Deepagent-code integration demo

demos/README.md

@@ -88,3 +88,4 @@ Integrations <ovms_demos_integrations>
 |[Integration with Open WebUI](integration_with_OpenWebUI/README.md)|Using Open WebUI with OVMS as inference provider. Shows text and image generation as well as usage with RAG and tools.|
 |[Visual Studio Code assistant](./code_local_assistant/README.md)|Use Continue extension in Visual Studio Code with local OVMS serving.|
 |[Cline coding agent integration](./cline_integration/README.md)|Use the Cline coding agent in Visual Studio Code with local OVMS serving, including MCP, image input and reasoning_effort usage.|
+|[Integration with DeepAgents Code](integration_with_deepagents_code/README.md)|Using DeepAgents Code with OVMS as inference provider. Shows MCP server creation, subagent-based validation and tool use.|

demos/continuous_batching/structured_output/README.md

@@ -15,9 +15,9 @@ There are no extra steps needed to use structured output. Whole behavior is trig
 **Required:** Docker Engine installed
 
 ```bash
-mkdir models
+mkdir -p ${HOME}/models
 export GPU_ARGS=$(if ls /dev/dri/render* >/dev/null 2>&1; then echo "--device /dev/dri --group-add $(stat -c '%g' /dev/dri/render* | head -n1)"; fi)
-docker run  ${GPU_ARGS} --user $(id -u):$(id -g) -d --rm -p 8000:8000 -v $(pwd)/models:/models:rw openvino/model_server:latest-gpu --source_model OpenVINO/Mistral-7B-Instruct-v0.3-int4-cw-ov --model_repository_path /models --rest_port 8000
+docker run  ${GPU_ARGS} --user $(id -u):$(id -g) -d --rm -p 8000:8000 -v ${HOME}/models:/models:rw openvino/model_server:latest-gpu --source_model OpenVINO/Mistral-7B-Instruct-v0.3-int4-cw-ov --model_repository_path /models --rest_port 8000
 ```
 :::
 :::{tab-item} On Baremetal Host Windows

demos/integration_with_deepagents_code/.deepagents/.mcp.json

@@ -0,0 +1,10 @@
+{
+  "mcpServers": {
+    "time-server": {
+      "command": "python",
+      "args": [
+        "${DEMO_DIR}/mcp_server/time_mcp_server.py"
+      ]
+    }
+  }
+}

demos/integration_with_deepagents_code/.deepagents/agents/mcp-tester/AGENTS.md

@@ -0,0 +1,49 @@
+---
+name: mcp-tester
+description: Verifies local Python MCP stdio servers by static checks and short runtime startup checks.
+model: openai:OpenVINO/Qwen3.8-27B-int4-ov
+---
+
+You are a focused subagent for MCP server verification.
+Be concise. Do not explain your reasoning.
+Do not produce plans or long commentary.
+
+Goal:
+- Verify that a local Python MCP stdio server is runnable.
+- Return a short pass/fail report with concrete evidence.
+
+What this subagent is responsible for:
+- Inspect the target server script.
+- Verify required elements exist: FastMCP usage, at least one @mcp.tool, and stdio run call.
+- Run a syntax/import check: `python -m py_compile <server_path>`.
+- Run a short startup check: `timeout 2s python <server_path>`.
+- Read `.deepagents/.mcp.json` and confirm there is a matching stdio entry.
+
+What this subagent is not responsible for:
+- Writing or changing MCP config files.
+- Refactoring unrelated project files.
+- Long prose explanations.
+
+Rules:
+- Resolve paths from the current working directory.
+- Treat `python -m py_compile <server_path>` as authoritative for syntax: any unterminated string, malformed docstring quote, or other parse error is a hard FAIL.
+- If `py_compile` fails, include the exact syntax error line and message in the Evidence section.
+- If `timeout 2s python <server_path>` exits with 124, treat that as expected for a long-running stdio server if no traceback appears.
+- If startup exits non-zero with traceback, report failure and include the key error line.
+- Do not claim success without command output evidence.
+- Never return placeholders or template tokens such as `%stdio*`, `%args%`, `<server_path>`, `<exit_code>`, or `PASS/FAIL` literals without resolved values.
+- If any evidence line contains unresolved placeholders, set Verdict to FAIL and report `placeholder_output` as the reason.
+- Keep output short.
+
+Output format:
+1. Verdict: PASS or FAIL.
+2. Checks:
+   - script_structure: PASS/FAIL
+   - py_compile: PASS/FAIL
+   - startup_timeout_run: PASS/FAIL
+   - mcp_config_match: PASS/FAIL
+3. Evidence: exactly 4 lines, one per check, each containing:
+   - the exact command run,
+   - exit code,
+   - one concrete output fragment (stdout/stderr snippet or "no output").
+4. If FAIL: a brief description of what failed, naming the failing check(s) and the key error or missing condition.

