vanzll

continuous-skill-learning

Distill verified lessons from user corrections, avoidable retries, reviews, incidents, and surprising task outcomes into concise reusable Codex skills. Use automatically after evidence exposes a wrong, missing, stale, duplicated, or overly specific skill rule, or when a genuinely reusable workflow has no existing skill. Correct and prune existing skills before creating new ones.

edit-overleaf-with-edge

Edit an open Overleaf paper through the user's Edge browser extension, process author review comments and tracked revisions, polish or translate manuscript text, compile and inspect the rendered PDF, upload paper assets, and preserve completed edit batches with timestamped Overleaf History labels. Use when the user asks Codex to modify, polish, review, compile, or visually refine an Overleaf project in Edge.

Code & Development Listed

github-paper-review-workflow

Run author-first academic paper revisions through GitHub pull requests. Use when creating section-specific paper PRs, maintaining a Chinese semantic Markdown source alongside formal English LaTeX, collecting GitHub review comments and unresolved threads, distinguishing human author instructions from AI replies made by the same account, revising a paper from review feedback, or auditing whether a paper PR is ready to merge.

long-task-relay

Monitor long-running commands with a token-free rule-based watcher that wakes the exact agent conversation only for milestones, failures, stalls, or completion. Use when an agent would otherwise sleep and poll a task for minutes or hours.

manage-paper-experiments

Manage formal ML paper experiments from planning through reproducible reporting. Use when building or auditing the paper-wide experiment ledger, assigning paper experiment IDs, freezing code and protocols, scheduling GPU-node queues or waiters, generating send-and-forget remote Codex prompts, monitoring and reconciling W&B runs, retaining best/final checkpoints, or preparing paper curves and result tables.

multinode-training

Design, implement, launch, audit, and debug reliable multi-node GPU training. Use for torchrun, Accelerate, DeepSpeed, DDP, FSDP, tensor or sequence parallel jobs; distributed RL pipelines with rollout or reward services; cluster launchers, topology changes, collective hangs, asset staging, checkpointing, and send-and-forget experiment queues.

plot-paper-experiments

Create publication-ready machine-learning experiment figures from W&B or local tabular data, especially RL training/evaluation curves, ablations, scaling plots, and benchmark bars. Use when Codex must retrieve run histories, compare methods, establish paper-facing plot semantics, render conference-quality figures, or preserve reproducible figure data and metadata.

shared-filesystem-agent-coordination

Coordinate coding agents through a reliable shared-filesystem message bus with one coordinator, persistent dispatchers, bounded stateless workers, optional exact-thread resumes, atomic request/ACK/result records, fencing, monitoring, and repair. Use when agents on different machines or sessions must collaborate without direct messaging.

write-insightful-topconf-paper

Plan, discuss, draft, revise, and audit insight-driven papers for top ML/AI/vision/NLP conferences such as ICLR, ICML, NeurIPS, CVPR, and ACL. Use when Codex must shape a paper story from observations, derive a method from evidence, formalize mathematical claims as labeled results with proofs, map claims to experiments, write or revise LaTeX/Overleaf sections, improve its writing rules from author feedback, create numbered Observation/Insight/key-equation callouts, simplify academic prose, or perform a skeptical top-conference review. Also use for paper outlines, abstracts, introductions, motivation, methods, experiments, related work, limitations, rebuttals, and submission-readiness checks.