Changelog

OpenADA follows semantic versioning for the Python package and agent plugins. Contract, operation-profile, and conformance identifiers have independent versions as described in the compatibility policy.

0.4.0 — Unreleased

Changed

• The driver owns the simulator's startup file, not just the deck. The collateral refusal read the deck, and the deck was never where the damage came from. ngspice reads a startup file before it reads any deck; IHP's PDK ships <pdk-root>/ihp-sg13g2/libs.tech/ngspice/.spiceinit, which preloads its own PSP103 modules from $PDK_ROOT/$PDK; the workstation image installs it in the user's home; and a binding exports PDK and PDK_ROOT. So every sky130A run in the container preloaded IHP's compact model out of SkyWater's tree, where it has never existed — one PDK's incantation applied to another, written by nobody, invisible in every deck.

  A --pdk run now writes its own openada-pdk.spiceinit, hands it to ngspice explicitly and thereby suppresses every ambient spinit and .spiceinit. The file is a retained, content-bound input, and pdk_startup renders it from the reviewed profile: near-empty by design, because everything a bound deck needs is in the bound deck. A startup file the caller names with --init-file or --system-init-file is held to the same collateral rules as a deck, and the osdi card spelling — IHP's — is recognised alongside pre_osdi.

• A model library that will not load is a fact about collateral, not about the output. simulation.native_error swept up ngspice's OSDI preload failures and simulation.result.malformed then reported them as "the retained result is unreadable". It fired on a run that converged over 181 points, wrote a valid 8947-byte raw file and printed ngspice-46 done. Preload failures are now simulation.collateral.unloadable, a warning when the analysis produced structurally valid evidence anyway — no device in the deck needed the module — and an error when it did not. More generally, anything ngspice prints above its own banner came from a startup file and cannot be evidence about a deck it had not yet read; a non-preload error there is simulation.startup.failed. A capture with no banner is classified exactly as before.

• A number is not measured until a typed envelope says so. Answers reported self-derived quantities as "Measured" four times, independently of every simulation fix, because the typed chain took the result envelope as a file and the envelope only ever went to stdout — so hand-parsing the .raw was always one step shorter than doing it properly. A completed simulate now retains its own envelope as simulate.result.json, writes a ready-to-run simulate.selection.json next to it — the raw's real vector names, units inferred from them, and the axis excluded, since selecting the axis is refused with a diagnostic that names neither — and prints the exact openada extract command as claim.measurement.typed_chain. A run that establishes nothing instead carries claim.measurement.unsupported (error), which states in words that no quantity from it may be called measured, simulated or verified.

• An AC sweep with nothing driving it is now named. simulation.stimulus.absent fires when a deck declares .AC and no independent source carries an AC magnitude. Such a run converges over every point, returns identically zero everywhere, and is a perfectly legitimate pass — execution and engineering are about the run, not about whether the experiment asked anything. A live job reported a 280 MOhm cascode output impedance off exactly this shape (its V_OUT_STIM had a DC value and no AC card), and nothing in the evidence contradicted it. Re-running that published deck now shows the warning, and its extracted series is zero at all 181 points.

• --analysis no longer requires --backend. It requires the semantic operation, which --pdk, --models and an artifact target already select; demanding the flag as well cost one observed session four turns. Naming an analysis a PDK-bound deck does not declare is now refused with simulation.request.invalid rather than silently ignored while the deck's own analysis ran.

• One simulation semantic. openada.operation/circuit.simulate/v1alpha2 is now the only simulation operation and openada simulate the only verb. Its target is a SPICE deck or a published Simra schematic.artifact.json, detected by reading the file; its model source is nothing, a flattened --models card file, or an installed PDK bound by --pdk; and an artifact declaring several analyses is split into one single-analysis deck per declaration and every one is run. Every simulation, from every target, with every model source, now returns exactly one circuit.simulate/v1alpha2 envelope — a PDK-bound run previously returned a raw native ngspice payload with no analysis or evidence block at all.

  openada.operation/testbench.simulate/v1alpha1 is retired: it was a second operation that meant "simulate", and two semantics for one engineering act is how a live request took the path that consulted no PDK binding. Its profile document is marked deprecated and is bound by no driver and no surface; the org.openada.driver.simra.testbench.* identities are removed, so there is one driver identity per backend. openada testbench-simulate survives as a deprecated alias that emits simulation.operation.deprecated and delegates — the result it returns is a circuit.simulate result.

