Read-only MCP for design verification
Sentinel DV¶
Give Codex, Claude, Copilot, and other MCP clients structured access to SystemVerilog, UVM, cocotb, assertion, coverage, regression, and waveform evidence.
From regression noise to traceable evidence¶
- Index exported logs, JUnit XML, assertion status, coverage reports, and waveform summaries.
- Query 28 read-only MCP tools with stable IDs, filters, pagination, schemas, and evidence references.
- Investigate with focused skills for regression triage, failure debugging, and coverage closure.
- Decide with observations, inference, missing-data warnings, and reproducible follow-up criteria kept separate.
Three engineering workflows¶
| Workflow | Engineering question | Evidence path |
|---|---|---|
| Regression triage | What changed, and which failures matter first? | Run summary → health quality → clusters → baseline diff → focused evidence |
| Failure debugging | What initiated this test failure? | Test context → chronological events → assertions → topology → bounded waveform window |
| Coverage closure | Which valid gap should we close next? | Summary → trend → run-scoped gaps → vacuity → reviewed constraint candidate |
Watch the real workflow¶
Open the scene index and verification commands
Verification data model¶
Execution¶
Runs, tests, framework, simulator metadata, seeds, duration, CI context, replay command generation, and live status snapshots.
Debug¶
Normalized failure categories, stable signatures, bounded evidence, UVM topology, assertion definitions and failures, VCD and JSON waveform summaries.
Closure¶
Functional, code, toggle, and FSM metrics; run diffs; vacuity and SVA status; trend analysis; gap ranking; protocol-aware candidate constraints.
Designed for controlled agent access¶
Sentinel DV exposes exported verification artifacts, not EDA control. Every MCP tool is annotated read-only. Paths are constrained to configured artifact roots; excerpts and response sizes are bounded; optional redaction protects credentials, email addresses, and local paths.
runs.submit and tests.replay are intentionally named workflow tools, but they only generate dry-run commands for engineer review.
Supported artifacts¶
| Source | Indexed evidence |
|---|---|
| UVM logs | Reports, components, topology hints, failures, phases |
| cocotb / JUnit XML | Test status, duration, exception and failure details |
| Assertion exports | Definitions, intent, runtime failures, SVA pass/fail/vacuity status |
| Coverage exports | Functional, code, toggle, FSM, per-scope metrics and gaps |
| Waveforms | Precomputed *.wave.json and bounded VCD summaries |
| CI metadata | Suite, run, build identifiers, status and timestamps |
Commercial waveform databases such as FSDB and WLF are not streamed. Export a bounded summary or VCD inside an allowed artifact root.
Verify before adoption¶
The repository includes deterministic checks for the complete MCP surface and the published skills:
The checked-in demo corpus covers UVM, cocotb, AXI4, assertions, coverage, JSON waveform summaries, VCD, and exported VCS, Xcelium, Questa, and Verilator artifacts.