fix(skills): strengthen DoubleClickjacking validation

feat(skills): improve DoubleClickjacking detection and coverage
fix(report): validate deduplication decisions
2026-11-09 05:19:39 -10:00 · 2026-11-05 15:30:05 +02:00 · 2026-11-05 02:46:15 +02:00 · 2026-11-04 21:57:34 +02:00 · 2026-11-04 10:26:08 -05:00 · 2026-10-29 04:30:48 +03:00
120 changed files with 5284 additions and 622 deletions
@@ -31,6 +31,8 @@ jobs:
          assert not any(name.startswith("strix/interface/tui/cmd/") for name in names)
          PY
      - name: Install into a clean environment and run the CLI
+        env:
+          UV_EXCLUDE_NEWER: 5 days
        run: |
          uv venv --python 3.12 /tmp/strix-smoke
          uv pip install --python /tmp/strix-smoke/bin/python dist/*.whl
@@ -28,7 +28,7 @@ Target-specific workflows built on the same engine:
 - **Open-source CLI (self-hosted):** free, fully local, BYO LLM key, needs Docker. Best for local dev loops, air-gapped/offline, and full control.
  ```bash
  curl -sSL https://strix.ai/install | bash        # install
-  export STRIX_LLM="openrouter/z-ai/glm-5.3"        # any LiteLLM model id
+  export STRIX_LLM="openrouter/z-ai/glm-5.3"        # any LiteLLM model id, or vercel_ai_gateway/zai/glm-5.3
  export LLM_API_KEY="<key>"
  strix -n -t ./ --scan-mode quick --max-budget 10  # headless scan; always use -n
  ```
@@ -31,7 +31,7 @@ Thank you for your interest in contributing to Strix! This guide will help you g

 3. **Configure your LLM provider**
   ```bash
-   export STRIX_LLM="openrouter/z-ai/glm-5.3"
+   export STRIX_LLM="openrouter/z-ai/glm-5.3"   # or vercel_ai_gateway/zai/glm-5.3
   export LLM_API_KEY="your-api-key"
   ```

@@ -84,7 +84,8 @@ Strix are autonomous AI penetration testing agents that act just like real hacke
 curl -sSL https://strix.ai/install | bash

 # Configure your AI provider
-export STRIX_LLM="openrouter/z-ai/glm-5.3"
+export STRIX_LLM="vercel_ai_gateway/zai/glm-5.3"   # Vercel AI Gateway
+# export STRIX_LLM="openrouter/z-ai/glm-5.3"       # OpenRouter
 export LLM_API_KEY="your-api-key"

 # Run your first security assessment
@@ -94,6 +95,18 @@ strix --target ./app-directory
 > [!NOTE]
 > First run automatically pulls the sandbox Docker image. Results are saved to `strix_runs/<run-name>`

+### Updating
+
+```bash
+# Self-update the standalone binary to the latest release
+strix --update
+
+# ...or re-run the installer
+curl -sSL https://strix.ai/install | bash
+```
+
+Installed with pipx, uv, or pip? `strix --update` prints the matching upgrade command instead (`pipx upgrade strix-agent`, `uv tool upgrade strix-agent`, `pip install --upgrade strix-agent`). Strix also shows a notice with the right command when a new version is available.
+
 ---

 ## Ways to Run Strix
@@ -284,7 +297,8 @@ jobs:
 ### Configuration

 ```bash
-export STRIX_LLM="openrouter/z-ai/glm-5.3"
+export STRIX_LLM="vercel_ai_gateway/zai/glm-5.3"   # Vercel AI Gateway
+# export STRIX_LLM="openrouter/z-ai/glm-5.3"       # OpenRouter
 export LLM_API_KEY="your-api-key"

 # Optional
@@ -340,12 +354,14 @@ Each server's tools are namespaced by `name`, for example `github_list_issues`.

 **Recommended models for best results:**

- [Z.ai GLM-5.3 on OpenRouter](https://openrouter.ai/z-ai/glm-5.3) - `openrouter/z-ai/glm-5.3` (the default pick)
- [OpenAI GPT-5.4](https://openai.com/api/) - `openai/gpt-5.4`
- [Anthropic Claude Sonnet 4.6](https://claude.com/platform/api) - `anthropic/claude-sonnet-4-6`
- [Google Gemini 3 Pro Preview](https://cloud.google.com/vertex-ai) - `vertex_ai/gemini-3-pro-preview`
+- [Z.ai GLM-5.3 on Vercel AI Gateway](https://vercel.com/ai-gateway/models/glm-5.3) - `vercel_ai_gateway/zai/glm-5.3` (recommended)
+- [Z.ai GLM-5.3 on OpenRouter](https://openrouter.ai/z-ai/glm-5.3) - `openrouter/z-ai/glm-5.3` (recommended)
+- [DeepSeek V4.1 Flash on OpenRouter](https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017) - `openrouter/deepseek/deepseek-v4.1-flash-1017`
+- [OpenAI GPT-6.1 Sol](https://openai.com/api/) - `openai/gpt-6.1-sol` (needs [Daybreak](https://openai.com/daybreak/) access)
+- [Anthropic Claude Opus 5.5](https://claude.com/platform/api) - `anthropic/claude-opus-5-5` (needs [CVP](https://support.claude.com/en/articles/14604842-cyber-verification-program) access)
+- [Google Gemini 4 Argon](https://cloud.google.com/vertex-ai) - `vertex_ai/gemini-4-argon` (needs [Fairwind](https://deepmind.google/fairwind-program/) access)
 - [DeepSeek V4 Pro](https://platform.deepseek.com) - `deepseek/deepseek-v4-pro`
- [Moonshot Kimi K3](https://platform.kimi.ai) - `moonshot/kimi-k3`
+- [Moonshot Kimi K3.1 Pro](https://platform.kimi.ai) - `moonshot/kimi-k3.1-pro`

 See the [LLM Providers documentation](https://docs.strix.ai/llm-providers/overview) for all supported providers including Vertex AI, Bedrock, Azure, and local models.

@@ -176,7 +176,7 @@ USER root
 # overwrite a root-owned binary under /usr/local/bin, which otherwise fails with
 # "cannot move binary" and aborts the scan. Pin the initial version for
 # reproducible builds; self-update then pulls fresh detectors at runtime.
-ARG TRUFFLEHOG_VERSION=3.95.9
+ARG TRUFFLEHOG_VERSION=3.99.2
 RUN curl -sSfL https://raw.githubusercontent.com/trufflesecurity/trufflehog/main/scripts/install.sh | sh -s -- -b /home/pentester/.local/bin "v${TRUFFLEHOG_VERSION}" && \
    chown -R pentester:pentester /home/pentester/.local
 RUN set -eux; \
@@ -186,20 +186,84 @@ RUN set -eux; \
        aarch64|arm64) GITLEAKS_ARCH="arm64" ;; \
        *) echo "Unsupported architecture: $ARCH" >&2; exit 1 ;; \
    esac; \
-    TAG="$(curl -fsSL https://api.github.com/repos/gitleaks/gitleaks/releases/latest | jq -r .tag_name)"; \
+    TAG="$(curl -fsSLI -o /dev/null -w '%{url_effective}' https://github.com/gitleaks/gitleaks/releases/latest | sed 's#.*/##')"; \
    curl -fsSL "https://github.com/gitleaks/gitleaks/releases/download/${TAG}/gitleaks_${TAG#v}_linux_${GITLEAKS_ARCH}.tar.gz" -o /tmp/gitleaks.tgz; \
    tar -xzf /tmp/gitleaks.tgz -C /tmp; \
    install -m 0755 /tmp/gitleaks /usr/local/bin/gitleaks; \
    rm -f /tmp/gitleaks /tmp/gitleaks.tgz

-RUN curl -sfL https://raw.githubusercontent.com/aquasecurity/trivy/main/contrib/install.sh | sh -s -- -b /usr/local/bin
+ARG TRIVY_VERSION=0.76.1
+ARG TRIVY_SHA256_X86=c7cbe3137fb8a48a17574bede3dd6beb16742d1e9772929940fc90082c39d9c0
+ARG TRIVY_SHA256_AARCH64=e174dff2a6dba5e3ae41f8d0537a6324405bd08132497340304c6b20e0f16a1f
+RUN set -eux; \
+    case "$(uname -m)" in \
+        x86_64) TRIVY_ARCH="64bit"; TRIVY_SHA256="${TRIVY_SHA256_X86}" ;; \
+        aarch64|arm64) TRIVY_ARCH="ARM64"; TRIVY_SHA256="${TRIVY_SHA256_AARCH64}" ;; \
+        *) echo "Unsupported architecture: $(uname -m)" >&2; exit 1 ;; \
+    esac; \
+    curl -fsSL "https://github.com/aquasecurity/trivy/releases/download/v${TRIVY_VERSION}/trivy_${TRIVY_VERSION}_Linux-${TRIVY_ARCH}.tar.gz" -o /tmp/trivy.tgz; \
+    echo "${TRIVY_SHA256}  /tmp/trivy.tgz" | sha256sum -c -; \
+    tar -xzf /tmp/trivy.tgz -C /tmp trivy; \
+    install -m 0755 /tmp/trivy /usr/local/bin/trivy; \
+    rm -f /tmp/trivy /tmp/trivy.tgz; \
+    trivy --version

 RUN apt-get install -y wapiti

+# CVE-2026-111497: out-of-bounds write in the brotli decoder, fixed in brotli
+# 1.2.1 / Debian 1.2.0-4.1. kali-rolling still ships 1.2.0-4, and node, curl and
+# python3 in this image all load the shared libbrotlidec, so every HTTP client here
+# would decode a malicious `Content-Encoding: br` response with the vulnerable
+# code. Install the fixed Debian packages until Kali picks them up.
+ARG BROTLI_DEB_VERSION=1.2.0-4.1
+RUN set -eux; \
+    ARCH="$(dpkg --print-architecture)"; \
+    for pkg in libbrotli1 python3-brotli; do \
+        curl -fsSL "https://deb.debian.org/debian/pool/main/b/brotli/${pkg}_${BROTLI_DEB_VERSION}_${ARCH}.deb" -o "/tmp/${pkg}.deb"; \
+    done; \
+    dpkg -i /tmp/libbrotli1.deb /tmp/python3-brotli.deb; \
+    rm -f /tmp/libbrotli1.deb /tmp/python3-brotli.deb; \
+    test "$(dpkg-query -W -f='${Version}' libbrotli1)" = "${BROTLI_DEB_VERSION}"; \
+    node -e 'const z=require("zlib");z.brotliDecompressSync(z.brotliCompressSync(Buffer.from("ok")))'
+
+ARG OPENGREP_VERSION=1.31.0
+ARG OPENGREP_SHA256_X86=dbbe3c2058dbe600c449a8ba22ea79a1643a6660658400090d0323a8d0d31147
+ARG OPENGREP_SHA256_AARCH64=2b713a29a00237f85408b83002fc111db5b94c7fa8d263f0a93184a960655d9d
+RUN set -eux; \
+    case "$(uname -m)" in \
+        x86_64) OPENGREP_ARCH="x86"; OPENGREP_SHA256="${OPENGREP_SHA256_X86}" ;; \
+        aarch64|arm64) OPENGREP_ARCH="aarch64"; OPENGREP_SHA256="${OPENGREP_SHA256_AARCH64}" ;; \
+        *) echo "Unsupported architecture: $(uname -m)" >&2; exit 1 ;; \
+    esac; \
+    curl -fsSL "https://github.com/opengrep/opengrep/releases/download/v${OPENGREP_VERSION}/opengrep_manylinux_${OPENGREP_ARCH}" -o /tmp/opengrep; \
+    echo "${OPENGREP_SHA256}  /tmp/opengrep" | sha256sum -c -; \
+    install -m 0755 /tmp/opengrep /usr/local/bin/opengrep; \
+    rm -f /tmp/opengrep; \
+    opengrep --version
+
+# Opengrep ships no rule registry of its own; bake the community ruleset so
+# scans run offline. Keep only rule YAML and LICENSE; drop tests, fixtures,
+# tooling, and rules marked with a proprietary license.
+ARG OPENGREP_RULES_REF=f1d2b562b414783763fd02a6ed2736eaed622efa
+RUN set -eux; \
+    mkdir -p /opt/opengrep-rules; \
+    curl -fsSL "https://codeload.github.com/opengrep/opengrep-rules/tar.gz/${OPENGREP_RULES_REF}" \
+        | tar -xz --strip-components=1 -C /opt/opengrep-rules; \
+    cd /opt/opengrep-rules; \
+    rm -rf .github scripts stats; \
+    find . -maxdepth 1 -type f ! -name LICENSE -delete; \
+    find . -mindepth 2 -type f \( ! \( -name '*.yaml' -o -name '*.yml' \) -o -name '*.test.yaml' -o -name '*.test.yml' -o -name '.*' \) -delete; \
+    grep -rlZ 'license: proprietary' . | xargs -0 -r rm -f; \
+    find . -type d -empty -delete; \
+    chmod -R a+rX /opt/opengrep-rules
+
+# Strix-authored rules that close gaps in the community ruleset.
+COPY containers/opengrep-rules/ /opt/opengrep-rules/strix/
+RUN chmod -R a+rX /opt/opengrep-rules/strix
+
 USER pentester

-RUN pipx install semgrep && \
-    pipx install bandit
+RUN pipx install bandit

 RUN npm install -g jshint

@@ -0,0 +1,37 @@
+rules:
+  - id: github-owned-action-not-pinned-to-commit-sha
+    patterns:
+      - pattern-inside: "{steps: ...}"
+      - pattern: |
+          uses: "$USES"
+      - metavariable-regex:
+          metavariable: $USES
+          regex: '^(actions|github)/'
+      - metavariable-pattern:
+          metavariable: $USES
+          language: generic
+          patterns:
+            - pattern-not-regex: "@[0-9a-f]{40}$"
+    message: >-
+      GitHub-owned action `$USES` is referenced by a mutable tag or branch
+      instead of a full-length commit SHA. Tags can be moved, so a compromised
+      or force-pushed release changes what this workflow runs. Pin to the
+      commit SHA of the release you reviewed and let a bot keep it current.
+    languages:
+      - yaml
+    severity: WARNING
+    metadata:
+      cwe:
+        - "CWE-1357: Reliance on Insufficiently Trustworthy Component"
+        - "CWE-353: Missing Support for Integrity Check"
+      owasp: A06:2021 - Vulnerable and Outdated Components
+      references:
+        - https://docs.github.com/en/actions/security-guides/security-hardening-for-github-actions#using-third-party-actions
+      category: security
+      technology:
+        - github-actions
+      subcategory:
+        - vuln
+      likelihood: LOW
+      impact: LOW
+      confidence: HIGH
@@ -8,7 +8,7 @@ Configure Strix using environment variables or a config file.
 ## LLM Configuration

 <ParamField path="STRIX_LLM" type="string" required>
-  Model name in LiteLLM format (e.g., `openrouter/z-ai/glm-5.3`, `openai/gpt-5.4`).
+  Model name in LiteLLM format (e.g., `openrouter/z-ai/glm-5.3`, `openai/gpt-6.1-sol`).
 </ParamField>

 <ParamField path="LLM_API_KEY" type="string">
@@ -19,10 +19,29 @@ 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="STRIX_LLM_INPUT_COST_PER_MTOK" type="number">
+  Price per 1M input tokens in USD for `STRIX_LLM`. Use it for gateway or self-hosted model ids that LiteLLM can't price.
+</ParamField>
+
+<ParamField path="STRIX_LLM_OUTPUT_COST_PER_MTOK" type="number">
+  Price per 1M output tokens in USD for `STRIX_LLM`.
+</ParamField>
+
+<ParamField path="STRIX_LLM_CACHED_INPUT_COST_PER_MTOK" type="number">
+  Price per 1M cached input tokens in USD for `STRIX_LLM`. If unset, cached tokens use the input price.
+</ParamField>
+
+<Note>
+  Setting any of these prices replaces the provider-reported and LiteLLM cost for `STRIX_LLM` in reports and `--max-budget`. An unset input or output price counts as `0`, so set both. A separate `STRIX_DEDUPE_MODEL` keeps its LiteLLM price.
+</Note>
+
 <ParamField path="STRIX_API_TYPE" type="string">
  Select the OpenAI API path for the model: `responses` or `chat_completions`.
-  By default, a custom `LLM_API_BASE` uses chat completions. Set this variable
-  when your gateway requires the other API. Also accepts `STRIX_FORCE_API`.
+  By default the model decides: Responses when LiteLLM's model catalog lists
+  `/v1/responses` for it (e.g. `gpt-5`), chat completions otherwise. A custom
+  `LLM_API_BASE` does not change this, so set `STRIX_API_TYPE=chat_completions`
+  when your gateway only serves chat completions. Applies to bare and `openai/`
+  model names. Also accepts `STRIX_FORCE_API`.
 </ParamField>

 <ParamField path="LLM_EXTRA_HEADERS" type="string">
@@ -33,6 +52,18 @@ Configure Strix using environment variables or a config file.
  the LiteLLM and native OpenAI routing paths.
 </ParamField>

+<ParamField path="LLM_EXTRA_BODY" type="string">
+  Extra fields merged into the JSON body of every request to the main model, as a JSON object
+  (e.g. `{"service_tier":"flex","safety_identifier":"9f86d081..."}`). Use it
+  for provider parameters Strix doesn't expose, such as OpenAI's `service_tier`
+  or `safety_identifier`. Applies to both the LiteLLM and native OpenAI routing
+  paths. Fields are not validated, so the provider or gateway must accept them;
+  unknown fields are usually rejected. Extra fields can override fields Strix
+  sends and break requests, so don't set request-shaping fields such as `model`, `messages`, `input`,
+  `tools` or `stream`, and use `STRIX_REASONING_EFFORT` rather than
+  `reasoning_effort`.
+</ParamField>
+
 <ParamField path="LLM_TIMEOUT" default="300" type="integer">
  Request timeout in seconds for LLM calls.
 </ParamField>
@@ -88,6 +119,12 @@ affecting the agents that do the actual testing.
  inherits `LLM_EXTRA_HEADERS`; set this when its endpoint needs custom headers.
 </ParamField>

+<ParamField path="DEDUPE_LLM_EXTRA_BODY" type="string">
+  Optional JSON object of extra request-body fields sent on every
+  deduplication-model request. A dedicated dedupe model never inherits
+  `LLM_EXTRA_BODY`; set this when its endpoint needs them too.
+</ParamField>
+
 <ParamField path="STRIX_DEDUPE_REASONING_EFFORT" type="string">
  Reasoning effort for the deduplication model. Defaults to the model's own
  baseline when unset.
@@ -123,6 +160,10 @@ affecting the agents that do the actual testing.
  Telemetry toggle. Set to `0`, `false`, `no`, or `off` to disable telemetry (PostHog, Scarf, OTEL).
 </ParamField>

+<ParamField path="STRIX_RUN_RECORD_TIMINGS" default="false" type="boolean">
+  Record every LLM request's and tool call's timing in `run.json`, under `llm_requests` and `tool_calls`. Adds about 0.5 MB to a typical scan's `run.json`.
+</ParamField>
+
 <ParamField path="TRACELOOP_BASE_URL" type="string">
  OTLP/Traceloop base URL for remote OpenTelemetry export. If unset, Strix keeps traces local only.
 </ParamField>
@@ -145,7 +186,7 @@ When remote vars are set, Strix dual-writes telemetry to both local JSONL and th

 ## Docker Configuration

-<ParamField path="STRIX_IMAGE" default="ghcr.io/usestrix/strix-sandbox:1.3.0" type="string">
+<ParamField path="STRIX_IMAGE" default="ghcr.io/usestrix/strix-sandbox:1.4.0" type="string">
  Docker image to use for the sandbox container.
 </ParamField>

@@ -35,8 +35,8 @@
              "llm-providers/overview",
              "llm-providers/openai",
              "llm-providers/anthropic",
-              "llm-providers/openrouter",
              "llm-providers/vercel-ai-gateway",
+              "llm-providers/openrouter",
              "llm-providers/vertex",
              "llm-providers/bedrock",
              "llm-providers/azure",
@@ -73,7 +73,20 @@ Strix uses a graph of specialized agents for comprehensive security testing:

 ## Quick Example

-```bash
+<CodeGroup>
+```bash Vercel AI Gateway
+# Install
+curl -sSL https://strix.ai/install | bash
+
+# Configure
+export STRIX_LLM="vercel_ai_gateway/zai/glm-5.3"
+export LLM_API_KEY="your-api-key"
+
+# Scan
+strix --target ./your-app
+```
+
+```bash OpenRouter
 # Install
 curl -sSL https://strix.ai/install | bash

@@ -84,6 +97,7 @@ export LLM_API_KEY="your-api-key"
 # Scan
 strix --target ./your-app
 ```
+</CodeGroup>

 ## Community

@@ -6,7 +6,7 @@ description: "Configure Strix with Claude models"
 ## Setup

 ```bash
-export STRIX_LLM="anthropic/claude-sonnet-4-6"
+export STRIX_LLM="anthropic/claude-sonnet-5-5"
 export LLM_API_KEY="sk-ant-..."
 ```

@@ -14,8 +14,12 @@ export LLM_API_KEY="sk-ant-..."

 | Model | Description |
 |-------|-------------|
-| `anthropic/claude-sonnet-4-6` | Best balance of intelligence and speed |
-| `anthropic/claude-opus-4-6` | Maximum capability for deep analysis |
+| `anthropic/claude-sonnet-5-5` | Best balance of intelligence and speed |
+| `anthropic/claude-opus-5-5` | Maximum capability for deep analysis |
+
+<Note>
+Security testing with Claude needs the [Cyber Verification Program](https://support.claude.com/en/articles/14604842-cyber-verification-program) (CVP). Without it the safety classifiers block most offensive tasks. Red Team Access covers authorized penetration testing.
+</Note>

 ## Get API Key

@@ -6,7 +6,7 @@ description: "Configure Strix with OpenAI models via Azure"
 ## Setup

 ```bash
-export STRIX_LLM="azure/your-gpt5-deployment"
+export STRIX_LLM="azure/your-gpt6-deployment"
 export AZURE_API_KEY="your-azure-api-key"
 export AZURE_API_BASE="https://your-resource.openai.azure.com"
 export AZURE_API_VERSION="2025-11-01-preview"
@@ -24,7 +24,7 @@ export AZURE_API_VERSION="2025-11-01-preview"
 ## Example

 ```bash
-export STRIX_LLM="azure/gpt-5.4-deployment"
+export STRIX_LLM="azure/gpt-6.1-sol-deployment"
 export AZURE_API_KEY="abc123..."
 export AZURE_API_BASE="https://mycompany.openai.azure.com"
 export AZURE_API_VERSION="2025-11-01-preview"
@@ -33,5 +33,9 @@ export AZURE_API_VERSION="2025-11-01-preview"
 ## Prerequisites

 1. Create an Azure OpenAI resource
-2. Deploy a model (e.g., GPT-5.4)
+2. Deploy a model such as GPT-6.1 Sol
 3. Get the endpoint URL and API key from the Azure portal
+
+<Note>
+Security testing with OpenAI models needs [Daybreak](https://openai.com/daybreak/) access, also on Azure.
+</Note>
@@ -6,7 +6,7 @@ description: "Configure Strix with models via AWS Bedrock"
 ## Setup

 ```bash
-export STRIX_LLM="bedrock/anthropic.claude-4-5-sonnet-20251022-v1:0"
+export STRIX_LLM="bedrock/us.anthropic.claude-sonnet-5-5"
 ```

 No API key required—uses AWS credentials from environment.
@@ -36,10 +36,13 @@ Automatically uses instance role credentials.

 | Model | Description |
 |-------|-------------|
-| `bedrock/anthropic.claude-4-5-sonnet-20251022-v1:0` | Claude 4.5 Sonnet |
-| `bedrock/anthropic.claude-4-5-opus-20251022-v1:0` | Claude 4.5 Opus |
-| `bedrock/anthropic.claude-4-5-haiku-20251022-v1:0` | Claude 4.5 Haiku |
-| `bedrock/amazon.titan-text-premier-v2:0` | Amazon Titan Premier v2 |
+| `bedrock/us.anthropic.claude-sonnet-5-5` | Claude Sonnet 5.5 |
+| `bedrock/us.anthropic.claude-opus-5-5` | Claude Opus 5.5 |
+| `bedrock/us.anthropic.claude-haiku-5-5` | Claude Haiku 5.5 |
+
+<Note>
+Security testing with Claude needs the [Cyber Verification Program](https://support.claude.com/en/articles/14604842-cyber-verification-program) (CVP), also on Bedrock.
+</Note>

 ## Prerequisites

@@ -68,6 +68,21 @@ export LLM_API_KEY="your-bearer-token"          # sent as Authorization: Bearer
 export LLM_EXTRA_HEADERS='{"X-Feature-Key":"value","X-Tenant":"acme"}'
 ```

+### Extra request-body fields
+
+Some provider parameters go in the request body rather than a header, and
+Strix has no dedicated setting for them. Set them with `LLM_EXTRA_BODY` as a
+JSON object — they are merged into the body of every request to the model. For example, to
+use OpenAI's cheaper flex processing and attribute requests to a user with a
+safety identifier (a hash, never a raw email):
+
+```bash
+export LLM_EXTRA_BODY='{"service_tier":"flex","safety_identifier":"9f86d081884c7d65"}'
+```
+
+Fields aren't validated, so use only fields your provider accepts, and don't
+set fields Strix already sends, such as `model`, `messages` or `stream`.
+
 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.
@@ -8,7 +8,7 @@ description: "Configure Strix with Novita AI models"
 ## Setup

 ```bash
-export STRIX_LLM="openai/moonshotai/kimi-k2.5"
+export STRIX_LLM="openai/zai-org/glm-5.3"
 export LLM_API_KEY="your-novita-api-key"
 export LLM_API_BASE="https://api.novita.ai/openai"
 ```
@@ -18,11 +18,12 @@ export LLM_API_BASE="https://api.novita.ai/openai"
 | Model | Configuration |
 |-------|---------------|
 | GLM-5.3 | `openai/zai-org/glm-5.3` |
+| Kimi K3.1 Pro | `openai/moonshotai/kimi-k3.1-pro` |
 | Kimi K3 | `openai/moonshotai/kimi-k3` |
+| DeepSeek V4.1 Flash | `openai/deepseek/deepseek-v4.1-flash-1017` |
 | DeepSeek V4 Pro | `openai/deepseek/deepseek-v4-pro` |
-| Kimi K2.5 | `openai/moonshotai/kimi-k2.5` |
-| GLM-5 | `openai/zai-org/glm-5` |
-| MiniMax M2.5 | `openai/minimax/minimax-m2.5` |
+| MiniMax M3.1 Flash | `openai/minimax/minimax-m3.1-flash` |
+| GLM-5.3 Flash | `openai/zai-org/glm-5.3-flash` |

 ## Get API Key

@@ -6,13 +6,23 @@ description: "Configure Strix with OpenAI models"
 ## Setup

 ```bash
-export STRIX_LLM="openai/gpt-5.4"
+export STRIX_LLM="openai/gpt-6.1-sol"
 export LLM_API_KEY="sk-..."
 ```

 ## Available Models

-See [OpenAI Models Documentation](https://platform.openai.com/docs/models) for the full list of available models.
+| Model | Description |
+|-------|-------------|
+| `openai/gpt-6.1-sol` | Best balance of capability and cost for agentic security work |
+| `openai/gpt-6-astra` | Flagship, maximum capability |
+| `openai/gpt-6-luna` | Fast and cheap |
+
+See [OpenAI Models Documentation](https://platform.openai.com/docs/models) for the full list.
+
+<Note>
+Security testing with OpenAI models needs [Daybreak](https://openai.com/daybreak/) (Trusted Access for Cyber). Without it the models refuse most offensive tasks. [Request access](https://openai.com/form/enterprise-trusted-access-for-cyber/).
+</Note>

 ## Get API Key

@@ -25,7 +35,7 @@ See [OpenAI Models Documentation](https://platform.openai.com/docs/models) for t
 For OpenAI-compatible APIs:

 ```bash
-export STRIX_LLM="openai/gpt-5.4"
+export STRIX_LLM="openai/gpt-6.1-sol"
 export LLM_API_KEY="your-key"
 export LLM_API_BASE="https://your-proxy.com/v1"
 ```
@@ -18,13 +18,19 @@ Access any model on OpenRouter using the format `openrouter/<provider>/<model>`:

 | Model | Configuration |
 |-------|---------------|
-| GLM-5.3 (default) | `openrouter/z-ai/glm-5.3` |
-| GPT-5.4 | `openrouter/openai/gpt-5.4` |
-| Claude Sonnet 4.6 | `openrouter/anthropic/claude-sonnet-4.6` |
-| Gemini 3 Pro | `openrouter/google/gemini-3-pro-preview` |
+| GLM-5.3 (recommended) | `openrouter/z-ai/glm-5.3` |
+| GPT-6.1 Sol | `openrouter/openai/gpt-6.1-sol` |
+| GPT-6 Astra | `openrouter/openai/gpt-6-astra` |
+| Claude Opus 5.5 | `openrouter/anthropic/claude-opus-5.5` |
+| Claude Sonnet 5.5 | `openrouter/anthropic/claude-sonnet-5.5` |
+| Gemini 4 Argon | `openrouter/google/gemini-4-argon` |
+| Gemini 3.8 Flash | `openrouter/google/gemini-3.8-flash` |
+| DeepSeek V4.1 Flash | `openrouter/deepseek/deepseek-v4.1-flash-1017` |
 | DeepSeek V4 Pro | `openrouter/deepseek/deepseek-v4-pro` |
-| Kimi K3 | `openrouter/moonshotai/kimi-k3` |
-| GLM-4.7 | `openrouter/z-ai/glm-4.7` |
+| Kimi K3.1 Pro | `openrouter/moonshotai/kimi-k3.1-pro` |
+| GLM-5.3 Flash | `openrouter/z-ai/glm-5.3-flash` |
+
+OpenAI, Anthropic and Gemini models keep their provider safeguards on OpenRouter. Security testing with them needs [Daybreak](https://openai.com/daybreak/), [CVP](https://support.claude.com/en/articles/14604842-cyber-verification-program) or [Fairwind](https://deepmind.google/fairwind-program/) access.

 ## Get API Key

@@ -9,19 +9,33 @@ Strix uses [LiteLLM](https://docs.litellm.ai/docs/providers) for model compatibi

 Set your model and API key:

-| Model                | Provider          | Configuration                    |
-| -------------------- | ----------------- | -------------------------------- |
-| GLM-5.3 (default)    | Z.ai (OpenRouter) | `openrouter/z-ai/glm-5.3`        |
-| GPT-5.4              | OpenAI            | `openai/gpt-5.4`                 |
-| Claude Sonnet 4.6    | Anthropic         | `anthropic/claude-sonnet-4-6`    |
-| Gemini 3 Pro         | Google Vertex     | `vertex_ai/gemini-3-pro-preview` |
-| DeepSeek V4 Pro      | DeepSeek          | `deepseek/deepseek-v4-pro`       |
-| Kimi K3              | Moonshot          | `moonshot/kimi-k3`               |
+| Model                 | Provider                  | Configuration                    |
+| --------------------- | ------------------------- | -------------------------------- |
+| GLM-5.3 (recommended) | Vercel AI Gateway         | `vercel_ai_gateway/zai/glm-5.3`  |
+| GLM-5.3 (recommended) | OpenRouter                | `openrouter/z-ai/glm-5.3`        |
+| GPT-6.1 Sol           | OpenAI                    | `openai/gpt-6.1-sol`             |
+| Claude Opus 5.5       | Anthropic                 | `anthropic/claude-opus-5-5`      |
+| Claude Sonnet 5.5     | Anthropic                 | `anthropic/claude-sonnet-5-5`    |
+| Gemini 4 Argon        | Google Vertex             | `vertex_ai/gemini-4-argon`       |
+| DeepSeek V4.1 Flash   | DeepSeek                  | `deepseek/deepseek-flash`        |
+| DeepSeek V4 Pro       | DeepSeek                  | `deepseek/deepseek-v4-pro`       |
+| Kimi K3.1 Pro         | Moonshot                  | `moonshot/kimi-k3.1-pro`         |

-```bash
+<Note>
+Frontier OpenAI, Anthropic and Gemini models need a cyber access program for security testing: [Daybreak](https://openai.com/daybreak/) for OpenAI, the [Cyber Verification Program](https://support.claude.com/en/articles/14604842-cyber-verification-program) for Anthropic, and the [Fairwind Program](https://deepmind.google/fairwind-program/) for Gemini.
+</Note>
+
+<CodeGroup>
+```bash Vercel AI Gateway
+export STRIX_LLM="vercel_ai_gateway/zai/glm-5.3"
+export LLM_API_KEY="your-api-key"
+```
+
+```bash OpenRouter
 export STRIX_LLM="openrouter/z-ai/glm-5.3"
 export LLM_API_KEY="your-api-key"
 ```
+</CodeGroup>

 ## Local Models

@@ -47,7 +61,7 @@ See the [Local Models guide](/llm-providers/local) for setup instructions and re
    Access 100+ models through a single API.
  </Card>
  <Card title="Vercel AI Gateway" href="/llm-providers/vercel-ai-gateway">
-    Access models from multiple providers through one endpoint.
+    Access 100+ models through one OpenAI-compatible endpoint.
  </Card>
  <Card title="Google Vertex AI" href="/llm-providers/vertex">
    Gemini 3 models via Google Cloud.
@@ -69,9 +83,9 @@ Use LiteLLM's `provider/model-name` format:

 ```
 openrouter/z-ai/glm-5.3
-openai/gpt-5.4
-anthropic/claude-sonnet-4-6
-vertex_ai/gemini-3-pro-preview
-bedrock/anthropic.claude-4-5-sonnet-20251022-v1:0
+openai/gpt-6.1-sol
+anthropic/claude-sonnet-5-5
+vertex_ai/gemini-4-argon
+bedrock/us.anthropic.claude-sonnet-5-5
 ollama/llama4
 ```
@@ -1,32 +1,47 @@
 ---
 title: "Vercel AI Gateway"
-description: "Configure Strix with models via Vercel AI Gateway"
+description: "Configure Strix with models through Vercel AI Gateway"
 ---

-[Vercel AI Gateway](https://vercel.com/docs/ai-gateway) provides an OpenAI-compatible API for models from multiple providers.
+[Vercel AI Gateway](https://vercel.com/docs/ai-gateway) provides access to 100+ models from multiple providers through one OpenAI-compatible API.

 ## Setup

 Create an [AI Gateway API key](https://vercel.com/docs/ai-gateway/authentication-and-byok), then configure Strix:

 ```bash
-export STRIX_LLM="openai/anthropic/claude-opus-5"
+export STRIX_LLM="vercel_ai_gateway/zai/glm-5.3"
 export LLM_API_KEY="your-ai-gateway-api-key"
-export LLM_API_BASE="https://ai-gateway.vercel.sh/v1"
 ```

-The first `openai/` segment tells Strix to use its OpenAI-compatible client. The remaining value is the [AI Gateway model ID](https://vercel.com/docs/ai-gateway/models-and-providers).
+The `vercel_ai_gateway/` prefix selects the Gateway. The remaining value is the [AI Gateway model ID](https://vercel.com/docs/ai-gateway/models-and-providers).

 ## Available Models

-Use any language model returned by the AI Gateway models endpoint:
+Access any model on the Gateway using the format `vercel_ai_gateway/<provider>/<model>`:
+
+| Model | Configuration |
+|-------|---------------|
+| GLM-5.3 (recommended) | `vercel_ai_gateway/zai/glm-5.3` |
+| GPT-6.1 Sol | `vercel_ai_gateway/openai/gpt-6.1-sol` |
+| GPT-6 Astra | `vercel_ai_gateway/openai/gpt-6-astra` |
+| Claude Opus 5.5 | `vercel_ai_gateway/anthropic/claude-opus-5.5` |
+| Claude Sonnet 5.5 | `vercel_ai_gateway/anthropic/claude-sonnet-5.5` |
+| Gemini 4 Argon | `vercel_ai_gateway/google/gemini-4-argon` |
+| Gemini 3.8 Flash | `vercel_ai_gateway/google/gemini-3.8-flash` |
+| DeepSeek V4.1 Flash | `vercel_ai_gateway/deepseek/deepseek-v4.1-flash-1017` |
+| DeepSeek V4 Pro | `vercel_ai_gateway/deepseek/deepseek-v4-pro` |
+| Kimi K3.1 Pro | `vercel_ai_gateway/moonshotai/kimi-k3.1-pro` |
+| GLM-5.3 Flash | `vercel_ai_gateway/zai/glm-5.3-flash` |
+
+OpenAI, Anthropic and Gemini models keep their provider safeguards on the Gateway. Security testing with them needs [Daybreak](https://openai.com/daybreak/), [CVP](https://support.claude.com/en/articles/14604842-cyber-verification-program) or [Fairwind](https://deepmind.google/fairwind-program/) access.
+
+The full list is at the [AI Gateway models page](https://vercel.com/ai-gateway/models) or the models endpoint:

 ```text
 https://ai-gateway.vercel.sh/v1/models
 ```

-Prefix its model ID with `openai/` when setting `STRIX_LLM`. For example, the Gateway model ID `anthropic/claude-opus-5` becomes `openai/anthropic/claude-opus-5` in Strix.
-
 ## Get API Key

 1. Open the [AI Gateway API key settings](https://vercel.com/docs/ai-gateway/authentication-and-byok)
@@ -35,5 +50,6 @@ Prefix its model ID with `openai/` when setting `STRIX_LLM`. For example, the Ga

 ## Benefits

- Access models from multiple providers through one endpoint
- Track Gateway usage and cost in Vercel
+- **Single API** — Access models from OpenAI, Anthropic, Google, Z.ai, and more
+- **Bring your own keys** — Route through your own provider accounts when you want
+- **Cost tracking** — Monitor Gateway usage and cost in Vercel
@@ -6,7 +6,7 @@ description: "Configure Strix with Gemini models via Google Cloud"
 ## Setup

 ```bash
-export STRIX_LLM="vertex_ai/gemini-3-pro-preview"
+export STRIX_LLM="vertex_ai/gemini-4-argon"
 ```

 No API key required—uses Google Cloud Application Default Credentials.
@@ -29,8 +29,12 @@ export GOOGLE_APPLICATION_CREDENTIALS="/path/to/service-account.json"

 | Model | Description |
 |-------|-------------|
-| `vertex_ai/gemini-3-pro-preview` | Best overall performance for security testing |
-| `vertex_ai/gemini-3-flash-preview` | Faster and cheaper |
+| `vertex_ai/gemini-4-argon` | Best overall performance for security testing |
+| `vertex_ai/gemini-3.8-flash` | Faster and cheaper |
+
+<Note>
+Security testing with Gemini needs the [Fairwind Program](https://deepmind.google/fairwind-program/). Without it the cyber guardrails block most offensive tasks.
+</Note>

 ## Project Configuration

@@ -25,15 +25,22 @@ description: "Install Strix and run your first security scan"

 ## Configuration

-Set your LLM provider:
+Set your LLM provider. We recommend GLM-5.3 through [Vercel AI Gateway](/llm-providers/vercel-ai-gateway) or [OpenRouter](/llm-providers/openrouter):

-```bash
-export STRIX_LLM="openrouter/z-ai/glm-5.3"
+<CodeGroup>
+```bash Vercel AI Gateway
+export STRIX_LLM="vercel_ai_gateway/zai/glm-5.3"
 export LLM_API_KEY="your-api-key"
 ```

+```bash OpenRouter
+export STRIX_LLM="openrouter/z-ai/glm-5.3"
+export LLM_API_KEY="your-api-key"
+```
+</CodeGroup>
+
 <Tip>
-For best results, use `openrouter/z-ai/glm-5.3` (the default pick), `openai/gpt-5.4`, `anthropic/claude-opus-4-6`, or `openai/gpt-5.2`.
+For best results, use `zai/glm-5.3`, `deepseek/deepseek-v4.1-flash-1017`, `anthropic/claude-opus-5-5` (needs [CVP](https://support.claude.com/en/articles/14604842-cyber-verification-program) access), or `openai/gpt-6.1-sol` (needs [Daybreak](https://openai.com/daybreak/) access).
 </Tip>

 ## Run Your First Scan
@@ -49,7 +49,7 @@ Strix runs inside a Kali Linux-based Docker container with a comprehensive set o

 | Tool                                                    | Description                                   |
 | ------------------------------------------------------- | --------------------------------------------- |
-| [Semgrep](https://github.com/semgrep/semgrep)          | Fast SAST and custom rule matching            |
+| [Opengrep](https://github.com/opengrep/opengrep)        | Fast SAST and custom rule matching            |
 | [ast-grep](https://ast-grep.github.io)                 | Structural AST/CST-aware code search (`sg`)   |
 | [Tree-sitter](https://tree-sitter.github.io/tree-sitter/) | Syntax tree parsing and symbol extraction (Java/JS/TS/Python/Go/Bash/JSON/YAML grammars pre-configured) |
 | [Bandit](https://bandit.readthedocs.io)                | Python security linter                        |
@@ -33,7 +33,7 @@ sqlmap -u "https://example.com/page?id=1"

 ```bash
 # Fast SAST triage
-semgrep --config auto ./src
+opengrep scan --config /opt/opengrep-rules ./src

 # Structural AST search
 sg scan ./src
@@ -46,7 +46,7 @@ Thorough penetration testing. Best for:

 **Duration**: 1-4 hours depending on target complexity

-**White-box behavior**: Runs broad source-aware triage (`semgrep`, AST structural search, secrets, supply-chain checks) and then systematically validates top candidates dynamically.
+**White-box behavior**: Runs broad source-aware triage (`opengrep`, AST structural search, secrets, supply-chain checks) and then systematically validates top candidates dynamically.

