:
+        return None
+    target = f"{os_name}-{arch}"
+    supported = {
+        "linux-x86_64",
+        "linux-arm64",
+        "macos-x86_64",
+        "macos-arm64",
+        "windows-x86_64",
+    }
+    return target if target in supported else None
+
+
+def _download_and_replace(version: str, target: str, console: Console) -> bool:
+    is_windows = target.startswith("windows")
+    archive_ext = ".zip" if is_windows else ".tar.gz"
+    filename = f"strix-{version}-{target}{archive_ext}"
+    url = f"https://github.com/{GITHUB_REPO}/releases/download/v{version}/{filename}"
+    binary_name = f"strix-{version}-{target}" + (".exe" if is_windows else "")
+    current_exe = Path(sys.executable).resolve()
+
+    with tempfile.TemporaryDirectory() as tmp:
+        tmp_dir = Path(tmp)
+        archive_path = tmp_dir / filename
+        console.print(f"[dim]Downloading[/] {url}")
+        with requests.get(  # nosec B113
+            url,
+            stream=True,
+            timeout=REQUEST_TIMEOUT_SECONDS * 12,
+        ) as response:
+            response.raise_for_status()
+            with archive_path.open("wb") as f:
+                for chunk in response.iter_content(chunk_size=1 << 20):
+                    f.write(chunk)
+
+        expected_digest = _fetch_asset_digest(version, filename)
+        if expected_digest:
+            actual_digest = _sha256_file(archive_path)
+            if actual_digest != expected_digest:
+                raise RuntimeError(
+                    f"checksum mismatch for {filename}: "
+                    f"expected sha256 {expected_digest}, got {actual_digest}"
+                )
+        else:
+            console.print("[dim yellow]No published checksum available; skipping verification[/]")
+
+        if is_windows:
+            with zipfile.ZipFile(archive_path) as zf:
+                zf.extract(binary_name, tmp_dir)
+        else:
+            with tarfile.open(archive_path, "r:gz") as tf:
+                tf.extract(binary_name, tmp_dir, filter="data")
+
+        new_binary = tmp_dir / binary_name
+        new_binary.chmod(new_binary.stat().st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH)
+
+        staged = current_exe.with_name(current_exe.name + ".new")
+        try:
+            shutil.copy2(new_binary, staged)
+            if is_windows:
+                # Windows can't replace a running executable in place; move it aside first.
+                old = current_exe.with_name(current_exe.name + ".old")
+                old.unlink(missing_ok=True)
+                current_exe.rename(old)
+                try:
+                    staged.replace(current_exe)
+                except Exception:
+                    old.rename(current_exe)
+                    raise
+            else:
+                staged.replace(current_exe)
+        except Exception:
+            staged.unlink(missing_ok=True)
+            raise
+    return True
+
+
+def self_update(console: Console | None = None, version: str | None = None) -> bool:
+    """Replace the running standalone binary with the latest release.
+
+    Returns True on success. For package-manager installs this only
+    prints the right upgrade command and returns False.
+    """
+    console = console or Console()
+
+    if not is_binary_install():
+        method = get_install_method()
+        console.print(
+            f"[#eab308]This strix was installed via {method};[/] "
+            f"upgrade it with: [#60a5fa]{get_upgrade_command(method)}[/]"
+        )
+        return False
+
+    latest = version or _fetch_latest_version()
+    if not latest:
+        console.print("[bold red]Could not determine the latest strix version.[/]")
+        return False
+
+    current = get_version()
+    if current != "unknown" and not _is_newer(latest, current):
+        console.print(f"[#22c55e]strix {current} is already the latest version.[/]")
+        return True
+
+    target = _release_target()
+    if not target:
+        console.print(
+            f"[bold red]No prebuilt binary for this platform "
+            f"({platform.system()}/{platform.machine()}).[/]"
+        )
+        return False
+
+    try:
+        _download_and_replace(latest, target, console)
+    except Exception as e:  # noqa: BLE001
+        logger.debug("self-update failed", exc_info=True)
+        console.print(f"[bold red]Update failed:[/] {e}")
+        console.print(
+            "[dim]You can reinstall manually with:[/] "
+            "[#60a5fa]curl -sSL https://strix.ai/install | bash[/]"
+        )
+        return False
+
+    _write_cache(latest_version=latest, checked_at=time.time())
+    console.print(f"[#22c55e]✓ Updated strix to {latest}[/]")
+    return True

