| name | ace-rtl |
| description | Consolidated ACE-RTL skill for running an iterative generator-reflector-coordinator RTL agent on CVDP. Use for CVDP official repo setup with native or Docker environments, no-cheating benchmark runs, role-based external LLM routing, unique outputs logging, task-driven CVDP prompt guidance, and real pass-rate reporting. |
ACE-RTL
Use this skill as the entry point for ACE-RTL work. It contains the agent
protocol, CVDP runbooks, role guidance, optional external LLM bridge, and
task-driven CVDP prompt guidance.
Core Rules
- Clone the official CVDP benchmark code from
https://github.com/NVlabs/cvdp_benchmark.git into this workspace and follow
that checkout's own setup instructions. Do not use a CVDP checkout, setup,
environment, or data outside the current workspace unless the user explicitly
authorizes it.
- Report final pass/fail only from the real CVDP evaluator.
- When Docker is unavailable or the daemon socket is inaccessible, use the
checked-in native runner
scripts/ace_cvdp_native_runner.py after
setup-native.md. Do not hand-roll Docker-to-native path translation.
- Native CVDP runs must execute under the repo-local
.cvdp-native Python
environment when it exists. Missing Python packages, pytest
collection/import failures, or missing harness env variables are
INFRA_SETUP, not candidate RTL failures.
- Before starting any native CVDP datapoint workers or LLM calls, verify the
selected rows' harness services have their required simulator tools on
PATH. Rows that request Xcelium/Cadence services require a configured
simulator command such as irun or xrun and usually imc; if either is
missing, use the eda_tool_setup skill to configure site-provided EDA tools
before classifying the run as setup-blocked.
- For native commercial/Cadence rows, copied harness command files must have
container paths, configured Cadence command names, and a default Cadence
timescale option rewritten before execution. The runner should inject
-timescale 1ns/1ps into xrun/irun commands unless the harness already
sets a timescale.
- Do not classify candidate RTL reset/register/output initialization failures
as
INFRA_SETUP just because a simulator or assertion line says "failed to
initialize". Setup classification is for missing tools, missing Python
modules, invalid harness environment variables, license/database/tool
initialization failures, collection failures, or native path/setup breakage.
- The native runner must post-process generated SystemVerilog before CVDP
evaluation, save bounded evaluator reports, and feed reports through the
reusable
FocusedDebugger reflector and FreshStartCoordinator coordinator
components.
- Preserve every generated target path declared by the selected CVDP row. For
multi-file targets, generator prompts must list every target file and require
one complete
// TARGET_FILE: <path> section per file.
- Native-runner generator prompts must be task-aware and data-driven from the
selected CVDP row, target paths, prompt, and harness services. Do not
hard-code a partial list of task IDs or inject one task shape's guidance into
unrelated CVDP task types.
- Reusable RTL-agent components live under
scripts/ace_rtl_agent/ and must be
launched through scripts/ace_agent_runner.py.
- Do not use hidden solutions, golden outputs, private reference code,
expected answer tables, or injected bug internals.
- Do not hardcode datapoint-specific answers or recognizable test vectors.
- Do not modify immutable harness, source, or helper files unless the task
explicitly lists them as generated targets.
- During native-runner iteration, role agents and external models may receive
public prompt/context, generated targets, current candidates, role history,
and bounded evaluator reports emitted by the real CVDP run. Do not separately
open, copy, or inject private scorer files, real test files,
expected-output code, mutation definitions, or harness internals.
- Default CVDP scheduling is 4 datapoints in parallel, with 30 max iterations
per process and 5 parallel independent processes per datapoint.
- Native evaluator subprocesses must be bounded and reaped after timeout or
service exit.
- Every CVDP run must create a unique directory under
outputs/, stream run
logs to run.log, and write a Markdown result report with overall pass rate
for the requested runtime filter, unique solved datapoint IDs, failed
datapoint IDs, and useful run metadata.
- By default, do not call external LLM APIs. Use coding-agent reasoning and
spawned role agents unless the user explicitly assigns a model to a role.
- If the user assigns a model to a role, use
scripts/llm_call.py and
scripts/extract_solution.py.
Role Agents
Spawn three role agents for non-trivial runs:
- generator: produces the next complete target RTL, testbench, assertion, or
optimization artifact.
- reflector: use the reusable
FocusedDebugger component to analyze the latest
evaluator report and create focused fix guidance.
- coordinator: use the reusable
FreshStartCoordinator component to maintain
history and decide restart.
Read references/agent-workflow.md for the full loop,
references/role-guidance.md for role prompts, and
references/agent-components.md when selecting or modifying the local reusable
implementation.
CVDP Setup
- Dataset setup:
references/benchmarks/cvdp/setup-datasets.md
- Native no-Docker setup:
references/benchmarks/cvdp/setup-native.md
- Docker setup:
references/benchmarks/cvdp/setup-docker.md
- Concrete run checklist:
references/benchmarks/cvdp/run-cvdp.md
- EDA tool setup skill:
../eda_tool_setup/SKILL.md
CVDP Task Guidance
Read references/benchmarks/cvdp/cid-index.md, inspect the JSONL row, target
paths, prompt, and harness services, then choose guidance based on task behavior
rather than hard-coded task IDs.
Optional LLM Calls
Read references/llm-routing.md only when the user explicitly assigns external
models to roles. The default hosted provider is NVIDIA Inference API with model
nvidia/nemotron-3-ultra-550b-a55b; users may replace that path with
ACE_RTL_LLM_SCRIPT.