 <Note>
 Deep mode is the default. It explores edge cases, chained vulnerabilities, and complex attack paths.
@@ -73,6 +73,12 @@ dev = [
 [tool.pytest.ini_options]
 asyncio_mode = "auto"

+[tool.uv]
+# Supply-chain cooldown: resolve only distributions that have been on the index for
+# at least 5 days, so a poisoned release (litellm 1.107.0, 2026-10-21) is never
+# picked up by `uv lock` / `uv add` before it can be caught and yanked.
+exclude-newer = "5 days"
+
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"
@@ -4,7 +4,7 @@ set -euo pipefail

 APP=strix
 REPO="usestrix/strix"
-STRIX_IMAGE="ghcr.io/usestrix/strix-sandbox:1.3.0"
+STRIX_IMAGE="ghcr.io/usestrix/strix-sandbox:1.4.0"

 MUTED='\033[0;2m'
 RED='\033[0;31m'
@@ -63,7 +63,7 @@ jobs:
          fi
 ```

-Then tell the user to add two repository secrets: `STRIX_LLM` (model id, for example `openai/gpt-5.4`) and `LLM_API_KEY` (the provider key). Do not create these values yourself.
+Then tell the user to add two repository secrets: `STRIX_LLM` (model id, for example `openrouter/z-ai/glm-5.3`) and `LLM_API_KEY` (the provider key). Do not create these values yourself.

 Notes:
 - In CI/headless runs Strix automatically scopes to the PR's changed files (`--scope-mode auto`). If diff resolution fails, keep `fetch-depth: 0` or set `--diff-base` to the PR's actual base branch — use `origin/${{ github.base_ref }}` in GitHub Actions rather than a hard-coded `origin/main`, since repos use different default branches.
@@ -47,7 +47,7 @@ If unsure and the user has (or will create) an app.strix.ai account, prefer **Cl
   ```
 3. **LLM configured** — two environment variables:
   ```bash
-   export STRIX_LLM="openai/gpt-5.4"      # any LiteLLM model id (openai/..., anthropic/..., openrouter/...)
+   export STRIX_LLM="openrouter/z-ai/glm-5.3"      # any LiteLLM model id (openai/..., anthropic/..., openrouter/...)
   export LLM_API_KEY="<provider api key>"
   ```
   Ask the user for these if unset. Never hardcode or commit keys.
@@ -13,11 +13,13 @@ from agents.agent import ToolsToFinalOutputResult
 from agents.sandbox import SandboxAgent
 from agents.sandbox.capabilities import Filesystem, Shell
 from agents.sandbox.errors import InvalidManifestPathError
+from agents.sandbox.session.pty_types import PTY_YIELD_TIME_MS_MAX
 from agents.tool import CustomTool, FunctionTool, Tool
 from pydantic import ValidationError

 from strix.agents.prompt import render_system_prompt
 from strix.config import load_settings
+from strix.config.tool_call_arguments import describe_malformed_arguments
 from strix.tools.agents_graph.tools import (
    agent_finish,
    create_agent,
@@ -267,6 +269,10 @@ def _with_coerced_arguments(tool: FunctionTool) -> FunctionTool:
    nullish = tool.name.startswith(_QUERY_TOOL_PREFIXES)

    async def invoke(ctx: Any, raw_input: str) -> Any:
+        malformed = describe_malformed_arguments(tool.name, raw_input)
+        if malformed is not None:
+            logger.debug("Tool %s got malformed arguments; asking the model to re-issue", tool.name)
+            return malformed
        return await invoke_tool(ctx, _coerce_arguments(raw_input, schema, nullish=nullish))

    tool.on_invoke_tool = invoke
@@ -375,11 +381,15 @@ def _configure_filesystem_tools(
            )


-def _make_filesystem_configurator(*, chat_completions: bool, strict_schemas: bool) -> Any:
+def _make_filesystem_configurator(
+    *, chat_completions: bool, strict_schemas: bool, supports_images: bool = True
+) -> Any:
    def configure(toolset: Any) -> None:
        _configure_filesystem_tools(
            toolset, chat_completions=chat_completions, strict_schemas=strict_schemas
        )
+        if not supports_images:
+            toolset.view_image.is_enabled = False

    return configure

@@ -432,8 +442,47 @@ def _apply_shell_output_cap(parsed: dict[str, Any]) -> None:
    )


+_EXEC_YIELD_CAP_S = PTY_YIELD_TIME_MS_MAX // 1000
+
+_EXEC_COMMAND_DESCRIPTION_NOTE = (
+    f"The wait for output is capped at {_EXEC_YIELD_CAP_S} s whatever yield_time_ms says. "
+    "A non-TTY command still running at the yield cannot be read later (write_stdin fails "
+    "on it and its output is lost): run anything longer in the background with its output "
+    "redirected to a file (`cmd > /tmp/out.log 2>&1 &`, then read the log), or start it with "
+    'tty=true and poll with write_stdin(session_id=<id>, chars="").'
+)
+
+_RUNNING_NON_TTY_NOTICE = (
+    "Note: this non-TTY process cannot be polled and its later output is lost. Rerun it in "
+    "the background (`cmd > /tmp/out.log 2>&1 &`, then read the log) or with tty=true."
+)
+
+_PROCESS_RUNNING_RE = re.compile(r"^Process running with session ID \d+$", re.MULTILINE)
+_OUTPUT_HEADER_RE = re.compile(r"^Output:\n", re.MULTILINE)
+
+
+def _annotate_running_non_tty(parsed: dict[str, Any], result: Any) -> Any:
+    if parsed.get("tty") or not isinstance(result, str):
+        return result
+    header = _OUTPUT_HEADER_RE.search(result)
+    head_end = len(result) if header is None else header.start()
+    if _PROCESS_RUNNING_RE.search(result, 0, head_end) is None:
+        return result
+    notice = _RUNNING_NON_TTY_NOTICE
+    requested = parsed.get("yield_time_ms")
+    if isinstance(requested, int | float) and requested > PTY_YIELD_TIME_MS_MAX:
+        notice = (
+            f"Note: yield_time_ms={int(requested)} was capped at {PTY_YIELD_TIME_MS_MAX} "
+            f"({_EXEC_YIELD_CAP_S} s). " + notice.removeprefix("Note: ")
+        )
+    if header is None:
+        return f"{result}\n{notice}"
+    return f"{result[: header.end()]}{notice}\n{result[header.end() :]}"
+
+
 def _wrap_exec_command(tool: FunctionTool) -> FunctionTool:
    invoke_tool = tool.on_invoke_tool
+    tool.description = f"{tool.description.rstrip()} {_EXEC_COMMAND_DESCRIPTION_NOTE}"

    async def invoke(ctx: Any, raw_input: str) -> Any:
        try:
@@ -446,7 +495,7 @@ def _wrap_exec_command(tool: FunctionTool) -> FunctionTool:
            _apply_shell_output_cap(parsed)
            raw_input = json.dumps(parsed)
        try:
-            return await invoke_tool(ctx, raw_input)
+            result = await invoke_tool(ctx, raw_input)
        except ValidationError as exc:
            return _format_validation_error(tool.name, exc)
        except InvalidManifestPathError as exc:
@@ -456,6 +505,9 @@ def _wrap_exec_command(tool: FunctionTool) -> FunctionTool:
                "(or omitted to use the turn's cwd). "
                f"Got: {rel!r}."
            )
+        if isinstance(parsed, dict):
+            return _annotate_running_non_tty(parsed, result)
+        return result

    tool.on_invoke_tool = invoke
    return tool
@@ -567,6 +619,17 @@ def _finish_tool_use_behavior(
    return ToolsToFinalOutputResult(is_final_output=False, final_output=None)


+_COVERAGE_TOOLS: tuple[Tool, ...] = (record_coverage, update_coverage, list_coverage)
+_FINISH_SCAN_COVERAGE_STEP = re.compile(
+    r"\n5\. \*\*Coverage reconciliation\.\*\*.*?(?=\n\n)", re.DOTALL
+)
+
+
+def _without_coverage_step(tool: FunctionTool) -> FunctionTool:
+    description = _FINISH_SCAN_COVERAGE_STEP.sub("", tool.description)
+    return dataclasses.replace(tool, description=description)
+
+
 _BASE_TOOLS: tuple[Tool, ...] = (
    think,
    load_skill,
@@ -665,11 +728,16 @@ def build_strix_agent(
    is_whitebox: bool = False,
    is_diff_scoped: bool = False,
    interactive: bool = False,
+    parallel_tool_calls: bool = False,
    chat_completions_tools: bool = False,
    strict_tool_schemas: bool = True,
    system_prompt_context: dict[str, Any] | None = None,
    extra_tools: Sequence[Tool] | None = None,
    instructions_override: str | None = None,
+    supports_images: bool = True,
+    finish_tool: Tool = finish_scan,
+    report_tool: Tool = create_vulnerability_report,
+    coverage_tools: bool = True,
 ) -> SandboxAgent[Any]:
    """Build a SandboxAgent for either root or child use.

@@ -682,6 +750,11 @@ def build_strix_agent(
            registered via ``register_agent_tools``.
        instructions_override: Use this verbatim as the system prompt instead
            of rendering the built-in scan prompt.
+        finish_tool: The tool that ends the run, given to the root agent.
+        report_tool: The tool every agent files findings with, in place of
+            ``create_vulnerability_report``.
+        coverage_tools: Give the agent the coverage ledger tools and the
+            prompt rules for them.
    """
    if instructions_override is not None:
        instructions = instructions_override
@@ -693,17 +766,26 @@ def build_strix_agent(
            is_root=is_root,
            is_diff_scoped=is_diff_scoped,
            interactive=interactive,
+            parallel_tool_calls=parallel_tool_calls,
            system_prompt_context=system_prompt_context,
+            supports_images=supports_images,
+            report_tool=report_tool.name,
+            coverage_tools=coverage_tools,
        )

    agent_tools = [*_EXTRA_TOOLS, *(extra_tools or [])]
    if interactive:
        # Yielding to the user is only meaningful when one is attached.
        agent_tools.append(wait_for_user)
+    base_tools = [report_tool if t is create_vulnerability_report else t for t in _BASE_TOOLS]
+    if not coverage_tools:
+        base_tools = [t for t in base_tools if t not in _COVERAGE_TOOLS]
+        if finish_tool is finish_scan:
+            finish_tool = _without_coverage_step(finish_scan)
    if is_root:
-        tools: list[Tool] = [*_BASE_TOOLS, *agent_tools, finish_scan]
+        tools: list[Tool] = [*base_tools, *agent_tools, finish_tool]
    else:
-        tools = [*_BASE_TOOLS, *agent_tools, agent_finish]
+        tools = [*base_tools, *agent_tools, agent_finish]
    _ensure_unique_tool_names(tools)
    tools = [
        _with_bounded_result(_with_strictness(_with_coerced_arguments(tool), strict_tool_schemas))
@@ -733,6 +815,7 @@ def build_strix_agent(
                configure_tools=_make_filesystem_configurator(
                    chat_completions=chat_completions_tools,
                    strict_schemas=strict_tool_schemas,
+                    supports_images=supports_images,
                ),
            ),
            Shell(
@@ -751,9 +834,13 @@ def make_child_factory(
    is_whitebox: bool = False,
    is_diff_scoped: bool = False,
    interactive: bool = False,
+    parallel_tool_calls: bool = False,
    chat_completions_tools: bool = False,
    strict_tool_schemas: bool = True,
    system_prompt_context: dict[str, Any] | None = None,
+    supports_images: bool = True,
+    report_tool: Tool = create_vulnerability_report,
+    coverage_tools: bool = True,
 ) -> Any:
    """Return the runner-owned builder used by ``spawn_child_agent``.

@@ -771,9 +858,13 @@ def make_child_factory(
            is_whitebox=is_whitebox,
            is_diff_scoped=is_diff_scoped,
            interactive=interactive,
+            parallel_tool_calls=parallel_tool_calls,
            chat_completions_tools=chat_completions_tools,
            strict_tool_schemas=strict_tool_schemas,
            system_prompt_context=system_prompt_context,
+            supports_images=supports_images,
+            report_tool=report_tool,
+            coverage_tools=coverage_tools,
        )

    return _factory
@@ -83,8 +83,12 @@ def render_system_prompt(
    is_root: bool = False,
    is_diff_scoped: bool = False,
    interactive: bool = False,
+    parallel_tool_calls: bool = False,
    system_prompt_context: dict[str, Any] | None = None,
    include_scope: bool = True,
+    supports_images: bool = True,
+    report_tool: str = "create_vulnerability_report",
+    coverage_tools: bool = True,
 ) -> str:
    """Render the system prompt. Returns empty string on template failure.

@@ -119,7 +123,9 @@ def render_system_prompt(
            is_root=is_root,
            is_diff_scoped=is_diff_scoped,
        )
-        skill_content = load_skills(skills_to_load)
+        skill_content = load_skills(
+            skills_to_load, supports_images=supports_images, coverage_tools=coverage_tools
+        )
        env.globals["get_skill"] = lambda name: skill_content.get(name, "")

        # Skills every agent of this kind loads come first, so siblings share them
@@ -129,9 +135,12 @@ def render_system_prompt(
            requested_skill_names=[name for name in skill_content if name not in shared],
            available_skills=get_available_skills(),
            interactive=interactive,
+            parallel_tool_calls=parallel_tool_calls,
            is_root=is_root,
            system_prompt_context=system_prompt_context or {},
            include_scope=include_scope,
+            report_tool=report_tool,
+            coverage_tools=coverage_tools,
            **skill_content,
        )
    except Exception:
@@ -21,7 +21,7 @@ MCP CONNECTIONS (available this run):
 {% if system_prompt_context.mcp_connections %}
 - Connected this run (search one to find relevant tools):
 {% for connection in system_prompt_context.mcp_connections %}
-  - {{ connection.name }} ({{ connection.tool_count }} tools){% if connection.purpose %}: {{ connection.purpose }}{% endif %}
+  - {{ connection.name }}{% if connection.state == "unavailable" %} (unavailable right now){% elif connection.tool_count is not none %} ({{ connection.tool_count }} tools){% endif %}{% if connection.purpose %}: {{ connection.purpose }}{% endif %}
 {% endfor %}
 {% endif %}
 - Reach for a connection whenever the target itself cannot give you information a connection could: its database schema and access policies, real deployment or infrastructure configuration, known issues or prior findings, or server logs. In those cases call list_mcps early to see what is available, and prefer a connection's authoritative data over inferring from the target's responses. Do not wait to be told a connection exists.
@@ -44,7 +44,11 @@ INTERACTIVE BEHAVIOR:
 - If you end a turn on plain text and the nudge arrives, your words already reached the user. Do not repeat them: call wait_for_user.
 - You can narrate while you work — plain text is shown to the user as you go. Narrating is free; wait_for_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.
+{%- if parallel_tool_calls %}
+- While actively working on a task, every turn should carry a tool call (independent calls can go in the same turn) — use think to plan, the appropriate tool to act, and wait_for_user only when you genuinely need the user.
+{%- else %}
 - While actively working on a task, every turn should carry exactly one tool call — use think to plan, the appropriate tool to act, and wait_for_user only when you genuinely need the user.
+{%- endif %}
 - Never loop through think or other tools just to prepare, polish, confirm, or announce an answer. Once you know the answer, write it and call wait_for_user.
 {% else %}
 AUTONOMOUS BEHAVIOR:
@@ -55,6 +59,9 @@ AUTONOMOUS BEHAVIOR:
 - 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.
+{%- if parallel_tool_calls %}
+- Read-only calls can go together in one turn: when several file reads or searches don't depend on each other's output, issue them as separate tool calls in the same turn rather than one per turn. Anything else — writing files, running commands that change state, sending requests, creating or messaging agents, reporting, or finishing — goes alone in its own turn.
+{%- endif %}
 {% endif %}
 </communication_rules>

@@ -108,18 +115,18 @@ BLACK-BOX TESTING (domain/subdomain only):

 WHITE-BOX TESTING (code provided):
 - MUST perform BOTH static AND dynamic analysis
- Static: Use source-aware triage first to map risk quickly (`semgrep`, `ast-grep`, Tree-sitter tooling, `gitleaks`, `trufflehog`, `trivy fs`). Then review code for vulnerabilities
+- Static: Use source-aware triage first to map risk quickly (`opengrep`, `ast-grep`, Tree-sitter tooling, `gitleaks`, `trufflehog`, `trivy fs`). Then review code for vulnerabilities
 - Static coverage floor: execute at least one structural AST mapping pass (`sg` and/or Tree-sitter) per repository and keep artifact output
- Static coverage target per repository: run one `semgrep` pass, one secrets pass (`gitleaks` and/or `trufflehog`), one `trivy fs` pass, and one AST-structural pass (`sg` and/or Tree-sitter)
+- Static coverage target per repository: run one `opengrep` pass, one secrets pass (`gitleaks` and/or `trufflehog`), one `trivy fs` pass, and one AST-structural pass (`sg` and/or Tree-sitter)
 - Keep AST artifacts bounded and high-signal: scope to relevant paths/hypotheses, avoid whole-repo generic function dumps
- AST target selection rule: build `sg-targets.txt` from `semgrep.json` scope first (`paths.scanned`, fallback to unique `results[].path`), then run `xargs ... sg run` against that file list. Only use path-heuristic fallback if semgrep scope is unavailable.
+- AST target selection rule: build `sg-targets.txt` from `opengrep.json` scope first (`paths.scanned`, fallback to unique `results[].path`), then run `xargs ... sg run` against that file list. Only use path-heuristic fallback if opengrep scope is unavailable.
 - Dynamic: Run the application and test live to validate exploitability
 - NEVER rely solely on static code analysis when dynamic validation is possible
 - Begin with fast source triage and dynamic run preparation in parallel; use static findings to prioritize live testing.
 - Local execution, unit/integration testing, patch verification, and HTTP requests against locally started in-scope services are normal authorized white-box validation
 - If dynamically running the code proves impossible after exhaustive attempts, pivot to comprehensive static analysis.
 - Try to infer how to run the code based on its structure and content.
- Derive the code fix as PART OF reporting, not as a separate later pass: create_vulnerability_report already requires the concrete patch inline (`code_locations` with verbatim `fix_before`/`fix_after` and `fix_pr_body`), so the reporting agent that analyzes the root cause is the one that produces the fix. Do NOT spawn a downstream agent afterwards to re-derive/re-apply the same patch.
+- Derive the code fix as PART OF reporting, not as a separate later pass: {{ report_tool }} already requires the concrete patch inline (`code_locations` with verbatim `fix_before`/`fix_after` and `fix_pr_body`), so the reporting agent that analyzes the root cause is the one that produces the fix. Do NOT spawn a downstream agent afterwards to re-derive/re-apply the same patch.
 - If you also apply and verify the patch in the repo (edit the file, re-test that the vulnerability is gone), do it in the same agent/turn while the analysis is fresh — right before or as part of filing the report — never as a second re-analysis pass.

 COMBINED MODE (code + deployed target present):
@@ -181,7 +188,7 @@ EFFICIENCY TACTICS:
 - For Caido proxy automation inside Python, explicitly import from
  `caido_api`:
  `from caido_api import list_requests, view_request, repeat_request, list_sitemap, view_sitemap_entry, scope_rules`
- Prefer established fuzzers/scanners where applicable: ffuf, sqlmap, nuclei, wapiti, arjun, httpx, katana, semgrep, bandit, trufflehog, nmap. Use scripts mainly to coordinate or validate around them, not to replace them without reason
+- Prefer established fuzzers/scanners where applicable: ffuf, sqlmap, nuclei, wapiti, arjun, httpx, katana, opengrep, bandit, trufflehog, nmap. Use scripts mainly to coordinate or validate around them, not to replace them without reason
 - For trial-heavy vectors (SQLi, XSS, XXE, SSRF, RCE, auth/JWT, deserialization), DO NOT iterate payloads manually in the browser. Always spray payloads via Python scripts through `exec_command` or terminal tools.
 - When using established fuzzers/scanners, use the proxy for inspection where helpful
 - Generate/adapt large payload corpora: combine encodings (URL, unicode, base64), comment styles, wrappers, time-based/differential probes. Expand with wordlists/templates
@@ -202,12 +209,14 @@ VALIDATION REQUIREMENTS:
 - Keep going until you find something that matters
 - CLOSURE DISCIPLINE: every candidate you open ends in exactly one explicit state — `confirmed` (working PoC, or a complete source→control→sink→impact trace that is reachable), `ruled_out` (you can name the SPECIFIC control, at a location, that runs on every attacker-reachable path before the sink), or `open_proof_gap` (plausible, unconfirmed, and you could NOT name such a control). "I moved on" is not a closure state. Silently dropping an uncertain candidate is mislabelling an `open_proof_gap` as `ruled_out` and is how real bugs get missed.
 - Missing information is NOT proof of safety: no caller found, can't tell if deployed/exposed, couldn't stand up the service, build failed — each is an `open_proof_gap`, never a reason to mark a candidate clean. Difficulty is a reason to defer, not to suppress.
- COVERAGE: record every surface you assess with `record_coverage` (surface + risk area + outcome + evidence), including the ones that came back clean — a report that only lists findings cannot say what was reviewed and cleared. Use the `needs_follow_up` outcome for anything left in an `open_proof_gap` state, and carry the same items up in `agent_finish(open_items=[...])`. The ledger is shared and mutable: when you resolve a surface another agent left open — or find that a closed one is not — move that entry with `update_coverage` instead of recording a second one for the same surface. The root agent reconciles all of it via `list_coverage` before `finish_scan`.
+{% if coverage_tools -%}
+- COVERAGE: record every surface you assess with `record_coverage` (surface + risk area + outcome + evidence), including the ones that came back clean — a report that only lists findings cannot say what was reviewed and cleared. Use the `needs_follow_up` outcome for anything left in an `open_proof_gap` state, and carry the same items up in `agent_finish(open_items=[...])`. The ledger is shared and mutable: when you resolve a surface another agent left open — or find that a closed one is not — move that entry with `update_coverage` instead of recording a second one for the same surface. Before `finish_scan`, the root agent calls `list_coverage(outcome="needs_follow_up")` and follows up every row it returns: it assigns the outstanding work to a subagent that moves the existing entry with `update_coverage` to its real outcome, and a row stays `needs_follow_up` only when it is genuinely blocked, with the blocker stated in its evidence.
+{% endif -%}
 - THREAT MODEL: before you start testing, call `get_threat_model` on the target you were pointed at — it is the scan's shared answer to who the attacker is, where the trust boundaries sit, and what counts as critical here. It is scoped to this scan and nothing carries over from an earlier run, so `found: false` means no agent on this run has derived one yet. Read it instead of re-deriving trust boundaries yourself; where your testing disproves it — a boundary it calls trusted turns out to be attacker-reachable, a role it did not know about, a host or endpoint it never listed — record that with `amend_threat_model` so the agents after you inherit the correction. Amending is not optional politeness: a model nobody corrects turns the first agent's guesses into everyone's assumptions.
 - Before filing any report, run the counterevidence pass: argue the strongest case AGAINST the finding, record what you found in the `counterevidence` field, set `confidence` honestly (a static-only trace you couldn't execute is at best `medium`), and state what evidence would change the severity. See the counterevidence and severity-calibration knowledge above.
- 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. If your evidence proves more than the finding it matched (a working exploit where that one had only a static trace, a chain that raises the impact), revise that finding with update_vulnerability_report using the duplicate_of id — never re-file it.
+- A vulnerability is ONLY considered reported when a reporting agent uses {{ report_tool }} (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 {{ report_tool }} (`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 {{ report_tool }} 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. If your evidence proves more than the finding it matched (a working exploit where that one had only a static trace, a chain that raises the impact), revise that finding with update_vulnerability_report using the duplicate_of id — never re-file it.
 - HTTP EVIDENCE: a finding you validated through the proxy is not fully filed until `http_exchange_ids` carries the proxy request ids of the exchanges that prove it — the request that triggers the vulnerability plus the baseline/control request it differs from (an unauthenticated success next to the authenticated one, the payload response next to the benign one). Copy the ids exactly as `list_requests`/`view_request` show them, never invent or guess one, and never omit the field to bypass validation. Leave it out only when there is no captured HTTP exchange at all (static-only code findings, dependency CVEs). If you filed before the proving exchanges existed, attach them afterwards with update_vulnerability_report. Without the ids, the finding ships as prose nobody can replay.
 - REVISING A FINDING: use update_vulnerability_report (report id + the fields you want to replace + update_reason) when you learn something a finding already on file does not carry — you built the PoC after filing it, a chain raised its impact, further testing weakened it, or its counterevidence/remediation/code locations were wrong. Editing a finding needs no duplicate verdict, and it is always better than filing a second report for the same issue. Read the finding first with get_report, and pass only the fields that change.
 - WITHDRAWING A FINDING: when re-testing proves a finding you filed is not a vulnerability at all — the exploit only worked because of your own test setup (a mixed-up session, a self-inflicted state change, a misread response), the behaviour is intended, or the control is in fact enforced — delete it with delete_vulnerability_report (report id + delete_reason). Never leave a disproved finding on file as a zero-CVSS, "retracted", or "false positive" report, and never rewrite its text into a note saying it was wrong: a report that exists is read as a finding. Delete is for a finding that does not exist; a finding that is real but weaker than filed is revised, not deleted.
@@ -220,14 +229,16 @@ Every one of these tools writes to state the rest of the scan reads. Reaching fo
 - TODOS — `create_todo` / `list_todos` / `update_todo` / `mark_todo_done` / `mark_todo_pending` / `delete_todo`: your own working checklist for a multi-step task. Create todos when your task has several distinct steps so nothing is dropped across a long run; mark them done as you finish. This is private working memory — use `notes` for anything another agent needs.
 - NOTES — `create_note` / `list_notes` / `get_note` / `update_note` / `delete_note`: the scan's shared scratchpad, visible to every agent. Write a note for a durable cross-agent fact that is not a finding and not coverage — a working credential set, a discovered endpoint inventory, an enumerated tenant list, a rate-limit quirk the next agent needs. `update_note` to keep a living inventory current; `delete_note` only for something now wrong or superseded. Check `list_notes`/`get_note` before recon work so you build on what is already mapped instead of redoing it.
 - THREAT MODEL — `get_threat_model` / `amend_threat_model` / `save_threat_model`: covered above. `save_threat_model` REPLACES the whole document and clears amendments, so it is for establishing the baseline or folding amendments in (normally root) — to correct part of an existing model, `amend_threat_model` instead.
+{% if coverage_tools -%}
 - COVERAGE — `record_coverage` / `update_coverage` / `list_coverage`: covered above. One row per surface+risk; correct an existing row with `update_coverage`, never a second `record_coverage`.
+{% endif -%}
 - RESEARCH — `web_search`: pull fresh, target-specific external knowledge — latest bypasses, WAF evasions, DB-/framework-specific syntax, CVE and advisory detail — before falling back to memorized payloads, and refresh payload corpora mid-spray.
 - SPAWN WORK — `create_agent`: delegate a focused subtask to a specialist child (see the multi-agent rules below for when to spawn and how to scope it). Give it the target to model against and what is already known.
 - TRACK CHILDREN — `view_agent_graph`: your live map of every agent and its status. Call it before spawning (to confirm no existing agent already covers the scope) and before finishing (to confirm no child is still running).
 - STEER CHILDREN — `send_message_to_agent`: send a running child new information, a course correction, or a request to wrap up, without killing it. Use it to answer a child's question or narrow its scope mid-run.
 - BLOCK ON CHILDREN — `wait_for_agents`: block until named children report back when your next move genuinely depends on their results. If you can keep making progress in parallel, keep working instead of waiting.
 - CANCEL CHILDREN — `stop_agent`: gracefully cancel a child whose work is redundant, misdirected, or no longer needed. Prefer `send_message_to_agent` to redirect a child that is merely off-track; reserve `stop_agent` for work that should not continue at all.
- FINISH — subagents call `agent_finish` (with `open_items=[...]` for anything left unresolved); the root agent calls `finish_scan` exactly once, only after every child is wrapped up and coverage is reconciled. `agent_finish`/`finish_scan` are handoffs, not reporting channels — a vulnerability is reported only via `create_vulnerability_report`/`create_dependency_report`.
+- FINISH — subagents call `agent_finish` (with `open_items=[...]` for anything left unresolved); the root agent calls `finish_scan` exactly once, only after every child is wrapped up and coverage is reconciled. `agent_finish`/`finish_scan` are handoffs, not reporting channels — a vulnerability is reported only via `{{ report_tool }}`/`create_dependency_report`.
 </execution_guidelines>

 <vulnerability_focus>
@@ -380,7 +391,7 @@ CRITICAL RULES:
 - **REALISTIC OUTCOMES** - Some tests find nothing, some validations fail
 - **ONE AGENT = ONE TASK** - Don't let agents do multiple unrelated jobs
 - **SPAWN REACTIVELY** - Create new agents based on what you discover
- **ONLY REPORTING AGENTS** can use create_vulnerability_report tool
+- **ONLY REPORTING AGENTS** can use {{ report_tool }} tool
 - **AGENT SPECIALIZATION MANDATORY** - Each agent must be highly specialized; prefer 1–3 skills, up to 5 for complex contexts
 - **NO GENERIC AGENTS** - Avoid creating broad, multi-purpose agents that dilute focus

@@ -447,7 +458,7 @@ JAVASCRIPT ANALYSIS:
 - js-beautify - JS beautifier/deobfuscator

 CODE ANALYSIS:
- semgrep - Static analysis/SAST
+- opengrep - Static analysis/SAST (bundled rules in /opt/opengrep-rules)
 - ast-grep (sg) - Structural AST/CST-aware code search
 - tree-sitter - Syntax-aware parsing and symbol extraction support
 - bandit - Python security linter
@@ -6,7 +6,8 @@ import json
 import logging
 import os
 from pathlib import Path
-from typing import TYPE_CHECKING, Any
+from types import UnionType
+from typing import TYPE_CHECKING, Any, Union, get_args, get_origin

 from pydantic import AliasChoices, BaseModel

@@ -126,13 +127,36 @@ def _read_json_overrides(path: Path) -> dict[str, dict[str, Any]]:
                continue  # env wins under some alias; skip the JSON file for this field
            for alias in aliases:
                if alias in env_block_upper:
-                    sub_data[fname] = env_block_upper[alias]
+                    sub_data[fname] = _decode_saved_value(finfo, env_block_upper[alias])
                    break
        if sub_data:
            nested[sub_name] = sub_data
    return nested


+def _decode_saved_value(finfo: FieldInfo, value: Any) -> Any:
+    """Decode a JSON string saved for a dict field, as pydantic does for env vars.
+
+    ``persist_current`` saves env values verbatim, so a dict setting such as
+    ``LLM_EXTRA_HEADERS`` lands in the file as JSON text. Values that are
+    already objects, or aren't valid JSON, are left for validation to judge.
+    """
+    if not isinstance(value, str) or not _is_dict_field(finfo):
+        return value
+    try:
+        return json.loads(value)
+    except ValueError:
+        return value
+
+
+def _is_dict_field(finfo: FieldInfo) -> bool:
+    annotation = finfo.annotation
+    candidates = (
+        get_args(annotation) if get_origin(annotation) in (Union, UnionType) else (annotation,)
+    )
+    return any(c is dict or get_origin(c) is dict for c in candidates)
+
+
 def _first_alias_value(aliases: list[str], source: Mapping[str, Any]) -> Any | None:
    return next((source[alias] for alias in aliases if alias in source), None)

@@ -5,6 +5,7 @@ from __future__ import annotations
 import asyncio
 import contextlib
 import inspect
+import json
 import logging
 import os
 import time
@@ -40,6 +41,7 @@ from openai.types.shared import Reasoning
 from strix.agents.prompt import CACHE_POINT
 from strix.config import codex
 from strix.config.loader import load_settings
+from strix.config.tool_call_arguments import repair_input
 from strix.config.tool_call_ids import TurnCallIdRewriter, dedupe_input
 from strix.config.tool_call_limits import TurnToolCallLimiter
 from strix.llm import request_log
@@ -254,6 +256,12 @@ class _TurnGuardModel(Model):
    Ids that collide with the history are rewritten before the turn is
    recorded, and already-corrupted histories are repaired on the way out.

