Back to the catalog

asic-superpowers

Evidence-first Superpowers methodology for ASIC RTL, DV, and Physical Design agents

Open source Open in the app JSON README (API)

About

Evidence-first Superpowers methodology for ASIC RTL, DV, and Physical Design agents

Details

Kind
Plugins
Topic
No topic detected
Publisher
ariklapid
Origin
gemini
Category
ferramentas
Version
5.1.1
Stars
2
Forks
2
Last push
2026-06-14T07:30:33Z
Repository state
ativo
Language
Shell
License
MIT
Added
2026-08-30 14:13:39
Updated
2026-08-30 14:13:39
Origin id
ariklapid/asic-superpowers

README

# ASIC Superpowers

ASIC Superpowers is a lean Superpowers-style methodology plugin for ASIC
engineering agents. It helps agents plan, debug, review, and verify claims
across:

- RTL design and implementation
- Design Verification (DV), assertions, coverage, scoreboards, and regressions
- Physical Design / Backend timing, constraints, reports, ECO, and signoff-boundary work

It is not an EDA tool, simulator, synthesis runner, signoff flow, or replacement
for specialist ASIC workflows. It makes the agent work like a careful hardware
engineer: clarify intent, name the evidence, avoid unsupported claims, and state
what remains unresolved.

ASIC Superpowers is also toolchain-aware. The agent should identify your
project's configured simulator, lint, synthesis, formal, CDC/RDC, STA, physical
implementation, power, regression, and signoff/report flow when that context is
available. It stays vendor-neutral: it should not assume a specific EDA vendor,
command syntax, report dialect, waiver format, or signoff recipe unless your
repo or reports provide it.

## Quickstart

Install the plugin in your agent harness, start a clean session, and ask a
hardware-specific prompt.

### Codex

For local development from this checkout, install or load the plugin from this
repo path:

```text
/plugins
```

Then choose the local `asic-superpowers` plugin if your Codex environment exposes
local plugin sources.

### Claude Code

Use the local plugin path while developing:

```text
/plugin install /path/to/asic-superpowers
```

### OpenCode

Add this checkout to `opencode.json`:

```json
{
  "plugin": ["/path/to/asic-superpowers"]
}
```

Restart OpenCode.

### Gemini

From this repo:

```bash
gemini extensions install /path/to/asic-superpowers
```

If your harness does not support local plugin installs, use the repo's `skills/`
directory as a local skill source and make sure `using-asic-superpowers` is
loaded at session start.

## First Prompts

Use one of these after install.

### RTL

```text
Help me modify this RTL block. Start by checking what evidence you need before editing.
Files: rtl/counter.sv, filelist.f
Available evidence: lint.rpt, sim.log
Toolchain context: use the repo Makefile targets if present; otherwise ask before inventing commands.
Goal: add a programmable threshold interrupt.
```

Expected behavior:

- asks about clock/reset, interrupt clear policy, latency, interface semantics,
  and DV acceptance criteria
- identifies compile/lint/sim commands or asks for the project's toolchain
- uses `hardware-evidence-first-development` before editing
- does not claim timing, CDC/RDC, or signoff closure without reports

### DV

```text
Debug this UVM scoreboard mismatch.
Failing test: packet_smoke
Seed: 12345
Files: tb/packet_scoreboard.sv, tb/packet_monitor.sv
Evidence: sim.log, fail.wdb
Toolchain context: use the existing regression launcher and seed syntax.
```

Expected behavior:

- uses systematic debugging before proposing a fix
- asks for observed vs expected transaction behavior
- identifies simulator/regression command, waveform format, and coverage/assertion report context when available
- distinguishes DUT, monitor, scoreboard, sequence, and reference-model root causes
- does not claim full regression or coverage closure from one passing seed

### Physical Design / Backend

```text
Review this setup timing issue and SDC.
Reports: reports/report_timing.rpt
Constraints: constraints/top.sdc
Toolchain context: use the report headers, run directory, and repo flow scripts; do not assume a vendor.
Goal: reduce setup WNS on the control path group.
```

