Skip to content

Agent Skills

Sentinel DV ships focused skills for the three recurring verification workflows that benefit most from a stable tool sequence.

Regression triage

Assess a run, compare a baseline, cluster failures, and prioritize investigations.

runs.* · regression.health · tests.cluster · failures.list

Failure debugging

Build a causal explanation for one failed test.

tests.* · failures.list · assertions.* · wave.*

Coverage closure

Rank gaps, assess trend and vacuity, and plan measurable closure work.

coverage.* · assertions.vacuity · runs.diff

Why skills and MCP are separate

The MCP server owns facts and controlled access. Skills own procedure:

flowchart TD
    A["Engineer request"] --> S["Sentinel DV skill"]
    S --> M["Read-only MCP tools"]
    M --> I["DuckDB index"]
    I --> R["UVM, cocotb, SVA, coverage, waveform artifacts"]
    M --> S
    S --> O["Evidence-backed report"]

This separation keeps tool outputs deterministic while allowing the workflow to evolve without embedding agent judgment in the server.

Quality contract

Every skill:

  • resolves stable IDs before detailed analysis;
  • pages through the intended result scope;
  • distinguishes observations from inference;
  • treats missing data as unavailable, not as success;
  • preserves evidence and truncation limits;
  • treats generated commands and constraints as reviewable candidates;
  • defines an explicit deliverable.

Cross-agent packaging

skills/ is canonical. scripts/sync_agent_skills.py copies it into .agents/skills, .claude/skills, and .github/skills. CI rejects drift between those locations.

.venv/bin/python scripts/sync_agent_skills.py --check
.venv/bin/python scripts/verify_skill_workflows.py

The workflow verifier indexes the checked-in demo corpus and exercises all three procedures. It is intentionally separate from scripts/verify_all_mcp_tools.py, which checks endpoint coverage for all 28 tools.

Start here

  1. Complete Agent setup.
  2. Verify runs.list from your MCP client.
  3. Invoke the skill that matches the engineering question.
  4. Include a run, suite, test name, baseline, protocol, or goal when known.