  The alias has a stated end. It is removed in 0.6.0, and not before 2026-11-01. The date and version are stated in the retired profile's extensions["org.openada"].deprecation, in openada.operations.testbench_simulate.REMOVAL_VERSION / REMOVAL_NOT_BEFORE, and in the warning text every alias call returns; a test asserts the three agree. openada.operations re-exports only the alias entry point the CLI dispatches to — TESTBENCH_SIMULATE_PROFILE and resolve_testbench_driver are no longer a second spelling there, since every re-export of a retired name is a fresh way for it to acquire a caller.

  profiles/circuit.simulate-v1alpha2.json now declares 25 of the 26 diagnostics the unified path could emit but had not published. simulation.operation.deprecated is deliberately not declared: it is a fact about the retired alias, not about circuit.simulate, and it disappears with the alias — declaring a code meant to disappear inside an identifier the compatibility policy treats as immutable is how a retired semantic becomes permanent.

  The binding facts moved from data.extensions["org.openada.pdk_binding"] to data.extensions["org.openada.pdk-binding"]: the underscore spelling is illegal under the circuit.simulate/v1alpha2 extension-name pattern, which is itself evidence the two operations had drifted apart.

• A netlist carries devices, not technologies. A deck that binds PDK collateral by hand — any pre_osdi, or a .lib/.include reaching into an installed PDK's tree — is refused with pdk.collateral.hand_bound, even when its paths are correct, because such a deck can only ask its question in one technology. pdk.collateral.foreign names the case where one PDK's incantation was applied to another (the observed live failure: IHP's psp103.osdi preloaded at a sky130 path); pdk.collateral.missing the case where a deck would bind nothing; pdk.collateral.conflict a hand-written prelude handed to --pdk. --unmanaged-collateral is the one way past and stamps the result with a permanent provenance limitation.

• nmos.svt/pmos.svt are accepted as synonyms of nmos.core/pmos.core.

Added

• The repository root is now a conformant Agent Plugins 1.0.0 package: a root plugin.json carries the standard $schema and the same metadata as the client-specific manifests, and the existing skills/ tree is the standard's fixed skill discovery location. Clients implementing the vendor-neutral standard can install OpenADA directly from the repository; the .claude-plugin/, .codex-plugin/, and .agents/ manifests are unchanged and stay version-locked with the root manifest.

• One canonical netlist, four PDKs. A published Simra deck now names a technology-independent device role (nmos.core, pmos.core, and .lvt/.hvt/.io variants) and SI geometry, and the binding profile translates it — so the same bytes bind to ihp-sg13g2, sky130A, gf180mcuD and freepdk45 with nothing but --pdk changing. A deck naming one PDK's own model is translated through the same role index, so every artifact published so far keeps working; each substitution is reported under data.extensions["org.openada.pdk-binding"].model_translations.

  PdkBinding now carries the four properties that previously had no representation and made every non-IHP PDK fail:

  • geometry_scale — the .option scale a PDK's own collateral installs. sky130A's libs.tech/ngspice/all.spice:2 sets scale=1.0u, reached from every corners/<corner>.spice, so instance geometry must be a plain micron number. An SI-valued card is scaled a second time, lands outside every model bin, and ngspice reports only could not find a valid modelname. Geometry is now converted at bind time and the bound deck states its own convention.
  • library_entries — an ordered prelude of .include and .lib <file> <section> cards rather than one path. gf180mcuD's corner library evaluates global switches (fnoicor, sw_stat_global) defined only in design.ngspice, which must be included first; without it every model card fails to evaluate. Both a path and a section may carry {corner}, so a corner can be a library section or a directory — FreePDK45 has no library sections at all.
  • device_geometry — the binning envelope the simulator enforces, read out of the PDK's own model cards. A device outside it is refused with pdk.device.geometry_out_of_range naming the dimension and the legal range, instead of surfacing as a missing model.
  • analog / unsupported_reason — asap7, nangate45 and gt2n ship LEF, Liberty and GDS and no transistor models for any simulator. They are registered so that requesting one fails with that reason rather than "unknown PDK", which an agent answers by trying another spelling.

