bootstrap-asic-project

name

bootstrap-asic-project

description

Bootstrap and coordinate a fresh open-source ASIC project when the PDK, EDA runtime, project configuration, or physical-design collateral is missing or ambiguous. Use for requests to choose an open PDK, provision compatible tools, create a reproducible ASIC workspace, turn RTL such as a CPU into a core or full-chip tapeout candidate, add an IO/pad ring, or stage-gate synthesis, place-and-route, timing, DRC, LVS, and final handoff without confusing open-tool evidence with foundry signoff.

Bootstrap ASIC Project

Turn an underspecified chip idea into a frozen project context and a sequence of honest engineering gates. Use $openada:openada for every implemented OpenADA assertion. Keep provisioning, design creation, native gap work, and foundry acceptance distinct from OpenADA-normalized evidence.

Define the finish line first

Record the intended deliverable before installing or generating anything:

• core: synthesized or routed hard macro without package-facing IO;
• full-chip candidate: core plus IO/ESD cells, power pads, bondpads, pad ring, seal ring, fill/density, final GDS and reference netlist; or
• submission candidate: a full-chip candidate checked against the current shuttle checklist and ready for the foundry's intake review.

Also record the function, interfaces, clock, reset, voltage domains, die or area limit, package/bonding assumption, license constraint, target shuttle, and required verification. Do not let “CPU” silently mean only synthesizable RTL. If the user authorizes autonomous choice, choose the smallest target consistent with these constraints and document the decision. Otherwise ask one choice that materially changes the process or padframe.

Inspect in the safe order

1. Establish one writable project root and a separate evidence root. Treat PDK, official template, and shared source checkouts as read-only.
2. Resolve the OpenADA CLI exactly as $openada:openada specifies. Record the plugin and CLI versions separately; a plugin marketplace installation does not install the Python entry point or EDA products.
3. Check free disk, memory, architecture, container/Nix availability, and an already-present compatible runtime before downloading tools. Prefer a pinned compatible runtime already on the machine over unrelated “latest” binaries.
4. Run openada capabilities, then one scoped preflight for only the smallest next implemented assertion. A preflight pass is tool readiness, not a design result.
5. Read configured PDK roots and explicit project context. Do not recursively crawl the host or a PDK tree to compensate for an empty catalog. Do not infer that an empty preflight PDK list means no PDK exists.

--profile iic-osic-tools recognizes an existing /foss layout; it does not launch Docker or install that layout. Treat OCI/Nix launch and mount policy as a separate provisioning decision.

Select one coherent target stack

Compare candidates using official current sources: process status, license, voltage and device needs, standard-cell and IO libraries, SRAM/macros, supported flow, public DRC/LVS decks, shuttle availability, package constraints, and the quality of a complete full-chip example. Never splice a CPU synthesized for one library into DRC/LVS evidence from another PDK.

For an IHP SG13G2 full-chip request, read references/ihp-sg13g2-full-chip.md before provisioning. It records the evaluated official template and its coupled PDK, tool, padframe, verification, and submission assumptions. Re-verify current submission dates and required decks with IHP; do not silently float one pin.

For another PDK, require the same classes of authoritative collateral. If a complete open padframe or final LVS path is unavailable, narrow the deliverable to a core or report the full-chip row as not evaluated.

Freeze a bootstrap manifest

Create .openada/bootstrap-manifest.json before the first expensive run. Use the bundled identity-ledger helper rather than inventing a project-wide ambient profile. First write a project specification covering the finish-line choices above and a hash-ordered source/firmware manifest:

Use a lowercase stable ledger key for --pdk-id, not necessarily the PDK's native display name. For example, use sky130a for native name sky130A, while retaining the native spelling and exact path in project/runtime context.

python3 /path/to/skill/scripts/bootstrap_manifest.py init \
  --output /absolute/project/.openada/bootstrap-manifest.json \
  --project-root /absolute/project --name cpu-chip --deliverable full-chip \
  --top chip_top --project-spec /absolute/project/PROJECT_SPEC.md \
  --source-manifest /absolute/project/SOURCE_MANIFEST.sha256 \
  --template-origin https://github.com/IHP-GmbH/ihp-sg13g2-librelane-template \
  --template-revision-scheme git-sha1 --template-revision TEMPLATE_COMMIT \
  --template-lock /absolute/project/flake.lock \
  --pdk-id ihp-sg13g2 --pdk-root /absolute/pdk-root \
  --pdk-revision-scheme git-sha1 --pdk-revision PDK_COMMIT \
  --runtime-kind oci --runtime-identity /absolute/project/OCI_RUNTIME.json \
  --runtime-profile iic-osic-tools \
  --image-reference registry/image@sha256:DIGEST \
  --image-platform linux/amd64 --flow-name librelane --flow-tool librelane \
  --flow-revision-scheme git-sha1 --flow-revision FLOW_COMMIT \
  --flow-config /absolute/project/librelane/config.yaml \
  --evidence-root /absolute/scratch/evidence

Use stable IDs with bind-file for every authoritative Liberty, LEF, GDS, CDL, DRC, LVS, revision attestation, padframe, package, and checklist input, and set-tool for each native tool path plus its separately observed version. Then run freeze followed by validate --require-frozen. See references/project-manifest.md for the deliverable-dependent requirements, lifecycle, closed roles, and claim boundary. When adding gaps, use only stages project, rtl, function, synthesis, physical, timing, padframe, drc, lvs, handoff, or submission, and kinds capability, collateral, compatibility, evidence, external-acceptance, or resource.