+    Tool-call arguments: a turn whose ``arguments`` are not valid JSON fails
+    that one call with a parse error, but the raw string is recorded and
+    strict providers then reject every request that replays it. Such
+    arguments are rewritten to a valid JSON object on the way out (see
+    :mod:`strix.config.tool_call_arguments`).
+
    Tool-call volume: a degenerate response can queue hundreds of calls that
    the run loop then honours one by one. Only the first
    ``LLM_MAX_TOOL_CALLS_PER_TURN`` calls of a response are kept.
@@ -262,6 +270,8 @@ class _TurnGuardModel(Model):
    not covered by the request timeout, which resets on any byte (keepalives
    included). ``LLM_STREAM_IDLE_TIMEOUT`` bounds the gap between events so the
    turn fails instead of hanging, and the existing retry path replays it.
+    ``LLM_STREAM_FIRST_EVENT_TIMEOUT`` and ``LLM_STREAM_TOTAL_TIMEOUT`` bound
+    the first event and the whole stream.
    """

    def __init__(
@@ -269,11 +279,15 @@ class _TurnGuardModel(Model):
        inner: Model,
        *,
        max_tool_calls_per_turn: int = 0,
-        stream_idle_timeout: float = 0.0,
+        stream_idle_timeout: float | None = None,
+        first_event_timeout: float | None = None,
+        total_timeout: float | None = None,
    ) -> None:
        self._inner = inner
        self._max_tool_calls_per_turn = max_tool_calls_per_turn
        self._stream_idle_timeout = stream_idle_timeout
+        self._first_event_timeout = first_event_timeout
+        self._total_timeout = total_timeout

    def _limiter(self) -> TurnToolCallLimiter:
        return TurnToolCallLimiter(self._max_tool_calls_per_turn)
@@ -306,7 +320,7 @@ class _TurnGuardModel(Model):
        conversation_id: str | None,
        prompt: ResponsePromptParam | None,
    ) -> ModelResponse:
-        sanitized = dedupe_input(input)
+        sanitized = _sanitize_input(input)
        rewriter = TurnCallIdRewriter(sanitized)
        system_instructions, sanitized = _split_cached_prefix(
            system_instructions, sanitized, model_settings
@@ -342,7 +356,7 @@ class _TurnGuardModel(Model):
        conversation_id: str | None,
        prompt: ResponsePromptParam | None,
    ) -> AsyncIterator[TResponseStreamEvent]:
-        sanitized = dedupe_input(input)
+        sanitized = _sanitize_input(input)
        rewriter = TurnCallIdRewriter(sanitized)
        system_instructions, sanitized = _split_cached_prefix(
            system_instructions, sanitized, model_settings
@@ -360,13 +374,22 @@ class _TurnGuardModel(Model):
            conversation_id=conversation_id,
            prompt=prompt,
        )
-        async for event in _with_idle_timeout(stream, self._stream_idle_timeout):
+        async for event in _with_timeouts(
+            stream,
+            idle=self._stream_idle_timeout,
+            first_event=self._first_event_timeout,
+            total=self._total_timeout,
+        ):
            guarded = _guard_event(event, rewriter, limiter)
            if guarded is not None:
                yield guarded
        self._log_dropped(limiter)


+def _sanitize_input(model_input: str | list[TResponseInputItem]) -> str | list[Any]:
+    return repair_input(dedupe_input(model_input))
+
+
 def _split_cached_prefix(
    system_instructions: str | None,
    model_input: str | list[Any],
@@ -394,28 +417,68 @@ async def _aclose(stream: AsyncIterator[TResponseStreamEvent]) -> None:
            await stream.aclose()


-async def _with_idle_timeout(
-    stream: AsyncIterator[TResponseStreamEvent], timeout: float
+async def _with_timeouts(
+    stream: AsyncIterator[TResponseStreamEvent],
+    *,
+    idle: float | None = None,
+    first_event: float | None = None,
+    total: float | None = None,
 ) -> AsyncIterator[TResponseStreamEvent]:
-    if timeout <= 0:
-        async for event in stream:
-            yield event
-        return
-
+    """Bound the first event, the gap between events and the whole stream; None is no bound."""
    iterator = stream.__aiter__()
+    started = time.monotonic()
    while True:
+        limits: list[tuple[float, float, str]] = []
+        # Wait until the soonest limit expires: first-event (else idle) until the
+        # first event arrives, then idle; plus whatever is left of the total.
+        if first_event is not None:
+            limits.append((first_event, first_event, "stream_first_event_timeout"))
+        elif idle is not None:
+            limits.append((idle, idle, "stream_idle_timeout"))
+        if total is not None:
+            left = max(0.0, started + total - time.monotonic())
+            limits.append((left, total, "stream_total_timeout"))
        try:
-            event = await asyncio.wait_for(iterator.__anext__(), timeout)
+            if limits:
+                event = await _next_event(iterator, *min(limits))
+            else:
+                event = await iterator.__anext__()
        except StopAsyncIteration:
            return
-        except TimeoutError:
+        except TimeoutError as exc:
            await _aclose(stream)
-            message = f"model stream produced no event for {timeout:.0f}s"
-            logger.warning("%s; abandoning the turn", message)
-            raise TimeoutError(message) from None
+            logger.warning("%s; abandoning the turn", exc)
+            raise
+        first_event = None
        yield event


+async def _next_event(
+    iterator: AsyncIterator[TResponseStreamEvent], wait: float, limit: float, reason: str
+) -> TResponseStreamEvent:
+    # asyncio.timeout() cancels without a message, so the request log could not
+    # tell this from a shutdown; this is asyncio.timeout() with a message.
+    task = asyncio.current_task()
+    assert task is not None
+    cancelling = task.cancelling()
+    expired = False
+
+    def expire() -> None:
+        nonlocal expired
+        expired = True
+        task.cancel(msg=f"{request_log.CANCEL_REASON_PREFIX}{reason}")
+
+    handle = asyncio.get_running_loop().call_later(wait, expire)
+    try:
+        return await iterator.__anext__()
+    except BaseException as exc:
+        if expired and task.uncancel() <= cancelling and isinstance(exc, asyncio.CancelledError):
+            raise TimeoutError(f"model stream hit {reason} ({limit:.0f}s)") from None
+        raise
+    finally:
+        handle.cancel()
+
+
 def _guard_event(
    event: TResponseStreamEvent, rewriter: TurnCallIdRewriter, limiter: TurnToolCallLimiter
 ) -> TResponseStreamEvent | None:
@@ -550,7 +613,10 @@ class StrixProvider(MultiProvider):
    def get_model(self, model_name: str | None) -> Model:
        llm = load_settings().llm
        slug = codex.subscription_model(model_name)
-        idle_timeout = float(llm.stream_idle_timeout)
+        # The settings use 0 for no bound.
+        idle_timeout = float(llm.stream_idle_timeout) or None
+        first_event_timeout = float(llm.stream_first_event_timeout) or None
+        total_timeout = float(llm.stream_total_timeout) or None
        if slug:
            # The ChatGPT subscription backend is always streamed; it has no
            # non-streaming mode to fall back to, so LLM_DISABLE_STREAMING
@@ -591,11 +657,13 @@ class StrixProvider(MultiProvider):
                # The wrapper emits its single event only once the whole request
                # is done, so an idle gap is meaningless here; the request
                # timeout bounds it instead.
-                idle_timeout = 0.0
+                idle_timeout = first_event_timeout = total_timeout = None
        return _TurnGuardModel(
            model,
            max_tool_calls_per_turn=llm.max_tool_calls_per_turn,
            stream_idle_timeout=idle_timeout,
+            first_event_timeout=first_event_timeout,
+            total_timeout=total_timeout,
        )


@@ -637,11 +705,15 @@ def configure_sdk_model_defaults(settings: Settings) -> None:
    """Apply Strix config to SDK-native defaults."""
    llm = settings.llm
    set_tracing_disabled(True)
+    from strix.report.state import record_llm_request
+
    request_log.install()
+    if settings.telemetry.run_record_timings:
+        request_log.register_sink(record_llm_request)
    if codex.subscription_model(llm.model):
        return
    _configure_litellm_compatibility()
-    _configure_openrouter_attribution(llm.model)
+    _configure_attribution(llm.model)
    if llm.api_key:
        set_default_openai_key(llm.api_key, use_for_tracing=False)
        _configure_litellm_default("api_key", llm.api_key)
@@ -745,6 +817,11 @@ def _install_openrouter_stream_cost_capture() -> None:

    class _StrixOpenRouterStreamingHandler(OpenRouterChatCompletionStreamingHandler):
        def chunk_parser(self, chunk: dict[str, Any]) -> Any:
+            # Before parsing: LiteLLM raises on an error chunk and drops its
+            # top-level ``provider``.
+            request_log.record_upstream_provider(
+                chunk.get("provider"), _openrouter_error_type(chunk.get("error"))
+            )
            stream = super().chunk_parser(chunk)
            usage = chunk.get("usage")
            response_id = chunk.get("id") or getattr(stream, "id", None)
@@ -763,12 +840,22 @@ def _install_openrouter_stream_cost_capture() -> None:
                json_mode=json_mode,
            )

+        def get_error_class(self, error_message: str, status_code: int, headers: Any) -> Any:
+            # A non-2xx reply names the provider in ``error.metadata``.
+            with contextlib.suppress(Exception):
+                error = json.loads(error_message)["error"]
+                request_log.record_upstream_provider(
+                    error["metadata"].get("provider_name"), _openrouter_error_type(error)
+                )
+            return super().get_error_class(error_message, status_code, headers)
+
        def transform_response(self, *args: Any, **kwargs: Any) -> Any:
            # Non-streamed replies (LLM_DISABLE_STREAMING) skip the chunk parser.
            response = super().transform_response(*args, **kwargs)
            raw_response = kwargs.get("raw_response", args[1] if len(args) > 1 else None)
            with contextlib.suppress(Exception):
                body = raw_response.json()  # type: ignore[union-attr]
+                request_log.record_upstream_provider(body.get("provider"))
                if body.get("usage"):
                    record_openrouter_provider(body.get("provider"), body["usage"])
            return response
@@ -789,9 +876,21 @@ def _install_openrouter_stream_cost_capture() -> None:
    litellm.OpenrouterConfig = _StrixOpenrouterConfig  # type: ignore[misc]


-OPENROUTER_ATTRIBUTION_HEADERS = {
+def _openrouter_error_type(error: object) -> object:
+    if isinstance(error, dict):
+        metadata = error.get("metadata")
+        if isinstance(metadata, dict):
+            return metadata.get("error_type")
+    return None
+
+
+APP_ATTRIBUTION_HEADERS = {
    "HTTP-Referer": "https://strix.ai",
    "X-Title": "Strix",
+}
+
+OPENROUTER_ATTRIBUTION_HEADERS = {
+    **APP_ATTRIBUTION_HEADERS,
    "X-OpenRouter-Categories": "cli-agent",
 }

@@ -800,20 +899,25 @@ def is_openrouter_model(model_name: str | None) -> bool:
    return bool(model_name) and "openrouter/" in (model_name or "").strip().lower()


-def _configure_openrouter_attribution(model_name: str | None) -> None:
+def is_vercel_ai_gateway_model(model_name: str | None) -> bool:
+    return (model_name or "").strip().lower().startswith("vercel_ai_gateway/")
+
+
+def attribution_headers(model_name: str | None) -> dict[str, str]:
+    if is_openrouter_model(model_name):
+        return OPENROUTER_ATTRIBUTION_HEADERS
+    if is_vercel_ai_gateway_model(model_name):
+        return APP_ATTRIBUTION_HEADERS
+    return {}
+
+
+def _configure_attribution(model_name: str | None) -> None:
    import litellm

    current: object = litellm.headers
    existing: dict[str, str] = current if isinstance(current, dict) else {}
-    if not is_openrouter_model(model_name):
-        if any(key in existing for key in OPENROUTER_ATTRIBUTION_HEADERS):
-            remaining = {
-                k: v for k, v in existing.items() if k not in OPENROUTER_ATTRIBUTION_HEADERS
-            }
-            litellm.headers = remaining or None  # type: ignore[assignment]
-        return
-
-    litellm.headers = {**existing, **OPENROUTER_ATTRIBUTION_HEADERS}  # type: ignore[assignment]
+    remaining = {k: v for k, v in existing.items() if k not in OPENROUTER_ATTRIBUTION_HEADERS}
+    litellm.headers = {**remaining, **attribution_headers(model_name)} or None  # type: ignore[assignment]


 def _configure_extra_headers(llm: LlmSettings) -> None:
@@ -905,6 +1009,12 @@ def model_supports_reasoning(model_name: str) -> bool:
    return bool(entry and entry.get("supports_reasoning"))


+def model_supports_images(model_name: str) -> bool:
+    """Return whether the model accepts image input. Assume yes until proven otherwise."""
+    entry = _catalog_entry(model_name)
+    return entry is None or bool(entry.get("supports_vision"))
+
+
 def _bare_openai_name(model_name: str) -> str:
    name = model_name.strip().lower()
    for prefix in ("litellm/", "any-llm/", "openai/"):
@@ -2,7 +2,7 @@

 from __future__ import annotations

-from typing import Literal
+from typing import Any, Literal

 from pydantic import AliasChoices, Field, field_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
@@ -55,11 +55,29 @@ class LlmSettings(BaseSettings):
        alias="LLM_EXTRA_HEADERS",
        repr=False,
    )
+    extra_body: dict[str, Any] | None = Field(
+        default=None,
+        alias="LLM_EXTRA_BODY",
+        repr=False,
+    )
    reasoning_effort: ReasoningEffort = Field(default="high", alias="STRIX_REASONING_EFFORT")
+    input_cost_per_mtok: float | None = Field(
+        default=None, ge=0, allow_inf_nan=False, alias="STRIX_LLM_INPUT_COST_PER_MTOK"
+    )
+    output_cost_per_mtok: float | None = Field(
+        default=None, ge=0, allow_inf_nan=False, alias="STRIX_LLM_OUTPUT_COST_PER_MTOK"
+    )
+    cached_input_cost_per_mtok: float | None = Field(
+        default=None, ge=0, allow_inf_nan=False, alias="STRIX_LLM_CACHED_INPUT_COST_PER_MTOK"
+    )
    force_required_tool_choice: bool = Field(
        default=False,
        alias="STRIX_FORCE_REQUIRED_TOOL_CHOICE",
    )
+    parallel_tool_calls: bool = Field(
+        default=False,
+        alias="STRIX_PARALLEL_TOOL_CALLS",
+    )
    prompt_cache: bool = Field(
        default=True,
        alias="STRIX_PROMPT_CACHE",
@@ -79,6 +97,10 @@ class LlmSettings(BaseSettings):
    timeout: int = Field(default=300, alias="LLM_TIMEOUT")
    preflight_timeout: int = Field(default=30, ge=1, alias="LLM_PREFLIGHT_TIMEOUT")
    stream_idle_timeout: int = Field(default=300, ge=0, alias="LLM_STREAM_IDLE_TIMEOUT")
+    stream_first_event_timeout: int = Field(
+        default=120, ge=0, alias="LLM_STREAM_FIRST_EVENT_TIMEOUT"
+    )
+    stream_total_timeout: int = Field(default=600, ge=0, alias="LLM_STREAM_TOTAL_TIMEOUT")
    max_tool_calls_per_turn: int = Field(
        default=32,
        ge=0,
@@ -106,6 +128,11 @@ class DedupeSettings(BaseSettings):
        alias="DEDUPE_LLM_EXTRA_HEADERS",
        repr=False,
    )
+    extra_body: dict[str, Any] | None = Field(
+        default=None,
+        alias="DEDUPE_LLM_EXTRA_BODY",
+        repr=False,
+    )

    @field_validator("reasoning_effort", mode="before")
    @classmethod
@@ -137,7 +164,7 @@ class RuntimeSettings(BaseSettings):
    model_config = _BASE_CONFIG

    image: str = Field(
-        default="ghcr.io/usestrix/strix-sandbox:1.3.0",
+        default="ghcr.io/usestrix/strix-sandbox:1.4.0",
        alias="STRIX_IMAGE",
    )
    backend: str = Field(default="docker", alias="STRIX_RUNTIME_BACKEND")
@@ -149,6 +176,7 @@ class TelemetrySettings(BaseSettings):
    model_config = _BASE_CONFIG

    enabled: bool = Field(default=True, alias="STRIX_TELEMETRY")
+    run_record_timings: bool = Field(default=False, alias="STRIX_RUN_RECORD_TIMINGS")


 WebSearchProvider = Literal["auto", "perplexity", "exa"]
@@ -0,0 +1,100 @@
+"""Keep replayed tool-call arguments valid JSON.
+
+A model occasionally emits a tool call whose ``arguments`` string is not a
+JSON object (truncated, unbalanced, empty). The call itself fails safely: the
+tool reports the parse error and the run continues. But the raw string is
+recorded in the session as-is, and strict OpenAI-compatible servers (vLLM,
+SGLang, ...) validate every assistant tool call in the request, so from then
+on each turn is rejected with ``Assistant tool call function.arguments must
+be valid JSON`` and the agent can never recover. Rewriting the replayed
+arguments to a JSON object that carries the original text keeps the history
+valid while the model still sees what it sent.
+"""
+
+from __future__ import annotations
+
+import json
+from typing import Any
+
+from openai.types.responses import ResponseFunctionToolCall
+
+
+MALFORMED_ARGUMENTS_KEY = "malformed_arguments"
+
+
+def _reject_constant(name: str) -> Any:
+    raise ValueError(f"non-standard JSON constant {name}")
+
+
+def _loads_strict(text: str) -> Any:
+    try:
+        return json.loads(text, parse_constant=_reject_constant)
+    except RecursionError as exc:
+        raise ValueError("arguments are nested too deeply to parse") from exc
+
+
+def describe_malformed_arguments(tool_name: str, arguments: str) -> str | None:
+    """Return a model-facing recovery message if ``arguments`` is not a JSON object.
+
+    A tool call whose arguments do not parse is almost always one the stream cut
+    off (the server flushed a prefix of the JSON, or the client closed early),
+    so instead of the SDK's generic parse-error result the model is told the
+    call never ran and must be re-issued whole.
+    """
+    if not arguments.strip():
+        return None
+    try:
+        parsed = _loads_strict(arguments)
+    except ValueError as exc:
+        detail = str(exc)
+    else:
+        if isinstance(parsed, dict):
+            return None
+        detail = f"expected a JSON object, got {type(parsed).__name__}"
+    return (
+        f"{tool_name}: the tool call was not executed because its arguments were "
+        f"truncated or otherwise not valid JSON ({detail}). The response was likely "
+        "cut off mid-call. Re-issue the call with complete, valid JSON arguments."
+    )
+
+
+def repair_arguments(arguments: object) -> str | None:
+    """Return replacement arguments a strict server accepts, or ``None`` if already valid."""
+    if not isinstance(arguments, str) or not arguments.strip():
+        return "{}"
+    try:
+        parsed = _loads_strict(arguments)
+    except ValueError:
+        parsed = None
+    if isinstance(parsed, dict):
+        return None
+    return json.dumps({MALFORMED_ARGUMENTS_KEY: arguments}, ensure_ascii=False)
+
+
+def repair_history_arguments(items: list[Any]) -> tuple[list[Any], bool]:
+    """Rewrite function calls in a conversation history whose arguments are not a JSON object."""
+    rebuilt: list[Any] = []
+    changed = False
+
+    for item in items:
+        if isinstance(item, dict):
+            if item.get("type") == "function_call":
+                repaired = repair_arguments(item.get("arguments"))
+                if repaired is not None:
+                    item = {**item, "arguments": repaired}  # noqa: PLW2901
+                    changed = True
+        elif isinstance(item, ResponseFunctionToolCall):
+            repaired = repair_arguments(item.arguments)
+            if repaired is not None:
+                item = item.model_copy(update={"arguments": repaired})  # noqa: PLW2901
+                changed = True
+        rebuilt.append(item)
+
+    return rebuilt, changed
+
+
+def repair_input(model_input: str | list[Any]) -> str | list[Any]:
+    if isinstance(model_input, str):
+        return model_input
+    rebuilt, changed = repair_history_arguments(model_input)
+    return rebuilt if changed else model_input
@@ -71,9 +71,11 @@ class AgentCoordinator:
        self._lock = asyncio.Lock()
        self._snapshot_path: Path | None = None
        self.is_shutting_down = False
+        self.root_finish_tool = "finish_scan"
        self._budget_stopped = False
        self._reserve_stopped = False
        self._budget_paused = False
+        self._operator_paused = False
        self._resume_epoch = 0
        self._budget_policy: BudgetPolicy = "stop"
        self._extend_budget: Callable[[], None] | None = None
@@ -104,6 +106,10 @@ class AgentCoordinator:
    def budget_paused(self) -> bool:
        return self._budget_paused

+    @property
+    def operator_paused(self) -> bool:
+        return self._operator_paused
+
    @property
    def resume_epoch(self) -> int:
        """Bumped by every ``resume_budget``; an agent parks against the value it read."""
@@ -149,6 +155,7 @@ class AgentCoordinator:
        """
        async with self._lock:
            self._budget_paused = True
+            self._operator_paused = True
        logger.info("scan paused by the operator")
        await self._maybe_snapshot()

@@ -164,6 +171,7 @@ class AgentCoordinator:
            self._set_budget_limit(max_budget_usd)
        async with self._lock:
            self._budget_paused = False
+            self._operator_paused = False
            self._resume_epoch += 1
            woken = [aid for aid, status in self.statuses.items() if status == "budget_paused"]
            for aid in woken:
@@ -231,7 +239,7 @@ class AgentCoordinator:
            self._budget_stopped = budget_stopped
            self._reserve_stopped = reserve_stopped
            if not budget_paused:
-                self._budget_paused = False
+                self._budget_paused = self._operator_paused
                for aid, status in self.statuses.items():
                    if status == "budget_paused":
                        self.statuses[aid] = "waiting"
@@ -423,7 +431,12 @@ class AgentCoordinator:
        unknown, or it is terminal and its loop does not park for wake-ups.
        """
        from_user = message.get("from") == "user"
-        if from_user and self._budget_paused and self._budget_policy != "pause":
+        if (
+            from_user
+            and self._budget_paused
+            and not self._operator_paused
+            and self._budget_policy != "pause"
+        ):
            await self.resume_from_budget_pause(exclude=target_agent_id)
        async with self._lock:
            if target_agent_id not in self.statuses:
@@ -446,7 +459,9 @@ class AgentCoordinator:
                self.recovery_counts.pop(target_agent_id, None)
                self.idle_resume_counts.pop(target_agent_id, None)
                self._parent_notified.discard(target_agent_id)
-                self.statuses[target_agent_id] = "waiting"
+                held = self._operator_paused and self.statuses[target_agent_id] == "budget_paused"
+                if not held:
+                    self.statuses[target_agent_id] = "waiting"
            runtime.wake.set()
            stream = runtime.stream
            interrupt_on_message = runtime.interrupt_on_message
@@ -639,6 +654,7 @@ class AgentCoordinator:
                "budget_stopped": self._budget_stopped,
                "reserve_stopped": self._reserve_stopped,
                "budget_paused": self._budget_paused,
+                "operator_paused": self._operator_paused,
            }

    async def restore(self, snap: dict[str, Any]) -> None:
@@ -661,6 +677,7 @@ class AgentCoordinator:
            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))
+            self._operator_paused = bool(snap.get("operator_paused", False))
            for aid in self.statuses:
                self.runtimes.setdefault(aid, AgentRuntime())

@@ -4,9 +4,12 @@ from __future__ import annotations

 import asyncio
 import contextlib
+import inspect
 import logging
+import re
 import uuid
 from collections.abc import Callable
+from dataclasses import replace
 from functools import cache
 from typing import TYPE_CHECKING, Any, cast

@@ -21,6 +24,8 @@ from openai import (

 from strix.config import codex
 from strix.core.hooks import (
+    SUBAGENT_TURN_LIMITS_KEY,
+    TURN_LIMIT_KEY,
    BudgetExceededError,
    BudgetPausedError,
    SubagentBudgetReservedError,
@@ -30,11 +35,13 @@ from strix.core.sessions import (
    enforce_image_budget,
    open_agent_session,
    replace_session_items,
+    scrub_images_from_items,
    seed_initial_input,
    strip_all_images_from_session,
 )
 from strix.llm import request_log
 from strix.llm.compaction import is_context_overflow, maybe_compact
+from strix.skills import skill_turn_limit


 if TYPE_CHECKING:
@@ -44,6 +51,7 @@ if TYPE_CHECKING:
    from agents.lifecycle import RunHooks
    from agents.memory import Session, SQLiteSession
    from agents.result import RunResultBase
+    from agents.run_config import CallModelData, ModelInputData

    from strix.core.agents import AgentCoordinator, Status

@@ -53,6 +61,18 @@ logger = logging.getLogger(__name__)
 StreamEventSink = Callable[[str, Any], None]

 _INPUT_REJECTION_CODES = frozenset({400, 404, 422})
+# Replies meaning "this model takes no images", not image errors in general: a
+# context overflow that counts "image/vision expansion" must not match.
+_IMAGE_REJECTION = re.compile(
+    r"no endpoints found that support image input"  # OpenRouter
+    r"|image_url is only supported by certain models"  # OpenAI
+    r"|is not a multimodal model|at most 0 image\(s\)"  # vLLM
+    r"|does not support image input"  # LiteLLM's Fireworks check
+    r"|doesn't support the image field"  # Bedrock Converse
+    r"|unknown variant `image_url`"  # DeepSeek, e.g. via Vercel AI Gateway
+    r"|'[^']*image[^']*' functionality not supported",  # Vercel AI Gateway (AI SDK), unverified
+    re.IGNORECASE,
+)
 _MAX_COMPACTIONS_PER_CYCLE = 2


@@ -120,6 +140,28 @@ async def _compact_session(
    )


+_TEXT_ONLY_IMAGE_TEXT = "[error: this model cannot view images; use `snapshot -i` instead]"
+
+
+def _with_image_scrub(run_config: RunConfig, context: dict[str, Any]) -> RunConfig:
+    if context.get("supports_images", True):
+        return run_config
+    # Chain any filter already set; it sees the scrubbed input.
+    inner = run_config.call_model_input_filter
+
+    async def _scrub(data: CallModelData[Any]) -> ModelInputData:
+        model_data = replace(
+            data.model_data,
+            input=scrub_images_from_items(data.model_data.input, text=_TEXT_ONLY_IMAGE_TEXT),
+        )
+        if inner is None:
+            return model_data
+        result = inner(replace(data, model_data=model_data))
+        return await result if inspect.isawaitable(result) else result
+
+    return replace(run_config, call_model_input_filter=_scrub)
+
+
 _MAX_TRANSIENT_MODEL_RETRIES = 5
 _TRANSIENT_MODEL_RETRY_BASE_DELAY_S = 2.0
 _TRANSIENT_MODEL_RETRY_MAX_DELAY_S = 90.0
@@ -130,6 +172,12 @@ def _model_error_status_code(exc: BaseException) -> int | None:
    return code if isinstance(code, int) else None


+def _is_image_rejection(exc: BaseException) -> bool:
+    return _model_error_status_code(exc) in _INPUT_REJECTION_CODES and bool(
+        _IMAGE_REJECTION.search(str(exc))
+    )
+
+
 def _is_transient_model_error(exc: BaseException) -> bool:
    if codex.is_content_guardrail_error(exc):
        return False
@@ -139,9 +187,7 @@ def _is_transient_model_error(exc: BaseException) -> bool:
        return True
    code = _model_error_status_code(exc)
    if code is not None:
-        import litellm
-
-        return bool(litellm._should_retry(code))
+        return code >= 400 and code not in (401, 402, 403, 404)
    return isinstance(exc, APIError)


@@ -259,7 +305,7 @@ async def _run_agent_loop(
        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())
+        await coordinator.send(agent_id, _reserve_notice(coordinator.root_finish_tool))

    if not (start_parked and interactive):
        with contextlib.suppress(BudgetPausedError):
@@ -384,6 +430,7 @@ async def spawn_child_agent(
        name=name,
        parent_id=parent_id,
        task=task,
+        skills=skills,
        initial_input=child_initial_input(
            name=name,
            child_id=child_id,
@@ -467,6 +514,7 @@ async def respawn_subagents(
                name=name,
                parent_id=parent_id,
                task=str(md.get("task", "")),
+                skills=child_skills,
                initial_input=[],
                start_parked=start_parked,
                event_sink=event_sink,
@@ -556,7 +604,7 @@ async def _run_until_lifecycle(
                    hooks=hooks,
                )
        except BudgetPausedError as exc:
-            if coordinator.budget_policy != "pause":
+            if coordinator.budget_policy != "pause" and not exc.operator_pause:
                raise
            # The agent parked right before an LLM call; everything up to that
            # point is already in its session. Once resumed, the same call goes
@@ -569,6 +617,7 @@ async def _run_until_lifecycle(
                # Stopped while parked (operator stop or a parent's stop_agent).
                await coordinator.reset_recovery(agent_id)
                return result
+            await coordinator.consume_pending(agent_id)
            if session is not None:
                input_data = []
            continue
@@ -599,6 +648,7 @@ async def _run_until_lifecycle(
        input_data = await _append_tool_required_message(
            session=session,
            context=context,
+            root_finish_tool=coordinator.root_finish_tool,
            attempt=recoveries,
            limit=recovery_limit,
            interactive=interactive,
@@ -747,7 +797,7 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
            stream = Runner.run_streamed(
                agent,
                input=input_data,
-                run_config=run_config,
+                run_config=_with_image_scrub(run_config, context),
                context=context,
                max_turns=max_turns,
                session=session,
@@ -787,7 +837,7 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
                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)
-            if coordinator.budget_policy == "pause":
+            if coordinator.budget_policy == "pause" or exc.operator_pause:
                await coordinator.park_for_budget(agent_id)
            else:
                await coordinator.pause_for_budget(agent_id)
@@ -805,11 +855,7 @@ async def _run_cycle(  # noqa: PLR0912, PLR0915
            await coordinator.trigger_budget_stop()
            raise
        except Exception as exc:
-            if (
-                image_strips < 3
-                and session is not None
-                and getattr(exc, "status_code", None) in _INPUT_REJECTION_CODES
-            ):
+            if image_strips < 3 and session is not None and _is_image_rejection(exc):
                try:
                    stripped = await strip_all_images_from_session(session)
                except Exception:
@@ -949,8 +995,9 @@ async def _append_tool_required_message(
    limit: int,
    interactive: bool,
    silent_yield: bool = False,
+    root_finish_tool: str = "finish_scan",
 ) -> list[dict[str, str]]:
-    finish_tool = "finish_scan" if context.get("parent_id") is None else "agent_finish"
+    finish_tool = root_finish_tool if context.get("parent_id") is None else "agent_finish"
    if silent_yield:
        message = (
            "You called wait_for_user without having written anything to the user since "
@@ -962,7 +1009,7 @@ async def _append_tool_required_message(
        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. "
+            "run continues. Continue immediately and call a tool. "
            "If you have nothing to do until the user replies, call wait_for_user; your "
            "text already reached them, so do not repeat it. "
            "If you are blocked waiting for another agent, call wait_for_agents. "
@@ -974,7 +1021,7 @@ async def _append_tool_required_message(
        message = (
            "Your previous response ended the autonomous 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. "
+            "ignored. Continue immediately and call a 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. "
@@ -1068,7 +1115,7 @@ async def notify_parent_on_terminal(
    )


-def _reserve_notice() -> dict[str, Any]:
+def _reserve_notice(finish_tool: str) -> dict[str, Any]:
    return {
        "from": "system",
        "type": "budget_reserve_stop",
@@ -1078,7 +1125,7 @@ def _reserve_notice() -> dict[str, Any]:
            "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."
+            f"spawn new ones — wrap up now and call {finish_tool}."
        ),
    }

@@ -1087,7 +1134,7 @@ 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())
+    await coordinator.send(root, _reserve_notice(coordinator.root_finish_tool))


 async def _notify_parent_on_exit(
@@ -1119,6 +1166,7 @@ async def _start_child_runner(
    name: str,
    parent_id: str | None,
    task: str,
+    skills: list[str],
    initial_input: Any,
    start_parked: bool = False,
    event_sink: StreamEventSink | None = None,
@@ -1132,6 +1180,9 @@ async def _start_child_runner(
    child_ctx["agent_id"] = child_id
    child_ctx["parent_id"] = parent_id
    child_ctx["task"] = task
+    child_ctx[TURN_LIMIT_KEY] = skill_turn_limit(
+        skills, parent_ctx.get(SUBAGENT_TURN_LIMITS_KEY) or {}
+    )

    async def _child_loop() -> None:
        # A budget stop is a clean scan-wide shutdown, not a child failure: the
@@ -4,24 +4,32 @@ from __future__ import annotations

 import logging
 import math
+import time
+from datetime import UTC, datetime
 from typing import TYPE_CHECKING, Any

 from agents.lifecycle import RunHooks

 from strix.core.agents import BudgetPolicy, coordinator_from_context
-from strix.report.state import get_global_report_state
+from strix.report.state import get_global_report_state, record_tool_call


 if TYPE_CHECKING:
    from agents import RunContextWrapper
    from agents.agent import Agent
    from agents.items import ModelResponse, TResponseInputItem
+    from agents.tool import Tool


 logger = logging.getLogger(__name__)


 LLM_TURN_KEY = "llm_turn"
+# ``run_strix_scan(subagent_turn_limits=...)``, and the limit it gives one sub-agent.
+SUBAGENT_TURN_LIMITS_KEY = "subagent_turn_limits"
+TURN_LIMIT_KEY = "turn_limit"
+_TURN_LIMIT_WARN_BEFORE = 10
+_TURN_LIMIT_REPEAT = 5

 # ``BudgetPolicy`` decides what happens when the accumulated LLM cost reaches
 # ``max_budget_usd``.
@@ -35,6 +43,8 @@ LLM_TURN_KEY = "llm_turn"
 # same call when the operator raises the limit or resumes.
 __all__ = [
    "LLM_TURN_KEY",
+    "SUBAGENT_TURN_LIMITS_KEY",
+    "TURN_LIMIT_KEY",
    "BudgetExceededError",
    "BudgetPausedError",
    "BudgetPolicy",
@@ -65,9 +75,12 @@ class BudgetPausedError(RuntimeError):
    decided to park; the agent waits for a resume newer than that.
    """

-    def __init__(self, message: str, *, resume_epoch: int = 0) -> None:
+    def __init__(
+        self, message: str, *, resume_epoch: int = 0, operator_pause: bool = False
+    ) -> None:
        super().__init__(message)
        self.resume_epoch = resume_epoch
+        self.operator_pause = operator_pause


 def _validate_budget(max_budget_usd: float | None) -> None:
@@ -104,16 +117,16 @@ _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."
+        "track so you can call {finish_tool} 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."
+        "{finish_tool} 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 "
+        "secure your findings and call {finish_tool} now — anything left unfinished when the "
        "limit is hit is discarded."
    ),
 )
@@ -138,8 +151,11 @@ _SUBAGENT_DIRECTIVES: tuple[str, ...] = (

 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]
+    if not is_root:
+        return _SUBAGENT_DIRECTIVES[stage]
+    coordinator = coordinator_from_context(context.context)
+    finish_tool = coordinator.root_finish_tool if coordinator is not None else "finish_scan"
+    return _ROOT_DIRECTIVES[stage].format(finish_tool=finish_tool)


 def _urgency(stage: int) -> str:
@@ -157,6 +173,7 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
        max_turns: int | None = None,
        interactive: bool = False,
        budget_policy: BudgetPolicy = "stop",
+        record_tool_timings: bool = False,
    ) -> None:
        _validate_budget(max_budget_usd)
        if max_turns is not None and max_turns <= 0:
@@ -169,6 +186,8 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
        self._max_turns = max_turns
        self._interactive = interactive
        self._budget_policy: BudgetPolicy = budget_policy
+        self._record_tool_timings = record_tool_timings
+        self._tool_starts: dict[str, tuple[datetime, float]] = {}

    @property
    def max_budget_usd(self) -> float | None:
@@ -195,15 +214,26 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
        system_prompt: str | None,  # noqa: ARG002
        input_items: list[TResponseInputItem],
    ) -> None:
+        self._pause_if_operator_paused(context)
        if self._budget_policy == "pause":
            self._pause_if_limited(context)
        context.context[LLM_TURN_KEY] = int(context.context.get(LLM_TURN_KEY, 0)) + 1
        try:
            self._maybe_warn_turns(context, input_items)
+            self._maybe_warn_soft_turn_limit(context, input_items)
            self._maybe_warn_budget(context, input_items)
        except Exception:
            logger.exception("budget/turn warning injection failed")

+    def _pause_if_operator_paused(self, context: RunContextWrapper[dict[str, Any]]) -> None:
+        coordinator = coordinator_from_context(context.context)
+        if coordinator is not None and coordinator.operator_paused:
+            raise BudgetPausedError(
+                "scan paused; waiting for the operator to resume",
+                resume_epoch=coordinator.resume_epoch,
+                operator_pause=True,
+            )
+
    def _pause_if_limited(self, context: RunContextWrapper[dict[str, Any]]) -> None:
        """Park the agent before a paid call when the scan is at its limit or paused.

@@ -254,6 +284,40 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
        )
        input_items.append({"role": "user", "content": content})

+    def _maybe_warn_soft_turn_limit(
+        self,
+        context: RunContextWrapper[dict[str, Any]],
+        input_items: list[TResponseInputItem],
+    ) -> None:
+        """A soft limit, never a stop: warn 10 turns before it, then every few turns past it."""
+        limit = context.context.get(TURN_LIMIT_KEY)
+        requests = getattr(getattr(context, "usage", None), "requests", None)
+        if not isinstance(limit, int) or not isinstance(requests, int):
+            return
+        turns_used = requests + 1
+        if turns_used == limit - _TURN_LIMIT_WARN_BEFORE:
+            content = (
+                f"[NOTICE] You have used {turns_used} of your {limit}-turn limit. Start "
+                "wrapping up: take only the steps you still need to answer your task, then "
+                "call agent_finish. You may go past the limit only while you are still "
+                "making real progress."
+            )
+        elif turns_used >= limit and (turns_used - limit) % _TURN_LIMIT_REPEAT == 0:
+            content = (
+                f"[NOTICE] You have used {turns_used} turns, past your {limit}-turn limit. "
+                "Call agent_finish now with what you have, listing anything unresolved in "
+                "open_items, unless your next steps are clearly making progress on your task."
+            )
+        else:
+            return
+        logger.info(
+            "Turn-limit notice for agent %s: turn %d, limit %d",
+            context.context.get("agent_id"),
+            turns_used,
+            limit,
+        )
+        input_items.append({"role": "user", "content": content})
+
    def _maybe_warn_budget(
        self,
        context: RunContextWrapper[dict[str, Any]],
@@ -299,6 +363,37 @@ class ReportUsageHooks(RunHooks[dict[str, Any]]):
            )
        input_items.append({"role": "user", "content": content})

+    async def on_tool_start(
+        self,
+        context: RunContextWrapper[dict[str, Any]],
+        agent: Agent[dict[str, Any]],  # noqa: ARG002
+        tool: Tool,  # noqa: ARG002
+    ) -> None:
+        call_id = getattr(context, "tool_call_id", None)
+        if self._record_tool_timings and isinstance(call_id, str):
+            self._tool_starts[call_id] = (datetime.now(UTC), time.monotonic())
+
+    async def on_tool_end(
+        self,
+        context: RunContextWrapper[dict[str, Any]],
+        agent: Agent[dict[str, Any]],  # noqa: ARG002
+        tool: Tool,
+        result: object,  # noqa: ARG002
+    ) -> None:
+        call_id = getattr(context, "tool_call_id", None)
+        start = self._tool_starts.pop(call_id, None) if isinstance(call_id, str) else None
+        if start is None:
+            return
+        started_at, started = start
+        ctx = context.context if isinstance(context.context, dict) else {}
+        agent_id = ctx.get("agent_id")
+        record_tool_call(
+            started_at=started_at,
+            agent_id=agent_id if isinstance(agent_id, str) else None,
+            tool=tool.name,
+            duration_ms=round((time.monotonic() - started) * 1000),
+        )
+
    async def on_llm_end(
        self,
        context: RunContextWrapper[dict[str, Any]],
@@ -10,12 +10,11 @@ from openai.types.shared import Reasoning

 from strix.config.models import (
    DEFAULT_MODEL_RETRY,
-    OPENROUTER_ATTRIBUTION_HEADERS,
+    attribution_headers,
    bedrock_route_supports_prompt_caching,
    is_bedrock_route,
    is_claude_model,
    is_known_openai_bare_model,
-    is_openrouter_model,
    model_supports_reasoning,
    request_timeout_extra_args,
    routes_through_litellm,
@@ -247,25 +246,30 @@ def make_model_settings(
    *,
    model_name: str,
    force_required_tool_choice: bool = False,
+    parallel_tool_calls: bool = False,
    request_timeout: float | None = None,
    prompt_cache: bool = True,
    extra_headers: dict[str, str] | None = None,
+    extra_body: dict[str, Any] | None = None,
    has_tools: bool = True,
 ) -> ModelSettings:
    headers = _request_headers(model_name, extra_headers)
    model_settings = ModelSettings(
-        parallel_tool_calls=False if has_tools else None,
+        parallel_tool_calls=parallel_tool_calls if has_tools else None,
        retry=DEFAULT_MODEL_RETRY,
        include_usage=True,
        extra_args=request_timeout_extra_args(request_timeout),
        extra_headers=headers,
+        extra_body=dict(extra_body) if extra_body else None,
    )
    if (
        reasoning_effort is not None
        and reasoning_effort != "none"
        and model_supports_reasoning(model_name)
    ):
-        model_settings = model_settings.resolve(_reasoning_settings(reasoning_effort))
+        model_settings = model_settings.resolve(
+            _reasoning_settings(reasoning_effort, extra_body),
+        )
    if force_required_tool_choice and _accepts_required_tool_choice(model_name):
        model_settings = model_settings.resolve(ModelSettings(tool_choice="required"))

@@ -282,15 +286,15 @@ def make_model_settings(
 def _request_headers(
    model_name: str, extra_headers: dict[str, str] | None
 ) -> dict[str, str] | None:
-    headers: dict[str, str] = {}
-    if is_openrouter_model(model_name):
-        headers.update(OPENROUTER_ATTRIBUTION_HEADERS)
+    headers: dict[str, str] = dict(attribution_headers(model_name))
    if extra_headers:
        headers.update(extra_headers)
    return headers or None


-def _reasoning_settings(effort: ReasoningEffort) -> ModelSettings:
+def _reasoning_settings(
+    effort: ReasoningEffort, extra_body: dict[str, Any] | None = None
+) -> ModelSettings:
    """``max`` is not in the OpenAI SDK's ``Reasoning.effort`` enum, so send it as
    a raw body field instead — also keeping it clear of LiteLLM's DeepSeek mapping,
    which collapses every ``reasoning_effort`` level to plain thinking-enabled.
@@ -298,11 +302,13 @@ def _reasoning_settings(effort: ReasoningEffort) -> ModelSettings:

    It goes in ``extra_body``, the field every model implementation forwards as the
    request's ``extra_body``; the same value under ``extra_args`` collides with that
-    keyword and raises before a request is ever sent.
+    keyword and raises before a request is ever sent. ``resolve`` replaces
+    ``extra_body`` wholesale, so the user's ``LLM_EXTRA_BODY`` fields are carried
+    over, with ``reasoning_effort`` taking precedence.
    """
    if effort != "max":
        return ModelSettings(reasoning=Reasoning(effort=effort))
-    return ModelSettings(extra_body={"reasoning_effort": "max"})
+    return ModelSettings(extra_body={**(extra_body or {}), "reasoning_effort": "max"})


 def _prompt_cache_extra_args(model_name: str) -> dict[str, Any] | None:
@@ -8,7 +8,7 @@ import io
 import json
 import logging
 import uuid
-from collections.abc import Callable
+from collections.abc import Awaitable, Callable, Mapping
 from pathlib import Path
 from typing import TYPE_CHECKING, Any

@@ -23,6 +23,7 @@ from strix.config.models import (
    StrixProvider,
    configure_sdk_api_route,
    configure_sdk_model_defaults,
+    model_supports_images,
    supports_strict_tool_schemas,
    uses_chat_completions_tool_schema,
 )
@@ -35,7 +36,12 @@ from strix.core.execution import (
 from strix.core.execution import (
    spawn_child_agent as start_child_agent,
 )
-from strix.core.hooks import BudgetExceededError, ReportUsageHooks, recomputed_budget_flags
+from strix.core.hooks import (
+    SUBAGENT_TURN_LIMITS_KEY,
+    BudgetExceededError,
+    ReportUsageHooks,
+    recomputed_budget_flags,
+)
 from strix.core.inputs import (
    build_root_task,
    build_scan_targets,
@@ -48,15 +54,18 @@ from strix.report.state import get_global_report_state
 from strix.runtime import session_manager
 from strix.telemetry import set_scan_phase
 from strix.telemetry.logging import set_scan_id, setup_scan_logging
+from strix.tools.finish.tool import finish_scan
 from strix.tools.output_store import (
    WORKSPACE_SPILL_DIR,
    configure_spill_writer,
 )
+from strix.tools.reporting.tool import create_vulnerability_report


 if TYPE_CHECKING:
    from agents.memory import SQLiteSession
    from agents.result import RunResultBase
+    from agents.tool import Tool

    from strix.runtime.status import StatusSink
    from strix.tools.mcp import (
@@ -67,6 +76,8 @@ if TYPE_CHECKING:

 logger = logging.getLogger(__name__)

+_MCP_PROMPT_WARMUP_TIMEOUT_SECONDS = 30.0
+
 StreamEventSink = Callable[[str, Any], None]

 # Receives the run's MCP connection roster as a list of non-secret status dicts
@@ -92,6 +103,19 @@ def _mcp_roster_payload(registry: McpRegistry) -> list[dict[str, Any]]:
    ]


+def _mcp_prompt_roster(registry: McpRegistry) -> list[dict[str, Any]]:
+    """The MCP roster rendered in agent prompts, with only verified counts."""
+    return [
+        {
+            "name": summary.name,
+            "purpose": summary.purpose,
+            "tool_count": summary.tool_count if summary.state == "catalog_ready" else None,
+            "state": summary.state,
+        }
+        for summary in registry.summaries()
+    ]
+
+
 def _record_mcp_connections(connection_names: list[str]) -> None:
    """Record which MCP servers this run configured, for the interfaces.