strix/interface/utils.py

@@ -11,11 +11,10 @@ import tempfile
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any
-from urllib.error import HTTPError, URLError
 from urllib.parse import urlparse
-from urllib.request import Request, urlopen
 
 import docker
+import requests
 from docker.errors import DockerException, ImageNotFound
 from rich.console import Console
 from rich.panel import Panel
@@ -253,6 +252,20 @@ def _llm_usage(report_state: Any) -> dict[str, Any]:
     return usage if isinstance(usage, dict) else {}
 
 
+def _is_subscription(report_state: Any) -> bool:
+    """Whether this run uses a model subscription (no metered cost).
+
+    Prefers the run record so it's correct for hydrated/resumed runs; falls back
+    to current settings.
+    """
+    record = getattr(report_state, "run_record", None)
+    if isinstance(record, dict) and record.get("auth_mode"):
+        return record.get("auth_mode") == "subscription"
+    from strix.config import codex
+
+    return codex.auth_mode(load_settings().llm.model) == "subscription"
+
+
 def _int_stat(usage: dict[str, Any], key: str) -> int:
     try:
         return max(0, int(usage.get(key) or 0))
@@ -283,11 +296,16 @@ def _build_llm_usage_stats(
     *,
     live: bool = False,
 ) -> None:
+    subscription = _is_subscription(report_state)
     usage = _llm_usage(report_state)
     if not usage or _int_stat(usage, "requests") <= 0:
         stats_text.append("\n")
         stats_text.append("Cost ", style="dim")
-        stats_text.append("$0.0000 ", style="#fbbf24")
+        if subscription:
+            stats_text.append("$0.00 ", style="#22c55e")
+            stats_text.append("(subscription) ", style="dim")
+        else:
+            stats_text.append("$0.0000 ", style="#fbbf24")
         stats_text.append("· ", style="dim white")
         stats_text.append("Tokens ", style="dim")
         stats_text.append("0", style="white")
@@ -312,7 +330,12 @@ def _build_llm_usage_stats(
     stats_text.append("Output Tokens ", style="dim")
     stats_text.append(format_token_count(output_tokens), style="white")
 
-    if live or cost > 0:
+    if subscription:
+        stats_text.append("  ·  ", style="dim white")
+        stats_text.append("Cost ", style="dim")
+        stats_text.append("$0.00", style="#22c55e")
+        stats_text.append(" (subscription)", style="dim")
+    elif live or cost > 0:
         stats_text.append("  ·  ", style="dim white")
         stats_text.append("Cost ", style="dim")
         stats_text.append(f"${cost:.4f}", style="#fbbf24")
@@ -337,6 +360,9 @@ def build_live_stats_text(report_state: Any) -> Text:
     model = load_settings().llm.model or "unknown"
     stats_text.append("Model ", style="dim")
     stats_text.append(str(model), style="white")
+    if _is_subscription(report_state):
+        stats_text.append("  ·  ", style="dim white")
+        stats_text.append("ChatGPT subscription", style="#22c55e")
     stats_text.append("\n")
 
     vuln_count = len(report_state.vulnerability_reports)
@@ -379,6 +405,10 @@ def build_tui_stats_text(report_state: Any) -> Text:
 
     model = load_settings().llm.model or "unknown"
     stats_text.append(str(model), style="white")
+    subscription = _is_subscription(report_state)
+    if subscription:
+        stats_text.append("\n")
+        stats_text.append("ChatGPT subscription", style="#22c55e")
 
     usage = _llm_usage(report_state)
     if usage and _int_stat(usage, "total_tokens") > 0:
@@ -388,7 +418,10 @@ def build_tui_stats_text(report_state: Any) -> Text:
             style="white",
         )
         cost = _float_stat(usage, "cost")
-        if cost > 0:
+        if subscription:
+            stats_text.append(" · ", style="white")
+            stats_text.append("$0.00", style="white")
+        elif cost > 0:
             stats_text.append(" · ", style="white")
             stats_text.append(f"${cost:.2f}", style="white")
 
