Skip to content

CHANGELOG.md

Latest commit

 

History

History
403 lines (353 loc) · 19.6 KB

CHANGELOG.md

File metadata and controls

403 lines (353 loc) · 19.6 KB
Raw

Changelog

All notable changes to this project are documented here. Format follows Keep a Changelog; versioning will start at 0.1.0 once the first public release ships. Until then, entries track the multi-session Arcadia-1 integration roadmap (~/.claude/plans/concurrent-beaming-bear.md) on branch feat/arcadia-integration.

Each session entry points at its integration commit. Earlier pre-roadmap history (IHP digital port, post-PnR GL sim gates, LibreLane template parity) is in git log; only roadmap sessions are summarized below.

Unreleased — branch feat/s9-residual-closure

Session S9-residual-closure — close the "caveats honestos" from S9-gap-closure

Follow-up session after feat/s9-gap-closure merged to main (merge 40026ac). Two gaps had closed with residual caveats that the original session acknowledged but could not close inside its scope:

  • Gap #6 residual (#6a + #6b): the SAR 11-bit ENOB task passed warm but was flagged as "cold-cache flaky" (SNDR=16.18 dB reported once on a cold full-bench run); thresholds were pinned to the defaults with zero architectural headroom.
  • Gap #4 residual: the real-mode path of digital_autoresearch_adapter (LibreLane + real LLM) was never exercised end-to-end; only the mock-metrics path had coverage.

Close-out commits on this branch:

  • 516d6b6 — Gap #6a cold-cache flakiness probe. Nine cold independent runs (5 SAR-only + 4 full-bench) all PASS with bit-exact identical numbers (ENOB=4.451, SNDR=28.56 dB). The 12 dB collapse previously observed did not reproduce. Evidence + README under bench/results/sar11_cold_flakiness_probe/. No code change; the previous observation stays as history with a "not reproducible, tracker-link if it reappears" conclusion.

  • 6c04ac8 — Gap #6b SAR 11-bit ceiling characterization. A 12-point Latin-square sweep over (comp_W_input_um, comp_L_input_um, cdac_C_unit_fF, bias_V) driven by the new scripts/characterize_sar11_ceiling.py measured the architectural ENOB ceiling at 5.64 bit / 35.70 dB at W=8, L=0.15, Cu=20 fF, Vb=0.5 (reproduced bit-exact across the replica pair). Applied the "floor(ceiling) − 0.5" rule: raised SARADC11BitTopology._SPEC_ENOB_MIN from 4.0 → 4.5 and _SPEC_SNDR_MIN from 25.0 → 28.0, and shifted default_params() to the ceiling point so the bench baseline stays above the new floor with 1.14 bit / 7.74 dB margin. New tests/test_sar_adc_11bit_ceiling.py locks the invariant against future regressions. TODO_calibration.md updated to cite the TSV.

  • f7f4c8f — Gap #4 live-mode digital_autoresearch. New bench/tasks/end-to-end/digital_autoresearch_counter_live.yaml drives the real pipeline: OpenRouter LLM proposes flow-config overrides, LibreLane v3 runs each eval to signoff (Checker.KLayoutDRC) on GF180MCU-D, FlowMetrics extracts WNS / cells / area / power, greedy keep/discard on FoM. Budget=2 runs in ~113 s wall-clock with all tools present; FAIL_INFRA (SKIPPED) otherwise. Adapter now snapshots and restores design_dir/config.yaml so repeated live runs don't mutate the committed baseline. New @pytest.mark.librelane integration test plus evidence at bench/results/gap_closure_digital_autoresearch_live/.

Bench surface after this session: 17 tasks (16 from S9-gap-closure

  • the new live autoresearch variant). Pass@1 = 100% on tool-complete hosts; SKIPPED (honestly) for tasks missing their tool chain.

Pre-merge history — branch feat/s9-gap-closure (merged 40026ac)

Session S9-gap-closure — close all 11 bench gaps

Dedicated session to close the 11 known-gap items listed in S10's README "In-tree gaps to close". Every gap is one commit; every commit is end-to-end verified on the local host, not just covered by mock tests. Split contingency authorized up-front but not needed — Wave 1 + Wave 2 + Wave 3 all land on feat/s9-gap-closure.