demos/integration_with_deepagents_code/.deepagents/skills/python-mcp-sdk-skill/SKILL.md

@@ -0,0 +1,89 @@
+---
+name: python-mcp-sdk-skill
+description: "Guidance for implementing Python MCP SDK stdio servers with FastMCP, robust validation, and deterministic tool output."
+license: MIT
+compatibility: designed for deepagents-code
+---
+
+# Python MCP SDK Server Builder
+
+Use this skill when the user asks to create or modify an MCP server in Python.
+
+## Goal
+Build a minimal, production-safe MCP stdio server using the Python MCP SDK package `mcp`.
+
+## Implementation pattern
+
+1. Create `mcp_server/<name>_mcp_server.py`.
+2. Use `FastMCP`.
+3. Register one or more `@mcp.tool()` functions.
+4. Add explicit docstrings and strict input validation.
+5. Start with `mcp.run(transport="stdio")`.
+
+## Required template
+
+```python
+from mcp.server.fastmcp import FastMCP
+from mcp.types import ToolAnnotations
+
+mcp = FastMCP("your-server-name")
+
+# Reusable annotations for tools that only read state and never mutate anything.
+READ_ONLY = ToolAnnotations(
+    readOnlyHint=True,
+    destructiveHint=False,
+    idempotentHint=True,
+    openWorldHint=False,
+)
+
+
+@mcp.tool(annotations=READ_ONLY)
+def your_tool_name(arg: str) -> str:
+    """Describe exactly what this tool returns."""
+    value = arg.strip()
+    if not value:
+        raise ValueError("arg cannot be empty")
+    return value
+
+
+if __name__ == "__main__":
+    mcp.run(transport="stdio")
+```
+
+## Tool annotations — MANDATORY
+
+Every `@mcp.tool()` MUST declare MCP tool annotations. Annotations tell hosts (like dcode) whether a tool is safe to call without user approval. Tools with no annotations are treated as "possibly mutating" and are rejected outright in dcode headless mode (`dcode -n ...`) with:
+
+> "This MCP action requires approval, but the current headless runtime has no approval UI."
+
+dcode's `HeadlessMCPGuardMiddleware` gates any tool whose metadata does not satisfy BOTH:
+- `readOnlyHint is True`
+- `destructiveHint is not True`
+
+Choose the annotation set that matches the tool's real behavior:
+
+| Tool behavior                                     | Annotations                                                                                                         |
+|---------------------------------------------------|---------------------------------------------------------------------------------------------------------------------|
+| Pure read (clock, config lookup, HTTP GET)        | `readOnlyHint=True, destructiveHint=False, idempotentHint=True, openWorldHint=False`                                |
+| Read that consults an external service (web/API)  | `readOnlyHint=True, destructiveHint=False, idempotentHint=True, openWorldHint=True`                                 |
+| Idempotent write (set a value; same input twice = same result) | `readOnlyHint=False, destructiveHint=False, idempotentHint=True, openWorldHint=<True/False>`             |
+| Destructive write (delete, drop, DROP TABLE, rm)  | `readOnlyHint=False, destructiveHint=True, idempotentHint=False, openWorldHint=<True/False>`                        |
+
+Never claim `readOnlyHint=True` for a tool that mutates state. Lying to skip the approval gate is a security issue.
+
+For tools that are NOT read-only, users must run them from the interactive TUI (approval required); document that limitation in the tool's docstring.
+
+## Time tool recipe
+
+When the user needs exact current time, expose:
+- `get_current_utc_iso8601()` returning UTC timestamp in ISO 8601 with trailing Z.
+- Optional timezone helper for named zones using `zoneinfo`.
+
+Both are pure reads → annotate with `READ_ONLY` (see template above).
+
+## Quality checks
+
+- Avoid third-party dependencies beyond `mcp`.
+- Do not use network calls unless required.
+- Keep tool outputs deterministic and machine-readable.
+- **Every tool must have `ToolAnnotations` matching its real behavior** — see the table above. Verify by running the tool from `dcode -n` after implementation; if it fails with the "approval, but the current headless runtime has no approval UI" message, the annotations are missing or wrong.