curryfromuestc

Run a short, reproducible benchmark of one optimizer step (forward + backward + optimizer step over N microbatches) using the project's registered runtime. Activate when the user asks to "benchmark a training step", "measure throughput", "time one optimizer step", or "smoke test the runtime". Wraps run_accumulated_step from curry_train.benchmark.

Diagnose a training failure or stall by inspecting recent logs, loss curves, OOM traces, NaN events, and config. Activate when the user asks "why did my training crash", "loss went to NaN", "OOM during step X", "training is not improving", or "help me debug this run". Delegates to the failure-diagnoser agent.

Lightning Fabric integration recipe — minimal 5-line setup that gives DDP / FSDP / mixed precision / mixed-precision while keeping a raw PyTorch training loop. Activate when the user asks "Lightning Fabric", "torchrun", "DDP setup", "FSDP setup", "mixed precision", or wires up the launch script.

Bootstrap a new curryTrain project by copying the framework Python template into the user's working directory. Use when the user runs /curry-train:init, asks to "start a new training project with curryTrain", "initialize curryTrain", "scaffold a curryTrain project", or wants the framework code copied locally for editing.

Scaffold a new model + dataset + config triple inside an existing curryTrain project. Activate when the user asks to "add a new model to curryTrain", "scaffold a new experiment", "generate model.py / dataset.py / config.yaml for X", or describes an idea they want to start training. Delegates the actual generation to the scaffolder agent.

Compare two training runs and produce a concise markdown diff covering config, key metrics, loss curves, and grad-norm trajectory. Activate when the user asks to "compare run A and run B", "diff two experiments", "did this change actually help", or "is this run better than the previous one". This is both the implementation of the action and the methodology guide for variance-aware comparison.

Define an abort condition before launching a run, so that a clearly broken or clearly under-performing run stops itself instead of consuming the full compute budget. Activate when the user asks "when should I kill a run", "abort condition", "early stop a bad run", "kill criterion", or before any expensive run.

Decide which parallelism primitive (DP, ZeRO, TP, PP, EP, CP) to introduce next based on what bottleneck appears at the current model size. Activate when the user asks "do I need tensor parallelism", "OOM at scale", "training too slow", "should I add pipeline parallel", "how to scale beyond N GPUs", or after capacity-sweep when single-GPU runs no longer fit.

Remove AI writing patterns from English text with academic-aware overrides. Auto-detects LaTeX context for passive voice, hyphenated compounds, title case, curly quotes handling.

Placeholder for Phase C bilingual annotator. Adds sentence-level Chinese %-comments to English LaTeX source files. Currently not implemented.

Process external reviewer comments on an academic paper submission. Parses comments, classifies Major/Minor/Editorial, builds RevisionRoadmap, drafts response letter in Reviewer-Action-Change format.

Simulate peer review on an academic LaTeX paper draft. Generates 5 reviewer personas, runs independent reviews with 5-dimension scoring, devils-advocate stress tests, editorial synthesis decision.

Manage references.bib in an academic LaTeX paper project. Add bibtex entries, validate natbib citation commands, check citation compliance (orphan citations, self-citation ratio, source currency).

Compile an academic LaTeX paper project with pdflatex and bibtex. Diagnoses LaTeX errors, checks page-limit compliance, runs bibtex, fixes undefined references and overfull-hbox warnings.

Generate publication-quality matplotlib or seaborn figures for an academic LaTeX paper. 11 EECS chart types (bar, line, heatmap, ROC, ablation, scaling). Output to figures/*.pdf.

Draft or revise a section of an academic LaTeX paper (abstract, intro, related work, method, experiments, results, discussion, conclusion). Uses TEEL paragraph framework and PaperConfig word budget.

Guides behavior model ↔ RTL equivalence checking — sampling point selection (commit/retire vs transaction-edge vs cycle), drive direction (DUT pushes ref), DPI-C probe interface, async event alignment, snapshot-based debugging, and DSL-driven probe generation. Activate when the user explicitly invokes /chipdev-method:align-and-difftest, or asks "怎么 diff cmodel 和 RTL", "behavior 对不上 RTL", "difftest 怎么搭", "what should I sample", "RVVI vs DPI", or "snapshot for difftest", or sets up a difftest infrastructure.

Routes a methodology compliance review of the user's current changes to Codex via the codex-plugin-cc adversarial-review command. Activate when the user explicitly invokes /chipdev-method:audit, or asks "audit my diff", "审一下我这版改动", "review my methodology compliance", "did I violate any of the principles", or wants an independent adversarial check that their work follows the plugin's six invariants and contract-first discipline.

Guides the construction of a functional behavior model that decouples timing but preserves architectural state for RTL difftest — what to keep, what to drop, where to embed probes, how to package the cross-language interface, and how to coordinate memory with the DUT. Activate when the user explicitly invokes /chipdev-method:build-behavior-model, or asks "什么是行为模型", "behavior model 边界", "difftest 用什么 ref 模型", "怎么做 functional simulator", or designs the reference half of a difftest setup.