@@ -1054,13 +1087,12 @@ def resolve_diff_scope_context(
 def _is_http_git_repo(url: str) -> bool:
     check_url = f"{url.rstrip('/')}/info/refs?service=git-upload-pack"
     try:
-        req = Request(check_url, headers={"User-Agent": "git/strix"})  # noqa: S310
-        with urlopen(req, timeout=10) as resp:  # noqa: S310  # nosec B310
-            return "x-git-upload-pack-advertisement" in resp.headers.get("Content-Type", "")
-    except HTTPError as e:
-        return e.code == 401
-    except (URLError, OSError, ValueError):
+        resp = requests.get(check_url, headers={"User-Agent": "git/strix"}, timeout=10)
+    except (requests.RequestException, ValueError):
         return False
+    if resp.status_code >= 400:
+        return resp.status_code == 401
+    return "x-git-upload-pack-advertisement" in resp.headers.get("Content-Type", "")
 
 
 def infer_target_type(target: str) -> tuple[str, dict[str, str]]:  # noqa: PLR0911
@@ -1147,9 +1179,7 @@ def read_target_list_file(path_str: str) -> list[str]:
             if (target := line.strip()) and not target.startswith("#")
         ]
     except UnicodeDecodeError as e:
-        raise ValueError(
-            f"Target list file '{path_str}' must be valid UTF-8 text: {e!s}"
-        ) from e
+        raise ValueError(f"Target list file '{path_str}' must be valid UTF-8 text: {e!s}") from e
     except OSError as e:
         raise ValueError(f"Failed to read target list file '{path_str}': {e!s}") from e
 

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

@@ -15,12 +15,12 @@ import base64
 import contextlib
 import json
 import logging
-import urllib.error
-import urllib.request
 from datetime import UTC, datetime
 from pathlib import Path
 from typing import Any
 
+import requests
+
 from strix.config.loader import load_settings
 
 
@@ -147,21 +147,17 @@ def _post_json(path: str, payload: dict[str, Any], *, timeout: int) -> tuple[int
     map, not raised.
     """
     url = f"{_app_url()}{path}"
-    body = json.dumps(payload).encode("utf-8")
-    request = urllib.request.Request(  # noqa: S310 - fixed https relay URL
-        url,
-        data=body,
-        headers={"Content-Type": "application/json", "Accept": "application/json"},
-        method="POST",
-    )
     try:
-        with urllib.request.urlopen(request, timeout=timeout) as response:  # noqa: S310
-            return response.status, _parse_body(response.read())
-    except urllib.error.HTTPError as exc:
-        return exc.code, _parse_body(exc.read())
-    except (urllib.error.URLError, TimeoutError, OSError) as exc:
+        response = requests.post(
+            url,
+            json=payload,
+            headers={"Accept": "application/json"},
+            timeout=timeout,
+        )
+    except requests.RequestException as exc:
         logger.warning("relay request to %s failed: %s", path, exc)
         raise RelayError("unavailable") from exc
+    return response.status_code, _parse_body(response.content)
 
 
 def _parse_body(raw: bytes) -> dict[str, Any]:

