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

refactor(tools): split wait_for_message into respond_to_user + wait_for_agents
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.
2026-08-16 17:27:26 +02:00 · 2026-08-01 22:16:00 +00:00 · 2026-08-01 22:13:07 +00:00 · 2026-08-01 21:36:54 +00:00 · 2026-08-01 21:30:44 +00:00 · 2026-08-01 21:25:09 +00:00
157 changed files with 7434 additions and 648 deletions
@@ -21,6 +21,8 @@ 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

@@ -43,6 +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

@@ -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__/
@@ -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

@@ -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!"
@@ -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>
@@ -55,6 +63,12 @@ affecting the agents that do the actual testing.
  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.
@@ -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.
@@ -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

@@ -1,6 +1,6 @@
 [project]
 name = "strix-agent"
-version = "1.3.1"
+version = "1.4.1"
 description = "Open-source AI Hackers for your apps"
 readme = "README.md"
 license = "Apache-2.0"
@@ -79,10 +79,10 @@ build-backend = "hatchling.build"

 [tool.hatch.build.targets.wheel]
 packages = ["strix"]
-# The prebuilt viewer bundle under strix/viewer/static/ ships automatically
+# The prebuilt viewer bundle under strix/interface/viewer/static/ ships automatically
 # (hatchling includes non-.py files under the package). The Vite SOURCE lives
-# under the package dir too (strix/viewer/frontend/) but must never ship in the wheel.
-exclude = ["strix/viewer/frontend", "strix/viewer/frontend/**"]
+# under the package dir too (strix/interface/viewer/frontend/) but must never ship in the wheel.
+exclude = ["strix/interface/viewer/frontend", "strix/interface/viewer/frontend/**"]

 # ============================================================================
 # Type Checking Configuration
@@ -220,12 +220,13 @@ ignore = [
 # 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"]
@@ -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}"
@@ -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',
@@ -23,7 +23,7 @@ from strix.tools.agents_graph.tools import (
    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
@@ -34,7 +34,7 @@ from strix.tools.notes.tools import (
    list_notes,
    update_note,
 )
-from strix.tools.output_store import bound_text
+from strix.tools.output_store import bound_and_store, bound_text
 from strix.tools.proxy.tools import (
    list_requests,
    list_sitemap,
@@ -43,7 +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,
@@ -110,11 +116,11 @@ def _tool_output_limits() -> tuple[int, int]:
    return context.tool_output_max_lines, context.tool_output_max_bytes


-def _bound_result(result: Any) -> Any:
+async def _bound_result(result: Any) -> Any:
    if not isinstance(result, str):
        return result
    max_lines, max_bytes = _tool_output_limits()
-    return bound_text(result, max_lines=max_lines, max_bytes=max_bytes)
+    return await bound_and_store(result, max_lines=max_lines, max_bytes=max_bytes)


 def _format_tool_error(exc: Exception) -> str:
@@ -130,7 +136,7 @@ def _with_bounded_result(tool: FunctionTool) -> FunctionTool:
    invoke_tool = tool.on_invoke_tool

    async def invoke(ctx: Any, raw_input: str) -> Any:
-        return _bound_result(await invoke_tool(ctx, raw_input))
+        return await _bound_result(await invoke_tool(ctx, raw_input))

    tool.on_invoke_tool = invoke
    tool._strix_bounded = True  # type: ignore[attr-defined]
@@ -142,7 +148,7 @@ def _function_tool_with_error_result(tool: FunctionTool) -> FunctionTool:

    async def invoke(ctx: Any, raw_input: str) -> Any:
        try:
-            return _bound_result(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)
@@ -157,7 +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 _bound_result(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)
@@ -194,7 +200,7 @@ def _bound_custom_tool(tool: CustomTool) -> CustomTool:
    invoke_tool = tool.on_invoke_tool

    async def invoke(ctx: Any, raw_input: str) -> Any:
-        return _bound_result(await invoke_tool(ctx, raw_input))
+        return await _bound_result(await invoke_tool(ctx, raw_input))

    tool.on_invoke_tool = invoke
    return tool
@@ -340,6 +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"
@@ -358,7 +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)
@@ -410,6 +420,8 @@ _BASE_TOOLS: tuple[Tool, ...] = (
    web_search,
    create_vulnerability_report,
    create_dependency_report,
+    list_reports,
+    get_report,
    list_requests,
    view_request,
    repeat_request,
@@ -418,7 +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,
 )
@@ -502,6 +514,9 @@ 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:
@@ -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>
@@ -18,12 +18,12 @@ import logging
 import secrets
 import threading
 import time
-import urllib.error
 import urllib.parse
-import urllib.request
 from pathlib import Path
 from typing import TYPE_CHECKING, Any

