env-precheck · git:20260907.781745f · 2026-09-07 · sha256 117cbac0e1e148c0
env-precheck git:20260907.781745fA
Immutable. This exact content never changes and is served at /api/v1/blob/117cbac0e1e148c0.
--- name: env-precheck description: Use when checking whether this machine's Python, EDA tools and licenses can actually run VeriPower's stages; not for any pipeline stage, module state, or design work. --- # Environment Pre-check Pre-pipeline, own session: you report, and change nothing — no module file, no event, no `kernel.py`. Nothing in the pipeline depends on this having run. `<skill>` is this skill's own base directory, named on the first line of this file. ## 1. Presence `<skill>/../../docs/eda-env.md` lists what must be true and what each miss costs. Probe every row of it, and `timeout` every probe: an unreachable license server hangs the tool. ## 2. Smoke Presence is not a checkout. Ask which stages to cover, then per row write the smallest input that forces the checkout and run it in a temp dir. A row passes iff it produces the file below — a tool that took the license and did nothing still exits 0. | Checkout | Hinges on | Produces | Gates | |---|---|---|---| | Design Compiler | `compile_ultra` (the DC-Ultra checkout — the flow runs no plain `compile`), then `write` | the netlist | synthesis | | PrimeTime | `report_timing` on that netlist | the timing report | timing-analysis | | PrimeTime-PX | `set power_enable_analysis TRUE` + `report_power` | the power report | power-analysis | | VCS + UVM | compiling `uvm_pkg.sv` and `uvm_dpi.cc` from `UVM_HOME`, then running simv | whatever the TB writes | simulation, power-analysis | | VCS coverage | `-cm line+cond+branch+tgl+fsm`, then `urg -report cov_merge -format text` | `cov_merge/dashboard.txt` | simulation coverage gate | | Verdi / FSDB | `-debug_access+all -kdb -lca` + a ucli `$fsdbDumpvars`, then `fsdbreport` | the fsdbreport output | simulation waveform, simulation-triage | | SpyGlass Lint | `current_goal lint/lint_rtl` + `run_goal` | that goal's `moresimple.rpt` | lint-cdc | | SpyGlass CDC | `cdc/cdc_setup`, `cdc/cdc_setup_check`, `cdc/cdc_verify_struct`, each `run_goal` | each goal's `moresimple.rpt` | lint-cdc | The three VCS rows each compile their own simv: coverage and FSDB instrumentation take effect at compile time. A row whose tool or variable already failed §1 is skipped, not failed. ## 3. Report Per row: pass or fail, and the stages it costs. Close with the stage list runnable now. - `lmstat` explains, the smoke decides: a feature the server lists can still fail to check out, and lmstat missing or timing out is not a failure. - Print `export` lines for the user; never edit their shell config.