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/package-lock.json

@@ -63,7 +63,6 @@
       "integrity": "sha512-RgHBCvtjbOK2gXSNBNIkNoEc9qoVEtau3hj8gEqKQuL3HZAibKarWFEI3Lfm6EYKkLalOh8eSrj9b+ch9H/VBA==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "@babel/code-frame": "^7.29.7",
         "@babel/generator": "^7.29.7",
@@ -1605,7 +1604,6 @@
       "resolved": "https://registry.npmjs.org/@types/react/-/react-19.2.17.tgz",
       "integrity": "sha512-MXfmqaVPEVgkBT/aY0aGCkRWWtByiYQXo3xdQ8r5RzuFrPiRn8Gar2tQdXSUQ2GKV3bkXckek89V8wQBY2Q/Aw==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "csstype": "^3.2.2"
       }
@@ -1616,7 +1614,6 @@
       "integrity": "sha512-jp2L/eY6fn+KgVVQAOqYItbF0VY/YApe5Mz2F0aykSO8gx31bYCZyvSeYxCHKvzHG5eZjc+zyaS5BrBWya2+kQ==",
       "devOptional": true,
       "license": "MIT",
-      "peer": true,
       "peerDependencies": {
         "@types/react": "^19.2.0"
       }
@@ -1739,7 +1736,6 @@
         }
       ],
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "baseline-browser-mapping": "^2.10.42",
         "caniuse-lite": "^1.0.30001803",
@@ -1920,7 +1916,6 @@
       "resolved": "https://registry.npmjs.org/d3-selection/-/d3-selection-3.0.0.tgz",
       "integrity": "sha512-fmTRWbNMmsmWq6xJV8D19U/gw/bwrHfNXxrIN+HfZgnzqTHp9jOmKMhsTUjXOJnZOdZY9Q28y4yebKzqDKlxlQ==",
       "license": "ISC",
-      "peer": true,
       "engines": {
         "node": ">=12"
       }
@@ -3571,7 +3566,6 @@
       "integrity": "sha512-RvwwcruNjI1ncT5xRakeyS9Lf8lcItv34KD+aif+VH9kduAyfYBipGh12274xtenIPZ119/R9BdTBa8gAwSh0A==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "engines": {
         "node": ">=12"
       },
@@ -3623,7 +3617,6 @@
       "resolved": "https://registry.npmjs.org/react/-/react-19.2.7.tgz",
       "integrity": "sha512-HNe9WslTbXmFK8o8cmwgAeJFSBvt1bPdHCVKtaaV+WlAN36mpT4hcRpwbf3fY56ar2oIXzsBpOAiIRHAdY0OlQ==",
       "license": "MIT",
-      "peer": true,
       "engines": {
         "node": ">=0.10.0"
       }
@@ -3633,7 +3626,6 @@
       "resolved": "https://registry.npmjs.org/react-dom/-/react-dom-19.2.7.tgz",
       "integrity": "sha512-t0BRVXvbiE/o20Hfw669rLbMCDWtYZLvmJigy2f0MxsXF+71pxhR3xOkspmsO8h3ZlNzyibAmtCa3l4lYKk6gQ==",
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "scheduler": "^0.27.0"
       },
@@ -4109,7 +4101,6 @@
       "integrity": "sha512-NTKlcQjlAK7MlQoyb6LgaqHc8sso/pVyUJYWMws3jg21uTJw/LddqIFPcPqP6PzpgbIcZyKI85sFE4HBrQDA8A==",
       "dev": true,
       "license": "MIT",
-      "peer": true,
       "dependencies": {
         "esbuild": "^0.25.0",
         "fdir": "^6.4.4",

strix/interface/viewer/frontend/src/components/RunDetails.tsx

@@ -100,6 +100,7 @@ export function RunDetails({
   const reasoning = num(rec(arr(usage.output_tokens_details)[0]).reasoning_tokens);
   const totalTokens = num(usage.total_tokens);
   const cost = num(usage.cost);
+  const subscription = str(raw.auth_mode) === "subscription";
 