  sky130A's previous entry was declared from the published layout and never exercised; it was wrong on two counts (its FETs are subcircuits, not .model cards, and it needs the micron convention) and is corrected here. --timeout defaults to 600 s because parsing sky130.lib.spice tt alone takes ~95 s.

• Per-PDK binding profiles for testbench-simulate, selected with --pdk, --pdk-root, and --corner. A published Simra testbench is deliberately model-free, so every MOS artifact previously required a hand-flattened model-card file; a binding profile now supplies the collateral from an installed PDK instead. The profile owns the parts that genuinely differ between PDKs and fail silently when guessed: the device prefix (IHP SG13G2 ships its MOS devices as subcircuits, so the emitted M card is rewritten to X), the parameter spelling (IHP's finger count is ng, while Simra emits NF, and an unmapped parameter is a hard ngspice error), the two-argument .lib <file> <section> corner entry point, and the Verilog-A modules ngspice must preload before a PSP103 device will bind. Every referenced PDK file is content-bound and reported under data.configuration; the binding itself is reported under data.extensions["org.openada.pdk_binding"]. --pdk and --models are mutually exclusive.

• An experimental drc-compare operation with explicit revision and deck modes. Revision mode requires different GDS content and reports persistent, resolved, and introduced bounded native markers. Deck mode requires the same GDS content and different generator scripts, then adds proximity-based cross-category correlations. Both modes recheck input stability and preserve explicit non-signoff limitations. Revision mode can additionally bind paired passing OpenADA LVS results that use the same reference netlist and setup, without claiming an unproven extraction-to-GDS relationship.

• A diagnostic drc-review CLI operation that consumes an existing validated KLayout LYRDB plus its exact GDS, deduplicates equivalent native cell variants, expands retained leaf-cell markers through the physical hierarchy, and emits hashed full-layout and ranked occurrence-level PNG views. Fresh output, input stability, bounded geometry, native PNG structure, dimensions, and renderer summaries are checked before the operation passes. The images remain representative diagnostic evidence, not a replacement for the native report, rule deck, or foundry signoff. Review results now also classify common rule families from native rule text, retain declared length constraints, measure marker bounds, and calculate coordinate-to-grid offsets for off-grid findings while explicitly avoiding automated-fix or reconstructed-rule claims.

• An experimental bootstrap-asic-project skill for blank open-ASIC workspaces. It defines core/full-chip/submission finish lines, selects one coherent PDK/flow/runtime stack, starts full chips from maintainer-owned padframe templates, and stage-gates RTL, function, synthesis, physical implementation, routed timing, DRC, LVS, and handoff. Its bounded standard library helper maintains a draft/frozen identity ledger for canonical project/collateral/tool paths and SHA-256 values, with deliverable-dependent requirements, explicit replacement/thaw, retained gap resolution, and machine-readable freeze-readiness/missing-requirement diagnostics. Pre-run assembly roles bind immutable generators such as seal-ring.config, not generated signoff outputs. Its successful freeze is structural/hash consistency only, not compatibility or engineering evidence. Missing OpenADA operations default to not evaluated; explicitly authorized native gap runs stay outside OpenADA result envelopes and foundry signoff is never inferred.

• result.series.extract/v1alpha1 and the extract CLI bridge. It consumes a complete passing circuit.simulate/v1alpha2 envelope plus that result's exact retained raw artifact, rechecks canonical path/bytes/SHA-256 and file stability, selects one request-bound padded plot within fixed bounds, and emits a canonical result.measure/v1alpha1-compatible real series. ngspice binary/ASCII and Xyce ASCII Spice3 evidence support explicit real/imaginary Cartesian voltage or current projections across each backend's advertised OP/DC/AC/TRAN rows.