@@ -152,10 +176,17 @@ def _compose_root_instructions_override(
    is_whitebox: bool,
    is_diff_scoped: bool,
    interactive: bool,
+    parallel_tool_calls: bool,
    system_prompt_context: dict[str, Any],
+    supports_images: bool,
+    report_tool: str,
+    coverage_tools: bool,
+    replace_scan_prompt: bool = False,
 ) -> str | None:
    if root_instructions_override is None:
        return None
+    if replace_scan_prompt:
+        return f"{root_instructions_override}\n\n{render_scope_prompt(system_prompt_context)}"

    base_instructions = render_system_prompt(
        skills=skills,
@@ -164,8 +195,12 @@ def _compose_root_instructions_override(
        is_root=True,
        is_diff_scoped=is_diff_scoped,
        interactive=interactive,
+        parallel_tool_calls=parallel_tool_calls,
        system_prompt_context=system_prompt_context,
        include_scope=False,
+        supports_images=supports_images,
+        report_tool=report_tool,
+        coverage_tools=coverage_tools,
    )
    return (
        f"{base_instructions}\n\n"
@@ -183,7 +218,7 @@ async def run_strix_scan(
    scan_id: str | None = None,
    image: str,
    local_sources: list[dict[str, Any]] | None = None,
-    extra_files: list[dict[str, Any]] | None = None,
+    extra_files: list[dict[str, Any]] | Awaitable[list[dict[str, Any]] | None] | None = None,
    coordinator: AgentCoordinator | None = None,
    interactive: bool = False,
    max_turns: int = DEFAULT_MAX_TURNS,
@@ -193,15 +228,22 @@ async def run_strix_scan(
    cleanup_on_exit: bool = True,
    event_sink: StreamEventSink | None = None,
    root_instructions_override: str | None = None,
+    replace_root_scan_prompt: bool = False,
    extra_system_prompt_context: dict[str, Any] | None = None,
    status_sink: StatusSink | None = None,
    mcp_connection_requests: list[McpConnectionRequest] | None = None,
    mcp_status_sink: McpStatusSink | None = None,
+    root_finish_tool: Tool = finish_scan,
+    report_tool: Tool = create_vulnerability_report,
+    coverage_tools: bool = True,
+    subagent_turn_limits: Mapping[str, int] | None = None,
 ) -> RunResultBase | None:
    """Run or resume one Strix scan against a sandbox.

    ``root_instructions_override`` adds root scan instructions to the rendered
    root prompt without replacing the system-verified scope block.
+    ``replace_root_scan_prompt=True`` makes it the root's whole prompt, followed
+    by the scope, instead of adding it to the rendered scan prompt.
    ``extra_files`` entries (``{"workspace_path", "content"}``) are placed into
    the sandbox workspace at session bring-up; see
    :func:`strix.runtime.session_manager.create_or_reuse`.
@@ -219,6 +261,14 @@ async def run_strix_scan(
    command-line default) it reads ``~/.strix/mcp-servers.json`` itself. Either
    way the engine does the connecting, so the caller passes inert configs plus
    metadata and never live sessions.
+    ``root_finish_tool`` is the tool the root agent ends the run with.
+    ``report_tool`` is the tool every agent files findings with, in place of
+    ``create_vulnerability_report``.
+    ``coverage_tools=False`` drops the coverage ledger tools and their prompt
+    rules from every agent.
+    ``subagent_turn_limits`` maps a skill name to a soft turn limit for sub-agents
+    spawned with that skill (the largest applies). Agents are warned 10 turns
+    before it and reminded past it, but never stopped.
    """

    def report(phase: str) -> None:
@@ -265,6 +315,9 @@ async def run_strix_scan(
    strict_tool_schemas = supports_strict_tool_schemas(resolved_model)
    if not strict_tool_schemas:
        logger.info("Sending non-strict tool schemas: %s caps strict tools", resolved_model)
+    supports_images = model_supports_images(resolved_model)
+    if not supports_images:
+        logger.info("Leaving out image tools: %s does not accept images", resolved_model)

    if budget_policy not in ("stop", "pause"):
        raise ValueError(f"unknown budget_policy: {budget_policy!r}")
@@ -272,6 +325,7 @@ async def run_strix_scan(
        coordinator = AgentCoordinator()
    coordinator.set_snapshot_path(agents_path)
    coordinator.set_budget_policy(budget_policy)
+    coordinator.root_finish_tool = root_finish_tool.name

    from strix.tools.coverage.tools import hydrate_coverage_from_disk
    from strix.tools.notes.tools import hydrate_notes_from_disk
@@ -372,9 +426,11 @@ async def run_strix_scan(
            settings.llm.reasoning_effort,
            model_name=resolved_model,
            force_required_tool_choice=settings.llm.force_required_tool_choice,
+            parallel_tool_calls=settings.llm.parallel_tool_calls,
            request_timeout=settings.llm.timeout,
            prompt_cache=settings.llm.prompt_cache,
            extra_headers=settings.llm.extra_headers,
+            extra_body=settings.llm.extra_body,
        )
        run_config = RunConfig(
            model=resolved_model,
@@ -392,6 +448,7 @@ async def run_strix_scan(
            max_turns=max_turns,
            interactive=interactive,
            budget_policy=budget_policy,
+            record_tool_timings=settings.telemetry.run_record_timings,
        )
        coordinator.set_budget_limit_setter(hooks.set_max_budget_usd)
        if interactive and budget_policy != "pause":
@@ -432,19 +489,13 @@ async def run_strix_scan(
                _record_mcp_connections(mcp_registry.names())
                report(
                    f"MCP: configured {len(mcp_registry)} connection(s); "
-                    "warming them in the background"
+                    "connecting and listing tools"
                )
                scope_context["mcp_available"] = True
-                scope_context["mcp_connections"] = [
-                    {
-                        "name": summary.name,
-                        "purpose": summary.purpose,
-                        "tool_count": summary.tool_count,
-                    }
-                    for summary in mcp_registry.summaries()
-                ]
+                scope_context["mcp_connections"] = _mcp_prompt_roster(mcp_registry)

                def _emit_mcp_status() -> None:
+                    scope_context["mcp_connections"] = _mcp_prompt_roster(mcp_registry)
                    roster = _mcp_roster_payload(mcp_registry)
                    _persist_mcp_status(roster)
                    if mcp_status_sink is not None:
@@ -455,7 +506,11 @@ async def run_strix_scan(

                mcp_registry.set_status_sink(_emit_mcp_status)
                _emit_mcp_status()
-                mcp_registry.start_warmup(max_concurrency=6)
+                warmup_task = mcp_registry.start_warmup(max_concurrency=6)
+                await asyncio.wait(
+                    {warmup_task},
+                    timeout=_MCP_PROMPT_WARMUP_TIMEOUT_SECONDS,
+                )
        except Exception:
            logger.exception("Failed to configure user MCP servers; continuing without them")

@@ -467,7 +522,12 @@ async def run_strix_scan(
            is_whitebox=is_whitebox,
            is_diff_scoped=is_diff_scoped,
            interactive=interactive,
+            parallel_tool_calls=settings.llm.parallel_tool_calls,
            system_prompt_context=root_context,
+            supports_images=supports_images,
+            report_tool=report_tool.name,
+            coverage_tools=coverage_tools,
+            replace_scan_prompt=replace_root_scan_prompt,
        )

        root_agent = build_strix_agent(
@@ -478,10 +538,15 @@ async def run_strix_scan(
            is_whitebox=is_whitebox,
            is_diff_scoped=is_diff_scoped,
            interactive=interactive,
+            parallel_tool_calls=settings.llm.parallel_tool_calls,
            chat_completions_tools=chat_completions_tools,
            strict_tool_schemas=strict_tool_schemas,
            system_prompt_context=root_context,
            instructions_override=root_instructions,
+            supports_images=supports_images,
+            finish_tool=root_finish_tool,
+            report_tool=report_tool,
+            coverage_tools=coverage_tools,
        )

        if not is_resume:
@@ -498,9 +563,13 @@ async def run_strix_scan(
            is_whitebox=is_whitebox,
            is_diff_scoped=is_diff_scoped,
            interactive=interactive,
+            parallel_tool_calls=settings.llm.parallel_tool_calls,
            chat_completions_tools=chat_completions_tools,
            strict_tool_schemas=strict_tool_schemas,
            system_prompt_context=scope_context,
+            supports_images=supports_images,
+            report_tool=report_tool,
+            coverage_tools=coverage_tools,
        )

        async def spawn_child_agent(**kwargs: Any) -> dict[str, Any]:
@@ -528,6 +597,9 @@ async def run_strix_scan(
            "spawn_child_agent": spawn_child_agent,
            "scan_targets": build_scan_targets(scan_config),
            "max_context_images": settings.runtime.max_context_images,
+            "supports_images": supports_images,
+            "coverage_tools": coverage_tools,
+            SUBAGENT_TURN_LIMITS_KEY: dict(subagent_turn_limits or {}),
        }

        root_session = open_agent_session(root_id, agents_db)
@@ -592,7 +664,7 @@ async def run_strix_scan(
        )
        if not interactive and result is not None:
            final = getattr(result, "final_output", None)
-            # Lifecycle tools mark the root completed. 
+            # Lifecycle tools mark the root completed.
            async with coordinator._lock:
                root_completed = coordinator.statuses.get(root_id) == "completed"
            if not root_completed:
@@ -199,13 +199,13 @@ async def enforce_image_budget(session: Session, max_images: int) -> bool:
    return await _rewrite_session(session, _transform)


-def scrub_images_from_items(items: list[Any]) -> list[Any]:
+def scrub_images_from_items(items: list[Any], *, text: str = _INHERITED_IMAGE_TEXT) -> list[Any]:
    """Return a copy of ``items`` with every image block replaced by text."""

    def _scrub(obj: Any) -> Any:
        if isinstance(obj, dict):
            if obj.get("type") == "input_image":
-                return {"type": "input_text", "text": _INHERITED_IMAGE_TEXT}
+                return {"type": "input_text", "text": text}
            return {k: _scrub(v) for k, v in obj.items()}
        if isinstance(obj, list):
            return [_scrub(v) for v in obj]
@@ -164,12 +164,13 @@ async def warm_up_llm() -> None:
            raw_model = dedupe_model
            deduper = resolve_dedupe_model(settings.dedupe, dedupe_model)
            # 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.
+            # never receive the main endpoint's headers or body fields; it has
+            # its own DEDUPE_LLM_EXTRA_HEADERS and DEDUPE_LLM_EXTRA_BODY.
            await preflight_request(
                deduper,
                model_name=dedupe_model,
                extra_headers=settings.dedupe.extra_headers,
+                extra_body=settings.dedupe.extra_body,
                timeout=llm.preflight_timeout,
                api_base_setting="DEDUPE_LLM_API_BASE",
            )
@@ -116,6 +116,7 @@ async def preflight_model_connection(
        model,
        model_name=model_name,
        extra_headers=resolved_settings.llm.extra_headers,
+        extra_body=resolved_settings.llm.extra_body,
        timeout=resolved_settings.llm.preflight_timeout,
        api_base_setting="LLM_API_BASE",
    )
@@ -126,6 +127,7 @@ async def preflight_request(
    *,
    model_name: str,
    extra_headers: dict[str, str] | None,
+    extra_body: dict[str, Any] | None,
    timeout: int,
    api_base_setting: str,
 ) -> None:
@@ -144,6 +146,7 @@ async def preflight_request(
        request_timeout=timeout,
        prompt_cache=False,
        extra_headers=extra_headers,
+        extra_body=extra_body,
        has_tools=False,
    )
    try:
@@ -66,7 +66,7 @@ func panelFocus(panel sidebarPanel) (focusMode, bool) {

 func (m Model) sidebarPanels() []panelRect {
 	statsHeight, vulnHeight, mcpHeight, agentHeight := m.sidebarHeights()
-	top := m.viewerHeight()
+	top := m.sidebarTopHeight()
 	rects := []panelRect{{panelAgents, top, agentHeight}}
 	top += agentHeight
 	if vulnHeight > 0 {
@@ -83,7 +83,7 @@ func (m Model) sidebarPanels() []panelRect {

 func (m Model) sidebarGap() int {
 	statsHeight, vulnHeight, mcpHeight, agentHeight := m.sidebarHeights()
-	return max(0, m.height-m.viewerHeight()-agentHeight-vulnHeight-mcpHeight-statsHeight)
+	return max(0, m.height-m.sidebarTopHeight()-agentHeight-vulnHeight-mcpHeight-statsHeight)
 }

 func (m Model) panelAt(y int) (panelRect, bool) {
@@ -355,3 +355,77 @@ func TestLonePanelHasNoControls(t *testing.T) {
 		t.Fatalf("controls should return once a second panel exists")
 	}
 }
+
+func TestTargetsBoxSitsUnderViewer(t *testing.T) {
+	m := panelsModel(t)
+	m.snapshot.Targets = []string{"https://example.com", "./app"}
+	m.snapshot.TargetCount = 2
+	m.resizeViewport()
+	_, sidebarWidth, _, _ := m.layout()
+	rows := strings.Split(ansi.Strip(m.sidebarView(sidebarWidth, m.height)), "\n")
+	if len(rows) != m.height {
+		t.Fatalf("sidebar renders %d rows, want %d", len(rows), m.height)
+	}
+	top := m.viewerHeight()
+	for i, want := range []string{"Targets 2", "▸ https://example.com", "▸ ./app"} {
+		if !strings.Contains(rows[top+1+i], want) {
+			t.Fatalf("row %d missing %q:\n%s", top+1+i, want, strings.Join(rows, "\n"))
+		}
+	}
+	if !strings.Contains(rows[panelRectOf(t, m, panelAgents).top+1], "Agents") {
+		t.Fatalf("agents panel not where its rect says:\n%s", strings.Join(rows, "\n"))
+	}
+}
+
+func TestTargetsBoxDropsOnShortTerminal(t *testing.T) {
+	m := panelsModel(t)
+	m.height = 19
+	m.snapshot.Targets = []string{"a", "b", "c", "d", "e"}
+	m.snapshot.TargetCount = 5
+	m.resizeViewport()
+	_, sidebarWidth, _, _ := m.layout()
+	if got := lipgloss.Height(m.sidebarView(sidebarWidth, m.height)); got > m.height {
+		t.Fatalf("sidebar renders %d rows on a %d-row terminal", got, m.height)
+	}
+}
+
+// The targets box only takes rows the agent tree would otherwise grow into, so
+// the stats panel keeps the Caido URL and the other panels keep their heights.
+func TestTargetsBoxNeverSqueezesPanels(t *testing.T) {
+	for height := 19; height <= 40; height++ {
+		m := panelsModel(t)
+		m.height = height
+		m.snapshot.Usage = map[string]any{"total_tokens": 1234, "cost": 0.5}
+		m.snapshot.CaidoURL = "http://127.0.0.1:8080"
+		m.resizeViewport()
+		stats, vuln, mcp, _ := m.sidebarHeights()
+		m.snapshot.Targets = []string{"a", "b", "c", "d", "e"}
+		m.snapshot.TargetCount = 5
+		m.resizeViewport()
+		gotStats, gotVuln, gotMcp, _ := m.sidebarHeights()
+		if gotStats != stats || gotVuln != vuln || gotMcp != mcp {
+			t.Fatalf("height %d: targets squeezed panels: stats %d->%d findings %d->%d mcp %d->%d",
+				height, stats, gotStats, vuln, gotVuln, mcp, gotMcp)
+		}
+		_, sidebarWidth, _, _ := m.layout()
+		if got := lipgloss.Height(m.sidebarView(sidebarWidth, m.height)); got != m.height {
+			t.Fatalf("height %d: sidebar renders %d rows", height, got)
+		}
+	}
+}
+
+func TestTargetsViewCountsTheTargetsItCannotShow(t *testing.T) {
+	m := panelsModel(t)
+	m.snapshot.Targets = []string{"a", "b", "c", "d", "e"}
+	m.snapshot.TargetCount = 5
+	for rows, want := range map[int]string{
+		2: "Targets 5\n+5 more",
+		3: "Targets 5\n▸ a\n+4 more",
+		6: "Targets 5\n▸ a\n▸ b\n▸ c\n▸ d\n+1 more",
+		9: "Targets 5\n▸ a\n▸ b\n▸ c\n▸ d\n+1 more",
+	} {
+		if got := ansi.Strip(m.targetsView(20, rows)); got != want {
+			t.Fatalf("%d rows:\n%s\nwant:\n%s", rows, got, want)
+		}
+	}
+}
@@ -421,7 +421,7 @@ func (m Model) setupComposer(width int) string {
 	inner := max(1, width-4)
 	body := m.highlightInputSelection(m.input.View())
 	body += "\n\n" + m.setupSummaryView(inner)
-	if targets := m.setupTargetsView(inner); targets != "" {
+	if targets := m.targetsView(inner, maxTargetRows); targets != "" {
 		body += "\n" + targets
 	}
 	return lipgloss.NewStyle().Width(max(1, width-2)).Padding(0, 1).
@@ -454,15 +454,22 @@ func (m Model) setupSummaryView(width int) string {
 	return truncate(strings.Join(chips, render.Dim().Render(" · ")), max(1, width))
 }

-// setupTargetsView lists what the scan is pointed at, once anything is queued.
-func (m Model) setupTargetsView(width int) string {
-	if len(m.snapshot.Targets) == 0 {
+// maxTargetRows fits the header, four targets and the count of the rest.
+const maxTargetRows = 6
+
+// targetsView lists what the scan is pointed at, once anything is queued, in at
+// most maxRows rows. Targets that do not fit are counted on the last row.
+func (m Model) targetsView(width, maxRows int) string {
+	if len(m.snapshot.Targets) == 0 || maxRows < 2 {
 		return ""
 	}
-	const visible = 4
 	total := max(m.snapshot.TargetCount, len(m.snapshot.Targets))
+	visible := min(len(m.snapshot.Targets), maxTargetRows-2, maxRows-1)
+	if visible < total {
+		visible = min(visible, maxRows-2)
+	}
 	rows := []string{render.Bold(green).Render("Targets") + render.Dim().Render(fmt.Sprintf(" %d", total))}
-	for _, target := range m.snapshot.Targets[:min(visible, len(m.snapshot.Targets))] {
+	for _, target := range m.snapshot.Targets[:visible] {
 		rows = append(rows, render.Col(dim).Render("▸ ")+render.Col(white).Render(truncate(target, max(1, width-2))))
 	}
 	if hidden := total - visible; hidden > 0 {
@@ -542,6 +542,15 @@ func (m Model) sidebarView(width, height int) string {
 		BorderForeground(dark).
 		Padding(0, 1).
 		Render(m.viewerBox(width - 4))}
+	if targetsHeight := m.targetsBoxHeight(); targetsHeight > 0 {
+		parts = append(parts, lipgloss.NewStyle().
+			Width(width-2).
+			Height(targetsHeight-2).
+			Border(lipgloss.RoundedBorder()).
+			BorderForeground(dark).
+			Padding(0, 1).
+			Render(m.targetsView(width-4, targetsHeight-2)))
+	}
 	agents := ""
 	if agentHeight > 1 {
 		agentRows := max(1, agentHeight-4)
@@ -592,6 +601,13 @@ func (m Model) sidebarView(width, height int) string {
 }

 func (m Model) sidebarHeights() (statsHeight, vulnHeight, mcpHeight, agentHeight int) {
+	_, statsHeight, vulnHeight, mcpHeight, agentHeight = m.sidebarLayout()
+	return
+}
+
+// sidebarLayout sizes the targets box and the panels. The targets box only
+// takes rows no panel needs, so it never squeezes the panels below it.
+func (m Model) sidebarLayout() (targetsHeight, statsHeight, vulnHeight, mcpHeight, agentHeight int) {
 	// Measure the stats panel the way its box will render it: a long model name
 	// wraps inside the sidebar, and counting only its newlines would size the
 	// box short and push the whole frame past the bottom of the terminal.
@@ -628,6 +644,14 @@ func (m Model) sidebarHeights() (statsHeight, vulnHeight, mcpHeight, agentHeight
 		}
 	}
 	spare := max(0, m.height-m.viewerHeight()-statsHeight-vulnHeight-mcpHeight-agentHeight)
+	// The box needs its border and two rows: the header and a target or the
+	// count of hidden ones. A zoomed panel keeps the whole sidebar.
+	if targets := m.targetsView(m.sidebarInnerWidth(), maxTargetRows); targets != "" && m.zoomedPanel == panelNone {
+		if fit := min(lipgloss.Height(targets)+2, spare); fit >= 4 {
+			targetsHeight = fit
+			spare -= fit
+		}
+	}
 	switch {
 	case m.zoomedPanel == panelFindings && vulnHeight > 1:
 		vulnHeight += spare
@@ -643,6 +667,17 @@ func (m Model) viewerHeight() int {
 	return strings.Count(m.viewerView(m.viewerContentWidth()), "\n") + 3
 }

+// sidebarTopHeight is the rows above the panels: the viewer box and, when the
+// run has targets, the targets box under it.
+func (m Model) sidebarTopHeight() int {
+	return m.viewerHeight() + m.targetsBoxHeight()
+}
+
+func (m Model) targetsBoxHeight() int {
+	targetsHeight, _, _, _, _ := m.sidebarLayout()
+	return targetsHeight
+}
+
 func (m Model) sidebarInnerWidth() int {
 	_, sidebarWidth, _, _ := m.layout()
 	if sidebarWidth == 0 {
@@ -0,0 +1,2 @@
+# Supply-chain cooldown: only install versions published at least 5 days ago (npm >= 11.10).
+min-release-age=5
@@ -29,7 +29,7 @@
        "@vitejs/plugin-react": "^5.0.0",
        "tailwindcss": "^4.1.0",
        "typescript": "^5.6.0",
-        "vite": "^6.0.0"
+        "vite": "^6.4.4"
      }
    },
    "node_modules/@babel/code-frame": {
@@ -1401,6 +1401,72 @@
        "node": ">=14.0.0"
      }
    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@emnapi/core": {
+      "version": "1.11.1",
+      "dev": true,
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "@emnapi/wasi-threads": "1.2.2",
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@emnapi/runtime": {
+      "version": "1.11.1",
+      "dev": true,
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@emnapi/wasi-threads": {
+      "version": "1.2.2",
+      "dev": true,
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@napi-rs/wasm-runtime": {
+      "version": "1.1.4",
+      "dev": true,
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "@tybys/wasm-util": "^0.10.1"
+      },
+      "funding": {
+        "type": "github",
+        "url": "https://github.com/sponsors/Brooooooklyn"
+      },
+      "peerDependencies": {
+        "@emnapi/core": "^1.7.1",
+        "@emnapi/runtime": "^1.7.1"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/@tybys/wasm-util": {
+      "version": "0.10.2",
+      "dev": true,
+      "inBundle": true,
+      "license": "MIT",
+      "optional": true,
+      "dependencies": {
+        "tslib": "^2.4.0"
+      }
+    },
+    "node_modules/@tailwindcss/oxide-wasm32-wasi/node_modules/tslib": {
+      "version": "2.8.1",
+      "dev": true,
+      "inBundle": true,
+      "license": "0BSD",
+      "optional": true
+    },
    "node_modules/@tailwindcss/oxide-win32-arm64-msvc": {
      "version": "4.3.3",
      "resolved": "https://registry.npmjs.org/@tailwindcss/oxide-win32-arm64-msvc/-/oxide-win32-arm64-msvc-4.3.3.tgz",
@@ -4096,9 +4162,9 @@
      }
    },
    "node_modules/vite": {
-      "version": "6.4.3",
-      "resolved": "https://registry.npmjs.org/vite/-/vite-6.4.3.tgz",
-      "integrity": "sha512-NTKlcQjlAK7MlQoyb6LgaqHc8sso/pVyUJYWMws3jg21uTJw/LddqIFPcPqP6PzpgbIcZyKI85sFE4HBrQDA8A==",
+      "version": "6.4.4",
+      "resolved": "https://registry.npmjs.org/vite/-/vite-6.4.4.tgz",
+      "integrity": "sha512-xGjU/SyjeNyb0E4KxXBvYz/1C+3JXuujvf2hXNOtnlcE+58o5ajd9N2/n2VhkiJvpCP2qQasnUpx5PXNeI+p1Q==",
      "dev": true,
      "license": "MIT",
      "dependencies": {
@@ -30,6 +30,6 @@
    "@vitejs/plugin-react": "^5.0.0",
    "tailwindcss": "^4.1.0",
    "typescript": "^5.6.0",
-    "vite": "^6.0.0"
+    "vite": "^6.4.4"
  }
 }
@@ -307,6 +307,7 @@ async def _summarize(model: str, prompt: str, max_tokens: int) -> str | None:
        request_timeout=llm.timeout,
        prompt_cache=False,
        extra_headers=llm.extra_headers,
+        extra_body=llm.extra_body,
        has_tools=False,
    ).resolve(ModelSettings(max_tokens=max_tokens))
    try:
@@ -159,11 +159,33 @@ class HttpReply:

    status_code: int | None = None
    headers: dict[str, str] | None = None
+    upstream_provider: str | None = None
+    upstream_error_type: str | None = None


 _http_reply: ContextVar[HttpReply | None] = ContextVar("strix_llm_http_reply", default=None)


+def record_upstream_provider(provider: object, error_type: object = None) -> None:
+    """Remember which provider behind a gateway served the attempt in flight."""
+    reply = _http_reply.get()
+    if reply is None:
+        return
+    if isinstance(provider, str) and provider:
+        reply.upstream_provider = provider
+    if isinstance(error_type, str) and error_type:
+        reply.upstream_error_type = error_type
+
+
+def _upstream_details(reply: HttpReply | None) -> list[tuple[str, object]]:
+    if reply is None:
+        return []
+    return [
+        ("upstream_provider", reply.upstream_provider),
+        ("upstream_error_type", reply.upstream_error_type),
+    ]
+
+
 async def record_http_reply(response: Response) -> None:
    """httpx ``response`` event hook: remember the reply for the attempt in flight.

@@ -772,6 +794,7 @@ def _litellm_success(
            ("request", _litellm_request_details(kwargs)),
            ("response", _litellm_response_details(response)),
            ("litellm", _litellm_hidden_details(hidden, slo)),
+            *_upstream_details(_http_reply.get()),
        ),
    )

@@ -821,6 +844,7 @@ def _litellm_failure(
            ("request", _litellm_request_details(kwargs)),
            ("error", error or None),
            ("litellm", _litellm_hidden_details(hidden, slo)),
+            *_upstream_details(_http_reply.get()),
        ),
    )

@@ -894,15 +918,18 @@ def _observe_sdk_shared_http_client() -> None:


 def _log_line_sink(event: LlmRequestEvent) -> None:
+    details = event.details or {}
    level = logging.DEBUG if event.outcome == "success" else logging.WARNING
    logger.log(
        level,
-        "llm_request route=%s provider=%s model=%s host=%s outcome=%s status=%s "
-        "request_id=%s response_id=%s stream=%s duration_ms=%d ttft_ms=%s "
+        "llm_request route=%s provider=%s upstream=%s upstream_error=%s model=%s host=%s "
+        "outcome=%s status=%s request_id=%s response_id=%s stream=%s duration_ms=%d ttft_ms=%s "
        "req_bytes=%s res_bytes=%s finish=%s "
        "in=%s out=%s cached=%s cost=%s agent=%s attempt=%d%s",
        event.route,
        event.provider or "-",
+        details.get("upstream_provider") or "-",
+        details.get("upstream_error_type") or "-",
        event.model,
        event.api_host or "-",
        event.outcome,
@@ -1070,6 +1097,7 @@ class RequestLoggingModel(Model):
                details=merge_details(
                    ("request", request.details),
                    ("response", _openai_response_details(response, raw_response)),
+                    *_upstream_details(reply),
                ),
            )
        status, request_id = _openai_error_fields(exc)
@@ -1082,13 +1110,14 @@ class RequestLoggingModel(Model):
            outcome="error",
            status_code=status,
            provider_request_id=request_id or base.provider_request_id,
-            error_type=type(exc).__name__,
+            error_type=_cancel_reason(exc) or type(exc).__name__,
            error_message=_abandonment_message(exc) or clean_error_message(exc),
            response_bytes=_exception_body_size(exc, status),
            response_headers=headers_from_response(error_headers) or base.response_headers,
            details=merge_details(
                ("request", request.details),
                ("error", _exception_details(exc)),
+                *_upstream_details(reply),
            ),
        )

@@ -1318,9 +1347,23 @@ def _is_abandonment(exc: BaseException | None) -> bool:
    return isinstance(exc, asyncio.CancelledError | GeneratorExit)


+# Prefixes the reason a stream timeout gives when it cancels an attempt.
+CANCEL_REASON_PREFIX = "strix:"
+
+
+def _cancel_reason(exc: BaseException) -> str | None:
+    """The stream timeout that cancelled the attempt, e.g. ``stream_idle_timeout``."""
+    if isinstance(exc, asyncio.CancelledError) and exc.args:
+        message = str(exc.args[0])
+        if message.startswith(CANCEL_REASON_PREFIX):
+            return message.removeprefix(CANCEL_REASON_PREFIX)
+    return None
+
+
 def _abandonment_message(exc: BaseException) -> str | None:
    if isinstance(exc, asyncio.CancelledError):
-        return "attempt cancelled before the reply was consumed (stream idle timeout or shutdown)"
+        reason = f": {exc.args[0]}" if exc.args else ""
+        return f"attempt cancelled before the reply was consumed{reason}"
    if isinstance(exc, GeneratorExit):
        return "stream closed by the caller before it finished"
    return None
@@ -39,11 +39,13 @@ def _dedupe_model_settings(
        model_name=model_name,
        force_required_tool_choice=False,
        request_timeout=request_timeout,
-        # The main model's headers apply only when dedupe falls back to the main
-        # model; a dedicated dedupe model may route to another provider, which
-        # must never receive the main endpoint's credentials. A dedicated model
-        # gets its own DEDUPE_LLM_EXTRA_HEADERS instead.
+        # The main model's headers and body fields apply only when dedupe falls
+        # back to the main model; a dedicated dedupe model may route to another
+        # provider, which must never receive the main endpoint's credentials or
+        # identifiers. A dedicated model gets its own DEDUPE_LLM_EXTRA_HEADERS
+        # and DEDUPE_LLM_EXTRA_BODY instead.
        extra_headers=dedupe.extra_headers if dedupe.model else llm.extra_headers,
+        extra_body=dedupe.extra_body if dedupe.model else llm.extra_body,
        has_tools=False,
    )

@@ -298,7 +300,15 @@ def _parse_dedupe_response(content: str) -> dict[str, Any]:
        raise ValueError(f"No JSON object found in dedupe response: {content[:500]}")
    parsed = json.loads(text[start : end + 1])

-    duplicate_id = str(parsed.get("duplicate_id") or "")[:64]
+    is_duplicate = parsed.get("is_duplicate")
+    if not isinstance(is_duplicate, bool):
+        raise TypeError("is_duplicate must be a boolean")
+    duplicate_id = parsed.get("duplicate_id") or ""
+    if is_duplicate and (not isinstance(duplicate_id, str) or not duplicate_id):
+        raise ValueError("A duplicate decision must name an existing report ID")
+    # Keep the ID verbatim: truncating or stripping it could turn an unknown ID
+    # into a different, existing one before the membership check below.
+    duplicate_id = str(duplicate_id)
    reason = str(parsed.get("reason") or "")[:500]
    try:
        confidence = float(parsed.get("confidence", 0.0))
@@ -306,13 +316,22 @@ def _parse_dedupe_response(content: str) -> dict[str, Any]:
        confidence = 0.0

    return {
-        "is_duplicate": bool(parsed.get("is_duplicate", False)),
+        "is_duplicate": is_duplicate,
        "duplicate_id": duplicate_id,
        "confidence": confidence,
        "reason": reason,
    }


+def _validate_duplicate_reference(
+    result: dict[str, Any], existing_reports: list[dict[str, Any]]
+) -> None:
+    if result["is_duplicate"] and not any(
+        report.get("id") == result["duplicate_id"] for report in existing_reports
+    ):
+        raise ValueError("Duplicate decision refers to an unknown report ID")
+
+
 def _extract_text(response: ModelResponse) -> str:
    parts: list[str] = []
    for item in response.output:
@@ -395,6 +414,7 @@ async def check_duplicate(
            }

        result = _parse_dedupe_response(content)
+        _validate_duplicate_reference(result, existing_reports)

        logger.info(
            "Deduplication check: is_duplicate=%s, confidence=%.2f, reason=%s",
@@ -28,6 +28,8 @@ from strix.telemetry import posthog, scarf
 if TYPE_CHECKING:
    from agents.usage import Usage

+    from strix.llm.request_log import LlmRequestEvent
+

 logger = logging.getLogger(__name__)

@@ -1019,6 +1021,47 @@ class StreamedOpenRouterCosts:
 streamed_openrouter_costs = StreamedOpenRouterCosts()


+def record_llm_request(event: "LlmRequestEvent") -> None:
+    """Request-log sink: each provider attempt's timing, kept in run.json's ``llm_requests``."""
+    report_state = get_global_report_state()
+    if report_state is None:
+        return
+    details = event.details or {}
+    report_state.run_record.setdefault("llm_requests", []).append(
+        {
+            "started_at": event.started_at.isoformat(),
+            "agent_id": event.agent_id,
+            "model": event.model,
+            "provider": details.get("upstream_provider") or event.provider,
+            "outcome": event.outcome,
+            "status": event.status_code,
+            "duration_ms": event.duration_ms,
+            "ttft_ms": event.time_to_first_token_ms,
+            "input_tokens": event.input_tokens,
+            "cached_tokens": event.cached_input_tokens,
+            "output_tokens": event.output_tokens,
+            "retry_attempt": event.retry_attempt,
+        }
+    )
+
+
+def record_tool_call(
+    *, started_at: datetime, agent_id: str | None, tool: str, duration_ms: int
+) -> None:
+    """Each tool call's timing, kept in run.json's ``tool_calls``."""
+    report_state = get_global_report_state()
+    if report_state is None:
+        return
+    report_state.run_record.setdefault("tool_calls", []).append(
+        {
+            "started_at": started_at.isoformat(),
+            "agent_id": agent_id,
+            "tool": tool,
+            "duration_ms": duration_ms,
+        }
+    )
+
+
 def record_openrouter_provider(provider: Any, usage: Any) -> None:
    """Tally which upstream provider served a stream, from its final usage chunk.