   const sub = (n: number, word: string) => (
     <span className="text-[#666]"> ({formatNumber(n)} {word})</span>
@@ -175,6 +176,15 @@ export function RunDetails({
           {hasUsage ? (
             <dl className="space-y-2.5 tabular-nums">
               <Field label="Model">{models.length ? models.join(", ") : "n/a"}</Field>
+              {subscription && (
+                <Field label="Provider">
+                  <span className="inline-flex items-center gap-1.5">
+                    <span className="rounded-full border border-[#22c55e]/40 bg-[#22c55e]/10 px-2 py-0.5 text-[11px] text-[#22c55e]">
+                      ChatGPT subscription
+                    </span>
+                  </span>
+                </Field>
+              )}
               <Field label="Run time">{fmtDuration(durationSeconds)}</Field>
               {requests != null && <Field label="Requests">{formatNumber(requests)}</Field>}
               {inputTokens != null && (
@@ -190,7 +200,14 @@ export function RunDetails({
                 </Field>
               )}
               {totalTokens != null && <Field label="Total tokens">{formatNumber(totalTokens)}</Field>}
-              {cost != null && <Field label="Cost">${cost.toFixed(2)}</Field>}
+              {subscription ? (
+                <Field label="Cost">
+                  <span className="text-[#22c55e]">$0.00</span>
+                  <span className="text-[#666]"> (subscription)</span>
+                </Field>
+              ) : (
+                cost != null && <Field label="Cost">${cost.toFixed(2)}</Field>
+              )}
               {agents.length > 0 && <Field label="Agents">{formatNumber(agents.length)}</Field>}
             </dl>
           ) : (

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

@@ -54,7 +54,7 @@ export default function AgentCommsRenderer({ toolName, args }: ToolRendererProps
     );
   }
 
-  if (toolName === "wait_for_message") {
+  if (toolName === "wait_for_agents") {
     const reason = (args.reason as string) ?? "";
     return (
       <div className="flex items-center gap-2">

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/RespondRenderer.tsx

@@ -0,0 +1,20 @@
+"use client";
+
+import type { ToolRendererProps } from "@/types/events";
+import Markdown from "./Markdown";
+
+/**
+ * `respond_to_user` carries the message the user is meant to read, so it renders
+ * as the agent's own prose rather than as a tool call.
+ */
+export default function RespondRenderer({ args }: ToolRendererProps) {
+  const message = (args.message as string) ?? "";
+  if (!message) return null;
+
+  return (
+    <div>
+      <Markdown text={message} />
+      <div className="mt-1.5 text-[#888] text-[13px]">waiting for your reply</div>
+    </div>
+  );
+}

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

@@ -3,6 +3,7 @@
 import type { ToolRendererProps } from "@/types/events";
 import { TruncatedText } from "./ToolCard";
 import { MdCodeBlock } from "@/components/vulnerability/MdCodeBlock";
+import { parseFencedCode } from "@/lib/fenced-code";
 import Markdown from "./Markdown";
 
 const SEVERITY_COLORS: Record<string, string> = {
@@ -19,7 +20,7 @@ export default function VulnReportRenderer({ args, result }: ToolRendererProps)
   const method = (args.method as string) ?? "";
   const technicalAnalysis = (args.technical_analysis as string) ?? "";
   const pocDescription = (args.poc_description as string) ?? "";
-  const pocCode = (args.poc_script_code as string) ?? "";
+  const { language: pocLang, code: pocCode } = parseFencedCode((args.poc_script_code as string) ?? "");
   const remediation = (args.remediation_steps as string) ?? "";
   const cve = (args.cve as string) ?? "";
   const cwe = (args.cwe as string) ?? "";
@@ -59,7 +60,7 @@ export default function VulnReportRenderer({ args, result }: ToolRendererProps)
         <div>
           <span className="text-emerald-400/60 text-sm font-semibold">Proof of Concept</span>
           {pocDescription && <div className="mt-1"><Markdown text={pocDescription} /></div>}
-          {pocCode && <MdCodeBlock>{pocCode}</MdCodeBlock>}
+          {pocCode && <MdCodeBlock className={pocLang ? `language-${pocLang}` : undefined}>{pocCode}</MdCodeBlock>}
         </div>
       )}
       {remediation && (

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";
@@ -23,6 +24,7 @@ import NotesRenderer from "./NotesRenderer";
 import TodoRenderer from "./TodoRenderer";
 import FallbackRenderer from "./FallbackRenderer";
 import LoadSkillRenderer from "./LoadSkillRenderer";
+import RespondRenderer from "./RespondRenderer";
 