+import requests
+

 if TYPE_CHECKING:
    from collections.abc import Iterator
@@ -221,26 +221,19 @@ def _first(query: dict[str, list[str]], key: str) -> str | None:


 def _post_form(payload: dict[str, str]) -> dict[str, Any]:
-    body = urllib.parse.urlencode(payload).encode("ascii")
-    request = urllib.request.Request(  # noqa: S310 - fixed https OAuth endpoint
-        TOKEN_URL,
-        data=body,
-        headers={
-            "Content-Type": "application/x-www-form-urlencoded",
-            "Accept": "application/json",
-        },
-        method="POST",
-    )
    try:
-        with urllib.request.urlopen(  # noqa: S310  # nosec B310 - fixed https endpoint
-            request, timeout=_TOKEN_TIMEOUT
-        ) as response:
-            data = json.loads(response.read() or b"{}")
-    except urllib.error.HTTPError as exc:
-        detail = exc.read().decode("utf-8", "replace")[:300]
-        raise CodexAuthError("token_http_error", f"HTTP {exc.code}: {detail}") from exc
-    except (urllib.error.URLError, TimeoutError, OSError) as exc:
+        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
@@ -5,6 +5,7 @@ from __future__ import annotations
 import contextlib
 import inspect
 import os
+import time
 from typing import TYPE_CHECKING, Any

 from agents import (
@@ -13,6 +14,8 @@ from agents import (
    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 (
@@ -21,6 +24,8 @@ from agents.retry import (
    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
@@ -30,10 +35,17 @@ from strix.config.loader import load_settings
 if TYPE_CHECKING:
    from collections.abc import AsyncIterator

-    from agents.models.interface import Model, ModelProvider
+    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 ReasoningEffort, Settings
+    from strix.config.settings import LlmSettings, ReasoningEffort, Settings


 def request_timeout_extra_args(timeout_s: float | None) -> dict[str, float] | None:
@@ -135,6 +147,124 @@ class _CodexResponsesModel(OpenAIResponsesModel):
                    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
@@ -159,14 +289,21 @@ class StrixProvider(MultiProvider):
        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=load_settings().llm.reasoning_effort,
+                reasoning_effort=llm.reasoning_effort,
            )
-        return super().get_model(model_name)
+        model = super().get_model(model_name)
+        if llm.disable_streaming:
+            return _NonStreamingModel(model)
+        return model


 DEFAULT_MODEL_RETRY = ModelRetrySettings(
@@ -243,6 +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:
@@ -277,6 +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 = {
@@ -302,6 +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

@@ -429,3 +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
@@ -35,11 +35,23 @@ 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")


@@ -53,6 +65,10 @@ class DedupeSettings(BaseSettings):
    )
    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):
@@ -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:
@@ -42,11 +51,17 @@ class AgentCoordinator:
        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
@@ -65,6 +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,
@@ -109,11 +189,58 @@ class AgentCoordinator:
            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 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:
@@ -126,51 +253,55 @@ class AgentCoordinator:
            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,
@@ -178,17 +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:
@@ -299,7 +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:
@@ -310,6 +473,18 @@ class AgentCoordinator:
            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())

@@ -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):
@@ -438,28 +790,10 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                status = "crashed"
            logger.exception("agent run failed for %s; parking as %s", agent_id, status)
            await coordinator.set_status(agent_id, status, error=str(exc) or type(exc).__name__)
-            await _notify_parent_on_crash(coordinator, agent_id, status)
+            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:
@@ -477,23 +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]
@@ -502,12 +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)
@@ -518,16 +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],
@@ -579,6 +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)
@@ -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."
+                    )
@@ -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,
@@ -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)
@@ -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()
@@ -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)."""