• result.spectral.measure/v1alpha1 and the spectral CLI operation for one closed coherent single-tone SNR, SINAD, signed-dB THD, or SFDR measurement. The method freezes uniform power-of-two sampling, rectangular window, mean removal, one-sided mean-square bin power, harmonic folding/collision rules, band membership, tie breaking, and a hashed component partition.

• result.transfer.measure/v1alpha1 and the transfer CLI operation for an explicit same-unit Cartesian AC output-over-input trace plus one first-positive-frequency gain, unique falling −3 dB bandwidth, unity-gain-frequency, or negative-feedback phase-margin scalar. The profile freezes phase unwrapping and log-frequency crossing interpolation, rejects ambiguous crossings, and explicitly excludes gain margin.

• A standards-scope map for IEEE 1241-2023 (ADC), IEEE 1658-2023 (DAC), IEEE 1057-2017 (waveform recorders), IEEE 2414-2020 (jitter/phase-noise terminology), and IEEE 181-2025 (transitions/pulses). Implemented spectral methods remain OpenADA definitions; converter/recorder references are explicitly candidate, not IEEE conformance claims.

• An explicit external-provider runtime over immutable openada.driver-manifest/v0alpha1 and openada.request/v0alpha1, exposed as provider validate, provider list, and provider invoke. It validates manifests and cross-references, resolves one active circuit-simulation local JSON-stdio wait capability, invokes without a shell under bounded I/O and timeout policy, and validates typed result data, provider/request correlation, truth-table execution status, requested artifact roles and limits, local file hashes, zero transport exit, and wait-process cleanup. Before launch it snapshots canonical regular target and configuration files, verifies declared SHA-256 identities, and enforces 16 MiB target, 256 MiB-per-configuration, and 512 MiB aggregate bounds; post-run mutation or replacement invalidates the evidence. It does not discover, install, rank, or trust providers; v0alpha1 does not digest-bind the complete request.

• A hash-bound ngspice PDK-control reference provider with a closed ordered save all → OP/DC/AC/TRAN → optional TRAN-only linearize → safe write grammar, exact analysis-feature matching, and sanitized native execution. Its pinned public-IHP chain retains real ngspice 46 evidence for all four analyses, independently reconstructs engineering facts and scoped agent decisions, and exercises request, native-error, and tamper boundaries. Its provider conformance claim is bound to the exact source-frozen chain receipt.

• Three complete digital semantic operations and their engineering workflows: strict rtl.lint/v1alpha1 through Verilator, flattened Liberty-mapped logic.synthesize/v1alpha1 through Yosys/ABC, and constraint-complete one-corner timing.analyze/v1alpha1 through OpenSTA. Results bind ordered RTL, conservative literal-include closure, stable input hashes, mapping policy, a version- and digest-bound external ABC executable, closed non-inheriting tool environments, fresh native artifacts, normalized inference/cell/area evidence, and setup/hold WNS/TNS in seconds without claiming equivalence, physical timing, or signoff. The plugin adds senior RTL-architecture, synthesis/inference, and ASIC-timing skills that stay inside those evidence boundaries.

• A closed semantic-surface catalog and non-waivable release ledger covering all 147 active rows through seven pinned public-design chains. Each accepted row now carries contract tests, a real native EDA run, independent artifact verification, normalized evidence, a downstream engineering decision, negative and tamper replays, agent-visible evidence, and clean-source attestation. CI mechanically checks manifest hashes, the seven-record index, provider receipt registration, offline verifiers, and zero release gaps.

• Public real-design chains for IHP DRC/LVS, IHP analog measurement and full agent workflows, IHP SAR RTL/lint, ORFS Ibex Nangate45 synthesis/timing, and ngspice/Xyce analysis portability. Native artifacts and public-design provenance are retained in the source distribution so downstream reviewers can rerun every offline oracle.