 /**
  * Tool-renderer mapping — data-driven, keyed by the engine's tool *family*.
@@ -101,12 +103,12 @@ 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"],
+  agents: ["create_agent", "agent_finish", "send_message_to_agent", "wait_for_agents", "view_agent_graph", "stop_agent"],
   search: ["web_search"],
   // scan_start_info / subagent_start_info are strix-app synthetic events; finish_scan is the engine's
-  lifecycle: ["scan_start_info", "subagent_start_info", "finish_scan"],
+  lifecycle: ["scan_start_info", "subagent_start_info", "finish_scan", "respond_to_user"],
   notes: ["create_note", "delete_note", "update_note", "list_notes", "get_note"],
   skills: ["load_skill"],
   todos: ["create_todo", "list_todos", "update_todo", "mark_todo_done", "mark_todo_pending", "delete_todo"],
@@ -126,8 +128,11 @@ const TOOL_CATEGORY: Record<string, ToolCategory> = Object.fromEntries(
  */
 const RENDERER_OVERRIDES: Partial<Record<string, ComponentType<ToolRendererProps>>> = {
   finish_scan: FinishRenderer,
+  respond_to_user: RespondRenderer,
   apply_patch: ApplyPatchRenderer,
   view_image: ViewImageRenderer,
+  list_reports: ReportListRenderer,
+  get_report: ReportListRenderer,
 };
 
 /**
@@ -137,7 +142,8 @@ const RENDERER_OVERRIDES: Partial<Record<string, ComponentType<ToolRendererProps
 const ICON_OVERRIDES: Partial<Record<string, ToolIconMeta>> = {
   agent_finish: { icon: Flag, color: "text-cyan-400" },
   send_message_to_agent: { icon: MessageCircle, color: "text-cyan-400" },
-  wait_for_message: { icon: MessageCircle, color: "text-cyan-400" },
+  wait_for_agents: { icon: MessageCircle, color: "text-cyan-400" },
+  respond_to_user: { icon: MessageCircle, color: "text-emerald-400" },
   view_agent_graph: { icon: Eye, color: "text-cyan-400" },
   stop_agent: { icon: Ban, color: "text-red-400" },
   scan_start_info: { icon: Crosshair, color: "text-emerald-400" },

strix/interface/viewer/frontend/src/components/vulnerability/MdCodeBlock.tsx

@@ -1,7 +1,7 @@
 "use client";
 
 import { useState } from "react";
-import hljs from "@/lib/hljs";
+import { highlightCode } from "@/lib/hljs";
 import "highlight.js/styles/github-dark.css";
 import { Copy, Check } from "lucide-react";
 import { copyToClipboard } from "@/lib/vulnerability-utils";
@@ -17,7 +17,7 @@ export function MdCodeBlock({
 }) {
   const [copied, setCopied] = useState(false);
   const raw = String(children).replace(/\n$/, "");
-  const match = /language-(\w+)/.exec(className || "");
+  const match = /language-(\S+)/.exec(className || "");
   const isBlock = raw.includes("\n") || match;
 
   if (!isBlock) {
@@ -35,16 +35,7 @@ export function MdCodeBlock({
       : fileName
     : null;
 
-  let highlighted: string;
-  if (match) {
-    try {
-      highlighted = hljs.highlight(raw, { language: match[1], ignoreIllegals: true }).value;
-    } catch {
-      highlighted = hljs.highlightAuto(raw).value;
-    }
-  } else {
-    highlighted = hljs.highlightAuto(raw).value;
-  }
+  const highlighted = highlightCode(raw, match?.[1]);
 
   const lines = highlighted.split("\n");
 

strix/interface/viewer/frontend/src/components/vulnerability/PocBlock.tsx

@@ -1,12 +1,13 @@
 "use client";
 
 import { useState } from "react";
-import hljs from "@/lib/hljs";
+import { highlightCode } from "@/lib/hljs";
 import "highlight.js/styles/github-dark.css";
 import ReactMarkdown from "react-markdown";
 import remarkGfm from "remark-gfm";
 import { Copy, Check } from "lucide-react";
 import { copyToClipboard } from "@/lib/vulnerability-utils";
+import { parseFencedCode } from "@/lib/fenced-code";
 import { rehypeCodeMeta, mdComponents } from "./MdCodeBlock";
 
 interface PocBlockProps {
@@ -20,9 +21,12 @@ export function PocBlock({ description, scriptCode, onCopy }: PocBlockProps) {
 
   if (!description && !scriptCode) return null;
 