@@ -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()
@@ -31,6 +31,7 @@ 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
@@ -210,7 +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()
@@ -223,7 +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")
@@ -381,7 +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=[],
@@ -403,7 +410,19 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
            # 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)
-            deduper_settings = ModelSettings(extra_args=deduper_extra or None)
+            # 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.",
@@ -422,7 +441,7 @@ async def warm_up_llm(show_model_warning: bool = True) -> None:
            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()
        sub_hint = _subscription_error_hint(e)
        if sub_hint is not None:
@@ -481,6 +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",
@@ -636,10 +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(
@@ -856,7 +902,7 @@ def display_completion_message(args: argparse.Namespace, results_path: Path) ->
    view_text = Text()
    view_text.append("\n")
    view_text.append("View", style="dim")
-    view_text.append("         ")
+    view_text.append("    ")
    view_text.append(f"strix view {args.run_name}", style="#22c55e")
    panel_parts.extend(["\n", view_text])

@@ -918,7 +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")
@@ -952,7 +998,7 @@ def main() -> None:
    # `strix view [<run>]` is a viewer-only subcommand, dispatched before the
    # scan argument parser (which requires a target) and before any scan setup.
    if len(sys.argv) > 1 and sys.argv[1] == "view":
-        from strix.viewer.cli import run_view
+        from strix.interface.viewer.cli import run_view

        run_view(sys.argv[2:])
        return
@@ -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
@@ -351,7 +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]
@@ -814,6 +816,7 @@ class StrixTUIApp(App):  # type: ignore[misc]
        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
@@ -1038,14 +1041,15 @@ class StrixTUIApp(App):  # type: ignore[misc]
                            name=names.get(agent_id, agent_id),
                            parent_id=parent_of.get(agent_id),
                            status=status,
-                            error_message=error,
+                            error_message=error or "",
                        )
-                        if status in {"failed", "crashed"} and error:
+                        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
@@ -1057,6 +1061,19 @@ class StrixTUIApp(App):  # type: ignore[misc]

        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
@@ -1069,6 +1086,7 @@ class StrixTUIApp(App):  # type: ignore[misc]
            status_indicators = {
                "running": "⚪",
                "waiting": "⏸",
+                "budget_paused": "⏸",
                "completed": "🟢",
                "failed": "🔴",
                "crashed": "🔴",
@@ -1266,10 +1284,21 @@ class StrixTUIApp(App):  # type: ignore[misc]
            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):
@@ -1494,6 +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,
                            ),
                        )
@@ -1501,10 +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
@@ -1559,6 +1586,7 @@ class StrixTUIApp(App):  # type: ignore[misc]
        status_indicators = {
            "running": "⚪",
            "waiting": "⏸",
+            "budget_paused": "⏸",
            "completed": "🟢",
            "failed": "🔴",
            "crashed": "🔴",
@@ -1605,6 +1633,7 @@ class StrixTUIApp(App):  # type: ignore[misc]
        status_indicators = {
            "running": "⚪",
            "waiting": "⏸",
+            "budget_paused": "⏸",
            "completed": "🟢",
            "failed": "🔴",
            "crashed": "🔴",
@@ -1729,7 +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()
@@ -1862,7 +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[/]")
@@ -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)
@@ -82,7 +82,7 @@ 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

@@ -223,7 +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"],
@@ -233,7 +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)
@@ -249,7 +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,
@@ -263,7 +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
@@ -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",
@@ -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
@@ -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")
@@ -431,3 +431,117 @@ class CreateDependencyReportRenderer(BaseToolRenderer):