This ledger proves only structurally declared, current hash identities. It does not parse role contents, probe compatibility, or establish engineering status. It is not an OpenADA operation profile, driver capability, conformance result, or signoff certificate. Keep operation result envelopes and native run directories beside it as separate evidence.

Provision without creating a version lottery

Provision only after the target stack is selected and the user has authorized installation or the original request already says to provision autonomously.

• Pin source repositories by full commit and runtime images by manifest digest.
• Keep one compatible tool/flow/PDK set. Do not combine a current flow with an old OpenROAD binary, then patch around command incompatibilities.
• Estimate required bytes before pulling an image, PDK, or large tool bundle. Put large disposable outputs on a filesystem with adequate headroom.
• Use the runtime's reviewed entry point and closed mounts. Mount the project writable; mount the PDK, flow source, and OpenADA source read-only; keep network disabled during EDA execution after setup.
• Run bounded version probes inside the execution environment and record exact paths and versions in the manifest.

If provisioning fails, retain the exact failed gate and diagnostic. Do not download several unrelated tool distributions as speculative recovery.

Start from the right design shell

For a full chip, begin from the PDK maintainer's reviewed full-chip template, not a blank RTL directory or a core-only P&R example. Freeze the template before editing a writable project copy. Preserve its IO/ESD, power-domain, bondpad, seal-ring, fill, density, extraction, and verification structure while replacing the example core.

Define CPU programmability and memory explicitly. A fixed self-test ROM can demonstrate execution but is not a generally programmable CPU. A serious CPU candidate needs a pinned core, memories or external bus, boot/loading path, reset sequencing, firmware identity, and tests that cover those boundaries. Audit every RTL, firmware, memory image, macro, and generated-artifact license.

Advance one evidence gate at a time

Gate	Minimum evidence	Stop condition
Project freeze	Frozen, hash-consistent identity ledger plus reviewed project specification	Any unresolved PDK, top, flow, deck, runtime, or mislabeled-input question
RTL structure	OpenADA rtl-check pass for exact sources/top	Engineering fail or unknown
Strict lint	openada.operation/rtl.lint/v1alpha1 pass	Any normalized warning/error
Function	openada.operation/rtl.test/v1alpha1 pass for a declared self-checking top plus architectural/firmware tests when the fixed Icarus or Verilator mapping fits	Missing, failing, unknown, or inadequately self-checking behavior
Synthesis	openada.operation/logic.synthesize/v1alpha1 pass when its request can express the stack	Unmapped cells, incomplete closure, or unsupported multi-library handoff
Physical implementation	Fresh floorplan/PDN/place/CTS/route/GDS run with stage reports	Congestion, route, antenna, clock, power, or stream-out failure
Timing	All declared corners and routed parasitics reviewed	OpenADA timing is currently one-corner ideal-interconnect only
Full-chip assembly	IO/ESD, power pads, bondpads, pad ring, seal ring, fill and density present	Core-only GDS or unresolved package/power assumptions
DRC	Official selected deck, fresh exact report, zero unexplained markers	fail, unknown, stale report, blanket waiver, or incomplete deck scope
LVS	Final GDS extraction versus an independently generated powered reference netlist	Layout-versus-layout comparison, mismatch, or incomplete hierarchy
Handoff	Hashed GDS/netlists/reports, pinout, package plan, waiver and limitation ledger	Any required artifact or foundry checklist row missing

Preflight only the next OpenADA-supported assertion immediately before that operation. Preserve execution.status separately from engineering.status. The scoped doctor catalog does not yet map the RTL self-test assertion. For the Function gate, inspect the selected rtl-test backend tools directly and retain that readiness as preflight evidence without claiming that the design assertion was evaluated. Route RTL review to $openada:review-rtl-architecture, mapping review to $openada:assess-synthesis-and-inference, and supported synthesis-stage timing to $openada:assess-asic-timing.

Handle missing OpenADA operations explicitly

OpenADA's rtl.test/v1alpha1 operation covers one bounded, declared self-checking HDL top through its fixed Icarus or Verilator mappings. It does not establish stimulus or assertion adequacy and does not expose arbitrary native flags or a general simulator workflow. OpenADA currently has no semantic operation for floorplanning, place/CTS/route, padframe generation, extraction/PEX, routed MCMM timing, power integrity, or final submission assembly.

The default action is not evaluated — capability unavailable. If the user explicitly requests an exploratory end-to-end run and authorizes native gap work, execute the reviewed project flow only as a labeled native operation and record:

• why no OpenADA profile applied;
• exact executable, version, argv, environment, cwd, timeout, and exit;
• all declared input and output hashes;
• native reports and the independent checks used to interpret them; and
• the missing semantic operation that should be added later.

Never place native output inside an openada.result envelope, call it OpenADA evidence, or let its success promote driver maturity. Return to OpenADA for DRC or LVS only when their exact caller-supplied inputs and driver contracts fit.

Repair and resume

Classify each failure before changing anything: design, constraint, collateral, tool/flow compatibility, environment, evidence contract, or resource capacity. Change one class at a time, preserve the failed evidence directory, and rerun into a fresh output path. Use thaw --reason, an explicit replacement command, and freeze when a frozen input intentionally changes; explain why prior evidence is no longer authoritative. Resolve a retained gap only with an explicit resolution—never by deleting its history.

Report the honest conclusion

Return a gate table with pass, fail, unknown, or not evaluated; exact tool, PDK, template and source revisions; artifact paths and hashes; fixes and retries; OpenADA gaps; native-gap evidence; and submission limitations.

Use open-PDK tapeout candidate only when every required pre-intake row is proved. Use submission candidate only when the current program checklist is complete. Foundry review and acceptance remain external; signoff: not claimed.