Expected behavior:

- asks for report provenance, design version, mode/corner, path group, and WNS/TNS
- identifies the backend flow stage, report dialect, SDC/MMMC setup, and run context when available
- separates structural timing-risk suggestions from STA-proven timing improvement
- preserves cycle behavior unless a latency/ECO change is explicitly approved
- does not claim signoff clean without signoff reports

## What The Plugin Adds

Core ASIC skills:

- `using-asic-superpowers` - session bootstrap and ASIC skill priority rules
- `hardware-evidence-first-development` - evidence-first workflow for RTL, DV,
  constraints, reports, and backend artifacts

ASIC reference lenses:

- `asic-engineering-contract.md` - intake contract for RTL, DV, and backend work
- `rtl-design-lens.md` - RTL implementation/review checklist
- `dv-verification-lens.md` - DV, assertion, coverage, and scoreboard checklist
- `physical-design-lens.md` - timing, SDC/MMMC, congestion, ECO, and signoff-boundary checklist
- `asic-review-checklist.md` - review rubric for hardware diffs
- `hardware-claim-discipline.md` - claim-to-evidence table
- `eda-toolchain-profile.md` - vendor-neutral toolchain discovery and context capture
- `tool-evidence.md` - how to treat tool, MCP, and report output

The plugin keeps the upstream Superpowers lifecycle: brainstorming, planning,
systematic debugging, code review, verification before completion, worktrees,
and branch finishing.

It routes by required reasoning, not by the user's job title or repo domain.
Plain software/tooling work by an ASIC engineer should use the inherited generic
Superpowers flow. Tasks that require RTL/DV/EDA interpretation, hardware-safe
behavior, or hardware completion claims use ASIC Superpowers. Mixed tasks split
the boundary: generic flow for the software mechanics, ASIC evidence discipline
for interpreting hardware artifacts or reports.

## Minimum Useful Inputs

You can start with partial context. The agent should ask for what is missing.

For RTL work, useful inputs are:

- RTL files, packages, interfaces, and filelist
- compile/elab/lint/sim targets or scripts
- clock/reset assumptions
- protocol or block spec
- compile, lint, sim, assertion, or formal logs
- waiver files or lint configuration when relevant

For DV work, useful inputs are:

- failing test and seed
- simulation log and waveform path
- UVM component scope: sequence, driver, monitor, scoreboard, reference model
- assertion or coverage report
- regression command
- simulator/regression launcher, seed syntax, coverage database format

For Physical Design / Backend work, useful inputs are:

- timing, congestion, utilization, power, DRC/LVS, or ECO reports
- SDC/MMMC files
- mode, corner, path group, startpoint/endpoint
- netlist/DEF/design version
- signoff boundary and ECO limits
- run directory, flow stage, report command, library/parasitic context when visible

## Evidence Discipline

ASIC Superpowers requires narrow, evidence-backed claims:

- Compile success is not design correctness.
- Lint clean is not simulation clean.
- Simulation pass is not formal proof.
- Structural timing-risk reduction is not STA timing closure.
- CDC/RDC review is not CDC/RDC signoff.
- SDC syntax is not constraint correctness.
- DRC/LVS/antenna/IR/EM claims require the relevant signoff reports.
- Vendor-specific interpretation requires visible toolchain evidence.

When evidence is missing, the agent should say what is unresolved instead of
upgrading the claim.

## Validation Status

Current status as of 2026-05-23:

- ASIC plugin metadata and bootstrap have been renamed from upstream Superpowers.
- `using-asic-superpowers` and `hardware-evidence-first-development` exist and validate.
- RTL, DV, Physical Design, EDA toolchain, review, evidence, and claim-discipline references exist.
- README, OpenCode docs, validation plan, and fixture provenance are current.
- Deterministic metadata, link, trigger, and fixture validation passes locally.
- Live harness transcript evals are still required before an industry-grade release claim.