        css_classes = cls.get_css_classes("completed")
        return Static(padded, classes=css_classes)
+
+
+_LIST_SEVERITY_COLORS = {
+    "critical": "#dc2626",
+    "high": "#ea580c",
+    "medium": "#d97706",
+    "low": "#65a30d",
+    "info": "#0284c7",
+    "none": "#6b7280",
+}
+
+
+def _severity_style(severity: Any) -> str:
+    return _LIST_SEVERITY_COLORS.get(str(severity or "").lower(), "#d97706")
+
+
+def _author_label(report: dict[str, Any]) -> str:
+    if report.get("by_you"):
+        return "you"
+    agent_name = report.get("agent_name")
+    return str(agent_name).strip() if agent_name else ""
+
+
+@register_tool_renderer
+class ListReportsRenderer(BaseToolRenderer):
+    tool_name: ClassVar[str] = "list_reports"
+    css_classes: ClassVar[list[str]] = ["tool-call", "reporting-tool"]
+
+    @classmethod
+    def render(cls, tool_data: dict[str, Any]) -> Static:
+        result = _coerce_dict(tool_data.get("result"))
+
+        text = Text()
+        text.append("◆ ", style="#ef4444")
+        text.append("reports", style="dim")
+
+        if isinstance(tool_data.get("result"), str) and str(tool_data["result"]).strip():
+            text.append("\n  ")
+            text.append(str(tool_data["result"]).strip(), style="dim")
+        elif result.get("success"):
+            total = result.get("total_count", 0)
+            reports = _coerce_list_of_dicts(result.get("reports"))
+            counts = _coerce_dict(result.get("severity_counts"))
+
+            text.append(f" ({total})", style="dim")
+            for sev, count in counts.items():
+                text.append("  ")
+                text.append(f"{sev} {count}", style=_severity_style(sev))
+
+            if not reports:
+                text.append("\n  ")
+                text.append("No reports filed yet", style="dim")
+            else:
+                for report in reports:
+                    rid = str(report.get("id", "")).strip()
+                    title = str(report.get("title", "")).strip() or "(untitled)"
+                    severity = str(report.get("severity", "")).strip()
+                    text.append("\n  - ")
+                    if severity:
+                        text.append(severity.upper(), style=f"bold {_severity_style(severity)}")
+                        text.append(" ")
+                    if rid:
+                        text.append(f"{rid} ", style="dim")
+                    text.append(title)
+                    author = _author_label(report)
+                    if author:
+                        text.append(f" ({author})", style="dim")
+        else:
+            text.append("\n  ")
+            text.append("Loading...", style="dim")
+
+        css_classes = cls.get_css_classes("completed")
+        return Static(text, classes=css_classes)
+
+
+@register_tool_renderer
+class GetReportRenderer(BaseToolRenderer):
+    tool_name: ClassVar[str] = "get_report"
+    css_classes: ClassVar[list[str]] = ["tool-call", "reporting-tool"]
+
+    @classmethod
+    def render(cls, tool_data: dict[str, Any]) -> Static:
+        result = _coerce_dict(tool_data.get("result"))
+
+        text = Text()
+        text.append("◆ ", style="#ef4444")
+        text.append("report read", style="dim")
+
+        report = _coerce_dict(result.get("report")) if result.get("success") else {}
+        if report:
+            rid = str(report.get("id", "")).strip()
+            title = str(report.get("title", "")).strip() or "(untitled)"
+            severity = str(report.get("severity", "")).strip()
+            text.append("\n  ")
+            if severity:
+                text.append(severity.upper(), style=f"bold {_severity_style(severity)}")
+                text.append(" ")
+            if rid:
+                text.append(f"{rid} ", style="dim")
+            text.append(title)
+            author = _author_label(report)
+            if author:
+                text.append(f" ({author})", style="dim")
+            target = str(report.get("target", "")).strip()
+            if target:
+                text.append("\n  ")
+                text.append(target, style="dim")
+        else:
+            text.append("\n  ")
+            detail = result.get("error") if result.get("success") is False else None
+            text.append(str(detail) if detail else "Loading...", style="dim")
+
+        css_classes = cls.get_css_classes("completed")
+        return Static(text, classes=css_classes)
@@ -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)
@@ -271,7 +271,13 @@ def _release_target() -> str | None:
    if os_name is None:
        return None
    target = f"{os_name}-{arch}"
-    supported = {"linux-x86_64", "macos-x86_64", "macos-arm64", "windows-x86_64"}
+    supported = {
+        "linux-x86_64",
+        "linux-arm64",
+        "macos-x86_64",
+        "macos-arm64",
+        "windows-x86_64",
+    }
    return target if target in supported else None


