Instructions file CodexOpenCode
AGENTS.md instructions for khamidov17/vpstack, covering agents.md — context for codex and other ai agents, what this repo is, critical domain facts — read before doing anything, extract and run b1 on your data and requires: pip install soundfile scipy numpy.
Instructions file
Claude Code instructions for khamidov17/vpstack, covering claude.md — context for ai agents working in this repo, what this repo is, architecture in one paragraph, non-negotiable rules and where things live.
Skill Claude CodeCodex
Run an ASV attacker against the user's anonymized speech to measure how well the anonymization hides the speaker. Default attacker: VP2024/2026 semi-informed ECAPA-TDNN trained on the user's anonymized train-clean-360. Reports per-gender EER (the official VPC privacy metric) plus linkability (ZEBRA Cllr) for context.…
Skill Claude CodeCodex
End-to-end voice-anonymization research lifecycle automation. Runs /vp-talk → /vp-hypothesis → /vp-plan-design-review → /vp-plan-eng-review → /vp-implement → /vp-qa → /vp-ship in sequence, with check-in gates between phases. Equivalent of gstack /autoplan but for voice-privacy research. Use when starting a new…
Skill Claude CodeCodex
Run B1 (McAdams) + B2 (neural) baselines on the same eval set as the user's anonymization system, return a delta table (EER / WER / linkability). Use when the user asks "how does my system compare to baseline?" or "is this better than B1?" or wants a quick sanity check of their anonymizer against canonical references.…
Skill Claude CodeCodex
Browse the local experiment store: list runs, rank by EER/WER, show one run, diff two runs, search by hypothesis text, see project stats, list recent skill activity, surface research learnings. Thin wrapper over the vpstack-brain (vpbrain) CLI. Use when the user asks "what experiments have I run?", "best run so far"…
Skill Claude CodeCodex
Restore working context from a previous session. Reads session-state.md and presents a 2-sentence briefing: what you were doing and what to do next. Run at the start of any session after a break. Pair with /vp-context-save. (vpstack) Voice triggers: "restore context", "resume", "where was I", "pick up where I left…
Skill Claude CodeCodex
Save current session state so any future session can resume without losing the thread. Captures: current hypothesis ID, last experiment ID, last skill run, what's pending, suggested next skill. Writes to /.vpstack/projects/{slug}/session-state.md. Run before ending a session, switching branches, or handing off to a…
Skill Claude CodeCodex
Run the full VP2026 evaluation pipeline on an anonymized output: EER per gender split (F-F / M-M / Mixed) via vpstack-score, WER via vpstack-wer (Whisper), naturalness PMOS via vpstack-utmos, and emit the official VP2026 submission CSV layout under exp/asvanon{suffix}/, exp/asr/, exp/ser/, exp/resultssummary/track1/.…
Skill Claude CodeCodex
Formalize a voice-anonymization experiment hypothesis before running anything. Captures what's being tested, expected delta vs baseline, evaluation criteria, and which components are being swapped. Use when starting a new experiment or before /vp-spike. Writes to /.vpstack/projects/{slug}/hypotheses/{id}.md. (vpstack)…
Skill Claude CodeCodex
Orchestrated implementation of a voice-anonymization recipe component (recipe runner, MCP tool, or hparams YAML). Reads the latest /vp-hypothesis, enforces the recipe-shape contract from CLAUDE.md (CLI args, JSON output, deterministic mode, lazy weight fetching), gates atomic commits with no placeholder hparams…
Skill Claude CodeCodex
Voice-privacy-domain debugging with VP2026-specific priors. When something looks wrong (EER weirdly high, WER tanked, repro fails, attacker output looks random), walks through the most-likely causes BEFORE diving into a full Python debugger session. Distinct from generic gstack /investigate (works fine on any code) by…
Skill Claude CodeCodex
Voice-privacy-domain design review of a recipe architecture, attacker spec, or reproducibility design BEFORE coding starts. Distinct from generic gstack-style design review (which is about visual UI). This reviews: recipe interface contract, attacker condition coverage, repro design (seed/splits/checkpoints), eval-set…
Skill Claude CodeCodex
Voice-privacy engineering review. Layers on top of the generic /plan-eng-review with 18 VP-specific quality gates: GPLv3 isolation, runtime model fetch, held-out test split safety, the 5 reproducibility checks, three attacker conditions, VP2026 submission format (CSV + Mixed gender + rank/zip), recipe interface…
Skill Claude CodeCodex
Voice-privacy QA pass on a project before commit / submission / handoff. Orchestrates vpstack's existing checks: /vp-repro-check, a fast /vp-attack lazyinformed smoke run, vpchecksubmission validation, and pytest of any project-level tests. Distinct from generic gstack /qa (which is web-app QA — vpstack has no web…
Skill Claude CodeCodex
Verify a VP2026 experiment is reproducible — checks seeds, dataset splits, model checkpoint hashes, and config completeness. Returns PASS or FAIL with specific reason. Use before submitting results, before merging changes, or when diagnosing why two runs produced different numbers. Catches silent drift early.…
Skill Claude CodeCodex
Ship a voice-anonymization change: run vp-domain quality gates first (vp-repro-check, attacker smoke, submission format if applicable), then bump VERSION + commit + push + optional PR. Distinct from generic gstack /ship by gating on vp-specific checks BEFORE pushing — a regression on EER or repro-check fail blocks the…
Skill Claude CodeCodex
Run 1-3 focused ablation experiments with given/when/then verdicts. Use when testing a specific hypothesis (paired with /vp-hypothesis) or comparing 2-3 component swaps quickly. Returns a verdict per spike (CONFIRMED / REFUTED / INCONCLUSIVE) with the metric deltas. Writes to /.vpstack/projects/{slug}/spikes/{id}.md.…
Skill Claude CodeCodex
Two-mode planning skill for voice anonymization work. Two-mode planning skill. Like gstack /office-hours — asks forcing questions, then writes a locked plan that downstream skills use. Mode R (Research): VP2026 benchmark — 8 forcing questions on open question, threat model, contribution claim, baseline, eval scope…
Skill Claude CodeCodex
Generate an INTERNAL experiment report from logs in /.vpstack/projects/{slug}/. Captures what was done, the method, hyperparameters, results table, and config hash. Does NOT generate citations or research-paper prose — that's the researcher's job. Use when a researcher wants a structured engineering log to share with…