+  const { language, code } = parseFencedCode(scriptCode);
+  const highlighted = highlightCode(code, language);
+
   const copy = () => {
-    if (!scriptCode) return;
-    copyToClipboard(scriptCode);
+    if (!code) return;
+    copyToClipboard(code);
     setCopied(true);
     setTimeout(() => setCopied(false), 2000);
     onCopy?.();
@@ -43,7 +47,7 @@ export function PocBlock({ description, scriptCode, onCopy }: PocBlockProps) {
             </ReactMarkdown>
           </div>
         )}
-        {scriptCode && (
+        {code && (
           <div className="group/poc relative rounded-md border border-[#2a2a2a] overflow-hidden">
             <div className="flex items-stretch">
               <span className="relative flex items-center text-[13px] text-[#999] font-mono px-4 py-2 border-r border-[#2a2a2a]">PoC Script<span className="absolute top-0 inset-x-0 h-0.5 bg-white/60 rounded-full" /></span>
@@ -64,7 +68,7 @@ export function PocBlock({ description, scriptCode, onCopy }: PocBlockProps) {
               <pre className="font-mono text-[12px] leading-[22px] whitespace-pre [font-variant-ligatures:none]">
                 <code
                   dangerouslySetInnerHTML={{
-                    __html: hljs.highlight(scriptCode, { language: "python" }).value,
+                    __html: highlighted,
                   }}
                 />
               </pre>

strix/interface/viewer/frontend/src/lib/fenced-code.ts

@@ -0,0 +1,27 @@
+export interface ParsedFencedCode {
+  language?: string;
+  code: string;
+}
+
+const FENCE_RE = /^```([^\n`]*)\r?\n([\s\S]*?)\r?\n?```$/;
+
+/**
+ * Agent-generated `poc_script_code` is stored wrapped in a markdown code fence
+ * that carries the language, e.g.
+ *
+ *     ```python
+ *     import requests
+ *     ```
+ *
+ * Renderers that show the value as bare code must not display the fence lines
+ * literally. This extracts the inner code and the fence's language tag. Returns
+ * the input unchanged (no language) when it isn't fenced.
+ */
+export function parseFencedCode(raw: string | null | undefined): ParsedFencedCode {
+  if (!raw) return { code: "" };
+  const match = FENCE_RE.exec(raw.trim());
+  if (!match) return { code: raw };
+  const info = match[1].trim();
+  const language = info ? info.split(/\s+/)[0] : undefined;
+  return { language: language || undefined, code: match[2] };
+}

strix/interface/viewer/frontend/src/lib/hljs.ts

@@ -0,0 +1,32 @@
+import hljs from "highlight.js/lib/common";
+import http from "highlight.js/lib/languages/http";
+import nginx from "highlight.js/lib/languages/nginx";
+import apache from "highlight.js/lib/languages/apache";
+import dockerfile from "highlight.js/lib/languages/dockerfile";
+import properties from "highlight.js/lib/languages/properties";
+
+hljs.registerLanguage("http", http);
+hljs.registerLanguage("nginx", nginx);
+hljs.registerLanguage("apache", apache);
+hljs.registerLanguage("dockerfile", dockerfile);
+hljs.registerLanguage("properties", properties);
+
+/**
+ * Highlight code, preferring an explicit language when it's recognized,
+ * otherwise auto-detecting. Falls back to Python when auto-detection is
+ * inconclusive, since legacy (unfenced) PoC scripts are Python.
+ */
+export function highlightCode(code: string, language?: string | null): string {
+  try {
+    if (language && hljs.getLanguage(language)) {
+      return hljs.highlight(code, { language, ignoreIllegals: true }).value;
+    }
+    const auto = hljs.highlightAuto(code);
+    if (auto.language) return auto.value;
+    return hljs.highlight(code, { language: "python", ignoreIllegals: true }).value;
+  } catch {
+    return hljs.highlight(code, { language: "python", ignoreIllegals: true }).value;
+  }
+}
+
+export default hljs;