@@ -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
@@ -1088,13 +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
@@ -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"]
@@ -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]:
@@ -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)

@@ -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">
@@ -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>
            ))}
@@ -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>
+  );
+}
@@ -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>
+  );
+}
@@ -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" },
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Ahmed Allam	d97dc1b2ef	docs(prompts): text-only turns no longer end an autonomous run	2026-08-01 22:16:00 +00:00
Ahmed Allam	8b464dae5d	refactor(tools): split wait_for_message into respond_to_user + wait_for_agents 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.	2026-08-01 22:13:07 +00:00
Ahmed Allam	b7bf52c468	docs(core): correct the rationale for notifying a stalled child's parent The user can message any agent from the TUI, not only the root, so the justification is that the parent is an agent with no other way to learn the child parked - not that the child has no human resumer.	2026-08-01 21:36:54 +00:00
Ahmed Allam	ceff3b5408	fix(core): tell the parent when an interactive subagent parks Parking is self-service only for the root, which the user is watching. A parked child owes its parent a report it can no longer send, so the parent would wait out its full timeout for nothing.	2026-08-01 21:30:44 +00:00
Ahmed Allam	44bb3abdf8	fix(tools): halve the wait_for_message ceiling to 300s A mutual wait between two agents resolves only when both hit their cap, so the ceiling is the worst-case idle burn. Name the constants instead of repeating the literal, and align the interactive auto-resume timeout.	2026-08-01 21:25:09 +00:00
Ahmed Allam	5726c2d4ef	fix(core): persist the tool-call recovery counter across resumes An exhausted agent parked in 'waiting' got a fresh nudge budget on every 600s auto-resume, so a wedged agent could nudge-park-nudge indefinitely. Track the count on the coordinator, snapshot it, and reset it only on real input or an explicit lifecycle tool.	2026-08-01 19:36:20 +00:00
Ahmed Allam	69a60f3b7a	fix(core): stop interactive runs stalling on a missing tool call Interactive turns ended by plain text left the agent parked in 'waiting' forever. Require an explicit lifecycle tool in both modes and nudge a text-only turn back into a tool call, bounded by a recovery limit.	2026-08-01 19:16:53 +00:00
devin-ai-integration[bot]andGitHub	5602bc23ca	fix: pre-v1-style lifecycle resilience — mailbox delivery, uniform revival, unexitable runner, waiting timeout, broader retries, crash-safe identity (#923 )	2026-08-01 11:17:08 -07:00
alex sandGitHub	a9deb84260	fix(llm): surface structured provider refusals (#944 ) * fix(llm): surface structured provider refusals * fix(llm): settle refused autonomous agents	2026-07-31 10:57:42 -04:00
chunguscodesandAhmed Allam	76e97e6a59	fix(llm): avoid auth during ChatGPT lookup LiteLLM treats provider-qualified metadata lookups as an auth path. Use the underlying model slug so context sizing cannot block the scan loop in a device-code poll.	2026-07-31 03:45:41 +03:00
Ahmed AllamandAhmed Allam	885b2ca5c5	test(llm): cover the full run loop against a non-streaming gateway; drop README note Adds an integration test that drives Runner.run_streamed against a non-streaming gateway through _NonStreamingModel: the synthetic terminal event feeds the runner, which executes the tool call and continues to a final answer over two non-streaming turns. Removes the README env-var note.	2026-07-30 08:30:06 +03:00