@@ -8,11 +8,38 @@ from typing import Any
 from agents.usage import Usage, deserialize_usage, serialize_usage
 from pydantic import BaseModel, TypeAdapter, ValidationError

+from strix.config.loader import load_settings
 from strix.report.pricing import resolve_litellm_model


 logger = logging.getLogger(__name__)

+Prices = tuple[float | None, float | None, float | None]
+
+
+def _configured_prices() -> tuple[str | None, Prices | None]:
+    llm = load_settings().llm
+    prices = (llm.input_cost_per_mtok, llm.output_cost_per_mtok, llm.cached_input_cost_per_mtok)
+    if all(price is None for price in prices):
+        return None, None
+    return (llm.model or "").strip() or None, prices
+
+
+def _configured_usage_cost(usage: Usage, prices: Prices) -> float:
+    input_rate, output_rate, cached_rate = prices
+    input_tokens = max(0, _int_or_zero(getattr(usage, "input_tokens", 0)))
+    cached_details = getattr(usage, "input_tokens_details", None)
+    cached_tokens = max(0, _int_or_zero(getattr(cached_details, "cached_tokens", 0)))
+    cached_tokens = min(cached_tokens, input_tokens)
+    output_tokens = max(0, _int_or_zero(getattr(usage, "output_tokens", 0)))
+    billable_input = input_tokens - cached_tokens
+    cost = (
+        billable_input * (input_rate or 0)
+        + cached_tokens * (cached_rate if cached_rate is not None else input_rate or 0)
+        + output_tokens * (output_rate or 0)
+    ) / 1_000_000
+    return _round_cost(cost)
+

 class ProviderUsage(BaseModel):
    """Running totals for one upstream provider OpenRouter routed calls to."""
@@ -37,6 +64,7 @@ class LLMUsageLedger:
        self._total_usage = Usage()
        self._agent_usage: dict[str, Usage] = {}
        self._agent_metadata: dict[str, dict[str, str]] = {}
+        self._agent_estimated_cost: dict[str, float] = {}
        self._observed_cost = 0.0
        self._estimated_cost = 0.0
        self._has_observed_cost = False
@@ -73,6 +101,9 @@ class LLMUsageLedger:
            estimated = _estimate_litellm_cost(usage, model)
            if estimated:
                self._estimated_cost += estimated
+                self._agent_estimated_cost[normalized_agent_id] = (
+                    self._agent_estimated_cost.get(normalized_agent_id, 0.0) + estimated
+                )

        return True

@@ -114,11 +145,23 @@ class LLMUsageLedger:
    def total_cost(self) -> float:
        if self.zero_cost:
            return 0.0
+        model, prices = _configured_prices()
+        if prices is not None:
+            return _round_cost(
+                sum(self._configured_agent_cost(aid, model, prices) for aid in self._agent_usage)
+            )
        return _round_cost(self._observed_cost if self._has_observed_cost else self._estimated_cost)

+    def _configured_agent_cost(self, agent_id: str, model: str | None, prices: Prices) -> float:
+        if self._agent_metadata.get(agent_id, {}).get("model") == model:
+            return _configured_usage_cost(self._agent_usage[agent_id], prices)
+        return self._agent_estimated_cost.get(agent_id, 0.0)
+
    def to_record(self) -> dict[str, Any]:
        record = serialize_usage(self._total_usage)
-        record["cost"] = self.total_cost
+        model, prices = _configured_prices()
+        total_cost = self.total_cost
+        record["cost"] = total_cost
        record["providers"] = {name: tally.model_dump() for name, tally in self._providers.items()}
        record["agents"] = []

@@ -128,7 +171,11 @@ class LLMUsageLedger:
            usage = self._agent_usage[agent_id]
            metadata = self._agent_metadata.get(agent_id, {})
            agent_cost = (
-                self.total_cost * (agent_tokens[agent_id] / total_tokens) if total_tokens else 0.0
+                self._configured_agent_cost(agent_id, model, prices)
+                if prices is not None and not self.zero_cost
+                else total_cost * (agent_tokens[agent_id] / total_tokens)
+                if total_tokens
+                else 0.0
            )

            agent_record = serialize_usage(usage)
@@ -148,6 +195,7 @@ class LLMUsageLedger:
        self._total_usage = Usage()
        self._agent_usage.clear()
        self._agent_metadata.clear()
+        self._agent_estimated_cost.clear()
        self._observed_cost = 0.0
        self._estimated_cost = 0.0
        self._has_observed_cost = False
@@ -183,6 +231,7 @@ class LLMUsageLedger:
                logger.exception("Failed to hydrate llm_usage for agent %s", agent_id)
                self._agent_usage[agent_id] = Usage()

+            self._agent_estimated_cost[agent_id] = _float_or_zero(raw_agent.get("cost"))
            metadata: dict[str, str] = {}
            agent_name = raw_agent.get("agent_name")
            model = raw_agent.get("model")
@@ -0,0 +1,255 @@
+from __future__ import annotations
+
+import asyncio
+import contextlib
+import logging
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+from agents.sandbox.session.manager import Instrumentation
+from agents.sandbox.session.sandbox_session import SandboxSession
+from agents.sandbox.session.sandbox_session_state import SandboxSessionState
+from agents.sandbox.snapshot import NoopSnapshot
+
+
+if TYPE_CHECKING:
+    import io
+    from collections.abc import Sequence
+
+    from agents.sandbox.entries import BaseEntry
+    from agents.sandbox.manifest import Manifest
+    from agents.sandbox.materialization import MaterializedFile
+    from agents.sandbox.session.base_sandbox_session import BaseSandboxSession
+    from agents.sandbox.types import ExecResult, ExposedPortEndpoint, User
+
+
+logger = logging.getLogger(__name__)
+
+
+class _PendingSessionState(SandboxSessionState):
+    type: str = "pending"
+
+
+class DeferredSandboxSession(SandboxSession):
+    def __init__(
+        self,
+        bring_up: asyncio.Task[tuple[Any, BaseSandboxSession]],
+        *,
+        manifest: Manifest,
+        exposed_ports: tuple[int, ...] = (),
+    ) -> None:
+        self._bring_up = bring_up
+        self._pending_state: SandboxSessionState = _PendingSessionState(
+            snapshot=NoopSnapshot(id="pending"),
+            manifest=manifest,
+            exposed_ports=exposed_ports,
+        )
+        self._pending_limits: Any = None
+        self._pending_archive_limits: Any = None
+        self._resolved: BaseSandboxSession | None = None
+        self._client: Any = None
+        self._instrumentation: Instrumentation = Instrumentation()
+        self._seq = 0
+
+    @property
+    def pending(self) -> bool:
+        return self._resolved is None
+
+    @property
+    def bring_up(self) -> asyncio.Task[tuple[Any, BaseSandboxSession]]:
+        return self._bring_up
+
+    async def resolve(self) -> BaseSandboxSession:
+        if self._resolved is not None:
+            return self._resolved
+        client, session = await asyncio.shield(self._bring_up)
+        self._adopt(client, session)
+        return session
+
+    def _adopt(self, client: Any, session: BaseSandboxSession) -> None:
+        if self._resolved is not None:
+            return
+        self._client = client
+        self._resolved = session
+        self._inner = session
+        if self._pending_limits is not None:
+            session._set_concurrency_limits(self._pending_limits)
+        if self._pending_archive_limits is not None:
+            session._set_archive_limits(self._pending_archive_limits)
+        if self._pending_state.manifest != session.state.manifest:
+            session.state = session.state.model_copy(
+                update={"manifest": self._pending_state.manifest}
+            )
+
+    @property
+    def client(self) -> Any:
+        return self._client
+
+    @property
+    def state(self) -> SandboxSessionState:
+        if self._resolved is None:
+            return self._pending_state
+        return self._resolved.state
+
+    @state.setter
+    def state(self, value: SandboxSessionState) -> None:
+        if self._resolved is None:
+            self._pending_state = value
+        else:
+            self._resolved.state = value
+
+    def _set_concurrency_limits(self, limits: Any) -> None:
+        if self._resolved is None:
+            self._pending_limits = limits
+        else:
+            self._resolved._set_concurrency_limits(limits)
+
+    def _set_archive_limits(self, limits: Any) -> None:
+        if self._resolved is None:
+            self._pending_archive_limits = limits
+        else:
+            self._resolved._set_archive_limits(limits)
+
+    def _manifest_base_dir(self) -> Path:
+        if self._resolved is None:
+            return Path.cwd()
+        return self._resolved._manifest_base_dir()
+
+    def normalize_path(self, path: Path | str, *, for_write: bool = False) -> Path:
+        if self._resolved is None:
+            return Path(path)
+        return self._resolved.normalize_path(path, for_write=for_write)
+
+    def register_persist_workspace_skip_path(self, path: Path | str) -> Path:
+        if self._resolved is None:
+            return Path(path)
+        return self._resolved.register_persist_workspace_skip_path(path)
+
+    def supports_pty(self) -> bool:
+        return self._resolved.supports_pty() if self._resolved is not None else False
+
+    async def running(self) -> bool:
+        if self._resolved is None:
+            if not self._bring_up.done():
+                return True
+            return not self._bring_up.cancelled() and self._bring_up.exception() is None
+        return await self._resolved.running()
+
+    async def start(self) -> None:
+        if self._resolved is None:
+            return
+        await self._resolved.start()
+
+    async def stop(self) -> None:
+        if self._resolved is None:
+            return
+        await self._resolved.stop()
+
+    async def shutdown(self) -> None:
+        if self._resolved is None:
+            return
+        await self._resolved.shutdown()
+
+    async def aclose(self) -> None:
+        if self._resolved is None:
+            return
+        await self._resolved.aclose()
+
+    async def _validate_manifest_application(self, *, only_ephemeral: bool = False) -> None:
+        session = await self.resolve()
+        await session._validate_manifest_application(only_ephemeral=only_ephemeral)
+
+    async def apply_manifest(self, *, only_ephemeral: bool = False) -> Any:
+        session = await self.resolve()
+        return await session.apply_manifest(only_ephemeral=only_ephemeral)
+
+    async def _apply_entry_batch(
+        self,
+        entries: Sequence[tuple[Path, BaseEntry]],
+        *,
+        base_dir: Path,
+    ) -> list[MaterializedFile]:
+        del base_dir
+        session = await self.resolve()
+        return await session._apply_entry_batch(entries, base_dir=session._manifest_base_dir())
+
+    async def exec(self, *command: str | Path, **kwargs: Any) -> ExecResult:
+        await self.resolve()
+        return await super().exec(*command, **kwargs)
+
+    async def ls(self, *args: Any, **kwargs: Any) -> Any:
+        await self.resolve()
+        return await super().ls(*args, **kwargs)
+
+    async def rm(self, *args: Any, **kwargs: Any) -> Any:
+        await self.resolve()
+        return await super().rm(*args, **kwargs)
+
+    async def mkdir(self, *args: Any, **kwargs: Any) -> Any:
+        await self.resolve()
+        return await super().mkdir(*args, **kwargs)
+
+    async def read(self, path: Path, *, user: str | User | None = None) -> io.IOBase:
+        await self.resolve()
+        return await super().read(path, user=user)
+
+    async def write(self, *args: Any, **kwargs: Any) -> Any:
+        await self.resolve()
+        return await super().write(*args, **kwargs)
+
+    async def resolve_exposed_port(self, port: int) -> ExposedPortEndpoint:
+        await self.resolve()
+        return await super().resolve_exposed_port(port)
+
+    async def _resolve_exposed_port(self, port: int) -> ExposedPortEndpoint:
+        session = await self.resolve()
+        return await session._resolve_exposed_port(port)
+
+    async def pty_exec_start(self, *args: Any, **kwargs: Any) -> Any:
+        await self.resolve()
+        return await super().pty_exec_start(*args, **kwargs)
+
+    async def pty_write_stdin(self, *args: Any, **kwargs: Any) -> Any:
+        await self.resolve()
+        return await super().pty_write_stdin(*args, **kwargs)
+
+    async def pty_terminate_all(self) -> None:
+        if self._resolved is None:
+            return
+        await super().pty_terminate_all()
+
+    async def persist_workspace(self) -> io.IOBase:
+        await self.resolve()
+        return await super().persist_workspace()
+
+    async def hydrate_workspace(self, data: io.IOBase) -> None:
+        await self.resolve()
+        await super().hydrate_workspace(data)
+
+    async def teardown(self) -> None:
+        if self._resolved is None:
+            self._bring_up.cancel()
+            with contextlib.suppress(BaseException):
+                client, session = await self._bring_up
+                self._adopt(client, session)
+        if self._resolved is None:
+            return
+        await self._client.delete(self._resolved)
+
+
+class DeferredSandboxClient:
+    def __init__(self, session: DeferredSandboxSession) -> None:
+        self._session = session
+
+    def __getattr__(self, name: str) -> Any:
+        client = self._session.client
+        if client is None:
+            raise AttributeError(f"sandbox client unavailable before bring-up finishes ({name})")
+        return getattr(client, name)
+
+    async def delete(self, session: Any) -> None:
+        if isinstance(session, DeferredSandboxSession):
+            await session.teardown()
+            return
+        real = await self._session.resolve()
+        await self._session.client.delete(real if session is self._session else session)
@@ -66,6 +66,33 @@ def _apply_sandbox_network(create_kwargs: dict[str, Any]) -> None:
        create_kwargs.pop("ports", None)


+_SANDBOX_DEVICES_ENV = "STRIX_DOCKER_SANDBOX_DEVICES"
+
+
+def _sandbox_devices() -> list[str]:
+    raw = os.environ.get(_SANDBOX_DEVICES_ENV, "")
+    return [item.strip() for item in raw.split(",") if item.strip()]
+
+
+def _apply_sandbox_devices(create_kwargs: dict[str, Any]) -> None:
+    """Pass the host devices listed in ``STRIX_DOCKER_SANDBOX_DEVICES`` through.
+
+    Comma-separated device paths, each mapped to the same path in the container
+    with ``rwm`` access. A tunnel brought up inside the sandbox (OpenVPN,
+    Tailscale) needs ``/dev/net/tun``: Docker's default device cgroup denies it
+    even with ``NET_ADMIN``. Unset means no extra devices.
+    """
+    devices = _sandbox_devices()
+    if not devices:
+        return
+    existing = [str(item) for item in (create_kwargs.get("devices") or [])]
+    for path in devices:
+        spec = f"{path}:{path}:rwm"
+        if spec not in existing and path not in existing:
+            existing.append(spec)
+    create_kwargs["devices"] = existing
+
+
 def _apply_resource_limits(create_kwargs: dict[str, Any]) -> None:
    """Apply optional cgroup resource caps from the environment. Unset/blank
    values leave docker's default (unbounded), so this is opt-in per host."""
@@ -233,6 +260,7 @@ class StrixDockerSandboxClient(DockerSandboxClient):
        extra_hosts["host.docker.internal"] = "host-gateway"

        _apply_sandbox_network(create_kwargs)
+        _apply_sandbox_devices(create_kwargs)
        _apply_resource_limits(create_kwargs)
        _apply_log_limits(create_kwargs)
        _apply_run_labels(create_kwargs)
@@ -3,6 +3,8 @@
 from __future__ import annotations

 import asyncio
+import contextlib
+import inspect
 import io
 import logging
 import os
@@ -19,9 +21,12 @@ from strix.config import load_settings
 from strix.runtime.backends import backend_supports_bind_mounts, get_backend
 from strix.runtime.caido_bootstrap import bootstrap_caido
 from strix.runtime.caido_handle import CaidoBootstrapHandle
+from strix.runtime.deferred_session import DeferredSandboxClient, DeferredSandboxSession


 if TYPE_CHECKING:
+    from collections.abc import Awaitable
+
    from agents.sandbox.session import BaseSandboxSession

    from strix.runtime.status import StatusSink
@@ -255,13 +260,22 @@ def _gitdir_from_pointer(git_file: Path) -> Path | None:
    return None


+async def _resolve_extra_file_archive(
+    extra_files: list[dict[str, Any]] | Awaitable[list[dict[str, Any]] | None] | None,
+    local_sources: list[dict[str, Any]],
+) -> bytes | None:
+    files = await extra_files if inspect.isawaitable(extra_files) else extra_files
+    return build_extra_file_archive(files, local_sources) if files else None
+
+
 async def create_or_reuse(
    scan_id: str,
    *,
    image: str,
    local_sources: list[dict[str, Any]],
-    extra_files: list[dict[str, Any]] | None = None,
+    extra_files: list[dict[str, Any]] | Awaitable[list[dict[str, Any]] | None] | None = None,
    status_sink: StatusSink | None = None,
+    defer: bool = False,
 ) -> dict[str, Any]:
    """Return the existing session bundle for ``scan_id`` or create a new one.

@@ -271,7 +285,9 @@ async def create_or_reuse(
    Each ``extra_files`` entry (``{"workspace_path": "/workspace/<rel>",
    "content": bytes | str}``) lands as a regular, agent-writable file at its
    ``workspace_path`` on every backend: the files are uploaded as one archive
-    and unpacked inside the sandbox right after bring-up.
+    and unpacked inside the sandbox right after bring-up. ``extra_files`` may
+    also be an awaitable resolving to that list: provisioning starts at once and
+    the files are awaited only when the sandbox is ready to receive them.
    """

    def report(phase: str) -> None:
@@ -292,9 +308,6 @@ async def create_or_reuse(
    else:
        bind_mounts = []
        entries = build_manifest_entries(local_sources)
-    extra_file_archive = (
-        build_extra_file_archive(extra_files, local_sources) if extra_files else None
-    )

    # Caido runs as an in-container sidecar; HTTP(S) traffic from any
    # process started via ``session.exec`` (the SDK's Shell tool, etc.)
@@ -323,41 +336,69 @@ async def create_or_reuse(
        backend_name,
        image,
    )
-    report("Starting sandbox container")
-    client, session = await backend(
-        image=image,
-        manifest=manifest,
-        exposed_ports=(_CONTAINER_CAIDO_PORT,),
-        bind_mounts=bind_mounts,
-    )

-    if extra_file_archive is not None:
-        report("Placing workspace files")
+    async def _bring_up() -> tuple[Any, Any]:
+        report("Starting sandbox container")
+        client, session = await _create_or_discard(
+            backend(
+                image=image,
+                manifest=manifest,
+                exposed_ports=(_CONTAINER_CAIDO_PORT,),
+                bind_mounts=bind_mounts,
+            )
+        )
        try:
-            await stage_extra_files(session, extra_file_archive)
+            extra_file_archive = await _resolve_extra_file_archive(extra_files, local_sources)
+            if extra_file_archive is not None:
+                report("Placing workspace files")
+                await stage_extra_files(session, extra_file_archive)
        except BaseException:
            await _discard_session(client, session)
            raise
+        return client, session

-    report("Setting up the proxy")
-    caido_endpoint = await session.resolve_exposed_port(_CONTAINER_CAIDO_PORT)
-    scheme = "https" if caido_endpoint.tls else "http"
-    host_caido_url = f"{scheme}://{caido_endpoint.host}:{caido_endpoint.port}"
-    logger.debug("Caido host endpoint resolved: %s", host_caido_url)
+    async def _caido_url(session: Any) -> str:
+        report("Setting up the proxy")
+        caido_endpoint = await session.resolve_exposed_port(_CONTAINER_CAIDO_PORT)
+        scheme = "https" if caido_endpoint.tls else "http"
+        host_caido_url = f"{scheme}://{caido_endpoint.host}:{caido_endpoint.port}"
+        logger.debug("Caido host endpoint resolved: %s", host_caido_url)
+        return host_caido_url

-    # The Caido login + project setup polls the guest for a couple of seconds
-    # and nothing needs the client before the first proxy tool call, so it
-    # runs concurrently with the rest of scan start; consumers resolve the
-    # handle at first use (see CaidoBootstrapHandle).
-    caido_client = CaidoBootstrapHandle(
-        asyncio.create_task(
-            bootstrap_caido(
-                session,
-                host_url=host_caido_url,
-                container_url=container_caido_url,
-            ),
-            name=f"caido-bootstrap-{scan_id}",
+    def _caido(session: Any, host_caido_url: str) -> Any:
+        return bootstrap_caido(
+            session,
+            host_url=host_caido_url,
+            container_url=container_caido_url,
        )
+
+    if defer:
+        bring_up = asyncio.create_task(_bring_up(), name=f"sandbox-bring-up-{scan_id}")
+        deferred = DeferredSandboxSession(
+            bring_up,
+            manifest=manifest,
+            exposed_ports=(_CONTAINER_CAIDO_PORT,),
+        )
+
+        async def _caido_when_up() -> Any:
+            session = await deferred.resolve()
+            return await _caido(session, await _caido_url(session))
+
+        bundle = {
+            "client": DeferredSandboxClient(deferred),
+            "session": deferred,
+            "caido_client": CaidoBootstrapHandle(
+                asyncio.create_task(_caido_when_up(), name=f"caido-bootstrap-{scan_id}")
+            ),
+        }
+        _SESSION_CACHE[scan_id] = bundle
+        logger.info("Sandbox session for scan %s deferred and cached", scan_id)
+        return bundle
+
+    client, session = await _bring_up()
+    host_caido_url = await _caido_url(session)
+    caido_client = CaidoBootstrapHandle(
+        asyncio.create_task(_caido(session, host_caido_url), name=f"caido-bootstrap-{scan_id}")
    )

    bundle = {
@@ -370,6 +411,26 @@ async def create_or_reuse(
    return bundle


+_DISCARD_TASKS: set[asyncio.Task[None]] = set()
+
+
+async def _create_or_discard(create: Awaitable[tuple[Any, Any]]) -> tuple[Any, Any]:
+    pending = asyncio.ensure_future(create)
+    try:
+        return await asyncio.shield(pending)
+    except asyncio.CancelledError:
+        discard = asyncio.create_task(_discard_when_created(pending))
+        _DISCARD_TASKS.add(discard)
+        discard.add_done_callback(_DISCARD_TASKS.discard)
+        raise
+
+
+async def _discard_when_created(create: asyncio.Future[tuple[Any, Any]]) -> None:
+    with contextlib.suppress(BaseException):
+        client, session = await create
+        await _discard_session(client, session)
+
+
 async def _discard_session(client: Any, session: Any) -> None:
    """Best-effort teardown of a session that never made it into the cache."""
    try:
@@ -40,7 +40,7 @@ The skills are dynamically injected into the agent's system prompt, allowing it

 Notable source-aware skills:
 - `source_aware_whitebox` (coordination): white-box orchestration playbook
- `source_aware_sast` (custom): semgrep/AST/secrets/supply-chain static triage workflow
+- `source_aware_sast` (custom): opengrep/AST/secrets/supply-chain static triage workflow
 - `dependency_cve_scanning` (custom): trivy-based SCA workflow for reporting known dependency CVEs via `create_dependency_report`
 - `npx_confusion` (custom): npx/npm exec/bunx fallback and adjacent package-runner identity confusion, with runner-specific registry and reporting gates
 - `semantic_confusion` (vulnerabilities): cross-boundary parser, normalization, and representation mismatch analysis
@@ -1,11 +1,12 @@
 import logging
 import re
 from collections import Counter
-from collections.abc import Iterator
+from collections.abc import Iterator, Mapping
 from pathlib import Path
 from typing import TypeGuard

 import yaml
+from jinja2 import Environment, StrictUndefined

 from strix.utils.resource_paths import get_strix_resource_path

@@ -13,6 +14,15 @@ from strix.utils.resource_paths import get_strix_resource_path
 logger = logging.getLogger(__name__)

 _FRONTMATTER_PATTERN = re.compile(r"^---\s*\n(?P<body>.*?)\n---\s*\n", re.DOTALL)
+# Only skills with `template: jinja` in their frontmatter are rendered; the rest hold
+# literal `{{ }}` payloads.
+_SKILL_TEMPLATES = Environment(
+    autoescape=False,  # noqa: S701  # nosec B701 - prompts, not HTML
+    undefined=StrictUndefined,
+    trim_blocks=True,
+    lstrip_blocks=True,
+    keep_trailing_newline=True,
+)

 _INTERNAL_SKILL_CATEGORIES: frozenset[str] = frozenset({"scan_modes", "coordination", "analysis"})
 _ROOT_SKILL_CATEGORY = "root"
@@ -261,7 +271,22 @@ def _candidate_skill_files(skill_name: str) -> list[Path]:
    return _bare_skill_files(skill_name)


-def load_skills(skill_names: list[str]) -> dict[str, str]:
+def skill_turn_limit(skill_names: list[str], turn_limits: Mapping[str, int]) -> int | None:
+    """Largest of ``turn_limits`` (keyed by skill name) among these skills, if any."""
+    files = {path for name in skill_names for path in _candidate_skill_files(name)}
+    return max(
+        (
+            limit
+            for name, limit in turn_limits.items()
+            if files.intersection(_candidate_skill_files(name))
+        ),
+        default=None,
+    )
+
+
+def load_skills(
+    skill_names: list[str], *, supports_images: bool = True, coverage_tools: bool = True
+) -> dict[str, str]:
    """Load skill markdown bodies (frontmatter stripped) by name.

    Skill files live at ``strix/skills/<category>/<name>.md`` (or any
@@ -291,7 +316,11 @@ def load_skills(skill_names: list[str]) -> dict[str, str]:
            continue

        var_name = skill_name.split("/")[-1]
-        _, skill_body = _parse_skill_content(content, file_path)
+        metadata, skill_body = _parse_skill_content(content, file_path)
+        if metadata.get("template") == "jinja":
+            skill_body = _SKILL_TEMPLATES.from_string(skill_body).render(
+                supports_images=supports_images, coverage_tools=coverage_tools
+            )
        skill_content[var_name] = skill_body
        logger.debug("Loaded skill: %s -> %s", skill_name, var_name)
        _track_skill_loaded(var_name, file_path)
@@ -1,6 +1,7 @@
 ---
 name: counterevidence
 description: Closure discipline for security findings — what counts as proof of safety, what does not, and how to record an unresolved candidate instead of silently dropping it
+template: jinja
 ---

 # Counterevidence and Closure Discipline
@@ -20,7 +21,7 @@ state, and "I moved on" is not one of them.

 **1. `confirmed`** — you have a working PoC or, in white-box, a complete
 source → control → sink → impact trace plus evidence the path is
-reachable. File it with `create_vulnerability_report`.
+reachable. File it as a vulnerability report.

 **2. `ruled_out`** — you can name the **specific control** that makes the
 code safe, at a specific location, and you have checked that the control
@@ -32,8 +33,12 @@ that sentence, you are not in `ruled_out`.

 **3. `open_proof_gap`** — the candidate is plausible, you could not
 confirm it, and you also could not name a control that rules it out. This
+{% if coverage_tools %}
 is a legitimate, expected outcome. Record it with
 `record_coverage(outcome="needs_follow_up")`, carry it up in
+{% else %}
+is a legitimate, expected outcome. Carry it up in
+{% endif %}
 `agent_finish(open_items=[...])`, and reflect it in `counterevidence` /
 `confidence_rationale` if you file a related report. Do **not** convert
 it to `ruled_out` to tidy up your worklist.
@@ -97,6 +102,7 @@ What ships and what is reachable is what matters.
 reduces severity — it does not make the finding unreal. Downgrade it;
 do not delete it.

+{% if coverage_tools %}
 ## Recording Closure

 Closure is only useful if it is written down. Every surface you assess
@@ -125,6 +131,7 @@ named control does not cover the path you just found goes back to
 previous state is kept as history, so correcting the record costs
 nothing and leaving it wrong costs a finding.

+{% endif %}
 ## What DOES Rule Out a Candidate

 - You executed the attack and it demonstrably failed, and you understand
@@ -144,8 +151,7 @@ endpoint is broken/unreachable for unrelated reasons".

 ## Before You File a Report

-Run this pass on every finding before calling
-`create_vulnerability_report`:
+Run this pass on every finding before filing it:

 1. **Argue the other side.** Spend real effort building the strongest
   case that this is *not* exploitable, or not as severe as you think.
@@ -1,6 +1,7 @@
 ---
 name: source_aware_discovery
 description: Enumeration discipline for reading code — which locations to keep as separate candidates, which safe siblings prove nothing, and the per-family sweeps that are routinely missed
+template: jinja
 ---

 # Source-Aware Discovery
@@ -206,6 +207,10 @@ Not worth it: "this could use more validation" with no path, style and
 maintainability complaints, and cosmetic variants of a candidate you
 already opened.

+{% if coverage_tools %}
 Keep reading until no distinct plausible candidate remains — then record
 what you swept with `record_coverage`, including the families that came
 back clean.
+{% else %}
+Keep reading until no distinct plausible candidate remains.
+{% endif %}
@@ -1,6 +1,7 @@
 ---
 name: root-agent
 description: Orchestration layer that coordinates specialized subagents for security assessments
+template: jinja
 ---

 # Root Agent
@@ -35,10 +36,12 @@ Every scan needs one shared answer to "who is the attacker here, and what are th

 Either way you write it with the least information anyone on this scan will ever have, so expect it to be wrong somewhere. Subagents correct it with `amend_threat_model`, which appends an attributed addendum instead of overwriting — expect many of these on a black-box run, as authenticating, pivoting between roles, and reaching internal surfaces is exactly what turns inference into fact. Read the amendments back before you write the final report: an agent telling you a boundary you called trusted is attacker-reachable is a finding about your model, not a note. Only call `save_threat_model` again to fold accumulated amendments into the body; it replaces the document and clears them.

+{% if coverage_tools %}
 ## Reconcile Coverage Before Finishing

 Coverage entries are shared and mutable. Before `finish_scan`, list the `needs_follow_up` rows: each one is either work you still owe or a row somebody already resolved without updating. Assign the former to a subagent and have it call `update_coverage` on the existing entry rather than recording a second one — a stale open item sitting next to its own resolution is worse than either alone.

+{% endif %}
 ## Agent Architecture

 Structure agents by function:
@@ -55,6 +58,13 @@ Structure agents by function:
 - Business logic flaws
 - Infrastructure vulnerabilities

+**Sensitive Browser Confirmations**
+- For every browser workflow containing a security-sensitive or consequential user action—whether authenticated or unauthenticated—assign `browser_security` to the agent testing that workflow or create a dedicated UI-redressing agent. This includes actions that change protected state, grant authority, submit credentials, expose sensitive data, or cross a trust boundary.
+- Its task must cover both iframe and cross-window/DoubleClickjacking paths; framing protections close only iframe coverage.
+- A workflow crossing origins or ports remains eligible when every involved origin is an explicit scan target. Same-origin policy blocks DOM reads, not top-level navigation, popups, opener relationships, focus changes, or window closing; cross-origin separation alone does not close UI-redressing coverage.
+- Never expand authorization from a link alone. If a required origin is not an explicit target, leave the path `needs_follow_up` and do not claim clickjacking or the related workflow is cleared; scope expansion must come from scan configuration.
+- Before accepting `no_issue_found` or `ruled_out` for cross-window UI redressing, reject closure evidence that cites popup-as-tab behavior, ignored geometry, unavailable OS-level input, or only a manual control click; keep the exact two-click path `needs_follow_up`. A successfully executed protected consent, approval, or grant must be impact-calibrated, not dismissed because it is idempotent, pre-seeded, or issues no immediate token or data.
+
 **Exploitation and Validation**
 - Proof-of-concept development
 - Impact demonstration
@@ -14,22 +14,22 @@ Increase white-box coverage by combining source-aware triage with dynamic valida
 ## Recommended Workflow

 1. Build a quick source map before deep exploitation, including at least one AST-structural pass (`sg` or `tree-sitter`) scoped to relevant paths.
-   - For `sg` baseline, derive `sg-targets.txt` from `semgrep.json` scope first (`paths.scanned`, fallback to unique `results[].path`) and run `xargs ... sg run` on that list.
-   - Only fall back to path heuristics when semgrep scope is unavailable.
+   - For `sg` baseline, derive `sg-targets.txt` from `opengrep.json` scope first (`paths.scanned`, fallback to unique `results[].path`) and run `xargs ... sg run` on that list.
+   - Only fall back to path heuristics when opengrep scope is unavailable.
 2. Run first-pass static triage to rank high-risk paths.
 3. Use triage outputs to prioritize dynamic PoC validation.
 4. Keep findings evidence-driven: no report without validation.

 ## Source-Aware Triage Stack

- `semgrep`: fast security-first triage and custom pattern scans
+- `opengrep`: fast security-first triage and custom pattern scans
 - `ast-grep` (`sg`): structural pattern hunting and targeted repo mapping
 - `tree-sitter`: syntax-aware parsing support for symbol and route extraction
 - `gitleaks` + `trufflehog`: complementary secret detection (working tree and history coverage)
 - `trivy fs`: dependency, misconfiguration, license, and secret checks

 Coverage target per repository:
- one `semgrep` pass
+- one `opengrep` pass
 - one AST structural pass (`sg` and/or `tree-sitter`)
 - one secrets pass (`gitleaks` and/or `trufflehog`)
 - one `trivy fs` pass
@@ -115,7 +115,8 @@ not just "upgrade <vulnerable pkg> to <fixed>".
 ### Usage / reachability analysis (required for every dependency CVE)

 For every CVE you are about to report, run a static usage analysis and record
-the result in the structured `reachability` + `reachability_evidence` fields.
+the result in the structured `reachability` + `reachability_evidence` fields,
+and report every usage site you find as a `code_locations` entry (see Reporting).
 The level is an **evidence ladder, never an exploitability verdict** — claim
 only what you proved, and cite the proof. It never changes severity (that is
 `advisory_cvss` alone); it exists so the reader can prioritize.
@@ -220,8 +221,18 @@ findings and rejects empty PoC fields):
  GHSA, look up the mapped CVE; if there is genuinely no CVE, do not report it
  with this tool.
 - There are no PoC fields — `create_dependency_report` does not take
-  `poc_description` / `poc_script_code` / `code_locations`. The proof lives in
-  `description` and `technical_analysis` (scanner output + advisory).
+  `poc_description` / `poc_script_code`. The proof lives in `description` and
+  `technical_analysis` (scanner output + advisory).
+- **You MUST report usage sites in `code_locations`.** Add one entry for each
+  import of the package and each call of an affected symbol that your usage
+  analysis found. Every `file:line` you cite in `reachability_evidence` must
+  also be a `code_locations` entry. Also add the line in `manifest_path` that
+  pins the vulnerable version. Each entry needs the repo-relative `file`, the
+  exact `start_line` / `end_line`, the verbatim `snippet` at those lines, and a
+  short `label` (e.g. "imports js-yaml", "calls yaml.load on CLI input"). Read
+  the real file to confirm the lines; never guess. When the package is
+  `not_imported`, the manifest line is the only entry. If an entry carries
+  `fix_before` / `fix_after`, also fill `fix_verification`.
 - **Always fill the structured dependency fields** (they power the dedicated
  dependency-report card; do not leave them only in free-text):
  - `package_name` — `PkgName` (required).
@@ -1,6 +1,6 @@
 ---
 name: source-aware-sast
-description: Practical source-aware SAST and AST playbook for semgrep, ast-grep, gitleaks, and trivy fs
+description: Practical source-aware SAST and AST playbook for opengrep, ast-grep, gitleaks, and trivy fs
 ---

 # Source-Aware SAST Playbook
@@ -23,19 +23,19 @@ Run this baseline once per repository before deep narrowing:
 ART=/workspace/.source-aware
 mkdir -p "$ART"

-semgrep scan --config p/default --config p/golang --config p/secrets \
-  --metrics=off --json --output "$ART/semgrep.json" .
-# Build deterministic AST targets from semgrep scope (no hardcoded path guessing)
+opengrep scan --config /opt/opengrep-rules --taint-intrafile \
+  --json --output "$ART/opengrep.json" .
+# Build deterministic AST targets from opengrep scope (no hardcoded path guessing)
 python3 - <<'PY'
 import json
 from pathlib import Path

 art = Path("/workspace/.source-aware")
-semgrep_json = art / "semgrep.json"
+opengrep_json = art / "opengrep.json"
 targets_file = art / "sg-targets.txt"

 try:
-    data = json.loads(semgrep_json.read_text(encoding="utf-8"))
+    data = json.loads(opengrep_json.read_text(encoding="utf-8"))
 except Exception:
    targets_file.write_text("", encoding="utf-8")
    raise
@@ -63,17 +63,18 @@ trivy fs --scanners vuln,misconfig --timeout 30m --offline-scan \
  --format json --output "$ART/trivy-fs.json" . || true
 ```

-## Semgrep First Pass
+## Opengrep First Pass

-Use Semgrep as the default static triage pass:
+Use Opengrep as the default static triage pass. Rules are bundled offline in `/opt/opengrep-rules` (one directory per language, plus `generic/secrets`):

 ```bash
-# Preferred deterministic profile set (works with --metrics=off)
-semgrep scan --config p/default --config p/golang --config p/secrets \
-  --metrics=off --json --output /workspace/.source-aware/semgrep.json .
+# Full bundled ruleset
+opengrep scan --config /opt/opengrep-rules --taint-intrafile \
+  --json --output /workspace/.source-aware/opengrep.json .

-# If you choose auto config, do not combine it with --metrics=off
-semgrep scan --config auto --json --output /workspace/.source-aware/semgrep-auto.json .
+# Narrow to the target stack when the full set is slow or noisy
+opengrep scan --config /opt/opengrep-rules/python --config /opt/opengrep-rules/generic/secrets \
+  --taint-intrafile --json --output /workspace/.source-aware/opengrep-python.json .
 ```

 If diff scope is active, restrict to changed files first, then expand only when needed.
@@ -205,10 +205,10 @@ Jinja2 backend without autoescape: `{{7*7}}`, RCE gadgets if sandbox misconfigur

 ## Tooling

-Static analysis is the fastest way to reach the sinks above in white-box scope. The sandbox ships `python`/`pipx`, `semgrep`, `bandit`, `ast-grep`, and `ripgrep`.
+Static analysis is the fastest way to reach the sinks above in white-box scope. The sandbox ships `python`/`pipx`, `opengrep`, `bandit`, `ast-grep`, and `ripgrep`.

 - **bandit** (preinstalled) — Python security linter; flags `mark_safe`, `extra()`, `RawSQL`, `subprocess`, weak crypto, hardcoded secrets: `bandit -r . -ll`
- **semgrep** (preinstalled) with the Django ruleset — higher-signal than bandit for framework-specific bugs (`.extra()`, `RawSQL`, `|safe`, `csrf_exempt`, `ALLOWED_HOSTS=['*']`): `semgrep --config p/django .`
+- **opengrep** (preinstalled) with the bundled Django rules — higher-signal than bandit for framework-specific bugs (`.extra()`, `RawSQL`, `|safe`, `csrf_exempt`, `ALLOWED_HOSTS=['*']`): `opengrep scan --config /opt/opengrep-rules/python/django .`
 - **pip-audit** (PyPA) — dependency CVE scanner for known-vuln Django/DRF/simplejwt versions: `pipx install pip-audit && pip-audit -r requirements.txt`
 - **ast-grep** (preinstalled) — quick structural grep for risky calls without a full SAST run: `ast-grep run -p 'mark_safe($X)' -l python`

@@ -15,7 +15,7 @@ Thorough understanding before exploitation. Test every parameter, every endpoint

 **Whitebox (source available)**
 - Map every file, module, and code path in the repository
- Start with broad source-aware triage (`semgrep`, `ast-grep`, `gitleaks`, `trufflehog`, `trivy fs`) and use outputs to drive deep review
+- Start with broad source-aware triage (`opengrep`, `ast-grep`, `gitleaks`, `trufflehog`, `trivy fs`) and use outputs to drive deep review
 - Execute at least one structural AST pass (`sg` and/or Tree-sitter) per repository and store artifacts for reuse
 - Keep AST artifacts bounded and query-driven (target relevant paths/sinks first; avoid whole-repo generic function dumps)
 - Use syntax-aware parsing (Tree-sitter tooling) to improve symbol, route, and sink extraction quality
@@ -1,6 +1,7 @@
 ---
 name: diff
 description: Methodology for diff-scoped review of a pull request, commit, or branch — what counts as in scope, how far to follow a change, and what not to report
+template: jinja
 ---

 # Diff-Scoped Review
@@ -69,9 +70,14 @@ When it is not, the closure rules apply unchanged: a complete
 source → control → sink → impact trace through the changed code is
 reportable at reduced confidence, with the missing runtime proof named in
 `confidence_rationale`. A candidate you can neither confirm nor rule out
+{% if coverage_tools %}
 with a named control is an `open_proof_gap` — record it as
 `needs_follow_up` coverage rather than dropping it because the
 environment was inconvenient.
+{% else %}
+with a named control is an `open_proof_gap` — carry it up as an open
+item rather than dropping it because the environment was inconvenient.
+{% endif %}

 ## Reporting

@@ -80,7 +86,9 @@ plainly which part of the diff introduced or exposed it. A reviewer
 reading your report next to the diff should be able to see the connection
 without re-deriving your analysis.

+{% if coverage_tools %}
 Record coverage per changed component, not per changed file — a
 formatting-only file and a rewritten auth module are not equal rows.
+{% endif %}
 State which changed areas you reviewed and cleared, so the author knows
 what a clean result actually covered.
@@ -15,7 +15,7 @@ Optimize for fast feedback on critical security issues. Skip exhaustive enumerat

 **Whitebox (source available)**
 - Focus on recent changes: git diffs, new commits, modified files—these are most likely to contain fresh bugs
- Run a fast static triage on changed files first (`semgrep`, then targeted `sg` queries)
+- Run a fast static triage on changed files first (`opengrep`, then targeted `sg` queries)
 - Run at least one lightweight AST pass (`sg` or Tree-sitter) so structural mapping is not skipped
 - Keep AST commands tightly scoped to changed or high-risk paths; avoid broad repository-wide pattern dumps
 - Run quick secret and dependency checks (`gitleaks`, `trufflehog`, `trivy fs`) scoped to changed areas when possible
@@ -15,7 +15,7 @@ Systematic testing across the full attack surface. Understand the application be

 **Whitebox (source available)**
 - Map codebase structure: modules, entry points, routing
- Run `semgrep` first-pass triage to prioritize risky flows before deep manual review
+- Run `opengrep` first-pass triage to prioritize risky flows before deep manual review
 - Run at least one AST-structural mapping pass (`sg` and/or Tree-sitter), then use outputs for route, sink, and trust-boundary mapping
 - Keep AST output bounded to relevant paths and hypotheses; avoid whole-repo generic function dumps
 - Identify architecture pattern (MVC, microservices, monolith)
@@ -1,6 +1,7 @@
 ---
 name: agent_browser
 description: agent-browser CLI for headless Chrome via shell. Snapshot-and-ref workflow, click/fill/extract, screenshots, multi-tab, multi-session, network mocking. Pre-installed in the sandbox; invoke via exec_command.
+template: jinja
 ---


@@ -260,8 +261,9 @@ only for simple expressions.

 ### Screenshot

-`agent-browser screenshot` writes a PNG to disk in the sandbox. The
-shell command alone does **not** put the image into your context —
+`agent-browser screenshot` writes a PNG to disk in the sandbox.
+{% if supports_images %}
+The shell command alone does **not** put the image into your context —
 chain it with the SDK ``view_image`` tool to actually see it:

 ```bash
@@ -275,6 +277,12 @@ the full path on stdout — pass that to ``view_image``). If you need a
 specific filename, keep it inside that directory or a sibling hidden
 dir under ``/workspace``. Never write screenshots to ``/tmp`` —
 ``view_image`` rejects anything outside the workspace root.
+{% else %}
+You are running on a text-only model and cannot view images. Screenshots
+are only useful to you as a saved file (for example, login evidence).
+Drive the page from `snapshot -i` refs, `eval` for any DOM/JS state you
+need to read, and `text @ref` / `get text` for content extraction.
+{% endif %}

 ```bash
 agent-browser screenshot                        # path printed on stdout
@@ -283,6 +291,7 @@ agent-browser screenshot --full                 # full scroll height
 agent-browser screenshot --annotate             # numbered labels + legend keyed to snapshot refs
 ```

+{% if supports_images %}
 `--annotate` is designed for multimodal models: each label `[N]` maps
 to ref `@eN`. Take the annotated screenshot, then ``view_image`` it,
 and you can correlate visual layout with snapshot refs.
@@ -294,10 +303,10 @@ layout questions, captchas, custom widgets where the accessibility
 tree is incomplete).