• profile list and profile show for cwd-independent inspection of all packaged operation, assertion, feature, parameter, and normalized-result schemas.

• Intent-ledger and implemented-routing references for the analog characterization coordinator, plus standards-aware spectral workflow guidance and concrete extraction/measurement/evaluation/provider commands in the execution skill.

Changed

• Updated the ASIC bootstrap coordinator to use the shipped rtl.test/v1alpha1 operation for compatible self-checking HDL tests while preserving its assertion-adequacy boundary and explicitly recording the current lack of a scoped RTL-test doctor mapping.
• Promoted jsonschema>=4.18 to a base dependency because operation-specific validation is part of the external-provider execution boundary.
• Extended evaluate to accept complete ordinary or spectral measurement envelopes, plus the new transfer measurement envelope, while preserving the unchanged specification.evaluate/v1alpha1 typed measurement input.
• Let measure, spectral, and transfer consume a complete passing result.series.extract envelope directly, removing an undocumented manual JSON handoff.
• Advanced the Python package, Codex plugin, Claude plugin, and built-in driver identity to 0.4.0; packaged wheels now include every current analog, measurement, digital, and specification operation profile.
• Made the source/plugin launcher import optional schema validation lazily, so dependency-free discovery commands still run and schema-backed commands emit a structured missing-dependency diagnostic. Plugin setup now states clearly that agent marketplaces install skills but not the Python runtime dependency.

Fixed

• result.series.extract accepts the envelope circuit.simulate writes. simulate records every file that bound the devices under data.extensions["org.openada"].configuration; extract refused that field as undeclared, so the only way to chain the two operations was to delete the record naming where the devices came from. A round trip broken inside one contract, paid for in provenance. The field is now declared and validated - role, path, digest and byte count - so accepting it opens no hole. The regression test round-trips a real simulate envelope with nothing stripped.

• openada simulate consults PDK_ROOT. --pdk refused with pdk.root.required even when the environment variable already named a valid root that openada doctor reports. --pdk-root still wins when given; the refusal now names both ways the root can be supplied.

• A named analysis takes its numbers from the deck. --analysis tran demanded --step-s and --stop-s from a deck that already states .TRAN 10p 8n. Since the request is checked against the deck's own directive anyway, those were the only values that could ever run. Any required field the caller omits is now taken from the deck's single top-level directive of the named type; an explicit flag always wins, and a deck stating no such directive is still refused, saying so.

• series.selector.missing names what was asked for and what is available. One sentence covered both a misspelled vector and naming the sweep axis alongside the signals, and gave a caller nothing to act on. The refusal now names the offending selector, which of the two mistakes it is, the plot's selectable signals and its axis vector.

• pdk.collateral.foreign is reachable for pre_osdi cards. The pre_osdi branch returned before the foreign check, so the module docstring's own headline example - IHP's psp103.osdi preloaded from the sky130A tree - classified as pdk.collateral.hand_bound. Both refuse before ngspice runs, so nothing unsafe ran; the diagnostic simply named the wrong mistake, and the two have different corrections. Foreign is now decided first, for every card kind; a preload naming its own tree's module is still hand_bound.

• Accepted Netgen 1.5.321 hierarchical JSON's exact, equivalent pin-only auxiliary records without obscuring the unique requested top-cell LVS comparison. Unequal pin lists, partial known-key records, duplicate requested tops, and other ambiguous shapes remain invalid and produce engineering unknown.

Limitations

• Spectral v1alpha1 intentionally rejects nonuniform or noncoherent records, non-rectangular windows, main-lobe integration, PSD/averaging, SNDR aliases, ENOB derivation, jitter, and phase noise. True zero-frequency DC gain, gain margin, phase-crossing search, poles/zeros, integrated noise, corners, and statistical campaigns remain future semantic operations.
• External provider execution is explicit-manifest, local CLI, JSON stdin/stdout, wait-only, and currently registered only for active circuit.simulate/v1alpha2. v0alpha1 still has no complete request digest, independent capability ID, per-feature maturity rows, normative MCP binding, catalog trust model, sessions, remote jobs, or artifact-transfer protocol.