Ahmed AllamandAhmed Allam	980216860e	feat(llm): opt-in LLM_DISABLE_STREAMING for non-streaming OpenAI-compatible endpoints Some OpenAI-compatible gateways don't support Server-Sent Events (or deliver them unreliably), but the SDK run loop Strix uses only issues streamed requests, so such a gateway fails every turn. Add an opt-in LLM_DISABLE_STREAMING setting that wraps the resolved model in _NonStreamingModel: each turn makes one non-streaming get_response and replays the completed result as a single terminal stream event, so tool calls, usage, and the rest of the agent loop are unchanged. Subscription (ChatGPT) models are always streamed and are not wrapped.	2026-07-30 08:30:06 +03:00
devin-ai-integration[bot]andGitHub	d4e58b2cd0	fix(llm): pass LLM_EXTRA_HEADERS through ModelSettings so they reach the agent loop (#937 )	2026-07-29 19:38:06 -07:00
Ahmed AllamandAhmed Allam	e9ebdc502f	fix(llm): apply LLM_EXTRA_HEADERS on native OpenAI route even without a custom base	2026-07-30 04:13:25 +03:00
Ahmed AllamandAhmed Allam	ebb3a62a99	feat(llm): custom request headers for OpenAI-compatible endpoints via LLM_EXTRA_HEADERS	2026-07-30 04:13:25 +03:00
devin-ai-integration[bot]GitHub Ahmed Allam	1a2fa89972	fix(runtime): label docker sandbox containers with the run id for teardown (#933 ) Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>	2026-07-29 08:05:42 -07:00
alex sandGitHub	9de747d135	fix(cost): capture OpenRouter streamed usage.cost (fixes $0 kimi-k3 c… (#929 ) * fix(cost): capture OpenRouter streamed usage.cost (fixes $0 kimi-k3 cost) * refactor(cost): encapsulate streamed OpenRouter cost cache, clear per run * test(cost): resolve OpenRouter handler via LiteLLM provider pipeline	2026-07-28 23:28:34 -04:00
yoni-at-strix GitHub Jonathan Singer	b313d78f60	Scope viewer session cookie to the bound port (#922 ) Co-authored-by: Jonathan Singer <jonathansinger@Mac-4078.lan>	2026-07-27 20:37:54 -04:00
devin-ai-integration[bot]GitHub Ahmed Allam	e037d8d727	fix: recoverable guardrail blocks and decoupled crash-notify (#919 ) Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>	2026-07-27 16:28:48 -07:00
alex sandGitHub	fade37025d	fix viewer tool call collisions across agents (#917 )	2026-07-27 18:51:02 -04:00
Ahmed AllamandAhmed Allam	f968f8e5a7	fix(cli): align View label spacing in final panel	2026-07-27 15:41:22 -07:00
Ahmed AllamandAhmed Allam	ac0014fe65	chore: release v1.4.1	2026-07-27 12:57:39 -07:00
yoni-at-strix GitHub Jonathan Singer Ahmed Allam	86282e83a8	fix(tls): replace raw urllib with requests for external HTTPS calls (frozen-build cert failures) (#903 ) Co-authored-by: Jonathan Singer <jonathansinger@Mac-4051.lan> Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>	2026-07-27 12:34:26 -07:00
Ahmed AllamandAhmed Allam	37c7f5a6ba	chore: release v1.4.0	2026-07-27 04:55:21 -07:00
Ahmed AllamandAhmed Allam	082d4ae62c	fix(runtime): wake parent when child hits a terminal state (MaxTurnsExceeded)	2026-07-27 04:27:21 -07:00
devin-ai-integration[bot]GitHub Ahmed Allam	c55a8fa4ba	feat(runtime): graduated wrap-up warnings, budget reserve, and interactive budget pause/continue (#893 ) Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>	2026-07-26 20:37:14 -07:00
devin-ai-integration[bot]andGitHub	47617969d3	fix(cli): don't dump raw warm-up traceback over the LLM error panel (#896 )	2026-07-26 20:01:20 -07:00
seanturner83 GitHub Sean Turner Ahmed Allam	27f9750cdc	feat(llm): enable Bedrock/Anthropic prompt caching for Claude models (#772 ) Co-authored-by: Sean Turner <sean.turner@zerohash.com> Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>	2026-07-26 17:12:57 -07:00
Matthew BrightandGitHub	427cdcd9d4	Add Linux ARM64 standalone release support (#886 )	2026-07-26 16:27:02 -07:00
devin-ai-integration[bot]GitHub Ahmed Allam	384338cf31	fix(runtime): retry transient mid-stream provider errors instead of crashing the scan (#891 ) Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>	2026-07-26 16:23:31 -07:00
devin-ai-integration[bot]GitHub Ahmed Allam	3b79e97f00	feat(context): spill oversized tool output into the sandbox workspace (#882 ) Co-authored-by: Ahmed Allam <ahmed39652003@gmail.com>	2026-07-26 14:42:22 -07:00
devin-ai-integration[bot]andGitHub	66a283b71b	feat(context): model-aware conversation compaction for long scans (#881 )	2026-07-26 14:39:41 -07:00