 If ``view_image`` errors back at you (rejected image, "vision not
-supported", or similar), you are running on a text-only model — stop
-calling it and stop taking screenshots. Drive the page entirely from
-`snapshot -i` refs, `eval` for any DOM/JS state you need to read, and
-`text @ref` / `get text` for content extraction.
+supported", or similar), you can drive the page from `snapshot -i`
+refs, `eval` for any DOM/JS state you need to read, and `text @ref`
+or `get text` for content extraction.
+{% endif %}

 ### Handle multiple pages via tabs

@@ -0,0 +1,69 @@
+---
+name: opengrep
+description: Exact Opengrep CLI structure, bundled offline rules, scoped rule selection, taint mode, and automation-safe output patterns.
+---
+
+# Opengrep CLI Playbook
+
+Official docs:
+- https://github.com/opengrep/opengrep
+- https://github.com/opengrep/opengrep-rules
+
+Canonical syntax:
+`opengrep scan [flags] [targets]`
+
+Bundled rules:
+- `/opt/opengrep-rules` holds the full ruleset, one directory per language (`python`, `javascript`, `typescript`, `go`, `java`, `php`, `ruby`, …) plus `generic/secrets`, `dockerfile`, `terraform`, `yaml`, and `strix` (Strix-authored rules, e.g. GitHub-owned actions not pinned to a commit SHA)
+- Framework rules sit under the language (`python/django`, `python/flask`, `javascript/express`, …)
+
+High-signal flags:
+- `--config <path>` rule file or directory; repeat for several
+- `--taint-intrafile` intra-file inter-procedural taint (follows data across functions in one file)
+- `--json` / `--sarif` output format
+- `--output <file>` write findings to file
+- `--severity <level>` filter by severity (`INFO`, `WARNING`, `ERROR`)
+- `--error` return non-zero exit when findings exist
+- `--quiet` suppress progress noise
+- `--jobs <n>` parallel workers
+- `--timeout <seconds>` per-rule, per-file timeout
+- `--exclude <pattern>` / `--include <pattern>` path filters
+- `--exclude-rule <rule_id>` suppress specific rule
+- `--baseline-commit <sha>` only report findings introduced after baseline
+
+Agent-safe baseline for automation:
+`opengrep scan --config /opt/opengrep-rules --taint-intrafile --json --output opengrep.json --quiet --jobs 4 --timeout 20 /workspace`
+
+Common patterns:
+- Full bundled scan:
+  `opengrep scan --config /opt/opengrep-rules --json --output opengrep.json --quiet /workspace`
+- High-severity focused pass:
+  `opengrep scan --config /opt/opengrep-rules --severity ERROR --json --output opengrep_high.json --quiet /workspace`
+- SARIF output:
+  `opengrep scan --config /opt/opengrep-rules --sarif --output opengrep.sarif --quiet /workspace`
+- Language- or framework-specific rules:
+  `opengrep scan --config /opt/opengrep-rules/python --config /opt/opengrep-rules/generic/secrets --json --output opengrep_python.json --quiet /workspace`
+- Scoped directory scan:
+  `opengrep scan --config /opt/opengrep-rules --json --output opengrep_api.json --quiet /workspace/services/api`
+- Custom rule:
+  `opengrep scan --config ./rules/my-rule.yaml --json --output opengrep_custom.json --quiet /workspace`
+
+Critical correctness rules:
+- Always provide an explicit `--config` from `/opt/opengrep-rules` or a local rule file. Registry shorthands (`p/...`, `auto`) fetch rules from the network; do not use them.
+- Opengrep has no telemetry; there is no `--metrics` flag, and passing one fails the scan.
+- Prefer `--json --output <file>` or `--sarif --output <file>` for machine-readable downstream processing. JSON keeps `paths.scanned` and `results[].path`.
+- Keep the target path explicit; use an absolute or clearly scoped workspace path instead of `.` when possible.
+
+Usage rules:
+- Start with the full `/opt/opengrep-rules` set unless the task clearly calls for a narrower language or framework directory.
+- Add `--taint-intrafile` for injection-class hunting (SQLi, command injection, SSRF, path traversal) in supported languages.
+- Use `--quiet` in automation to reduce noisy logs.
+- Use `--jobs` and `--timeout` explicitly for reproducible runtime behavior.
+- `PartialParsing` errors on templates are normal; they do not invalidate the run.
+
+Failure recovery:
+- If scans are too slow, narrow the target path and the rule directories before changing engine settings.
+- If scans time out, increase `--timeout` modestly or lower `--jobs`.
+- If output is too broad, scope `--config`, add `--severity`, or exclude known irrelevant paths.
+
+If uncertain, query web_search with:
+`site:github.com/opengrep/opengrep <flag>`
@@ -1,72 +0,0 @@
---
-name: semgrep
-description: Exact Semgrep CLI structure, metrics-off scanning, scoped ruleset selection, and automation-safe output patterns.
---
-
-# Semgrep CLI Playbook
-
-Official docs:
- https://semgrep.dev/docs/cli-reference
- https://semgrep.dev/docs/getting-started/cli
- https://semgrep.dev/docs/semgrep-code/semgrep-pro-engine-intro
-
-Canonical syntax:
-`semgrep scan [flags]`
-
-High-signal flags:
- `--config <rule_or_ruleset>` ruleset, registry pack, local rule file, or directory
- `--metrics=off` disable telemetry and metrics reporting
- `--json` JSON output
- `--sarif` SARIF output
- `--output <file>` write findings to file
- `--severity <level>` filter by severity
- `--error` return non-zero exit when findings exist
- `--quiet` suppress progress noise
- `--jobs <n>` parallel workers
- `--timeout <seconds>` per-file timeout
- `--exclude <pattern>` exclude path pattern
- `--include <pattern>` include path pattern
- `--exclude-rule <rule_id>` suppress specific rule
- `--baseline-commit <sha>` only report findings introduced after baseline
- `--pro` enable Pro engine if available
- `--oss-only` force OSS engine only
-
-Agent-safe baseline for automation:
-`semgrep scan --config p/default --metrics=off --json --output semgrep.json --quiet --jobs 4 --timeout 20 /workspace`
-
-Common patterns:
- Default security scan:
-  `semgrep scan --config p/default --metrics=off --json --output semgrep.json --quiet /workspace`
- High-severity focused pass:
-  `semgrep scan --config p/default --severity ERROR --metrics=off --json --output semgrep_high.json --quiet /workspace`
- OWASP-oriented scan:
-  `semgrep scan --config p/owasp-top-ten --metrics=off --sarif --output semgrep.sarif --quiet /workspace`
- Language- or framework-specific rules:
-  `semgrep scan --config p/python --config p/secrets --metrics=off --json --output semgrep_python.json --quiet /workspace`
- Scoped directory scan:
-  `semgrep scan --config p/default --metrics=off --json --output semgrep_api.json --quiet /workspace/services/api`
- Pro engine check or run:
-  `semgrep scan --config p/default --pro --metrics=off --json --output semgrep_pro.json --quiet /workspace`
-
-Critical correctness rules:
- Always include `--metrics=off`; Semgrep sends telemetry by default.
- Always provide an explicit `--config`; do not rely on vague or implied defaults.
- Prefer `--json --output <file>` or `--sarif --output <file>` for machine-readable downstream processing.
- Keep the target path explicit; use an absolute or clearly scoped workspace path instead of `.` when possible.
- If Pro availability matters, check it explicitly with a bounded command before assuming cross-file analysis exists.
-
-Usage rules:
- Start with `p/default` unless the task clearly calls for a narrower pack.
- Add focused packs such as `p/secrets`, `p/python`, or `p/javascript` only when they match the target stack.
- Use `--quiet` in automation to reduce noisy logs.
- Use `--jobs` and `--timeout` explicitly for reproducible runtime behavior.
- Do not use `-h`/`--help` for routine operation unless absolutely necessary.
-
-Failure recovery:
- If scans are too slow, narrow the target path and reduce the active rulesets before changing engine settings.
- If scans time out, increase `--timeout` modestly or lower `--jobs`.
- If output is too broad, scope `--config`, add `--severity`, or exclude known irrelevant paths.
- If Pro mode fails, rerun with `--oss-only` or without `--pro` and note the loss of cross-file coverage.
-
-If uncertain, query web_search with:
-`site:semgrep.dev semgrep <flag> cli`
@@ -1,6 +1,6 @@
 ---
 name: browser-security
-description: Browser-internals security testing for browsing-context relationships, postMessage, client-side path traversal, XS-Leaks, service workers, Web Workers, navigation behavior, CSP interactions, caches, and cross-origin state machines
+description: Browser-internals security testing for clickjacking, UI redressing, DoubleClickjacking, browsing-context relationships, postMessage, client-side path traversal, XS-Leaks, service workers, Web Workers, navigation behavior, CSP interactions, caches, and cross-origin state machines
 ---

 # Browser Security
@@ -43,6 +43,15 @@ Draw the context graph. Security checks on `event.origin`, `event.source`, or a
 - Check whether a blocked intermediate frame leaves a useful browsing-context relationship intact.
 - Use random per-flow names or `_blank` with `noopener` where an opener relationship is unnecessary.

+### Clickjacking and UI Redressing
+
+- **UI-redressing closure gate:** Never mark clickjacking as ruled out from `X-Frame-Options` or CSP `frame-ancestors` alone. Before closure, test both frame embedding and cross-window/popup retargeting.
+- Do not treat a different origin or port as protection against UI redressing. Same-origin policy blocks DOM access, not top-level navigation, `window.open()`, opener relationships, focus changes, or window closing. Test only explicitly authorized targets; if a required origin is outside scope, keep the path `needs_follow_up` and do not claim the vulnerability class is cleared.
+- For framed cases, prove the overlay and sensitive click. For DoubleClickjacking, use two real browser clicks and trace click 1 → window/focus/navigation/positioning change → click 2 → sensitive control. Never substitute scripted `.click()` calls.
+- Reproduce in the victim state required by the workflow—authenticated when necessary—with realistic user activation and stable geometry; capture window bounds and timing.
+- **Evidence-limited closure:** Verify a repeatable sensitive server-side state change with a recording or deterministic event trace and a negative control. If the exact sequence cannot be tested, record `needs_follow_up`, not `ruled_out` or `no_issue_found`. Popup-as-tab behavior, ignored geometry, unavailable OS-level input, or a manual control click are not counterevidence. Close the path only when a named mitigation blocks it, or when geometry is honored and both physical clicks are delivered without a sensitive action.
+- **Impact floor:** Successful execution of a protected consent, approval, or grant is an integrity impact even when idempotent, pre-seeded, or followed by no immediate token or data access. Use downstream capability to calibrate severity, not erase the demonstrated protected action.
+
 ### Client-Side Path Traversal

 Trace the complete source-to-request pipeline:
@@ -146,7 +155,7 @@ Also inspect DevTools network initiators, service workers, storage, CSP violatio

 ### Source Review

- Search for `postMessage`, message listeners, `window.open`, named targets, opener/parent access, frame creation, and sandbox attributes.
+- Search for `postMessage`, message listeners, `window.open`, named targets, opener/parent access, frame creation, `dblclick`, `mousedown`/`mouseup`, focus/blur, move/resize, navigation handlers, and sandbox attributes.
 - Search for router parameter APIs flowing into `fetch`, Axios, navigation, or HTML rendering.
 - Search for service-worker registration, Cache API writes, worker constructors, Blob URLs, and dynamic imports.
 - Search for raw HTML sinks and trust escape hatches in every supported frontend framework.
@@ -161,7 +170,7 @@ Also inspect DevTools network initiators, service workers, storage, CSP violatio
 5. **Prove the primitive** - Data transfer, path change, state oracle, cache modification, or context capture.
 6. **Escalate deliberately** - Chain to a privileged action, sensitive disclosure, SSRF, or executable DOM sink.
 7. **Cross-browser check** - At minimum record Chromium/Firefox/Safari applicability when the primitive is browser-specific.
-8. **State interaction requirements** - Click, drag, popup permission, timing window, login state, and visual deception.
+8. **State interaction requirements** - Record each physical click or drag, popup permission state, active context, focus, geometry, navigation, timing window, login state, and visual deception.

 ## Validation

@@ -171,7 +180,7 @@ Also inspect DevTools network initiators, service workers, storage, CSP violatio
 4. For XS-Leaks, repeat randomized success/failure trials and quantify separation and noise.
 5. For workers/caches, show which later context consumes the modified data.
 6. For client-side traversal, capture the final network request and the security-relevant response/action.
-7. For interaction-dependent chains, provide a screen recording or deterministic event trace.
+7. For interaction-dependent chains, show both click targets, intermediate context changes, and the resulting sensitive state change in a recording or deterministic event trace.

 ## False Positives

@@ -1,6 +1,6 @@
 ---
 name: xss
-description: XSS testing covering reflected, stored, and DOM-based vectors with CSP bypass techniques
+description: XSS and HTML injection testing covering reflected, stored, DOM-based, and normalization-dependent vectors
 ---

 # XSS
@@ -11,6 +11,7 @@ Cross-site scripting persists because context, parser, and framework edges are c

 **Types**
 - Reflected, stored, and DOM-based XSS across web/mobile/desktop shells
+- Scriptless HTML injection and second-order render paths

 **Contexts**
 - HTML, attribute, URL, JS, CSS, SVG/MathML, Markdown, PDF
@@ -20,6 +21,7 @@ Cross-site scripting persists because context, parser, and framework edges are c

 **Defenses to Bypass**
 - CSP/Trusted Types, DOMPurify, framework auto-escaping
+- Decode, normalization, sanitization, and encoding order

 ## Injection Points

@@ -57,6 +59,24 @@ Cross-site scripting persists because context, parser, and framework edges are c

 Distinguish HTML-string input from live DOM objects sanitized with `IN_PLACE`. Resolve the deployed DOMPurify build and check relevant upstream advisories; a controlled observable `nodeName` is one live-node bypass mechanism. Trace same-origin foreign nodes or `adoptNode()` into this mode before selecting the bypass ([example advisory](https://github.com/cure53/DOMPurify/security/advisories/GHSA-x4vx-rjvf-j5p4)).

+### Scriptless HTML Injection
+
+Do not require JavaScript execution before you report HTML injection. Prove that the browser creates attacker-controlled markup and a concrete security effect.
+
+Test base or form target changes, external requests, deceptive content, DOM clobbering through `id` or `name`, and dangling markup. Report XSS only when you prove a script-capable execution path. Do not dismiss HTML injection only because CSP blocks scripts.
+
+### Unicode Normalization
+
+Test normalization only when the target performs it. NFKC can fold fullwidth delimiters such as `＜` and `＞` into HTML syntax.
+
+Map validation, normalization, sanitization, encoding, and parsing in order. Start with the harmless canary `＜h1 id="strix-nfkc"＞marker＜/h1＞`.
+
+Report a bypass only when normalization creates dangerous syntax that reaches an unprotected sink. Load `semantic_confusion` when components interpret the value differently.
+
+### Second-Order Rendering
+
+Follow stored values into previews, administrative pages, email, exports, and hydration. Record which stage transforms and parses the value.
+
 ### DOM XSS

 **Sources**
@@ -166,7 +186,7 @@ Keep a compact set tuned per context:
 ## Testing Methodology

 1. **Identify sources** - URL/query/hash/referrer, postMessage, storage, WebSocket, server JSON
-2. **Trace to sinks** - Map data flow from source to sink
+2. **Trace transformations and sinks** - Record decoding, normalization, sanitization, reparsing, and the final sink
 3. **Classify context** - HTML node, attribute, URL, script block, event handler, JS eval-like, CSS, SVG
 4. **Assess defenses** - Output encoding, sanitizer, CSP, Trusted Types, DOMPurify config
 5. **Craft payloads** - Minimal payloads per context with encoding/whitespace/casing variants
@@ -175,9 +195,11 @@ Keep a compact set tuned per context:
 ## Validation

 1. Provide minimal payload and context (sink type) with before/after DOM or network evidence
-2. Demonstrate cross-browser execution where relevant or explain parser-specific behavior
-3. Show bypass of stated defenses (sanitizer settings, CSP/Trusted Types) with proof
-4. Quantify impact beyond alert: data accessed, action performed, persistence achieved
+2. Record the original and transformed value when decoding or normalization affects the result
+3. Demonstrate cross-browser execution where relevant or explain parser-specific behavior
+4. Show bypass of stated defenses (sanitizer settings, CSP/Trusted Types) with proof
+5. Separate HTML injection from XSS and state the proven capability
+6. Quantify impact beyond alert: data accessed, action performed, persistence achieved

 ## False Positives

@@ -185,6 +207,8 @@ Keep a compact set tuned per context:
 - CSP with nonces/hashes and no inline/event handlers
 - Trusted Types enforced on sinks; DOMPurify in strict mode with URI allowlists
 - Scriptable contexts disabled (no HTML pass-through, safe URL schemes enforced)
+- The target does not perform the claimed normalization or decode
+- A normalized value is safely encoded before the final sink

 ## Impact

@@ -207,4 +231,4 @@ Keep a compact set tuned per context:

 ## Summary

-Context + sink decide execution. Encode for the exact context, verify at runtime with CSP/Trusted Types, and validate every alternative render path. Small payloads with strong evidence beat payload catalogs.
+Context, transformation order, and sink decide whether input becomes text, markup, or executable code. Prove HTML injection separately from XSS, and validate every render path. Encode for the exact context and verify at runtime with CSP/Trusted Types. Small payloads with strong evidence beat payload catalogs.
@@ -3,6 +3,7 @@
 from __future__ import annotations

 import contextlib
+import functools
 import logging
 import os
 import sys
@@ -78,6 +79,16 @@ class _StdoutQuietFilter(logging.Filter):
        )


+class _PtyThresholdFilter(logging.Filter):
+    def filter(self, record: logging.LogRecord) -> bool:
+        return "PTY process count reached warning threshold" not in record.getMessage()
+
+
+@functools.cache  # install once; tests reset via ``cache_clear()``
+def _silence_pty_threshold_warning() -> None:
+    logging.getLogger("agents.sandbox.sandboxes.docker").addFilter(_PtyThresholdFilter())
+
+
 def configure_dependency_logging() -> None:
    """Quiet dependency logging/warnings that obscure Strix scan logs."""
    litellm = sys.modules.get("litellm")
@@ -89,6 +100,7 @@ def configure_dependency_logging() -> None:
    logging.getLogger("asyncio").propagate = False
    warnings.filterwarnings("ignore", category=RuntimeWarning, module="asyncio")
    _route_unraisable_to_log()
+    _silence_pty_threshold_warning()


 def _route_unraisable_to_log() -> None:
@@ -535,9 +535,8 @@ async def create_agent(
            success looks like, any constraints. Name the target the
            child should call ``get_threat_model`` on, and any shared
            state it should build on rather than rediscover — what
-            recon already mapped, which surfaces are already covered,
-            which coverage entry it is picking up. A child that is not
-            told what is already known repeats it.
+            recon already mapped, which surfaces are already covered.
+            A child that is not told what is already known repeats it.
        inherit_context: Default ``True``. The child receives the
            parent's input history as background; only set ``False``
            when starting a clean-slate task.
@@ -637,9 +636,8 @@ async def agent_finish(
    doing: what did you test, what did you find/confirm/rule out,
    what's still open.

-    **Close out honestly.** Before calling this, every surface you
-    assessed should have a ``record_coverage`` entry, and anything you
-    could neither confirm nor rule out belongs in ``open_items`` — an
+    **Close out honestly.** Before calling this, anything you could
+    neither confirm nor rule out belongs in ``open_items`` — an
    unresolved candidate handed up to the parent is useful, a silently
    dropped one is a missed vulnerability. Reporting nothing and
    listing no open items asserts the area is clean; only say that if
@@ -683,7 +681,8 @@ async def agent_finish(
            {
                "success": False,
                "error": (
-                    "agent_finish is for subagents. Root/main agents must call finish_scan instead"
+                    "agent_finish is for subagents. Root/main agents must call "
+                    f"{coordinator.root_finish_tool} instead"
                ),
            },
            ensure_ascii=False,
@@ -24,12 +24,16 @@ async def load_skill(ctx: RunContextWrapper, skills: list[str]) -> str:
            Max 5. Names match the bare files under
            ``strix/skills/<category>/<name>.md``.
    """
-    del ctx
    requested = list(skills or [])
    err = validate_requested_skills(requested)
    if err:
        return f"load_skill: {err}"
-    contents = load_skills(requested)
+    context = ctx.context if isinstance(ctx.context, dict) else {}
+    contents = load_skills(
+        requested,
+        supports_images=context.get("supports_images", True),
+        coverage_tools=context.get("coverage_tools", True),
+    )
    if not contents:
        return "load_skill: no content loaded for requested skills."
    sections = [f"## Skill: {name}\n\n{body}" for name, body in contents.items()]
@@ -24,6 +24,7 @@ automatically.

 from __future__ import annotations

+import asyncio
 import json
 from typing import TYPE_CHECKING, Any

@@ -46,6 +47,7 @@ def _registry_from_ctx(ctx: RunContextWrapper) -> McpRegistry | None:


 _NO_CONNECTIONS = "No MCP connections are configured for this run."
+_MCP_DISCOVERY_WAIT_SECONDS = 30.0


 def _unknown_connection(connection: str, registry: McpRegistry) -> str:
@@ -61,11 +63,13 @@ def _format_tool(tool: MCPTool) -> str:

@function_tool(timeout=60)
 async def list_mcps(ctx: RunContextWrapper) -> dict[str, Any]:
-    """List the MCP connections available this run, so you can discover them.
+    """List available MCP connections, loading each connection's catalog on first use.

    Read-only. Returns one entry per connection with its ``id`` (the exact name
    you pass to the other MCP tools), ``name``, ``description``, and
-    ``tool_count``. The response does not include tool schemas. Call
+    ``tool_count`` (null while the catalog is unknown). Waits at most 30 seconds
+    for catalogs; unfinished catalogs keep loading in the background.
+    The response does not include tool schemas. Call
    ``search_mcp_tools`` next. Then call ``get_mcp_tool_schema`` for one selected
    tool before you call ``call_mcp``. Returns an empty ``connections`` list when
    the run has no MCP connections.
@@ -73,6 +77,23 @@ async def list_mcps(ctx: RunContextWrapper) -> dict[str, Any]:
    registry = _registry_from_ctx(ctx)
    if registry is None or not registry:
        return {"connections": []}
+    pending = [
+        entry
+        for name in registry.names()
+        if (entry := registry.get(name)) is not None
+        and entry.state != "catalog_ready"
+        and (entry.session is None or not entry.session.is_dead)
+    ]
+    if pending:
+        await asyncio.wait(
+            {
+                asyncio.gather(
+                    *(entry.ensure_catalog() for entry in pending),
+                    return_exceptions=True,
+                )
+            },
+            timeout=_MCP_DISCOVERY_WAIT_SECONDS,
+        )
    dead_by_name = {status.name: status.dead for status in registry.statuses()}
    return {
        "connections": [
@@ -80,7 +101,7 @@ async def list_mcps(ctx: RunContextWrapper) -> dict[str, Any]:
                "id": summary.name,
                "name": summary.name,
                "description": summary.purpose,
-                "tool_count": summary.tool_count,
+                "tool_count": summary.tool_count if summary.state == "catalog_ready" else None,
                "dead": dead_by_name.get(summary.name, False),
                "state": summary.state,
            }
@@ -288,9 +309,11 @@ async def call_mcp(
            "Call search_mcp_tools, then get_mcp_tool_schema."
        )
    session = await entry.ensure_connected()
+    context = ctx.context if isinstance(ctx.context, dict) else {}
    return await session.dispatch(
        tool,
        arguments or {},
        label=namespaced_tool_name(connection, tool),
        result_transform=entry.result_transform,
+        supports_images=context.get("supports_images", True),
    )
@@ -32,6 +32,7 @@ from agents.mcp import (
 )
 from mcp.client.stdio import stdio_client
 from mcp.shared._httpx_utils import create_mcp_http_client
+from mcp.types import TextContent

 from strix.tools.mcp.failures import HttpStatusRecorder
 from strix.tools.mcp.session import McpConnectionUnavailableError, SupervisedMcpSession
@@ -183,6 +184,18 @@ def _build_server(config: McpConnectionConfig) -> BuiltMcpServer:
    )


+def _without_images(result: Any) -> Any:
+    content = [
+        TextContent(
+            type="text", text=f"[{item.mimeType} image omitted: this model cannot view images]"
+        )
+        if item.type == "image"
+        else item
+        for item in result.content
+    ]
+    return result.model_copy(update={"content": content})
+
+
 def _mcp_result_to_tool_output(server: MCPServer, result: Any) -> Any:
    """Serialize a ``CallToolResult`` to a tool output, mirroring the agents SDK.

@@ -217,6 +230,7 @@ async def dispatch_mcp_call(
    *,
    label: str,
    result_transform: ResultTransform | None = None,
+    supports_images: bool = True,
 ) -> Any:
    """Run one MCP tool call and convert its result to a tool output.

@@ -233,6 +247,8 @@ async def dispatch_mcp_call(
      corrupt the content the agent receives).
    """
    result = await server.call_tool(tool_name, arguments)
+    if not supports_images:
+        result = _without_images(result)
    if result_transform is not None:
        return result_transform(label, result.model_dump(mode="json"))
    tool_output = _mcp_result_to_tool_output(server, result)
@@ -432,19 +432,19 @@ class McpRegistry:
            entry.set_status_sink(sink)

    def start_warmup(self, *, max_concurrency: int = 6) -> asyncio.Task[None]:
-        """Connect every configured entry in the background with a fixed bound."""
+        """Connect and list every configured entry with a fixed concurrency bound."""
        if self._warmup_task is not None:
            return self._warmup_task

        async def warm() -> None:
            semaphore = asyncio.Semaphore(max(1, max_concurrency))

-            async def connect(entry: McpConnectionEntry) -> None:
+            async def load_catalog(entry: McpConnectionEntry) -> None:
                async with semaphore:
-                    with contextlib.suppress(McpConnectionUnavailableError):
-                        await entry.ensure_connected()
+                    with contextlib.suppress(Exception):
+                        await entry.ensure_catalog()

-            await asyncio.gather(*(connect(entry) for entry in self._entries.values()))
+            await asyncio.gather(*(load_catalog(entry) for entry in self._entries.values()))

        self._warmup_task = asyncio.create_task(warm(), name="mcp-warmup")
        return self._warmup_task
@@ -365,6 +365,7 @@ class SupervisedMcpSession:
        *,
        label: str,
        result_transform: ResultTransform | None = None,
+        supports_images: bool = True,
    ) -> Any:
        """Run one tool call with bounded retries for transient session failures.

@@ -382,6 +383,7 @@ class SupervisedMcpSession:
                arguments,
                label=label,
                result_transform=result_transform,
+                supports_images=supports_images,
            )

        outcome = await self._run_job(job, phase="call")
@@ -13,6 +13,7 @@ from datetime import datetime
 from typing import TYPE_CHECKING, Any, Literal

 from agents import RunContextWrapper, function_tool
+from caido_sdk_client import BaseError

 from strix.runtime.caido_handle import CaidoBootstrapHandle
 from strix.tools.nullish import clean_optional
@@ -68,6 +69,15 @@ async def _call[T](client: Client, fn: Callable[[Client], Awaitable[T]]) -> T:
        return await fn(client)


+_REQUEST_EXISTS_QUERY = "query RequestExists($id: ID!) { request(id: $id) { id } }"
+_REQUEST_EXISTS_ATTEMPTS = 2
+
+
+async def _request_exists(client: Client, request_id: str) -> bool:
+    data = await client.graphql.query(_REQUEST_EXISTS_QUERY, {"id": request_id})
+    return bool(data) and data.get("request") is not None
+
+
 async def existing_request_ids(
    ctx: RunContextWrapper,
    request_ids: list[str],
@@ -80,16 +90,24 @@ async def existing_request_ids(
    if client is None:
        raise RuntimeError("Caido client is not available")

-    # Request IDs are not an HTTPQL field. Resolve each ID through the same
-    # project-bound lookup as view_request rather than constructing a filter.
+    # Request IDs are not an HTTPQL field, so each one is resolved by id. Only
+    # the id is selected: pulling the raw request and response bodies just to
+    # learn whether a row exists is what made this time out on busy projects.
    existing: set[str] = set()
    for request_id in request_ids:
-        result = await _call(
-            client,
-            functools.partial(caido_api.get_request_with_client, request_id=request_id),
-        )
-        if result is not None:
-            existing.add(str(result.request.id))
+        for attempt in range(1, _REQUEST_EXISTS_ATTEMPTS + 1):
+            try:
+                exists = await _call(
+                    client,
+                    functools.partial(_request_exists, request_id=request_id),
+                )
+            except Exception:
+                if attempt == _REQUEST_EXISTS_ATTEMPTS:
+                    raise
+                continue
+            if exists:
+                existing.add(request_id)
+            break
    return existing


@@ -124,7 +142,11 @@ def _no_client() -> str:


 def _err(name: str, exc: Exception) -> str:
-    logger.exception("%s failed", name)
+    # Caido-side failures (timeouts, bad filters, ...) are returned to the model, not our bugs.
+    if isinstance(exc, BaseError):
+        logger.warning("%s failed", name, exc_info=True)
+    else:
+        logger.exception("%s failed", name)
    return json.dumps(
        {"success": False, "error": f"{name} failed: {exc}"},
        ensure_ascii=False,
@@ -222,9 +222,9 @@ def _normalize_http_exchange_ids(raw: Any) -> tuple[list[str] | None, list[str]]
    return normalized, errors


-_HTTP_EXCHANGE_DROPPED_WARNING = (
-    "http_exchange_ids were not stored: the proxy project could not be reached to verify "
-    "them. Attach them with update_vulnerability_report when the proxy responds again."
+_HTTP_EXCHANGE_UNVERIFIED_WARNING = (
+    "http_exchange_ids were stored without verification because the proxy project did "
+    "not respond in time. Ids that do not exist in the project will not render as evidence."
 )


@@ -235,9 +235,9 @@ async def _verify_http_exchange_ids(
    """Verify proxy exchange IDs against the current Caido project.