0.3.0 — 2026-07-15

Added

• Backend-independent result.measure/v1alpha1 and specification.evaluate/v1alpha1 operations, exposed as measure and evaluate. They operate on canonical-digest-bound normalized real inline series and typed measurement records; supported scalar algorithms, exact units, limits, and condition bindings are closed and explicit. A public openada.operations.normalized_series_sha256(...) helper computes the input digest, while results retain measurement-request and complete-specification digests and normalized evidence. These bindings detect changed content but are not signatures or authentication.
• A deterministic, network-free typed-evidence conformance bundle covering all nine measurement algorithms plus specification pass, fail, unknown, and tampered-binding cases.
• The additive immutable openada.operation-profile/v0alpha2 schema for deterministic semantic implementations, plus the active circuit.simulate/v1alpha2 profile with explicit OP/DC/AC/TRAN mappings. The published v0alpha1 schema and historical circuit.simulate/v1alpha1 profile remain unchanged.
• Experimental characterize-analog-block, analyze-feedback-stability, analyze-spectral-linearity, and assess-pvt-and-yield engineering skills. The skills inspect installed capabilities and leave unsupported metrics not evaluated; fresh-agent forward tests do not promote them beyond experimental.
• A provider, marketplace, connector-mining, and MCP boundary proposal. MCP is described as a future transport adapter for unchanged OpenADA requests/results, and a future marketplace catalogs conforming capability providers rather than raw executables.
• Namespaced plugin skill entry points for Codex ($openada:<skill>) and Claude Code (/openada:<skill>), while retaining skill-only installation through the standard ~/.agents/skills user directory.

Changed

• Moved the active shared simulation bridge from historical immutable circuit.simulate/v1alpha1 to additive circuit.simulate/v1alpha2, expanding the typed CLI flags to OP, DC, AC, and transient analyses. The ngspice mapping is structured for OP/DC/AC and workflow-validated for transient; Xyce is structured for DC/AC, workflow-validated for transient, and explicitly rejects OP as unsupported.
• Extended the circuit-simulation portability fixtures and independent native evidence checks by analysis in the new v0alpha2 conformance bundle. Its pinned success replay now binds the exact active operation-profile digest and supports structured maturity for the new analysis rows; the historical v0alpha1 transient bundle remains byte-stable. The shared subset still rejects includes, control blocks, native measurements, print directives, FFT, noise, Monte Carlo, and multiple analyses.
• Bounded each top-level simulation deck and explicit ngspice init input to 16 MiB before native launch or over-limit hashing. Conflicting generic native errors now prevent a terminal non-convergence observation from becoming an engineering fail.

Limitations

• result.measure consumes caller-supplied normalized inline series; OpenADA does not yet extract those series from native ngspice or Xyce waveform files. Optional upstream native-artifact lineage is recorded only as unverified.
• Runtime external-manifest discovery, generic request dispatch, and MCP provider invocation remain unimplemented.
• openada.driver-manifest/v0alpha1 has no normative MCP transport binding, independent capability IDs, or per-feature maturity/conformance rows; those require a future additive manifest revision rather than overloaded v0alpha1 values.

0.2.0 — 2026-07-15

Added

• A tool-independent review-circuit-simulation engineering skill above the OpenADA execution-and-evidence adapter.
• A shared typed circuit.simulate/v1alpha1 profile for ngspice and Xyce.
• A pinned, network-disabled native ngspice/Xyce portability replay with an independent verifier for both native waveform formats.

Changed

• Promoted the bounded Xyce transient mapping to workflow-validated shared alpha maturity alongside ngspice.
• Expanded plugin metadata so Codex and Claude Code expose both the execution skill and backend-independent engineering workflow.
• Kept the emitted result envelope at openada.result/v0alpha1; execution and engineering status semantics are unchanged.

0.1.0 — 2026-07-14

• Initial public preview of the semantic CLI, six open-source EDA drivers, normalized evidence contract, plugin packaging, schemas, and conformance workflows.