Bench summary before / after: 9/11 PASS → 16/16 PASS on scripts/run_bench.py with OPENROUTER_API_KEY sourced. Pass@1 100%.

Wave 1 — analog correctness + contract

  • Gap #11 (e9d7335) — Typed Pydantic schemas for every adapter's BenchTask.inputs. Typos in YAML now fail loudly as FAIL_INFRA with the Pydantic message surfaced instead of silently using defaults. 17 new unit tests in test_bench_adapter_inputs.py.
  • Gap #1 (0256407) — GF180MCU-D process parameters ported into a new topologies/process_params.py registry. MillerOTADesigner resolves ProcessParams from pdk.name; the IHP registry is bit-identical to the pre-fix defaults (pinned by test_miller_ota_gf180.py::test_ihp_designer_bit_identical_widths). spec_miller_ota_gf180_easy flips FAIL_SIM → PASS without touching the task YAML.
  • Gap #6 (64c9aa6) — End-to-end task e2e_sar11b_enob_ihp exercises SARADC11BitTopology on ngspice + PSP103 OSDI + Verilator. Audit thresholds anchor on measured defaults (ENOB=4.45, SNDR=28.56 dB) rather than aspirational 6.0/38.0; calibration story spelled out in the task notes. See gap #3 for the topology-level _SPEC_* anchor.
  • Gap #7 (ef9e4c0) — New check_vds_polarity pre-sim gate catches drain/source-swapped MOSFETs by inspecting source-node power-rail adjacency. Companion task bugfix_strongarm_vds_inversion.yaml injects the bug and asserts detection. 3 new unit tests.
  • Gap #8 (b6def97) — Real LLM adapter against OpenRouter (google/gemini-2.5-flash). Canned live-run evidence committed under bench/results/gap_closure_llm_proof/: Adc=36.3 dB, GBW=1.63 MHz. Without key the adapter SKIPS cleanly (FAIL_INFRA). Out-of-range LLM JSON funnels to FAIL_AUDIT, not FAIL_INFRA, so the LLM is scored against its own output.

Wave 2 — digital coverage (dependency-ordered)

  • Gap #5 (a0beb1e) — bench/designs/counter_bench/ + new run_librelane_flow_task adapter hardens a 4-bit counter through LibreLane v3 on GF180MCU-D, takes ~55s to signoff DRC, publishes DRC_violations=0. Hardened run symlinked under bench/cache/librelane_runs/counter/runs/<tag>/ so downstream tasks (gap #2) don't re-harden. Nix EDA tools auto-detected via detect_nix_eda_tool_dirs(); PDK resolved via new _resolve_librelane_pdk_root (GF180 wafer-space / IHP). SKIPS cleanly when the PDK or LibreLane Python interpreter is absent.
  • Gap #4 (1e9c867) — digital_autoresearch harness becomes a real wrapper over DigitalAutoresearchRunner. Mock-mode fixture at bench/designs/counter_bench/mock_flow_metrics.json keeps the task offline-clean (no LLM, no LibreLane, no API key needed). Real mode requires OPENROUTER_API_KEY + a PDK and SKIPs otherwise. Old NOT_IMPLEMENTED stub removed; its test becomes 3 new tests.
  • Gap #2 (77da7d5) — run_gl_sim_post_synth now auto-discovers the counter cache published by gap #5 when no explicit run_dir is given, wraps the counter in a GenericDesign, and exercises GlSimRunner.run_post_synth against the hardened netlist. e2e_gl_sim_post_synth_counter flips SKIPPED → PASS. Also fixes two latent defects in the adapter (abstract ToolEnvironment, stale StageResult attribute names) + repairs GenericDesign.testbench() to return a single TB path instead of RTL+TB concatenated into one string. 2 new regression tests.