    IDs the project does not know are rejected. When the proxy itself cannot be
-    queried the IDs are dropped and a warning is returned instead, so a proxy
-    outage never blocks a finding and unverified IDs are never recorded as
-    evidence.
+    queried the IDs are kept and a warning is returned instead: a slow or
+    unreachable proxy must never strip evidence from a finding, and consumers
+    resolve each id against the recorded traffic before rendering it.
    """
    request_ids, errors = _normalize_http_exchange_ids(raw)
    if request_ids is None or errors or not request_ids:
@@ -250,7 +250,7 @@ async def _verify_http_exchange_ids(
            "Could not verify HTTP exchange IDs against the current Caido project",
            exc_info=True,
        )
-        return None, [], _HTTP_EXCHANGE_DROPPED_WARNING
+        return request_ids, [], _HTTP_EXCHANGE_UNVERIFIED_WARNING

    missing_ids = [request_id for request_id in request_ids if request_id not in existing_ids]
    if missing_ids:
@@ -1876,14 +1876,10 @@ def _build_dependency_evidence(
        evidence += f"\n\n**Usage analysis:** {label}."
        if reachability_evidence and reachability_evidence.strip():
            evidence += f" {reachability_evidence.strip()}"
-        evidence += (
-            " This is a prioritization signal from static analysis, not a"
-            " proof of exploitability or of safety."
-        )
    return evidence


-async def _do_create_dependency(  # noqa: PLR0912
+async def _do_create_dependency(  # noqa: PLR0912, PLR0915
    *,
    title: str,
    description: str,
@@ -1905,6 +1901,8 @@ async def _do_create_dependency(  # noqa: PLR0912
    manifest_path: str | None = None,
    reachability: str = "unknown",
    reachability_evidence: str | None = None,
+    code_locations: list[dict[str, Any]] | None = None,
+    fix_verification: str | None = None,
    contextual_cvss_breakdown: dict[str, str] | None = None,
    contextual_cvss_reasoning: str | None = None,
    agent_id: str | None = None,
@@ -1947,6 +1945,16 @@ async def _do_create_dependency(  # noqa: PLR0912
    if manifest_err:
        errors.append(manifest_err)

+    parsed_locations = _normalize_code_locations(code_locations)
+    if parsed_locations:
+        errors.extend(_validate_code_locations(parsed_locations))
+    elif code_locations:
+        errors.append(
+            "code_locations were dropped as unusable - every location needs a relative "
+            "'file' and an integer 'start_line'"
+        )
+    errors.extend(_validate_fix_verification(parsed_locations, fix_verification))
+
    reachability = (reachability or "unknown").strip().lower()
    if reachability not in _VALID_REACHABILITY:
        errors.append(
@@ -2053,6 +2061,8 @@ async def _do_create_dependency(  # noqa: PLR0912
            cvss=cvss_score if advisory_cvss is not None else None,
            cve=parsed_cve,
            cwe=cwe,
+            code_locations=parsed_locations,
+            fix_verification=fix_verification,
            finding_class="dependency_cve",
            dependency_metadata=dependency_metadata,
            agent_id=agent_id if isinstance(agent_id, str) else None,
@@ -2107,6 +2117,8 @@ async def create_dependency_report(
    dependency_path: str | None = None,
    reachability: str = "unknown",
    reachability_evidence: str | None = None,
+    code_locations: list[dict[str, Any]] | None = None,
+    fix_verification: str | None = None,
    contextual_cvss_breakdown: dict[str, str] | None = None,
    contextual_cvss_reasoning: str | None = None,
 ) -> str:
@@ -2217,6 +2229,23 @@ async def create_dependency_report(
            is off in production), and say who controls the input. State
            it plainly when no entry point reaches the sink — that is the
            most useful result a reader can get.
+        code_locations: Where application code imports or calls the package —
+            the evidence behind your ``reachability`` claim. List of dicts,
+            same shape as ``create_vulnerability_report``::
+
+                {"file": "src/api/client.ts", "start_line": 14, "end_line": 14,
+                 "snippet": "const x = require('pkg')", "label": "imports it"}
+
+            Cite the repo-relative ``file`` and the exact 1-based
+            ``start_line`` where ``snippet`` actually sits — verify against
+            the real file, never guess. A location may also propose
+            the fix itself: ``fix_before`` (verbatim current code, e.g. the
+            manifest pin line) + ``fix_after`` (same lines with the fixed
+            version) — that combination requires ``fix_verification``.
+        fix_verification: Required whenever any ``code_locations`` entry
+            carries ``fix_after``. Same bar as a normal finding: security
+            closure, bypass review, preserved behavior, and how each was
+            checked.
        contextual_cvss_breakdown: **Required.** Full CVSS v3.1 rating of this
            CVE **in this codebase** — the same 8-metric object as
            ``create_vulnerability_report``'s ``cvss_breakdown``:
@@ -2277,6 +2306,8 @@ async def create_dependency_report(
        manifest_path=manifest_path,
        reachability=reachability,
        reachability_evidence=reachability_evidence,
+        code_locations=code_locations,
+        fix_verification=fix_verification,
        contextual_cvss_breakdown=contextual_cvss_breakdown,
        contextual_cvss_reasoning=contextual_cvss_reasoning,
        agent_id=agent_id,
@@ -80,6 +80,94 @@ async def test_wrap_exec_command_preserves_explicit_shell(shell: str) -> None:
    assert json.loads(captured["raw_input"])["shell"] == shell


+def _running_exec_tool(result: str) -> FunctionTool:
+    async def invoke(_ctx: Any, _raw_input: str) -> str:
+        return result
+
+    return FunctionTool(
+        name="exec_command",
+        description="Runs a command.",
+        params_json_schema={"type": "object", "properties": {}},
+        on_invoke_tool=invoke,
+    )
+
+
+_RUNNING_RESULT = (
+    "Chunk ID: abc123\nWall time: 30.0012 seconds\nProcess running with session ID 7\nOutput:\n"
+)
+
+
+def test_wrap_exec_command_describes_the_yield_cap() -> None:
+    wrapped = factory._wrap_exec_command(_running_exec_tool("ok"))
+
+    assert wrapped.description.startswith("Runs a command. The wait for output is capped at 30 s")
+    assert "cannot be read later" in wrapped.description
+    assert "tty=true" in wrapped.description
+
+
+@pytest.mark.asyncio
+async def test_wrap_exec_command_flags_a_running_non_tty_process() -> None:
+    wrapped = factory._wrap_exec_command(_running_exec_tool(_RUNNING_RESULT))
+
+    result = await wrapped.on_invoke_tool(
+        cast("Any", None), json.dumps({"cmd": "sleep 60; curl x", "yield_time_ms": 90000})
+    )
+
+    assert result == (
+        "Chunk ID: abc123\nWall time: 30.0012 seconds\n"
+        "Process running with session ID 7\nOutput:\n"
+        "Note: yield_time_ms=90000 was capped at 30000 (30 s). this non-TTY process cannot be "
+        "polled and its later output is lost. Rerun it in the background "
+        "(`cmd > /tmp/out.log 2>&1 &`, then read the log) or with tty=true.\n"
+    )
+
+
+@pytest.mark.asyncio
+async def test_wrap_exec_command_flags_a_running_non_tty_process_within_the_cap() -> None:
+    wrapped = factory._wrap_exec_command(_running_exec_tool(_RUNNING_RESULT))
+
+    result = await wrapped.on_invoke_tool(cast("Any", None), json.dumps({"cmd": "sleep 60"}))
+
+    assert "Note: this non-TTY process cannot be polled" in result
+    assert "was capped" not in result
+
+
+@pytest.mark.asyncio
+async def test_wrap_exec_command_leaves_a_running_tty_process_alone() -> None:
+    wrapped = factory._wrap_exec_command(_running_exec_tool(_RUNNING_RESULT))
+
+    result = await wrapped.on_invoke_tool(
+        cast("Any", None), json.dumps({"cmd": "python3 -m http.server", "tty": True})
+    )
+
+    assert result == _RUNNING_RESULT
+
+
+@pytest.mark.asyncio
+async def test_wrap_exec_command_leaves_an_exited_process_alone() -> None:
+    exited = "Chunk ID: abc123\nWall time: 0.1 seconds\nProcess exited with code 0\nOutput:\nhi"
+    wrapped = factory._wrap_exec_command(_running_exec_tool(exited))
+
+    result = await wrapped.on_invoke_tool(cast("Any", None), json.dumps({"cmd": "echo hi"}))
+
+    assert result == exited
+
+
+@pytest.mark.asyncio
+async def test_wrap_exec_command_ignores_the_running_marker_inside_command_output() -> None:
+    exited = (
+        "Chunk ID: abc123\nWall time: 0.2 seconds\nProcess exited with code 0\nOutput:\n"
+        "Process running with session ID 7\nOutput:\n"
+    )
+    wrapped = factory._wrap_exec_command(_running_exec_tool(exited))
+
+    result = await wrapped.on_invoke_tool(
+        cast("Any", None), json.dumps({"cmd": "cat transcript.txt"})
+    )
+
+    assert result == exited
+
+
@pytest.mark.asyncio
 async def test_responses_filesystem_custom_tool_output_is_bounded() -> None:
    async def invoke(_ctx: Any, _inp: str) -> str:
@@ -115,3 +203,15 @@ def test_function_tools_are_result_bounded() -> None:
    by_name = {t.name: t for t in agent.tools}

    assert getattr(by_name["think"], "_strix_bounded", False) is True
+
+
+@pytest.mark.parametrize(("supports_images", "has_view_image"), [(True, True), (False, False)])
+def test_view_image_is_left_out_for_text_only_models(
+    supports_images: bool, has_view_image: bool
+) -> None:
+    agent = factory.build_strix_agent(is_root=True, supports_images=supports_images)
+    filesystem = agent.capabilities[0]
+    filesystem.bind(cast("Any", object()))
+
+    view_image = next(tool for tool in filesystem.tools() if tool.name == "view_image")
+    assert cast("FunctionTool", view_image).is_enabled is has_view_image
@@ -38,11 +38,26 @@ async def _roundtrip(


 _STRING = {"todos": {"type": "string"}}
+
+
+@pytest.mark.asyncio
+async def test_truncated_arguments_short_circuit_with_a_reissue_message() -> None:
+    captured: dict[str, str] = {}
+    wrapped = factory._with_coerced_arguments(_capturing_tool(captured, _STRING))
+
+    result = await wrapped.on_invoke_tool(cast("Any", None), '{"todos": "a, b')
+
+    assert "not executed" in result
+    assert "Re-issue the call" in result
+    assert "raw_input" not in captured
+
+
 _ARRAY = {"tags": {"type": "array", "items": {"type": "string"}}}
 _NULLABLE_ARRAY = {
    "tags": {"anyOf": [{"type": "array", "items": {"type": "string"}}, {"type": "null"}]}
 }
 _OBJECT = {"modifications": {"type": "object"}}
+OVER_NESTED = '{"a":' + "[" * 100_000


@pytest.mark.asyncio
@@ -136,12 +151,15 @@ async def test_unknown_and_null_arguments_are_untouched() -> None:


@pytest.mark.asyncio
-async def test_non_object_payloads_pass_through_unchanged() -> None:
+@pytest.mark.parametrize("arguments", ["not json", '{"n": NaN}', OVER_NESTED])
+async def test_non_object_payloads_are_reported_instead_of_invoked(arguments: str) -> None:
    captured: dict[str, str] = {}
    wrapped = factory._with_coerced_arguments(_capturing_tool(captured, _ARRAY))

-    assert await wrapped.on_invoke_tool(cast("Any", None), "not json") == "ok"
-    assert captured["raw_input"] == "not json"
+    result = await wrapped.on_invoke_tool(cast("Any", None), arguments)
+
+    assert result.startswith("probe: the tool call was not executed")
+    assert "raw_input" not in captured


@pytest.mark.asyncio
@@ -59,6 +59,15 @@ def test_registered_tools_appear_before_lifecycle_tool() -> None:
    assert child_names[-2:] == ["extra", "agent_finish"]


+def test_root_ends_with_the_given_finish_tool() -> None:
+    root = factory.build_strix_agent(is_root=True, finish_tool=_tool("finish_pr_review"))
+
+    root_names = [t.name for t in root.tools]
+
+    assert root_names[-1] == "finish_pr_review"
+    assert "finish_scan" not in root_names
+
+
 def test_per_call_extra_tools_stack_with_registry() -> None:
    factory.register_agent_tools(_tool("registered"))

@@ -84,6 +93,51 @@ def test_per_call_extra_tools_reject_duplicate_registered_names() -> None:
        factory.build_strix_agent(is_root=True, extra_tools=[_tool("same_name")])


+def test_report_tool_replaces_create_vulnerability_report() -> None:
+    for is_root in (True, False):
+        agent = factory.build_strix_agent(is_root=is_root, report_tool=_tool("file_pr_finding"))
+        names = [t.name for t in agent.tools]
+        assert "file_pr_finding" in names
+        assert "create_vulnerability_report" not in names
+        assert isinstance(agent.instructions, str)
+        assert "file_pr_finding" in agent.instructions
+        assert "create_vulnerability_report" not in agent.instructions
+
+
+def _descriptions(tools: list[Any]) -> list[str]:
+    return [t.description for t in tools if isinstance(t, FunctionTool)]
+
+
+def test_coverage_tools_can_be_turned_off() -> None:
+    coverage = {"record_coverage", "update_coverage", "list_coverage"}
+    for is_root in (True, False):
+        on = factory.build_strix_agent(is_root=is_root, is_whitebox=True)
+        off = factory.build_strix_agent(is_root=is_root, is_whitebox=True, coverage_tools=False)
+        assert coverage <= {t.name for t in on.tools}
+        assert not coverage & {t.name for t in off.tools}
+        assert isinstance(on.instructions, str)
+        assert isinstance(off.instructions, str)
+        assert "- COVERAGE" in on.instructions
+        assert "## Recording Closure" in on.instructions
+        assert not any(name in off.instructions for name in coverage)
+        on_descriptions = "\n".join(_descriptions(on.tools))
+        off_descriptions = "\n".join(_descriptions(off.tools))
+        if is_root:
+            assert "Coverage reconciliation" in on_descriptions
+            assert "Coverage reconciliation" not in off_descriptions
+        assert not any(name in off_descriptions for name in coverage)
+
+    diff_on = factory.build_strix_agent(is_root=True, is_diff_scoped=True)
+    diff_off = factory.build_strix_agent(is_root=True, is_diff_scoped=True, coverage_tools=False)
+    assert isinstance(diff_on.instructions, str)
+    assert isinstance(diff_off.instructions, str)
+    assert "Record coverage per changed component" in diff_on.instructions
+    assert "`needs_follow_up` coverage" in diff_on.instructions
+    assert "Record coverage per changed component" not in diff_off.instructions
+    assert "`needs_follow_up` coverage" not in diff_off.instructions
+    assert not any(name in diff_off.instructions for name in coverage)
+
+
 def test_instructions_override_is_used_verbatim() -> None:
    custom = "You are a scan agent. Follow the provided scope."

@@ -21,7 +21,7 @@ from strix.core.sessions import open_agent_session


 if TYPE_CHECKING:
-    from collections.abc import AsyncIterator, Callable
+    from collections.abc import AsyncIterator, Awaitable, Callable
    from pathlib import Path


@@ -31,6 +31,7 @@ _CALL_LATENCY_S = 0.005

 class _FakeLedger:
    def __init__(self) -> None:
+        self.on_pause: Callable[[], Awaitable[None]] | None = None
        self.cost = 0.0
        self.calls: list[str] = []
        self.remaining: dict[str, int] = {}
@@ -80,6 +81,8 @@ class _FakeStream:
                await self._hooks.on_llm_start(ctx_wrapper, self._agent, None, input_items)
            except BudgetPausedError as exc:
                self.run_loop_exception = exc
+                if self._ledger.on_pause is not None:
+                    await self._ledger.on_pause()
                return
            self._ledger.warned_inputs.append(input_items)
            if self._ledger.gate is not None:
@@ -150,13 +153,16 @@ class _Scan:
        monkeypatch: pytest.MonkeyPatch,
        *,
        max_budget_usd: float | None,
+        budget_policy: str = "pause",
    ) -> None:
        self.ledger = _FakeLedger()
        self.hooks = ReportUsageHooks(
-            model="test-model", max_budget_usd=max_budget_usd, budget_policy="pause"
+            model="test-model",
+            max_budget_usd=max_budget_usd,
+            budget_policy=budget_policy,  # type: ignore[arg-type]
        )
        self.coordinator = AgentCoordinator()
-        self.coordinator.set_budget_policy("pause")
+        self.coordinator.set_budget_policy(budget_policy)  # type: ignore[arg-type]
        self.coordinator.set_budget_limit_setter(self.hooks.set_max_budget_usd)
        monkeypatch.setattr(execution, "Runner", _fake_runner(self.ledger, self.coordinator))
        monkeypatch.setattr(execution, "_compact_session", _noop_compact)
@@ -538,3 +544,140 @@ async def test_stop_policy_is_unchanged() -> None:
 def test_budget_policy_is_validated() -> None:
    with pytest.raises(ValueError, match="budget_policy"):
        ReportUsageHooks(model="m", budget_policy="later")  # type: ignore[arg-type]
+
+
+def _pause_flags(coordinator: AgentCoordinator) -> tuple[bool, bool]:
+    return coordinator.budget_paused, coordinator.operator_paused
+
+
+@pytest.mark.asyncio
+async def test_operator_pause_holds_under_the_stop_policy(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    scan = _Scan(tmp_path, monkeypatch, max_budget_usd=None, budget_policy="stop")
+    agents = ("root", "child-a", "child-b")
+
+    with patch("strix.core.hooks.get_global_report_state", return_value=scan.ledger):
+        await scan.start_root(calls=6)
+        await scan.start_child("child-a", calls=6)
+        await scan.start_child("child-b", calls=6)
+
+        await _wait_until(lambda: scan.ledger.cost >= 3.0)
+        await scan.coordinator.pause_budget()
+        spent_at_pause = scan.ledger.cost
+        await _wait_until(lambda: _all_parked(scan.coordinator, *agents))
+        assert _pause_flags(scan.coordinator) == (True, True)
+        assert all(not task.done() for task in scan.tasks())
+
+        await scan.coordinator.send("child-a", {"from": "root", "content": "status?"})
+        await scan.coordinator.send("root", {"from": "user", "content": "keep going"})
+        await asyncio.sleep(0.05)
+        assert _pause_flags(scan.coordinator) == (True, True)
+        assert scan.ledger.cost == pytest.approx(spent_at_pause)
+        assert _all_parked(scan.coordinator, *agents)
+
+        woken = await scan.coordinator.resume_budget()
+        assert sorted(woken) == sorted(agents)
+        assert _pause_flags(scan.coordinator) == (False, False)
+        await asyncio.wait_for(asyncio.gather(*scan.tasks(), return_exceptions=True), timeout=5.0)
+        assert scan.ledger.cost == pytest.approx(18.0)
+        assert {aid: str(s) for aid, s in scan.coordinator.statuses.items()} == {
+            "root": "completed",
+            "child-a": "completed",
+            "child-b": "completed",
+        }
+        root_items = await scan.sessions[0].get_items()
+        assert any("keep going" in str(item) for item in root_items)
+
+    for session in scan.sessions:
+        session.close()
+
+
+@pytest.mark.asyncio
+async def test_a_resume_that_lands_before_the_park_still_continues(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    scan = _Scan(tmp_path, monkeypatch, max_budget_usd=None, budget_policy="stop")
+    resumed = 0
+
+    async def resume_at_once() -> None:
+        nonlocal resumed
+        resumed += 1
+        await scan.coordinator.resume_budget()
+
+    scan.ledger.on_pause = resume_at_once
+    with patch("strix.core.hooks.get_global_report_state", return_value=scan.ledger):
+        await scan.start_root(calls=6)
+        await scan.start_child("child-a", calls=6)
+        await _wait_until(lambda: scan.ledger.cost >= 2.0)
+        await scan.coordinator.pause_budget()
+        await asyncio.wait_for(asyncio.gather(*scan.tasks(), return_exceptions=True), timeout=5.0)
+        assert resumed >= 1
+        assert _pause_flags(scan.coordinator) == (False, False)
+        assert scan.ledger.cost == pytest.approx(12.0)
+        assert {aid: str(s) for aid, s in scan.coordinator.statuses.items()} == {
+            "root": "completed",
+            "child-a": "completed",
+        }
+
+    for session in scan.sessions:
+        session.close()
+
+
+@pytest.mark.asyncio
+async def test_operator_pause_gates_the_stop_policy_hooks() -> None:
+    coordinator = AgentCoordinator()
+    hooks = ReportUsageHooks(model="m", max_budget_usd=None)
+    assert hooks.budget_policy == "stop"
+
+    await coordinator.pause_budget()
+    with pytest.raises(BudgetPausedError):
+        await hooks.on_llm_start(_ctx(coordinator), MagicMock(), None, [])
+
+    await coordinator.resume_budget()
+    await hooks.on_llm_start(_ctx(coordinator), MagicMock(), None, [])
+
+
+@pytest.mark.asyncio
+async def test_a_user_message_lifts_only_the_limit_pause(tmp_path: Path) -> None:
+    coordinator = AgentCoordinator()
+    await coordinator.register("root", "strix", parent_id=None)
+    session = open_agent_session("root", tmp_path / "agents.db")
+    await coordinator.attach_runtime("root", session=session)
+    extensions: list[int] = []
+    coordinator.set_budget_extender(lambda: extensions.append(1))
+
+    await coordinator.pause_budget()
+    await coordinator.park_for_budget("root")
+    assert coordinator.statuses["root"] == "budget_paused"
+    await coordinator.send("root", {"from": "user", "content": "keep going"})
+    assert _pause_flags(coordinator) == (True, True)
+    assert extensions == []
+    assert coordinator.pending_counts["root"] == 1
+
+    await coordinator.resume_budget()
+    await coordinator.pause_for_budget("root")
+    await coordinator.send("root", {"from": "user", "content": "keep going"})
+    assert _pause_flags(coordinator) == (False, False)
+    assert extensions == [1]
+    session.close()
+
+
+@pytest.mark.asyncio
+async def test_snapshot_and_reset_keep_an_operator_pause() -> None:
+    coordinator = AgentCoordinator()
+    await coordinator.register("root", "strix", parent_id=None)
+    await coordinator.pause_budget()
+    snap = await coordinator.snapshot()
+
+    restored = AgentCoordinator()
+    await restored.restore(snap)
+    assert _pause_flags(restored) == (True, True)
+    await restored.reset_budget_stops(budget_stopped=False, reserve_stopped=False)
+    assert _pause_flags(restored) == (True, True)
+
+    await restored.resume_budget()
+    await restored.pause_for_budget("root")
+    await restored.reset_budget_stops(budget_stopped=False, reserve_stopped=False)
+    assert _pause_flags(restored) == (False, False)
+    assert restored.statuses["root"] == "waiting"
@@ -147,7 +147,9 @@ def _patch_budget(monkeypatch: pytest.MonkeyPatch, *, keep_tokens: int, window:
    context.auto_compact = True
    settings = SimpleNamespace(
        context=context,
-        llm=SimpleNamespace(api_key=None, api_base=None, timeout=1, extra_headers=None),
+        llm=SimpleNamespace(
+            api_key=None, api_base=None, timeout=1, extra_headers=None, extra_body=None
+        ),
    )
    monkeypatch.setattr(compaction, "load_settings", lambda: settings)

@@ -228,7 +230,11 @@ async def test_summarize_routes_through_provider_with_settings(
        "load_settings",
        lambda: SimpleNamespace(
            llm=SimpleNamespace(
-                api_key=None, api_base=None, timeout=1, extra_headers={"X-Feature-Key": "svc"}
+                api_key=None,
+                api_base=None,
+                timeout=1,
+                extra_headers={"X-Feature-Key": "svc"},
+                extra_body={"service_tier": "flex"},
            )
        ),
    )
@@ -239,6 +245,7 @@ async def test_summarize_routes_through_provider_with_settings(
    assert captured["model"] == "litellm/openai/some-model"
    settings = captured["model_settings"]
    assert settings.extra_headers == {"X-Feature-Key": "svc"}
+    assert settings.extra_body == {"service_tier": "flex"}
    assert settings.max_tokens == 64


@@ -37,6 +37,12 @@ _LLM_ENV_KEYS = [
    "STRIX_RUNTIME_BACKEND",
    # TelemetrySettings
    "STRIX_TELEMETRY",
+    # JSON-object settings
+    "LLM_EXTRA_HEADERS",
+    "LLM_EXTRA_BODY",
+    "STRIX_DEDUPE_MODEL",
+    "DEDUPE_LLM_EXTRA_HEADERS",
+    "DEDUPE_LLM_EXTRA_BODY",
 ]


@@ -406,6 +412,57 @@ def test_persist_current_replaces_corrupt_file(
    assert json.loads(target.read_text(encoding="utf-8")) == {"env": {"STRIX_LLM": "env-model"}}


+@pytest.mark.parametrize(
+    ("env_var", "section", "field"),
+    [
+        ("LLM_EXTRA_HEADERS", "llm", "extra_headers"),
+        ("LLM_EXTRA_BODY", "llm", "extra_body"),
+        ("DEDUPE_LLM_EXTRA_HEADERS", "dedupe", "extra_headers"),
+        ("DEDUPE_LLM_EXTRA_BODY", "dedupe", "extra_body"),
+    ],
+)
+def test_persisted_json_object_setting_reloads_without_env(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch, env_var: str, section: str, field: str
+) -> None:
+    loader.apply_config_override(tmp_path / "cli-config.json")
+    monkeypatch.setenv(env_var, '{"X-Tenant": "acme"}')
+    loader.load_settings()
+    loader.persist_current()
+    monkeypatch.delenv(env_var)
+    loader.apply_config_override(tmp_path / "cli-config.json")
+
+    settings = loader.load_settings()
+
+    assert getattr(getattr(settings, section), field) == {"X-Tenant": "acme"}
+
+
+def test_json_object_setting_saved_as_object_still_loads(tmp_path: Path) -> None:
+    target = tmp_path / "cli-config.json"
+    target.write_text(
+        json.dumps({"env": {"LLM_EXTRA_BODY": {"service_tier": "flex"}}}), encoding="utf-8"
+    )
+    loader.apply_config_override(target)
+
+    assert loader.load_settings().llm.extra_body == {"service_tier": "flex"}
+
+
+def test_invalid_saved_json_object_setting_is_rejected(tmp_path: Path) -> None:
+    target = tmp_path / "cli-config.json"
+    target.write_text(json.dumps({"env": {"LLM_EXTRA_BODY": "not json"}}), encoding="utf-8")
+    loader.apply_config_override(target)
+
+    with pytest.raises(ValidationError):
+        loader.load_settings()
+
+
+def test_saved_string_setting_is_not_json_decoded(tmp_path: Path) -> None:
+    target = tmp_path / "cli-config.json"
+    target.write_text(json.dumps({"env": {"STRIX_LLM": '{"not": "decoded"}'}}), encoding="utf-8")
+    loader.apply_config_override(target)
+
+    assert loader.load_settings().llm.model == '{"not": "decoded"}'
+
+
 def test_persist_current_sets_0600_mode(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
    monkeypatch.setenv("STRIX_LLM", "persisted-model")
    target = tmp_path / "cli-config.json"
@@ -2,7 +2,9 @@

 from __future__ import annotations

+import json
 import uuid
+from datetime import UTC, datetime, timedelta
 from types import SimpleNamespace
 from typing import TYPE_CHECKING, Any
 from unittest.mock import MagicMock, call, patch
@@ -17,11 +19,13 @@ from strix.config.models import (
    _configure_litellm_compatibility,
    _install_openrouter_stream_cost_capture,
 )
+from strix.core.hooks import ReportUsageHooks
 from strix.llm import request_log
 from strix.report.state import (
    ReportState,
    litellm_cost_callback,
    openrouter_stream_cost,
+    record_llm_request,
    set_global_report_state,
    streamed_openrouter_costs,
 )
@@ -135,6 +139,7 @@ def test_cost_callback_estimates_cost_with_provider_prefixed_model() -> None:

    with (
        patch("strix.report.state.get_global_report_state", return_value=report_state),
+        patch("strix.report.state.resolve_litellm_model", side_effect=lambda model: model),
        patch("litellm.completion_cost", side_effect=fake_completion_cost),
    ):
        litellm_cost_callback(kwargs, response)
@@ -389,3 +394,137 @@ def test_openrouter_request_carries_agent_session_id() -> None:
            assert body()["session_id"] == session_id
    finally:
        request_log.reset_call_context(token)
+
+
+def _openrouter_config() -> Any:
+    _install_openrouter_stream_cost_capture()
+    config = ProviderConfigManager.get_provider_chat_config(
+        model="z-ai/glm-5.3", provider=LlmProviders.OPENROUTER
+    )
+    assert config is not None
+    return config
+
+
+def test_openrouter_records_upstream_of_stream_and_mid_stream_error() -> None:
+    handler = _openrouter_config().get_model_response_iterator(
+        streaming_response=iter([]), sync_stream=True
+    )
+    reply = request_log.HttpReply()
+    token = request_log._http_reply.set(reply)
+    try:
+        handler.chunk_parser(
+            {
+                "id": "gen-a",
+                "created": 1,
+                "model": "z-ai/glm-5.3",
+                "provider": "Relace",
+                "choices": [{"index": 0, "delta": {"content": "x"}}],
+            }
+        )
+        assert reply.upstream_provider == "Relace"
+        assert reply.upstream_error_type is None
+        with pytest.raises(Exception, match="incomplete tool call"):
+            handler.chunk_parser(
+                {
+                    "id": "gen-a",
+                    "provider": "Relace",
+                    "error": {
+                        "code": 400,
+                        "message": "Generation stopped with an incomplete tool call.",
+                        "metadata": {"error_type": "invalid_request"},
+                    },
+                }
+            )
+    finally:
+        request_log._http_reply.reset(token)
+    assert reply.upstream_provider == "Relace"
+    assert reply.upstream_error_type == "invalid_request"
+
+
+def test_openrouter_records_upstream_of_pre_stream_error() -> None:
+    body = {
+        "error": {
+            "message": "Provider returned error",
+            "code": 400,
+            "metadata": {"raw": "...", "provider_name": "InferenceNet"},
+        }
+    }
+    reply = request_log.HttpReply()
+    token = request_log._http_reply.set(reply)
+    try:
+        _openrouter_config().get_error_class(json.dumps(body), 400, {})
+        _openrouter_config().get_error_class("<html>502 Bad Gateway</html>", 502, {})
+    finally:
+        request_log._http_reply.reset(token)
+    assert reply.upstream_provider == "InferenceNet"
+    assert reply.upstream_error_type is None
+
+
+def test_request_timings_land_in_run_record() -> None:
+    started = datetime(2026, 10, 19, 12, 0, tzinfo=UTC)
+    event = request_log.LlmRequestEvent(
+        call_id="call-1",
+        route="litellm",
+        provider="openrouter",
+        model="openrouter/deepseek/deepseek-v4.1-flash",
+        api_host="openrouter.ai",
+        streaming=True,
+        outcome="success",
+        status_code=200,
+        provider_request_id=None,
+        response_id="gen-1",
+        error_type=None,
+        error_message=None,
+        started_at=started,
+        finished_at=started + timedelta(milliseconds=4200),
+        duration_ms=4200,
+        input_tokens=90_000,
+        output_tokens=600,
+        cached_input_tokens=85_000,
+        agent_id="agent-1",
+        time_to_first_token_ms=900,
+        details={"upstream_provider": "Together"},
+    )
+    state = ReportState.__new__(ReportState)
+    state.run_record = {}
+    try:
+        set_global_report_state(state)
+        record_llm_request(event)
+    finally:
+        set_global_report_state(None)
+
+    assert state.run_record["llm_requests"] == [
+        {
+            "started_at": "2026-10-19T12:00:00+00:00",
+            "agent_id": "agent-1",
+            "model": "openrouter/deepseek/deepseek-v4.1-flash",
+            "provider": "Together",
+            "outcome": "success",
+            "status": 200,
+            "duration_ms": 4200,
+            "ttft_ms": 900,
+            "input_tokens": 90_000,
+            "cached_tokens": 85_000,
+            "output_tokens": 600,
+            "retry_attempt": 0,
+        }
+    ]
+
+
+async def test_tool_call_timings_land_in_run_record() -> None:
+    hooks = ReportUsageHooks(model="m", record_tool_timings=True)
+    context = SimpleNamespace(tool_call_id="call-1", context={"agent_id": "agent-1"})
+    tool = SimpleNamespace(name="terminal_execute")
+    state = ReportState.__new__(ReportState)
+    state.run_record = {}
+    try:
+        set_global_report_state(state)
+        await hooks.on_tool_start(context, MagicMock(), tool)  # type: ignore[arg-type]
+        await hooks.on_tool_end(context, MagicMock(), tool, "ok")  # type: ignore[arg-type]
+    finally:
+        set_global_report_state(None)
+
+    [entry] = state.run_record["tool_calls"]
+    assert entry["agent_id"] == "agent-1"
+    assert entry["tool"] == "terminal_execute"
+    assert entry["duration_ms"] >= 0
@@ -0,0 +1,22 @@
+"""The root agent must follow up open coverage rows before finishing."""
+
+from __future__ import annotations
+
+import pytest
+
+from strix.agents.prompt import render_system_prompt
+
+
+_FOLLOW_UP = 'calls `list_coverage(outcome="needs_follow_up")` and follows up every row it returns'
+_KEEP_GENUINE_GAPS = "stays `needs_follow_up` only when it is genuinely blocked"
+
+
+@pytest.mark.parametrize("scan_mode", ["quick", "standard", "deep"])
+@pytest.mark.parametrize("is_whitebox", [False, True])
+@pytest.mark.parametrize("is_root", [False, True])
+def test_prompt_follows_up_open_coverage(scan_mode: str, is_whitebox: bool, is_root: bool) -> None:
+    prompt = render_system_prompt(scan_mode=scan_mode, is_whitebox=is_whitebox, is_root=is_root)
+
+    assert prompt
+    assert _FOLLOW_UP in prompt
+    assert _KEEP_GENUINE_GAPS in prompt
@@ -0,0 +1,159 @@
+"""Only validated model decisions may suppress a finding."""
+
+from __future__ import annotations
+
+import json
+from typing import TYPE_CHECKING, Any
+from unittest.mock import AsyncMock
+
+import pytest
+from agents.items import ModelResponse
+from agents.usage import Usage
+from openai.types.responses import ResponseOutputMessage, ResponseOutputText
+
+from strix.config import loader
+from strix.report import dedupe
+from strix.report.state import ReportState, get_global_report_state, set_global_report_state
+from strix.tools.reporting.tool import _do_create
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+    from pathlib import Path
+
+
+@pytest.fixture
+def report_state(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> Iterator[ReportState]:
+    monkeypatch.chdir(tmp_path)
+    previous = get_global_report_state()
+    state = ReportState(run_name="dedupe-decisions")
+    set_global_report_state(state)
+    state.add_vulnerability_report(title="Existing finding", severity="low")
+    try:
+        yield state
+    finally:
+        set_global_report_state(previous)
+
+
+def _mock_judge(monkeypatch: pytest.MonkeyPatch, decision: dict[str, Any]) -> AsyncMock:
+    settings = loader.load_settings().model_copy(deep=True)
+    settings.llm.model = "openai/synthetic-model"
+    model = AsyncMock()
+    model.get_response.return_value = ModelResponse(
+        output=[
+            ResponseOutputMessage(
+                id="msg_synthetic",
+                type="message",
+                role="assistant",
+                status="completed",
+                content=[
+                    ResponseOutputText(
+                        type="output_text",
+                        annotations=[],
+                        text=json.dumps({"confidence": 0.95, "reason": "Synthetic", **decision}),
+                    )
+                ],
+            )
+        ],
+        usage=Usage(),
+        response_id="synthetic",
+    )
+    monkeypatch.setattr(dedupe, "load_settings", lambda: settings)
+    monkeypatch.setattr(dedupe, "configure_sdk_model_defaults", lambda _: None)
+    monkeypatch.setattr(dedupe, "resolve_dedupe_model", lambda *_: model)
+    return model
+
+
+async def _file_finding() -> dict[str, Any]:
+    return await _do_create(
+        title="New finding",
+        description="The endpoint reflects unencoded input.",
+        impact="The reflected script executes in a victim's browser.",
+        target="https://example.invalid",
+        technical_analysis="Input is interpolated into HTML without encoding.",
+        poc_description="Open the endpoint with the reflected parameter.",
+        poc_script_code="print('synthetic reproduction')",
+        remediation_steps="Encode output for its HTML context.",
+        evidence="The synthetic response contains the unencoded parameter.",
+        assumptions="A victim follows the crafted link.",
+        counterevidence="No encoding or content security policy is present.",
+        confidence="high",
+        severity_change_conditions="A strict content security policy would lower the impact.",
+        fix_effort="low",
+        cvss_breakdown={
+            "attack_vector": "N",
+            "attack_complexity": "L",
+            "privileges_required": "N",
+            "user_interaction": "R",
+            "scope": "U",
+            "confidentiality": "L",
+            "integrity": "L",
+            "availability": "N",
+        },
+        endpoint="/search",
+        method="GET",
+        cve=None,
+        cwe="CWE-79",
+        code_locations=None,
+    )
+
+
+@pytest.mark.parametrize("value", ["false", "true", 0, 1, None, [], {}])
+def test_dedupe_parser_requires_a_boolean(value: Any) -> None:
+    with pytest.raises(TypeError, match="is_duplicate"):
+        dedupe._parse_dedupe_response(json.dumps({"is_duplicate": value}))
+
+
+@pytest.mark.parametrize(
+    "decision",
+    [
+        {"is_duplicate": "false", "duplicate_id": "vuln-0001"},
+        {"is_duplicate": "true", "duplicate_id": "vuln-0001"},
+        {"is_duplicate": 1, "duplicate_id": "vuln-0001"},
+        {"is_duplicate": True, "duplicate_id": "vuln-9999"},
+        {"is_duplicate": True, "duplicate_id": ""},
+        {"is_duplicate": True, "duplicate_id": None},
+        {"is_duplicate": True, "duplicate_id": ["vuln-0001"]},
+        {"is_duplicate": True},
+    ],
+)
+async def test_invalid_dedupe_decision_does_not_suppress_finding(
+    report_state: ReportState, monkeypatch: pytest.MonkeyPatch, decision: dict[str, Any]
+) -> None:
+    model = _mock_judge(monkeypatch, decision)
+
+    result = await _file_finding()
+
+    assert result["success"] is True
+    model.get_response.assert_awaited_once()
+    saved = json.loads((report_state.get_run_dir() / "vulnerabilities.json").read_text())
+    assert [report["title"] for report in saved] == ["Existing finding", "New finding"]
+    assert (report_state.get_run_dir() / "vulnerabilities" / "vuln-0002.md").is_file()
+
+
+@pytest.mark.parametrize("is_duplicate", [False, True])
+async def test_valid_dedupe_decisions_keep_their_meaning(
+    report_state: ReportState, monkeypatch: pytest.MonkeyPatch, is_duplicate: bool
+) -> None:
+    # An extraneous ID must not turn a valid false decision into a duplicate.
+    _mock_judge(monkeypatch, {"is_duplicate": is_duplicate, "duplicate_id": "vuln-0001"})
+
+    result = await _file_finding()
+
+    assert result["success"] is not is_duplicate
+    saved = json.loads((report_state.get_run_dir() / "vulnerabilities.json").read_text())
+    assert len(saved) == (1 if is_duplicate else 2)
+    if is_duplicate:
+        assert result["duplicate_of"] == "vuln-0001"
+
+
+async def test_dedupe_id_is_matched_without_truncation(monkeypatch: pytest.MonkeyPatch) -> None:
+    existing_id = "x" * 64
+    _mock_judge(monkeypatch, {"is_duplicate": True, "duplicate_id": existing_id + "-unknown"})
+
+    result = await dedupe.check_duplicate(
+        {"title": "New finding"}, [{"id": existing_id, "title": "Existing finding"}]
+    )
+
+    assert result["is_duplicate"] is False
+    assert "error" in result
@@ -148,3 +148,42 @@ def test_config_file_loads_dedupe_model(
    assert settings.dedupe.reasoning_effort == "minimal"
    # Main model stays independent of the dedupe override.
    assert settings.llm.model == "openai/root"
+
+
+def test_dedicated_dedupe_model_uses_own_body_fields_not_main(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LLM_EXTRA_BODY", json.dumps({"safety_identifier": "main"}))
+    loader._cached = None
+    try:
+        dedupe = DedupeSettings(
+            STRIX_DEDUPE_MODEL="deepseek/cheap",
+            DEDUPE_LLM_EXTRA_BODY={"service_tier": "flex"},
+        )
+        settings = _dedupe_model_settings(dedupe, "deepseek/cheap", 300)
+        assert settings.extra_body == {"service_tier": "flex"}
+    finally:
+        loader._cached = None
+
+
+def test_dedicated_dedupe_model_gets_no_main_body_fields_by_default(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    monkeypatch.setenv("LLM_EXTRA_BODY", json.dumps({"safety_identifier": "main"}))
+    loader._cached = None
+    try:
+        dedupe = DedupeSettings(STRIX_DEDUPE_MODEL="deepseek/cheap")
+        settings = _dedupe_model_settings(dedupe, "deepseek/cheap", 300)
+        assert settings.extra_body is None
+    finally:
+        loader._cached = None
+
+
+def test_fallback_dedupe_inherits_main_body_fields(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("LLM_EXTRA_BODY", json.dumps({"safety_identifier": "main"}))
+    loader._cached = None
+    try:
+        settings = _dedupe_model_settings(DedupeSettings(), "openai/main-model", 300)
+        assert settings.extra_body == {"safety_identifier": "main"}
+    finally:
+        loader._cached = None
@@ -0,0 +1,316 @@
+from __future__ import annotations
+
+import asyncio
+import io
+from dataclasses import dataclass
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+
+import pytest
+from agents.run_config import SandboxRunConfig
+from agents.sandbox import SandboxAgent
+from agents.sandbox.capabilities import Filesystem, Shell
+from agents.sandbox.runtime_session_manager import SandboxRuntimeSessionManager
+
+from strix.runtime import session_manager
+from strix.runtime.backends import _BACKENDS, _BIND_MOUNT_BACKENDS, register_backend
+from strix.runtime.deferred_session import DeferredSandboxClient, DeferredSandboxSession
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+
+
+@dataclass
+class _RuntimeSettings:
+    backend: str
+
+
+@dataclass
+class _Settings:
+    runtime: _RuntimeSettings
+
+
+@dataclass
+class _Endpoint:
+    host: str = "127.0.0.1"
+    port: int = 8080
+    tls: bool = False
+
+
+@dataclass
+class _ExecResult:
+    exit_code: int = 0
+    stdout: bytes = b""
+    stderr: bytes = b""
+
+    def ok(self) -> bool:
+        return self.exit_code == 0
+
+
+class _State:
+    def __init__(self) -> None:
+        self.manifest: Any = None
+        self.session_id = "00000000-0000-4000-8000-000000000001"
+
+    def model_copy(self, *, update: dict[str, Any]) -> _State:
+        copy = _State()
+        copy.manifest = update.get("manifest", self.manifest)
+        return copy
+
+
+class _Session:
+    def __init__(self) -> None:
+        self.state = _State()
+        self.execs: list[tuple[str, ...]] = []
+        self.writes: list[tuple[Path, bytes]] = []
+        self.limits: list[Any] = []
+
+    def _set_concurrency_limits(self, limits: Any) -> None:
+        self.limits.append(limits)
+
+    def _set_archive_limits(self, limits: Any) -> None:
+        self.limits.append(limits)
+
+    async def running(self) -> bool:
+        return True
+
+    async def resolve_exposed_port(self, _port: int) -> _Endpoint:
+        return _Endpoint()
+
+    async def exec(self, *argv: str, timeout: float | None = None, **_kw: Any) -> _ExecResult:
+        del timeout
+        self.execs.append(argv)
+        return _ExecResult()
+
+    async def write(self, path: Path, data: io.IOBase, **_kw: Any) -> None:
+        self.writes.append((path, data.read()))
+
+
+class _Client:
+    def __init__(self) -> None:
+        self.deleted: list[Any] = []
+
+    async def delete(self, session: Any) -> None:
+        self.deleted.append(session)
+
+
+async def _no_caido(*_args: Any, **_kwargs: Any) -> None:
+    return None
+
+
+class _Backend:
+    def __init__(self, *, fail: bool = False) -> None:
+        self.gate = asyncio.Event()
+        self.calls = 0
+        self.fail = fail
+        self.client = _Client()
+        self.session = _Session()
+
+    async def __call__(self, **_kwargs: Any) -> tuple[Any, Any]:
+        self.calls += 1
+        await self.gate.wait()
+        if self.fail:
+            raise RuntimeError("bring-up exploded")
+        return self.client, self.session
+
+
+@pytest.fixture
+def backend(monkeypatch: pytest.MonkeyPatch, request: pytest.FixtureRequest) -> Iterator[_Backend]:
+    name = f"deferred-{request.node.name}"
+    fake = _Backend(fail="fail" in request.node.name)
+    register_backend(name, fake, supports_bind_mounts=True)
+    monkeypatch.setattr(session_manager, "bootstrap_caido", _no_caido)
+    monkeypatch.setattr(
+        session_manager, "load_settings", lambda: _Settings(runtime=_RuntimeSettings(backend=name))
+    )
+    yield fake
+    _BACKENDS.pop(name, None)
+    _BIND_MOUNT_BACKENDS.discard(name)
+
+
+async def _settle() -> None:
+    for _ in range(10):
+        await asyncio.sleep(0)
+
+
+async def _create(scan_id: str, **kwargs: Any) -> dict[str, Any]:
+    return await session_manager.create_or_reuse(
+        scan_id, image="img", local_sources=[], defer=True, **kwargs
+    )
+
+
+@pytest.mark.asyncio
+async def test_deferred_bundle_returns_before_the_backend_finishes(backend: _Backend) -> None:
+    scan_id = "deferred-returns"
+    try:
+        bundle = await asyncio.wait_for(_create(scan_id), timeout=1)
+        session = bundle["session"]
+        assert isinstance(session, DeferredSandboxSession)
+        assert isinstance(bundle["client"], DeferredSandboxClient)
+        assert session.pending
+        await _settle()
+        assert backend.calls == 1
+        assert await session.running()
+        assert session.state.manifest.environment.value["PYTHONUNBUFFERED"] == "1"
+        again = await session_manager.create_or_reuse(scan_id, image="img", local_sources=[])
+        assert again is bundle
+        assert backend.calls == 1
+    finally:
+        await session_manager.cleanup(scan_id)
+
+
+@pytest.mark.asyncio
+async def test_first_io_resolves_the_sandbox_then_delegates(backend: _Backend) -> None:
+    scan_id = "deferred-io"
+    try:
+        bundle = await _create(scan_id)
+        session = bundle["session"]
+        session._set_concurrency_limits("limits")
+        exec_task = asyncio.create_task(session.exec("echo", "hi"))
+        await asyncio.sleep(0.05)
+        assert not exec_task.done()
+        backend.gate.set()
+        result = await asyncio.wait_for(exec_task, timeout=1)
+        assert result.ok()
+        assert backend.session.execs == [("echo", "hi")]
+        assert not session.pending
+        assert backend.session.limits == ["limits"]
+        assert backend.session.state.manifest is session.state.manifest
+        await session.write(Path("/workspace/x"), io.BytesIO(b"data"))
+        assert backend.session.writes == [(Path("/workspace/x"), b"data")]
+    finally:
+        await session_manager.cleanup(scan_id)
+    assert backend.client.deleted == [backend.session]
+
+
+@pytest.mark.asyncio
+async def test_failed_bring_up_surfaces_on_first_io(backend: _Backend) -> None:
+    scan_id = "deferred-fail"
+    try:
+        bundle = await _create(scan_id)
+        session = bundle["session"]
+        backend.gate.set()
+        with pytest.raises(RuntimeError, match="bring-up exploded"):
+            await asyncio.wait_for(session.exec("true"), timeout=1)
+        assert not await session.running()
+        with pytest.raises(RuntimeError, match="bring-up exploded"):
+            await bundle["caido_client"].get()
+    finally:
+        await session_manager.cleanup(scan_id)
+    assert backend.client.deleted == []
+
+
+@pytest.mark.asyncio
+async def test_cleanup_while_pending_cancels_the_bring_up(backend: _Backend) -> None:
+    scan_id = "deferred-cancel"
+    bundle = await _create(scan_id)
+    session = bundle["session"]
+    await asyncio.wait_for(session_manager.cleanup(scan_id), timeout=1)
+    assert session.bring_up.cancelled()
+    assert backend.client.deleted == []
+    assert session_manager._SESSION_CACHE.get(scan_id) is None
+    backend.gate.set()
+    await _settle()
+    assert backend.client.deleted == [backend.session]
+
+
+@pytest.mark.asyncio
+async def test_cleanup_while_pending_ignores_a_failed_late_bring_up(backend: _Backend) -> None:
+    scan_id = "deferred-cancel-fail"
+    await _create(scan_id)
+    await asyncio.wait_for(session_manager.cleanup(scan_id), timeout=1)
+    backend.gate.set()
+    await _settle()
+    assert backend.client.deleted == []
+
+
+@pytest.mark.asyncio
+async def test_extra_files_are_staged_inside_the_deferred_bring_up(backend: _Backend) -> None:
+    scan_id = "deferred-extra"
+    try:
+        bundle = await _create(
+            scan_id,
+            extra_files=[{"workspace_path": "/workspace/.strix/notes.md", "content": "hi"}],
+        )
+        assert backend.session.writes == []
+        backend.gate.set()
+        await asyncio.wait_for(bundle["session"].resolve(), timeout=1)
+        assert len(backend.session.writes) == 1
+        assert backend.session.execs
+    finally:
+        await session_manager.cleanup(scan_id)
+
+
+@pytest.mark.asyncio
+async def test_awaitable_extra_files_do_not_delay_the_backend_call(backend: _Backend) -> None:
+    scan_id = "deferred-extra-awaitable"
+    files: asyncio.Future[list[dict[str, Any]] | None] = asyncio.get_running_loop().create_future()
+    try:
+        bundle = await _create(scan_id, extra_files=files)
+        backend.gate.set()
+        await _settle()
+        assert backend.calls == 1
+        assert backend.session.writes == []
+        files.set_result([{"workspace_path": "/workspace/.strix/notes.md", "content": "hi"}])
+        await asyncio.wait_for(bundle["session"].resolve(), timeout=1)
+        assert len(backend.session.writes) == 1
+    finally:
+        await session_manager.cleanup(scan_id)
+
+
+@pytest.mark.asyncio
+async def test_awaitable_extra_files_resolving_to_none_stage_nothing(backend: _Backend) -> None:
+    scan_id = "deferred-extra-none"
+
+    async def _nothing() -> list[dict[str, Any]] | None:
+        return None
+
+    try:
+        bundle = await _create(scan_id, extra_files=_nothing())
+        backend.gate.set()
+        await asyncio.wait_for(bundle["session"].resolve(), timeout=1)
+        assert backend.session.writes == []
+    finally:
+        await session_manager.cleanup(scan_id)
+
+
+@pytest.mark.asyncio
+async def test_default_path_still_blocks_until_the_sandbox_is_up(backend: _Backend) -> None:
+    scan_id = "deferred-off"
+    backend.gate.set()
+    try:
+        bundle = await session_manager.create_or_reuse(scan_id, image="img", local_sources=[])
+        assert bundle["session"] is backend.session
+        assert bundle["client"] is backend.client
+    finally:
+        await session_manager.cleanup(scan_id)
+    assert backend.client.deleted == [backend.session]
+
+
+@pytest.mark.asyncio
+async def test_sdk_session_prep_does_not_wait_for_the_bring_up(backend: _Backend) -> None:
+    scan_id = "deferred-sdk-prep"
+    try:
+        bundle = await _create(scan_id)
+        agent = SandboxAgent(
+            name="root", instructions="x", tools=[], capabilities=[Filesystem(), Shell()]
+        )
+        manager = SandboxRuntimeSessionManager(
+            starting_agent=agent,
+            sandbox_config=SandboxRunConfig(client=bundle["client"], session=bundle["session"]),
+            run_state=None,
+        )
+        session = await asyncio.wait_for(
+            manager.ensure_session(
+                agent=agent, capabilities=list(agent.capabilities), is_resumed_state=False
+            ),
+            timeout=1,
+        )
+        assert session is bundle["session"]
+        assert session.pending
+        backend.gate.set()
+        await asyncio.wait_for(session.resolve(), timeout=1)
+        assert not session.pending
+    finally:
+        await session_manager.cleanup(scan_id)
@@ -0,0 +1,32 @@
+"""``STRIX_DOCKER_SANDBOX_DEVICES`` passes host devices through to the sandbox container."""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+from strix.runtime import docker_client
+
+
+if TYPE_CHECKING:
+    import pytest
+
+
+def test_unset_adds_no_devices(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.delenv(docker_client._SANDBOX_DEVICES_ENV, raising=False)
+    kwargs: dict[str, object] = {}
+    docker_client._apply_sandbox_devices(kwargs)
+    assert kwargs == {}
+
+
+def test_devices_are_mapped_to_the_same_path_rwm(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv(docker_client._SANDBOX_DEVICES_ENV, " /dev/net/tun, /dev/kvm ,,")
+    kwargs: dict[str, object] = {}
+    docker_client._apply_sandbox_devices(kwargs)
+    assert kwargs["devices"] == ["/dev/net/tun:/dev/net/tun:rwm", "/dev/kvm:/dev/kvm:rwm"]
+
+
+def test_existing_devices_are_kept_and_not_duplicated(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv(docker_client._SANDBOX_DEVICES_ENV, "/dev/fuse,/dev/net/tun")
+    kwargs: dict[str, object] = {"devices": ["/dev/fuse"]}
+    docker_client._apply_sandbox_devices(kwargs)
+    assert kwargs["devices"] == ["/dev/fuse", "/dev/net/tun:/dev/net/tun:rwm"]
@@ -9,9 +9,11 @@ from typing import Any, cast
 from unittest.mock import MagicMock