Ahmed AllamandAhmed Allam	74f334cb93	refactor(context): trim verbose comments	2026-07-26 14:38:12 -07:00
Ahmed AllamandAhmed Allam	8e9a6bf903	fix(context): reject tool-output byte ceilings below the notice size A configured tool_output_max_bytes smaller than the truncation notice itself can't fit a bounded preview, so a persisted result could exceed the ceiling. Enforce a config floor (ge=1024) so nonsensical values are rejected at load time instead of being worked around at runtime.	2026-07-26 14:38:12 -07:00
Ahmed AllamandAhmed Allam	0ebd3c6230	fix(context): reserve notice budget so bounded output honors max_bytes The head+tail slices could each take half of max_bytes, then the truncation notice and its separators were appended on top, so the value persisted to history could exceed the configured maximum. Reserve an upper bound for the notice (and separators) out of the byte budget before slicing so the whole joined result stays within max_bytes.	2026-07-26 14:38:12 -07:00
Ahmed AllamandAhmed Allam	6bda366065	fix(context): bound native filesystem tool output in Responses mode Chat-completions mode converts filesystem CustomTools to FunctionTools (which bounds their result), but the Responses-API path kept them native and unbounded, so a large read_file could still exhaust the context window. Always configure the Filesystem capability to head+tail bound tool output in both modes.	2026-07-26 14:38:12 -07:00
Ahmed AllamandAhmed Allam	1f36f5d401	fix(context): clamp shell output cap and count byte-trimmed dropped lines Treat tool_output_max_tokens as a ceiling so an explicit model-supplied cap can't exceed it, and derive the truncation notice's dropped-line count from the lines actually kept after the byte-trim pass. Also cast the pygments fallback lexer so it satisfies the resolve_lexer return type under the pre-commit mypy hook.	2026-07-26 14:38:12 -07:00
Ahmed AllamandAhmed Allam	a70a87f272	feat(context): bound per-tool output before it enters agent history Cap the size of every tool result so a single verbose command (recursive find, noisy scanner, full page dump) can't pin the conversation near the model's context window for the rest of a scan. - New ContextSettings config group with env-tunable caps. - Default the SDK shell tools' max_output_tokens so exec_command / write_stdin truncate head+tail instead of returning unbounded output. - Bound Strix's own FunctionTool/CustomTool results (line + UTF-8 byte head+tail preview with a truncation notice) and cap error strings.	2026-07-26 14:38:12 -07:00
devin-ai-integration[bot]andGitHub	d2fbcb726d	feat(reporting): add read-only list_reports + get_report tools (#889 )	2026-07-26 14:05:53 -07:00
Ahmed AllamandAhmed Allam	8169e177de	docs(skills): remove references to tools not installed in the sandbox Skills and the agent system prompt referenced external CLIs that are not present in containers/Dockerfile, which could lead the agent to invoke missing binaries. Replace them with installed equivalents: - asset_discovery: drop amass/cero and the projectdiscovery tools that are not installed (tlsx/dnsx/asnmap/mapcidr/uncover); rewrite around the installed subfinder/httpx/naabu plus curl+jq (crt.sh), openssl s_client, dig, and whois. Stop claiming the full projectdiscovery suite is available. - subdomain_takeover: replace dnsx with dig in the pipeline example. - weak_password_detection: drop hydra/cewl/patator; use ffuf for web logins and nmap NSE *-brute scripts for services; fix dead /usr/share/wordlists and /usr/share/seclists paths (nothing ships by default -> download to /home/pentester/tools/wordlists at runtime). - system_prompt: replace msfconsole with sqlmap in the interactive-process example. active_directory skill is left as-is: it already ships an explicit install block for its tools.	2026-07-26 14:00:07 -07:00
Ahmed AllamandAhmed Allam	589bade39a	fix: restore viewer-auth.json path in auth module docstring	2026-07-26 13:11:14 -07:00
Ahmed AllamandAhmed Allam	d1e8225d5f	refactor: move strix/viewer under strix/interface	2026-07-26 13:11:14 -07:00