Wave 3 — housekeeping

  • Gap #3 (b753ed4) — SAR rename cascade: canonical SAR7BitTopology + SAR7BitBehavioralTopology in sar_adc_7bit.py / sar_adc_7bit_behavioral.py; old sar_adc_8bit.py / sar_adc_8bit_behavioral.py become thin deprecation shims emitting DeprecationWarning on instantiation (not import). SARADC11BitTopology._SPEC_ENOB_MIN recalibrated 6.0 → 4.0 / SNDR 38 → 25 dB to match measured defaults (gap #6 anchor). TODO_naming.md marked RESOLVED; TODO_calibration.md item 1 RESOLVED, items 2-5 (tau_regen, LDO, bootstrap, corner sweep) DEFERRED — each its own session.
  • Gap #9 (d17bd71) — test_run_batch_workers_consistent verifies bench JobRegistry parity (workers=1 vs 2 same per-task statuses). Real --workers 4 wall-clock captured under bench/results/gap_closure_parallel/ with summary + README: 1.7x speedup (2m18s → 1m21s on this host, dominated by the two ~60-80s tasks).
  • Gap #10 (0fd0f0d) — .github/workflows/bench.yml runs a three-step bench-offline job on ubuntu-22.04 + Python 3.12: (1) 125-test adapter + bridge suite, (2) 896-test offline smoke with the tool-marker filter, (3) scripts/run_bench.py --no-real-tools.

Close-out

  • This CHANGELOG entry.
  • README known-limitations table cleaned of resolved rows.
  • Final bench --run-id gap_closure_final with 16/16 PASS persisted.
  • SESSION_HANDOFF.md rewritten for the next session (MCP spike).

Session 10 — Publication + cross-linking

Merged into main as commit 12e4214 (merge of feat/arcadia-integration). Kept below for session history.

Meta-session. No new runtime features; the deliverable is documentation honest about what works and what does not.

  • Bucket A hygiene
    • bench/results/* is now gitignored except the frozen s9_initial_smoke/ baseline. Local re-runs of scripts/run_bench.py no longer dirty the working tree. New bench/results/README.md explains the contract.
    • digital_autoresearch adapter stub landed in HARNESS_DISPATCH. Tasks using that harness return BenchStatus.SKIPPED with an explicit NOT_IMPLEMENTED note (never the ambiguous FAIL_INFRA). Covered by tests/test_bench_runner.py::test_digital_autoresearch_stub_returns_skipped_with_explicit_note.
  • CHANGELOG.md (this file), docs/architecture.md (new consolidated overview), and an expanded top-level README.md that lists every open gap and the scheduled gap-closure session, not a vague "future work" bullet.
  • docs/external_pr_drafts/awesome_ams_skills.md — PR draft for the upstream Arcadia-1/awesome-ams-skills catalog. The PR is not opened by this session; the user decides when to file it.

Session 9 — Bench suite (commit 04dfc91)

First iteration of the eda-agents benchmark. Schemas were reimplemented (not copied) from behavioral-veriloga-eval; runtime is fully in-tree and licensed Apache-2.0.

  • New package src/eda_agents/bench/: Pydantic v2 frozen models (BenchTask, BenchResult, BenchScores, MetricBound), harness adapter layer, and a runner with explicit PASS -> FAIL_AUDIT downgrade logic (a sim pass with a failing scoring criterion is never painted as PASS).
  • Adapters wrap S6/S7/S8 code read-only: analog_roles_adapter invokes the Session 6 DAG with DryRunExecutor; analytical_miller_design wraps MillerOTADesigner; run_gl_sim_post_synth wraps GlSimRunner and returns FAIL_INFRA (mapped to SKIPPED) rather than faking a pass when no hardened LibreLane run is available.
  • scripts/run_bench.py CLI with --family, --task, --workers, --dry-run, --no-real-tools, --list-tasks; writes per-task JSON + summary.json + report.md + a latest.md pointer.
  • 11 seed tasks across spec-to-topology, bugfix, tb-generation, and end-to-end families.
  • Honesty about the smoke run (bench/results/s9_initial_smoke/): 9/11 PASS, Pass@1 = 90% excluding skipped. The two non-PASS results are kept in the suite deliberately:
    • spec_miller_ota_gf180_easy is FAIL_SIM. Root cause diagnosed and documented in docs/upstream_issues/miller_ota_gf180_process_params.md: MillerOTADesigner.ProcessParams is hardcoded to IHP SG13G2, so the GF180 sizing produces sub-Wmin transistors that the BSIM4 binner rejects. Fix path requires a GF180 sEKV port, scheduled for the post-S10 gap-closure session.
    • e2e_gl_sim_post_synth_systolic is SKIPPED because the bench intentionally does not harden a fresh LibreLane run; the adapter's wiring is covered by tests but not ejected against a real run_dir yet.
  • 44 new tests under tests/test_bench_*.py, marker bench added to pyproject.toml.