Guides the construction of a cycle-accurate performance model — scheduler model, module lifecycle, idle/advance contract, pipeline primitives, sparse memory, transaction memory pool, configuration system, and dump infrastructure. Activate when the user explicitly invokes /chipdev-method:build-perf-model, or asks "怎么写 cycle-accurate 模拟器", "性能模拟器怎么写", "how to model timing", "should I use SC_THREAD or advance loop", or is implementing a microarchitecture timing model.

Guides RTL frontend design discipline — module boundary rules, synthesizability constraints, observability instrumentation, lint cleanliness as a hard gate, reset strategy, and clock-domain crossing. Activate when the user explicitly invokes /chipdev-method:build-rtl, or asks "RTL 怎么分层", "should I use always_ff or always", "怎么做 lint clean", "reset strategy", "CDC how to handle", or starts writing SystemVerilog for a new module.

Helps the user decide whether to build a performance model, a behavior model, or RTL for a chip frontend task, and in what order. Activate when the user explicitly invokes /chipdev-method:choose-artifact, or asks variants like "should I write cmodel or RTL first", "怎么选模型", "what artifact do I need", or stands at the start of a new chip / module / subsystem and has not yet committed to an artifact path.

Designs the contracts that hold across performance model, behavior model, and RTL — single-source-of-truth interface DSL, hierarchy DSL, global constants, protocol annotations, state contracts, observability hooks, and alignment points. Activate when the user explicitly invokes /chipdev-method:define-contracts, or when they describe defining new module boundaries, designing an interface DSL, planning difftest probes, asking "怎么定义接口", "interface DSL", "single source of truth", or starting any chip frontend project that will have more than one artifact.

Routes a build, simulation, test, or difftest failure to Codex via the codex-plugin-cc rescue command for root-cause investigation. Activate when the user explicitly invokes /chipdev-method:diagnose, or describes a failure ("build failed", "sim hangs", "difftest mismatch on cycle X", "loss went NaN", "OOM at step Y", "为什么挂了", "为什么对不上") and wants an independent investigation, or has been staring at a stack trace / log / waveform without progress.

极简并发开发协调器:把任务派给在 herdr workspace 里长跑的 codex worker,用 /loop 心跳 定期巡检防跑偏。协调器每 tick 自主决策(看什么、驭什么、何时合),验证通过自动合本地 main,只在自己裁不动或对外不可逆时通知人类。触发于:"起 curryflows"、"用 curryflows 跑并发开发"、"派个 codex 长跑任务"、"监督 codex 别跑飞"。

Incrementally builds a recursive tree of implementation-plan files under docs/plans/<root-topic>/, one node at a time, fully user-paced. Each node's plan is produced by an internal Claude<->Codex cross-review and written in the full gen-plan format (Goal, Acceptance Criteria with TDD tests, Path Boundaries, Feasibility Hints, Dependencies, Task Breakdown, Claude-Codex Deliberation, Pending User Decisions). A node can be generated from a spec/design file of any text format, a draft co-written in discussion, or pure discussion with no file. The root stays high-level; detail lives in children. There is no loop, no Stop hook, no cron, no budget, no auto-progression -- the user pushes each step. Use to grow a structured plan hierarchy where every node is independently cross-reviewed. Triggers on "plan tree", "plan-tree", "build a plan tree", "decompose this plan into sub-plans", "add a sub-plan", "grow the plan tree", or requests to recursively break a design into cross-reviewed plan files.

Bootstrap or refresh a schema-first, agent-retrievable knowledge base (sources/ + wiki/ layers, YAML frontmatter, id cross-references) for a technical domain, generalized from the KernelWiki pattern. Triggers on "build a wiki for X", "do what KernelWiki does for Y", "帮我做一个X领域的agent知识库", "把这个领域的资料整理成可检索的wiki", "用workflow构建<domain>wiki", "refresh the domain wiki". Do NOT use for a single one-off research question (use deep-research instead) or to query a wiki this skill already built (that wiki's own generated SKILL.md/scripts handle queries, not this one).

Read-only overseer for a Codex CLI session running /goal autonomously. A dedicated Claude Code session can either LAUNCH a fresh, fully-autonomous Codex as a subagent (codex --dangerously-bypass-approvals-and-sandbox, optionally seeded with a /goal) or ATTACH to one already running, then audits it on a self-managed cron tick using read-only Explore subagents (model opus). It derives the project's drift criteria from the target's own transcript and, on a finding, sends a single Escape to soft-stop the run, lands it idle, logs the finding to the terminal and a log file, and PAUSES -- handing off to the human. After the human reviews and directs a new instruction, the overseer types and submits it into the interrupted /goal so it resumes, then optionally auto-resumes the audit. The only writes to the target are the Esc and the human-directed instruction; everything else is read-only. Use to supervise an autonomous Codex /goal run for direction drift, fabricated data, fake/stub implementations, and broken project

Academic paper PDF reading and analysis tool. Supports two commands: (1) /read-paper [pdf-path] - Deep analysis of a single paper with structured Chinese output covering motivation, methods, experiments, and critical evaluation. (2) /compare-papers [pdf-path1] [pdf-path2] ... - Comparative analysis of multiple papers with side-by-side tables and recommendations. Use when the user wants to read, understand, summarize, or compare academic papers (PDF files).