AGENT SKILL

analyze-publish

Inspect, compare, regrade, export, share, upload, and publish Harbor results, trajectories, artifacts, Hub records, and versioned task or dataset packages with mutation and credential gates.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
analyze-publish
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Inspect, compare, regrade, export, share, upload, and publish Harbor results, trajectories, artifacts, Hub records, and versioned task or dataset packages with mutation and credential gates.
disable-model-invocation
true

02 资源文件 6

03 SKILL.md 正文

Analyze and publish Harbor outcomes

Use this skill when a Harbor run already exists and the task is to inspect its results, understand trajectories or artifacts, compare runs, regrade a recorded execution, browse the local viewer or Hub, export traces, upload/share job results, or publish and version task/dataset packages. Keep the source record immutable unless a command below explicitly creates a derived result or asks for confirmation.

Route by intent

Route task/schema/verifier authoring to author-benchmarks, job/provider/agent execution failures to run-evaluate, and custom framework or plugin code to integrations. Do not relaunch an evaluation merely to inspect a recorded run.

Safety and mutation gates

  1. Identify whether the input is a local trial directory, local job directory, Hub job/trial UUID, task package, or dataset package before selecting a command. Check harbor --version when command behavior may be version-sensitive.
  2. Prefer local parsing, --json, harbor hub ... show/list/tasks/trials/shares, harbor version list/show, and harbor view for read-only inspection.
  3. Treat harbor analyze and harbor check as source-preserving but not cost-free: they create derived Harbor jobs, may run an evaluator agent and environment, and write reports. Show rubric, agent/model, environment, attempts, filters, and output directory before running them.
  4. Never run harbor upload, Hub copy/share/retry/cancel/delete, visibility or access changes, harbor publish, or harbor version tag without explicit approval for the exact target, credentials, visibility, recipients, and mutation. harbor traces export is local by default; --push or OTLP endpoint upload is an external mutation and needs the same gate.
  5. Keep credentials in the CLI's auth flow or environment. Do not print tokens, copy private result contents into prompts, or infer that a local login grants access to another organization.

Safe operating sequence

  1. Inventory the local tree (config.json, lock.json, result.json, job.log/trial.log, artifacts/manifest.json, agent/trajectory.json, native session files, retry folders) and preserve a digest or copy when an audit needs reproducibility.
  2. Read result and manifest JSON before opening large logs. Separate the recorded agent/model/configuration from verifier rewards, exceptions, costs, retries, and post-hoc analysis. A retry is a new attempt; a regrade is a new derived trial that does not rerun the agent.
  3. Use the narrowest read-only command or local parser. For a Hub record, inspect with harbor hub job show, tasks, trials, shares, compare, or hub trial show; use --json for machine processing and --include-retries when retry history changes the conclusion.
  4. For a regrade, verify the source has readable result.json, a trustworthy artifact manifest and bytes, and that the replacement verifier is single-step and separate-mode. Use harbor job regrade or harbor trial regrade; record the new output path and compare the printed reward delta.
  5. For any external operation, preview the exact command and expected side effects, obtain confirmation, check authentication, then verify the returned ID, visibility, revision, digest, or URL. Record partial failures and retry only the failed/idempotent portion.

What this skill does not claim

Help/parser checks and mocked/unit evidence validate command contracts only. Docker, cloud providers, model APIs, Hub/registry credentials, Hugging Face, OTLP endpoints, GPU, and frontend build toolchains are optional capabilities and must remain explicitly unverified unless actually exercised with approval.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351