Session 8 — Bridge open-source (commit cd232e3)

Reimplementation of the virtuoso-bridge-lite patterns under Apache-2.0 for the open-source EDA tool chain.

  • New package src/eda_agents/bridge/: models.py (Pydantic v2 frozen BridgeResult, SimulationResult, ExecutionStatus enum with .ok / .save_json / .load_json); jobs.py (JobRegistry with UUID -> JSON persistence under ~/.cache/eda_agents/jobs/, ThreadPoolExecutor, serialization- error -> ERROR rather than hang); ssh.py (SSHRunner Linux-only OpenSSH wrapper with jump-host and ControlMaster); xschem.py (headless XschemRunner.export_netlist with infra_error to separate tool-bug from setup-bug); klayout_ops.py (facade delegating to the existing core/klayout_*.py runners — no rewrite).
  • eda-bridge CLI: init / status / jobs / cancel / stop / start xschem-netlist.
  • Fixed a pre-existing parser bug in core/spice_runner.py: _MEAS_LINE_RE is now anchored to ^token = number so status messages no longer smear into parsed measurements.
  • examples/14_bridge_e2e.py — IHP SG13G2 demo passes end-to-end (Adc = 32.5 dB, GBW = 1.39 MHz). GF180 path surfaces the spec_miller_ota_gf180_easy bug S9 later diagnosed as the designer's IHP-only process params.
  • 64 bridge tests, all mocked (zero network).

Session 7 — SAR ADC portfolio (commits d85bcc5, e11edc4, 4e25adc, 6a999f9)

Four-commit arc covering the SAR ADC deliverables deferred from S5 plus the 11-bit design reference and the skill / docs bundle.

  • topologies/sar_adc_netlist.py refactored to accept optional comparator_section / nand_section / extra_model_lines. Default 8-bit output stays bit-identical to S6.
  • topologies/sar_adc_8bit_behavioral.py — SARADC8BitBehavioralTopology by composition over the parent SAR, with the XSPICE comparator replacing the StrongARM. The class docstring opens with the "7-bit-effective" caveat (AnalogAcademy naming).
  • topologies/sar_adc_11bit.py — SARADC11BitTopology flagged DESIGN_REFERENCE = True. The CDAC is verified 11-bit (test_cdac_is_true_11bit), dummy cap permanently tied to VCM. check_system_validity implements PVT margin (Pelgrom), metastability BER bound, supply-ripple units (no more "5400 mA" bug), and reference settling.
  • src/eda_agents/data/sar_logic_11bit.v — 11-cycle SAR FSM with B/BN/D [10:0] and MSB-first decode (D[0]=MSB).
  • core/pdk.py::IHP_SG13G2.osdi_files extended with psp103.osdi after the demo surfaced the missing psp103va binding for sg13_lv_* subcircuits.
  • Skill analog.sar_adc_design registered; 9 markdown documents under docs/skills/sar_adc/ (reviewed architecture, comparator, bootstrap switch, SAR logic, LDO, integration, sim/verification) plus TODO_calibration.md and TODO_naming.md for the deferred pieces.
  • evals/sar_adc_arch.json — 20 architectural prompts.
  • examples/13_sar_adc_11bit.py and examples/13b_sar_adc_8bit_behavioral.py — end-to-end demos with skip-on-missing-tool logic for ngspice / openvaf / Verilator / XSPICE.
  • 30 new tests across test_sar_adc_11bit.py, test_sar_adc_8bit_behavioral.py, and an added test_skills.py case.

Session 6 — Analog 4-role DAG (commit 89da1c6)

Reimplementation of the analog-agents DAG under Apache-2.0.

  • src/eda_agents/specs/spec_yaml.py — Pydantic v2 frozen spec loader.
  • src/eda_agents/checks/pre_sim/ — inline SPICE subcircuit parser plus four gates (floating_nodes, bulk_connections, mirror_ratio, bias_source). These are what the bench's bugfix family exercises.
  • src/eda_agents/agents/iteration_log.py — Pydantic v2 iteration log with YAML round-trip and EscalationError.
  • src/eda_agents/agents/analog_roles/ — AnalogRolesHarness and DryRunExecutor. The bench invokes this harness verbatim to prove the DAG wiring without paying a model.
  • src/eda_agents/skills/analog_roles.py — four role skills (librarian, architect, designer, verifier).
  • examples/12_analog_roles_demo.py.
  • 40 new tests.