 import pytest
+from agents import RunConfig
 from agents.exceptions import MaxTurnsExceeded
 from agents.items import MessageOutputItem
 from agents.memory import SQLiteSession
+from agents.run_config import CallModelData, ModelInputData
 from agents.tool_context import ToolContext
 from openai.types.responses import ResponseOutputMessage, ResponseOutputRefusal, ResponseOutputText

@@ -130,6 +132,27 @@ async def test_reserve_stop_notifies_root_once(monkeypatch: pytest.MonkeyPatch)
    assert "finish_scan" in str(message["content"])


+@pytest.mark.asyncio
+async def test_reserve_notice_names_the_root_finish_tool(monkeypatch: pytest.MonkeyPatch) -> None:
+    coordinator = AgentCoordinator()
+    coordinator.root_finish_tool = "finish_pr_review"
+    await coordinator.register("root", "strix", parent_id=None)
+    await coordinator.register("child-a", "recon", parent_id="root")
+
+    sent: list[dict[str, Any]] = []
+
+    async def _record(_target_agent_id: str, message: dict[str, Any]) -> bool:
+        sent.append(message)
+        return True
+
+    monkeypatch.setattr(coordinator, "send", _record)
+
+    await _notify_root_on_budget_reserve(coordinator)
+
+    assert "call finish_pr_review" in str(sent[0]["content"])
+    assert "finish_scan" not in str(sent[0]["content"])
+
+
@pytest.mark.asyncio
 async def test_concurrent_reserve_claims_yield_single_root() -> None:
    coordinator = AgentCoordinator()
@@ -1293,6 +1316,21 @@ async def test_interactive_nudge_offers_waiting_without_repeating() -> None:
    assert "do not repeat it" in items[0]["content"]


+@pytest.mark.asyncio
+async def test_nudge_names_the_root_finish_tool() -> None:
+    items = await execution._append_tool_required_message(
+        session=None,
+        context={"parent_id": None},
+        attempt=1,
+        limit=3,
+        interactive=False,
+        root_finish_tool="finish_pr_review",
+    )
+
+    assert "call finish_pr_review" in items[0]["content"]
+    assert "finish_scan" not in items[0]["content"]
+
+
 def _cycle_with_items(
    coordinator: AgentCoordinator,
    agent_id: str,
@@ -1531,3 +1569,32 @@ async def test_autonomous_nudge_does_not_offer_the_user() -> None:
    )

    assert "wait_for_user" not in items[0]["content"]
+
+
+@pytest.mark.asyncio
+async def test_text_only_filter_scrubs_images_then_chains_existing_filter() -> None:
+    run_config = RunConfig(model="m")
+    assert execution._with_image_scrub(run_config, {"supports_images": True}) is run_config
+
+    seen: list[ModelInputData] = []
+
+    def _inner(data: CallModelData[Any]) -> ModelInputData:
+        seen.append(data.model_data)
+        return data.model_data
+
+    run_config = RunConfig(model="m", call_model_input_filter=_inner)
+    scrub: Any = execution._with_image_scrub(run_config, {"supports_images": False})
+    image = {"type": "input_image", "image_url": "data:image/png;base64,aGk="}
+    output = {"type": "function_call_output", "call_id": "c1", "output": [image]}
+    data = CallModelData(
+        model_data=ModelInputData(input=[cast("Any", output)], instructions="sys"),
+        agent=MagicMock(),
+        context=None,
+    )
+
+    result = await scrub.call_model_input_filter(data)
+    assert result is seen[0]
+    assert result.input[0]["output"] == [
+        {"type": "input_text", "text": execution._TEXT_ONLY_IMAGE_TEXT}
+    ]
+    assert data.model_data.input[0]["output"] == [image]
@@ -79,12 +79,13 @@ def test_content_guardrail_is_not_retried() -> None:
    assert execution._is_transient_model_error(guardrail) is False


-def test_client_errors_are_not_transient() -> None:
+def test_client_errors_are_transient() -> None:
    bad_request = BadRequestError(
        "bad", response=httpx.Response(400, request=_request()), body=None
    )
-    assert execution._is_transient_model_error(bad_request) is False
-    assert execution._is_transient_model_error(_status_error(404)) is False
+    assert execution._is_transient_model_error(bad_request) is True
+    for status in (401, 402, 403, 404):
+        assert execution._is_transient_model_error(_status_error(status)) is False
    assert execution._is_transient_model_error(ValueError("nope")) is False


@@ -109,6 +110,7 @@ def _patch_fast_backoff(monkeypatch: pytest.MonkeyPatch) -> None:
 async def _run_once(
    monkeypatch: pytest.MonkeyPatch,
    streams: list[_FakeStream],
+    session: Any = None,
 ) -> Any:
    _patch_fast_backoff(monkeypatch)
    calls = {"n": 0}
@@ -131,7 +133,7 @@ async def _run_once(
        run_config=cast("RunConfig", object()),
        context={},
        max_turns=5,
-        session=None,
+        session=session,
        interactive=False,
        event_sink=None,
        hooks=None,
@@ -166,9 +168,78 @@ async def test_run_cycle_gives_up_after_max_retries(
 async def test_run_cycle_does_not_retry_permanent_error(
    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
-    bad_request = BadRequestError(
-        "bad", response=httpx.Response(400, request=_request()), body=None
-    )
-    streams = [_FakeStream(exc=bad_request), _FakeStream()]
-    with pytest.raises(BadRequestError):
+    streams = [_FakeStream(exc=ValueError("nope")), _FakeStream()]
+    with pytest.raises(ValueError, match="nope"):
        await _run_once(monkeypatch, streams)
+
+
+@pytest.mark.parametrize(
+    ("status", "message", "expected"),
+    [
+        (
+            404,
+            "litellm.NotFoundError: NotFoundError: OpenrouterException - "
+            '{"error":{"message":"No endpoints found that support image input","code":404,'
+            '"metadata":{"failed_routing_step":"Filter by Image Support"}}}',
+            True,
+        ),
+        (400, "Invalid content type. image_url is only supported by certain models.", True),
+        (400, "my-model is not a multimodal model", True),
+        (400, "This model doesn't support the image field for user messages.", True),
+        (422, "messages[3]: unknown variant `image_url`, expected `text`", True),
+        (400, "'Image URLs in user messages' functionality not supported.", True),
+        (
+            400,
+            "OpenRouterException: Message: This model's maximum context length is 1048576 "
+            "tokens, but the request requires 1073484 tokens (942412 input including "
+            "image/vision expansion + 131072 for the completion). Reduce the input length "
+            "or max_tokens.",
+            False,
+        ),
+        (400, "At most 5 image(s) may be provided in one prompt.", False),
+        (
+            400,
+            "Upstream error from Relace: Generation stopped with an incomplete tool call",
+            False,
+        ),
+        (500, "No endpoints found that support image input", False),
+    ],
+)
+def test_is_image_rejection(status: int, message: str, expected: bool) -> None:
+    exc = APIStatusError(message, response=httpx.Response(status, request=_request()), body=None)
+    assert execution._is_image_rejection(exc) is expected
+
+
+async def _run_with_image_session(
+    monkeypatch: pytest.MonkeyPatch, streams: list[_FakeStream]
+) -> tuple[Any, int, int]:
+    strips = {"n": 0}
+
+    async def _fake_strip(_session: Any) -> bool:
+        strips["n"] += 1
+        return True
+
+    async def _no_compact(*_args: Any, **_kwargs: Any) -> bool:
+        return False
+
+    monkeypatch.setattr(execution, "strip_all_images_from_session", _fake_strip)
+    monkeypatch.setattr(execution, "_compact_session", _no_compact)
+    result, attempts, _coordinator = await _run_once(monkeypatch, streams, session=object())
+    return result, attempts, strips["n"]
+
+
+@pytest.mark.asyncio
+async def test_run_cycle_strips_images_on_image_rejection(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    rejection = APIStatusError(
+        "No endpoints found that support image input",
+        response=httpx.Response(status_code=404, request=_request()),
+        body=None,
+    )
+    streams = [_FakeStream(exc=rejection), _FakeStream()]
+    result, attempts, strips = await _run_with_image_session(monkeypatch, streams)
+
+    assert result is streams[1]
+    assert attempts == 2
+    assert strips == 1
@@ -2,11 +2,15 @@

 from __future__ import annotations

-from typing import Any
+from types import SimpleNamespace
+from typing import TYPE_CHECKING, Any
 from unittest.mock import MagicMock, patch

 import pytest
+from agents.usage import Usage

+from strix.config import loader
+from strix.core.agents import AgentCoordinator
 from strix.core.hooks import (
    BudgetExceededError,
    BudgetPausedError,
@@ -14,6 +18,11 @@ from strix.core.hooks import (
    SubagentBudgetReservedError,
    recomputed_budget_flags,
 )
+from strix.report.usage import LLMUsageLedger
+
+
+if TYPE_CHECKING:
+    from pathlib import Path


 def _make_hooks(max_budget: float | None) -> ReportUsageHooks:
@@ -84,6 +93,38 @@ async def test_over_budget_raises() -> None:
        await hooks.on_llm_end(_make_context(), MagicMock(), MagicMock())


+@pytest.mark.asyncio
+async def test_manual_rates_make_unpriced_model_trip_budget(
+    monkeypatch: pytest.MonkeyPatch, tmp_path: Path
+) -> None:
+    monkeypatch.setenv("STRIX_LLM", "test-model")
+    monkeypatch.setenv("STRIX_LLM_INPUT_COST_PER_MTOK", "10")
+    monkeypatch.delenv("STRIX_LLM_OUTPUT_COST_PER_MTOK", raising=False)
+    monkeypatch.delenv("STRIX_LLM_CACHED_INPUT_COST_PER_MTOK", raising=False)
+    monkeypatch.setattr(loader, "_override", tmp_path / "missing.json")
+    monkeypatch.setattr(loader, "_cached", None)
+    usage = Usage()
+    usage.requests = 1
+    usage.input_tokens = 1000
+    usage.total_tokens = 1000
+    ledger = LLMUsageLedger()
+
+    class ReportState:
+        def record_sdk_usage(self, **kwargs: Any) -> None:
+            ledger.record(**kwargs)
+
+        def get_total_llm_cost(self) -> float:
+            return ledger.total_cost
+
+    hooks = _make_hooks(0.005)
+    response: Any = SimpleNamespace(usage=usage)
+    with (
+        patch("strix.core.hooks.get_global_report_state", return_value=ReportState()),
+        pytest.raises(BudgetExceededError),
+    ):
+        await hooks.on_llm_end(_make_context(), MagicMock(), response)
+
+
@pytest.mark.asyncio
 async def test_budget_check_uses_live_cost_accessor() -> None:
    # The check must read the live ledger, not the persisted run-record snapshot,
@@ -335,6 +376,21 @@ async def test_budget_warning_root_directive_distinct_from_subagent() -> None:
    assert "confirmed" in sub


+@pytest.mark.asyncio
+async def test_root_directive_names_the_coordinators_finish_tool() -> None:
+    hooks = ReportUsageHooks(model="test-model", max_turns=100)
+    coordinator = AgentCoordinator()
+    coordinator.root_finish_tool = "finish_pr_review"
+    ctx = _make_warn_context(requests=85, parent_id=None)
+    ctx.context["coordinator"] = coordinator
+
+    items: list[Any] = []
+    await hooks.on_llm_start(ctx, MagicMock(), None, items)
+
+    assert "finish_pr_review" in items[0]["content"]
+    assert "finish_scan" not in items[0]["content"]
+
+
@pytest.mark.parametrize("parent_id", [None, "root-1"])
@pytest.mark.asyncio
 async def test_turn_warning_directive_escalates_per_stage(parent_id: str | None) -> None:
@@ -491,3 +547,30 @@ def test_recomputed_budget_flags(
    expected: tuple[bool, bool],
 ) -> None:
    assert recomputed_budget_flags(cost, max_budget, interactive=interactive) == expected
+
+
+@pytest.mark.asyncio
+async def test_turn_limit_warns_ten_before_and_repeats_past_it() -> None:
+    hooks = ReportUsageHooks(model="test-model")
+    noticed = []
+    for requests in range(40):
+        ctx = _make_warn_context(requests=requests, parent_id="root-1")
+        ctx.context["turn_limit"] = 25
+        items: list[Any] = []
+        await hooks.on_llm_start(ctx, MagicMock(), None, items)
+        if items:
+            noticed.append((requests + 1, items[0]["content"]))
+
+    assert [turn for turn, _ in noticed] == [15, 25, 30, 35, 40]
+    assert "15 of your 25-turn limit" in noticed[0][1]
+    assert "past your 25-turn limit" in noticed[1][1]
+
+
+@pytest.mark.asyncio
+async def test_no_turn_limit_no_notice() -> None:
+    hooks = ReportUsageHooks(model="test-model")
+    items: list[Any] = []
+    await hooks.on_llm_start(
+        _make_warn_context(requests=14, parent_id="root-1"), MagicMock(), None, items
+    )
+    assert items == []
Author	SHA1	Message	Date
theo-at-strix	37b8ee1bd4	fix(skills): strengthen DoubleClickjacking validation	2026-11-05 15:30:05 +02:00
theo-at-strix	aae7c97461	feat(skills): improve DoubleClickjacking detection and coverage	2026-11-05 02:46:15 +02:00
Samir Hanna Verza	09536cf148	fix(report): validate deduplication decisions	2026-11-04 21:57:34 +02:00
alex s	cb581bf93d	Feat/str 1646 opengrep (#1625 ) * feat(sandbox): replace Semgrep with Opengrep and bundle offline rules * fix(sandbox): keep opengrep-rules LICENSE and drop proprietary-licensed rules * feat(sandbox): add Strix rule for GitHub-owned actions not pinned to a commit SHA * chore: bump sandbox image to 1.4.0 (opengrep) * fix(sandbox): resolve the gitleaks release tag without the GitHub API * fix(sandbox): install a pinned, sha256-verified trivy release	2026-11-04 10:26:08 -05:00
Ahmed Allam	6b7ae2ebe4	fix(STR-1654): read manual LLM prices from settings and apply them to STRIX_LLM only	2026-10-29 04:30:48 +03:00
Alex Schapiro	92b30bcad5	fix(STR-1654): price LLM usage from model config	2026-10-29 04:30:48 +03:00
alex s	dfdbe4b2ab	fix(image): patch brotli decoder (CVE-2026-111497) in sandbox, bump tag 1.3.0 -> 1.3.1 (#1617 ) kali-rolling still ships libbrotli 1.2.0-4; node, curl and python3 in the sandbox all load the shared libbrotlidec, so a malicious Content-Encoding: br response would hit the vulnerable decoder. Install Debian's fixed 1.2.0-4.1 packages over Kali's and fail the build if the version is not applied.	2026-10-28 19:58:04 -04:00
Ahmed Allam	98e3df1425	feat(llm): send app attribution headers to Vercel AI Gateway	2026-10-28 06:04:41 +03:00
Thibault Jaigu	ffa35e22d3	docs(config): describe the current STRIX_API_TYPE default Since `99c0711` the API route follows the model's catalog entry, not LLM_API_BASE, so the doc saying a custom base URL defaults to chat completions was out of date.	2026-10-27 22:24:15 +03:00
Ian	ef5d810047	refactor(runner): default replace_scan_prompt to False in _compose_root_instructions_override	2026-10-26 22:56:57 +03:00
ian-at-strix	d6f4bc5ced	feat(runner): let root_instructions_override replace the root's scan prompt (#1592 )	2026-10-26 15:49:51 -04:00
Ahmed Allam	4ac5dc65e4	docs: show Vercel AI Gateway and OpenRouter in the README setup	2026-10-26 19:44:33 +03:00
Ahmed Allam	4bf353fb7b	docs: use the vercel_ai_gateway provider for the Vercel AI Gateway setup	2026-10-26 18:58:20 +03:00
Ahmed Allam	f121ab03fb	docs: update the recommended models in the repo markdown	2026-10-26 18:48:14 +03:00
Ahmed Allam	2c9e9f3444	docs: recommend Vercel AI Gateway alongside OpenRouter and update model ids Quickstart, providers overview and docs index show GLM-5.3 through Vercel AI Gateway and OpenRouter side by side, Vercel first. Model ids across the provider pages move to the current releases, including DeepSeek V4.1 Flash 1017, and the OpenAI, Anthropic and Gemini pages note the Daybreak, CVP and Fairwind access programs. Docs only, README unchanged. Closes STR-1686	2026-10-26 18:48:14 +03:00
Ahmed Allam	200a501f15	feat(runtime): pass host devices from STRIX_DOCKER_SANDBOX_DEVICES through to the sandbox container	2026-10-25 01:59:21 +03:00
Ahmed Allam	dc40345e26	test(agents): check the diff skill's own coverage lines, not the word	2026-10-24 21:02:27 +03:00
Ahmed Allam	fa3539cb75	fix(skills): the diff skill drops its coverage ledger instructions when the coverage tools are off	2026-10-24 21:02:27 +03:00
Ahmed Allam	aaf1f8c823	fix(agents): finish_scan drops its coverage reconciliation step when the coverage tools are off	2026-10-24 21:02:27 +03:00
Ian	bac386e509	refactor(skills): put the coverage conditionals on their own lines	2026-10-24 21:02:27 +03:00
Ian	898bbb4741	feat(agents): drop the coverage sections from skills and tool descriptions when the tools are off	2026-10-24 21:02:27 +03:00
Ian	5418c767a2	feat(agents): let callers turn off the coverage ledger tools	2026-10-24 21:02:27 +03:00
Ian	fff22045ae	refactor(prompts): put the parallel-tool-calls conditionals on their own lines	2026-10-24 20:32:27 +03:00
Ian	dca9482c82	feat(core): opt-in parallel tool calls via STRIX_PARALLEL_TOOL_CALLS	2026-10-24 20:32:27 +03:00
Ian	f46bed6740	refactor(hooks): name the soft turn-limit warning after the soft limit	2026-10-24 20:19:38 +03:00
Ian	b8d586a39b	fix(agents): an agent with several limited skills gets the largest turn limit	2026-10-24 20:19:38 +03:00
Ian	b17c0275a9	feat(agents): soft per-skill turn limits for sub-agents run_strix_scan(subagent_turn_limits={skill: turns}) gives each sub-agent the smallest limit among its skills. The agent is warned 10 turns before the limit and reminded every 5 turns past it, but never stopped.	2026-10-24 20:19:38 +03:00
alex s	14cc948c36	Make the root agent follow up open coverage rows before finishing (#1573 )	2026-10-23 10:48:00 -04:00
alex s	594398876f	Bump aiohttp, graphql-core, vite and trufflehog past known CVEs (#1569 ) - aiohttp 3.14.3 -> 3.14.5 and graphql-core 3.2.11 -> 3.2.13 in uv.lock - viewer frontend: vite 6.4.3 -> 6.4.4 - sandbox image: trufflehog 3.95.9 -> 3.99.2 (updates its embedded go-git, go-billy, amqp091-go and mongo-driver)	2026-10-22 23:28:10 -04:00
theo-at-strixandalex s	0f81f40765	feat(skills): expand XSS coverage for HTML injection and normalization (#1559 ) * feat(skills): expand XSS coverage for HTML injection and normalization * Update xss.md --------- Co-authored-by: alex s <[email protected]>	2026-10-22 01:09:33 -04:00
alex s	a3c31236bd	build(deps): enforce a 5-day release cooldown for uv and npm resolution (#1556 )	2026-10-21 11:30:08 -04:00
alex s	fde37a149f	fix(reporting): keep HTTP exchange evidence when the proxy is slow to verify ids (#1551 ) Verifying http_exchange_ids fetched every request with its full raw request and response bodies, serially, and timed out on busy proxy projects; the ids were then dropped and the finding lost its evidence. Check existence with an id-only GraphQL query, retry once, and when the proxy still does not answer keep the ids with a warning instead of discarding them.	2026-10-20 23:36:38 -04:00
ian-at-strix	3a3855fc4f	feat(agents): let callers choose the report tool (#1550 ) * feat(agents): let callers remove base tools from scan agents * feat(agents): let callers choose the report tool instead of removing tools * refactor(agents): name the report tool via the prompt template, not string replacement	2026-10-20 20:21:31 -04:00
Ian	aa2036f229	fix(llm): keep request_log.install(); record tool call timings; both only with STRIX_RUN_RECORD_TIMINGS	2026-10-20 18:26:50 +03:00
Ian	f4a7413963	feat(llm): run.json records every provider request's timing, provider and tokens in llm_requests	2026-10-20 18:26:50 +03:00
Ahmed Allam	a0f8ea2cd3	fix(core): a parked agent keeps its pause reason and drains queued messages on resume BudgetPausedError carries whether the operator pause caused it, so a resume that lands between the hook's refusal and the loop's handler still parks and continues instead of exiting under the stop policy. After the resume the agent drains its mailbox into the session before retrying the call, so a message queued while it was held is not skipped.	2026-10-18 14:58:16 +03:00
Ahmed Allam	6bd1cead1a	fix(core): an operator pause holds under every budget policy and only resume_budget lifts it	2026-10-18 14:58:16 +03:00
Ahmed Allam	ee22e2b8ed	fix(agents): the running-process marker is matched only in the result header, never in command output	2026-10-18 02:48:44 +03:00
Ahmed Allam	2b98deb3d2	fix(agents): the non-TTY note goes into the Output section so renderers keep stripping the metadata block	2026-10-18 02:48:44 +03:00
Ahmed Allam	5e9e69a07e	fix(agents): exec_command states the 30 s yield cap and that a running non-TTY process cannot be polled	2026-10-18 02:48:44 +03:00
Szymon NastalyandSzymon Nastaly	972f5097c2	feat(llm): extra request-body fields via LLM_EXTRA_BODY (#1495 ) * feat(llm): extra request-body fields via LLM_EXTRA_BODY Some provider parameters, such as OpenAI's service_tier and safety_identifier, are read only from the request body, so LLM_EXTRA_HEADERS can't set them. LLM_EXTRA_BODY merges a JSON object into the body of every request to the configured model, on both routing paths. A dedicated dedupe model gets its own DEDUPE_LLM_EXTRA_BODY instead, as with headers. * fix(config): decode saved JSON-object settings when loading persist_current() saves env values verbatim, so LLM_EXTRA_HEADERS, LLM_EXTRA_BODY and their DEDUPE_ variants land in cli-config.json as JSON text. The loader passed that text unchanged to the dict fields, so a later launch without the env var failed validation. Decode saved strings for dict fields, as pydantic-settings does for env vars. --------- Co-authored-by: Szymon Nastaly <[email protected]>	2026-10-15 16:37:51 -04:00
alex s	ca20626f84	fix(models): replay malformed tool-call arguments as valid JSON (#1327 ) * fix(models): replay malformed tool-call arguments as valid JSON A model that emits a tool call with truncated or otherwise invalid JSON arguments fails that one call, but the raw string is recorded in the session. Strict OpenAI-compatible servers (vLLM, SGLang) validate every assistant tool call in the request, so each later turn is rejected with "Assistant tool call function.arguments must be valid JSON" and the agent never recovers. The turn guard now rewrites such arguments on the way out to a JSON object carrying the original text, alongside the existing duplicate call-id repair. * fix(agents): tell the model to re-issue tool calls whose arguments are not valid JSON A streamed turn that ends mid-tool-call leaves the client with a JSON prefix as the arguments. Instead of the SDK's generic parse-error result, the tool wrapper now short-circuits with a clear 'not executed, re-issue with complete arguments' message so the model recovers in one turn. * fix(models): treat NaN/Infinity tool-call arguments as malformed * test(cost): pin model resolution in the provider-prefixed cost test * fix(models): treat over-nested tool-call arguments as malformed	2026-10-14 18:17:20 -04:00
ian-at-strix	e21f9ce4cc	test(resume-picker): put the pty in cbreak so the key-reader test can't hang on macOS (#1527 )	2026-10-14 15:55:22 -04:00
yoni-at-strix	87f2d783d6	fix(reporting): drop the static-analysis disclaimer from dependency evidence (#1520 )	2026-10-14 10:59:11 -04:00
Ahmed Allam	956a26715b	feat(runtime): accept an awaitable extra_files in create_or_reuse so the sandbox create does not wait for it	2026-10-14 03:32:29 +03:00
Ahmed Allam	0f8db1fca8	fix(runtime): keep the default path blocking on the proxy endpoint and discard a sandbox created after a cancelled deferred bring-up	2026-10-14 00:48:53 +03:00
Ahmed Allam	4bf7541599	feat(runtime): defer the sandbox bring-up behind a DeferredSandboxSession when create_or_reuse is called with defer=True	2026-10-14 00:48:53 +03:00
Ian	d8c00a8f4e	fix(proxy): log all Caido SDK errors as warnings	2026-10-12 22:18:42 +03:00
Ian	c5854e7be2	fix(proxy): log Caido timeouts as warnings	2026-10-12 22:18:42 +03:00
Ian	442b3c3df5	fix(tui): give the targets box only rows the agents panel would get, and always count hidden targets	2026-10-12 20:17:29 +03:00
Ian	710fe7b3b6	feat(tui): move the run's targets into their own sidebar box under the viewer	2026-10-12 20:17:29 +03:00
Ian	b68cdb8edc	feat(tui): list the run's targets in the sidebar stats panel	2026-10-12 20:17:29 +03:00
yoni-at-strix	429f6218f5	feat(reporting): accept code_locations on dependency reports (#1366 )	2026-10-12 08:54:36 -07:00
yoni-at-strix	f7a32ae62d	docs(readme): add update instructions (#1501 )	2026-10-10 20:31:50 -04:00
ian-at-strix	62b496430d	feat(agents): let callers choose the top-level agent's finish tool (#1496 )	2026-10-08 19:11:50 -04:00
alex s	f5a900416b	fix(mcp): load catalogs before agents discover connections (#1492 ) * fix(mcp): load catalogs before agents discover connections * fix(mcp): bound discovery waits without cancelling catalog loads * fix(mcp): keep discovery waiters alive across deadlines	2026-10-08 12:42:39 -04:00
ian-at-strix	4695290682	feat(llm): bound the first stream event and the whole stream (#1487 )	2026-10-07 18:35:45 -07:00
ian-at-strix	928f053610	feat(llm): retry 4xx model errors (#1486 ) * feat(llm): log the upstream provider behind OpenRouter LiteLLM drops OpenRouter's top-level `provider` from stream chunks (including error chunks) and the `metadata.provider_name` of non-2xx replies, so the request log could not say which provider served or failed an attempt. Record both, plus OpenRouter's `metadata.error_type`, per attempt and emit them as `upstream=` / `upstream_error=` on the llm_request line. * feat(llm): retry 4xx model errors * test: let the blocked-provider test run out of 4xx retries * refactor: retry any 4xx or 5xx status * fix: don't retry 401, 402 or 403 * fix: don't retry 404	2026-10-07 18:35:33 -07:00
ian-at-strix	39792c9d61	feat(llm): log the upstream provider behind OpenRouter (#1459 ) LiteLLM drops OpenRouter's top-level `provider` from stream chunks (including error chunks) and the `metadata.provider_name` of non-2xx replies, so the request log could not say which provider served or failed an attempt. Record both, plus OpenRouter's `metadata.error_type`, per attempt and emit them as `upstream=` / `upstream_error=` on the llm_request line.	2026-10-07 18:35:12 -07:00
ian-at-strix	278b6a280b	fix(execution): only strip images on image-rejection errors (#1461 ) * fix(execution): only strip images on image-rejection errors The image-strip recovery fired on any 400/404/422, wiping a session's screenshots and replaying the turn on unrelated errors (e.g. an upstream incomplete-tool-call 400, a context overflow). Require an image-related message too. * fix(execution): match only replies that mean the model takes no images * fix(execution): match DeepSeek's image rejection * fix(execution): guess at Vercel AI Gateway's image rejection	2026-10-07 23:59:43 +03:00
ian-at-strix	5a30961852	feat(llm): scrub images from every request to a text-only model (#1466 ) * feat(llm): scrub images from every request to a text-only model Replace images at the MCP call before any result_transform, give load_skill the same agent_browser screenshot swap as the system prompt, and add a call_model_input_filter that turns any image left in the input into a tool error while chaining any filter already set. * chore: fix the ruff and bandit failures on main	2026-10-07 23:59:19 +03:00
ian-at-strix	7f1d148778	feat(llm): keep images away from text-only models (#1465 ) * feat(llm): keep images away from text-only models Leave view_image out of the tools, replace MCP image blocks with a text note, and swap agent_browser's screenshot section for text-only guidance when the run's model does not accept images. * refactor(llm): disable view_image via is_enabled and gate skill image guidance by modality * refactor(skills): render opt-in Jinja skills instead of comment markers * Tweaked skill	2026-10-07 23:11:58 +03:00
ian-at-strix	acd64e929c	feat(llm): detect whether a model accepts image input (#1460 ) * feat(llm): detect whether a model accepts image input Add model_supports_images(): OpenRouter models are checked against architecture.input_modalities from OpenRouter's public model list (fetched once per process), others against LiteLLM's supports_vision. Unknown models and lookup failures count as image-capable. Not wired in yet. * refactor(llm): read image support from litellm's catalog, assuming it for unknown models * style: drop a trailing space in the image-support docstring	2026-10-07 22:58:43 +03:00
ian-at-strix	11976ca119	fix(logging): suppress Agents SDK PTY warning that interferes with TUI (#1392 )	2026-10-07 11:32:42 -07:00
Ian	6f2803295f	feat(tui): set the terminal title to the scan target	2026-10-07 20:41:34 +03:00