Agent Setup¶

Sentinel DV has two complementary layers:

1. The MCP server exposes 28 read-only tools backed by indexed artifacts.
2. The agent skills define repeatable regression triage, failure debugging, and coverage closure procedures.

Install or configure both. A skill without the MCP server has no live verification data; the MCP server without a skill still works, but the agent must plan each investigation from scratch.

Build the index first¶

Create config.yaml, then build the DuckDB index:

uvx --from sentinel-dv@2.3.1 \
  sentinel-dv-index --config /absolute/path/to/config.yaml --index-all

Use the same config path for the server. Relative artifact_roots and index.path entries are resolved from the config file's directory.

MCP configuration¶

Codex¶

Codex CLI, the IDE extension, and the desktop app share MCP configuration:

codex mcp add sentinel-dv \
  --env SENTINEL_DV_CONFIG=/absolute/path/to/config.yaml \
  -- uvx --from sentinel-dv@2.3.1 sentinel-dv-server

Equivalent ~/.codex/config.toml:

[mcp_servers.sentinel-dv]
command = "uvx"
args = ["--from", "sentinel-dv@2.3.1", "sentinel-dv-server"]

[mcp_servers.sentinel-dv.env]
SENTINEL_DV_CONFIG = "/absolute/path/to/config.yaml"

Verify with /mcp, then call runs.list.

Claude Code¶

Add a local server:

claude mcp add \
  --env SENTINEL_DV_CONFIG=/absolute/path/to/config.yaml \
  --transport stdio --scope local sentinel-dv \
  -- uvx --from sentinel-dv@2.3.1 sentinel-dv-server

For team-shared project configuration, add .mcp.json at the repository root and keep the machine-specific path in an environment variable:

{
  "mcpServers": {
    "sentinel-dv": {
      "command": "uvx",
      "args": ["--from", "sentinel-dv@2.3.1", "sentinel-dv-server"],
      "env": {
        "SENTINEL_DV_CONFIG": "${SENTINEL_DV_CONFIG}"
      }
    }
  }
}

Run /mcp or claude mcp list. Project MCP servers require workspace approval.

GitHub Copilot CLI¶

copilot mcp add sentinel-dv \
  --env SENTINEL_DV_CONFIG=/absolute/path/to/config.yaml \
  -- uvx --from sentinel-dv@2.3.1 sentinel-dv-server

Equivalent ~/.copilot/mcp-config.json:

{
  "mcpServers": {
    "sentinel-dv": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "sentinel-dv@2.3.1", "sentinel-dv-server"],
      "env": {
        "SENTINEL_DV_CONFIG": "/absolute/path/to/config.yaml"
      },
      "tools": ["*"]
    }
  }
}

Verify with /mcp list or copilot mcp list.

Other MCP clients¶

Use a stdio entry with uvx:

{
  "mcpServers": {
    "sentinel-dv": {
      "command": "uvx",
      "args": [
        "--from",
        "sentinel-dv@2.3.1",
        "sentinel-dv-server",
        "--config",
        "/absolute/path/to/config.yaml"
      ]
    }
  }
}

Skill discovery¶

The canonical source is skills/. Mirrors are committed for each host:

Available names:

• sentinel-dv-regression-triage
• sentinel-dv-failure-debugging
• sentinel-dv-coverage-closure

Codex and Claude Code can also load the repository as a plugin. The plugin packages the canonical skills/ directory; the Codex manifest also points at the bundled .mcp.json.

Verify skill loading¶

Test with a direct prompt:

Use the sentinel-dv-regression-triage skill on the latest failing run.
Report data-quality warnings and keep every run_id, test_id, and signature_id.

Security boundary¶

• All 28 MCP tools are annotated read-only.
• runs.submit and tests.replay generate commands; they do not execute simulations.
• The server reads only configured artifact roots and its index.
• Evidence excerpts are bounded and redacted according to config.yaml.
• An agent's permission system is separate from Sentinel DV. Keep normal host approvals enabled.

Troubleshooting¶

Server failed to start

Confirm uvx is on the host application's PATH and use an absolute config path.

INDEX_NOT_READY

Run sentinel-dv-index --index-all with the same config used by the server.

Skill is visible but tools are not

The skill loaded correctly, but the MCP server did not. Check the host's MCP status view and server stderr.

Tools work but the skill is not selected

Invoke the skill explicitly once and confirm the client started inside the repository containing the host-specific skill directory.

Health score looks incomplete

Inspect component_scores, effective_weights, and data_quality. Missing coverage, assertion, repeated-history, or multi-simulator cohorts are reported as unavailable rather than scored as perfect.