Session 5 — Behavioural primitives + Docker toolchain (commit 650bd25)

  • core/flow_stage.py gains FlowStage.XSPICE_COMPILE.
  • core/stages/xspice_compile.py (XSpiceCompiler, CodeModelSource, XSpiceToolchain) and core/spice_runner.py::SpiceRunner(extra_codemodel=..., preload_pdk_osdi=...) for loading compiled .cm code models and extra OSDI at simulation time.
  • src/eda_agents/veriloga/voltage_domain/ — comparator_ideal, clock_gen, opamp_ideal, edge_sampler (XSPICE C sources).
  • src/eda_agents/veriloga/current_domain/ — filter_1st.va, opamp_1p.va, ldo_beh.va (Verilog-A for OpenVAF).
  • src/eda_agents/topologies/sar_adc_8bit_behavioral.py — initial kit (the full topology class landed in S7).
  • Skill analog.behavioral_primitives; Docker image for the XSPICE toolchain.
  • Demo examples/11_ams_mixed_domain_demo.py.

Session 4 — gm/ID sizing API on top of the LUT (commit ca9a36e)

  • core/gmid_lookup.py::GmIdLookup gained size, size_from_ft, size_from_gmro, and operating_range. Semantics match the gm/ID methodology without reusing any gmoverid-skill code (repo has no LICENSE — reimplement, do not copy).
  • scripts/generate_gmid_lut.py — unified IHP + GF180 LUT generator.
  • src/eda_agents/tools/gmid_json_adapter.py — npz <-> JSON interop (gmoverid.v1 envelope).
  • Skill analog.gmid_sizing registered.
  • 18 new tests across test_gmid_sizing.py and test_gmid_json_adapter.py.

Session 3 — ADCToolbox integration (commit 0ea5eb2)

  • Extra [adc] (adctoolbox>=0.6.4, MIT). Pulled via PyPI rather than vendored.
  • src/eda_agents/tools/adc_metrics.py — compute_adc_metrics(...) and calculate_walden_fom(...) wrappers.
  • topologies/sar_adc_8bit.py::extract_enob now delegates to the toolbox rather than the in-tree 64-sample FFT.
  • Skill analog.adc_metrics.
  • 7 tests.

Session 2 — OpenVAF -> OSDI pipeline (commit 34c95f0)

  • core/stages/veriloga_compile.py + FlowStage.VERILOGA_COMPILE.
  • core/pdk.py::netlist_osdi_lines(pdk, extra_osdi=None) accepts user-side OSDI paths; core/spice_runner.py propagates them into the run directory's .spiceinit.
  • Marker veriloga added for the new pipeline tests.

Session 1 — Skill registry (commit 06e21f4)

  • New package src/eda_agents/skills/: Skill dataclass, registry, and 15 initial skills (analog / digital / flow / tools). More skills added by later sessions (S3, S4, S5, S6, S7) reach ~23 by the end of S9.
  • agents/adk_prompts.py and agents/digital_adk_prompts.py were converted to thin shims delegating to get_skill(...).render(...).
  • 25 tests (grew to 26 with the S7 skill addition).

Session 0 — Legal + tooling matrix (commits 33007fc, 989a772)

  • scripts/check_tools.sh — baseline tooling check (ngspice, openvaf, yosys, magic, klayout, netgen, PDK_ROOT for IHP + GF180).
  • docs/license_status.md — license matrix for the eight Arcadia-1 repos reviewed during the deep-dive. Five of the six unlicensed repos are treated as "study architecture, reimplement under Apache-2.0"; the sixth is quote-able but not copy-able.
  • Upstream licensing-clarification issues were drafted but deliberately not filed yet; reopen the decision only if a specific repo becomes a blocker.

Baseline (upstream main)

Prior work includes the IHP digital RTL-to-GDS port (merge 2362553), post-PnR GL sim gates with SDF annotation (aaa7c2a, 8e6ac66, 09ad783), LibreLane template parity CI (407843a, 96780bc), and IHP Magic step skips (4d20ee9, ed5af21). See git log 2362553...04dfc91 for details.