Run:

```bash
scripts/validate.sh
```

This checks:

- skill metadata and required skills
- ASIC bootstrap wiring
- local markdown links
- RTL/DV/Physical Design trigger scenarios
- eval fixture presence and provenance
- vendor-neutral EDA toolchain-awareness references

`npm run validate` is a convenience wrapper when npm is installed. The shell
launcher selects Python 3.7+ automatically, so workstations where `python3` is
older can set `PYTHON=/path/to/python3.11` or rely on a versioned
`python3.11`/`python3.10` binary in `PATH`.

The repo also validates the new skills with the system skill validator and the
plugin with the system plugin validator. Live harness transcript evals are still
required before calling a release industry-grade.

## Upstream Release Monitoring And Sync

A weekly GitHub Actions workflow checks published, non-prerelease releases from
`obra/superpowers`. Every stable release after the `v5.1.0` baseline gets one
issue titled `Upstream obra/superpowers <tag> baseline review`, including
releases with no baseline-file changes. Open and closed issues both suppress
duplicates.

The issue compares adjacent stable releases and classifies changed upstream
paths as:

- `candidate-generic`: inherited generic behavior eligible for selective review
- `asic-owned`: local ASIC behavior that upstream must not overwrite
- `mixed-manual`: integration or identity files that require a hand merge

Local-only ASIC files do not appear in the comparison because the monitor diffs
two upstream release tags. Unknown upstream paths default to
`candidate-generic`, but nothing is applied automatically.

To test discovery without creating issues, manually dispatch the
`Upstream obra/superpowers release monitor` workflow with `dry_run` enabled, or
run locally with a GitHub token:

```bash
GITHUB_TOKEN="$(gh auth token)" npm run monitor:upstream-releases -- \
  --destination-repo ariklapid/asic-superpowers \
  --dry-run
```

After an issue is reviewed, ask your coding agent to use
`syncing-upstream-superpowers` and prepare the existing selective sync report:

```bash
npm run sync:upstream
```

The report writes `candidate-generic.patch` plus a combined
`protected-manual.patch`. After the reviewed sync passes validation, update the
tracked upstream marker:

```bash
npm run sync:upstream:mark
```

## Future Plan

Next release work:

1. Run clean live transcript evals in at least one target harness for RTL, DV,
   Physical Design, and vendor-neutral EDA toolchain prompts.
2. Add transcript artifacts or summaries under `evals/` without exposing private
   project data.
3. Pressure-test unsupported claim prompts: lint clean, CDC clean, timing
   closed, signoff clean, scoreboard fixed, and generic EDA report assumptions.
4. Iterate only on failures observed in transcripts, then re-run `scripts/validate.sh`.
5. Add optional deeper workflow bridges to `asic-ai-workflows` only when a real
   repeated task needs them.
6. Keep the plugin lean and vendor-neutral; do not add mandatory EDA
   dependencies or vendor-specific rule packs to core.

## Fixtures And Evals

The repo includes compact eval fixtures under `evals/fixtures/`:

- MIT-licensed local fixtures from `asic-ai-workflows`
- permissively licensed public SDC, lint, timing-summary, power, and DV examples
- one hand-authored synthetic `report_timing.rpt` fixture for deterministic tests

See `evals/fixtures/PROVENANCE.md` for sources, licenses, and limitations.

Trigger scenarios live in `evals/trigger-scenarios/scenarios.json`.

## Scope Boundaries

Use specialist tools and repos as escalation paths when needed:

- `asic-ai-workflows` for deeper structured handoff/report workflows
- HDL MCPs for compiler-backed source context
- UPF/low-power MCPs when available
- EDA tools only when the user has provided the flow context and the active
  workflow supports that phase
- project-local EDA commands should be preferred over invented commands, and
  mutating implementation/signoff commands require explicit approval

ASIC Superpowers stays zero-runtime-dependency and methodology-focused.

## Full Validation Plan

See `docs/ASIC_PLUGIN_VALIDATION_PLAN.md` for the complete release validation
flow, including deterministic checks, live harness transcript evals, and
adversarial claim-pressure prompts.

More