AGENT SKILL

extension-patterns

Add and inspect Habitat-Lab registry extensions for tasks, sensors, measures, actions, datasets, simulators, and baselines components.

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

01 前言 Front Matter

name
extension-patterns
license
MIT
metadata
{ "disco-role": "operating" }
description
Add and inspect Habitat-Lab registry extensions for tasks, sensors, measures, actions, datasets, simulators, and baselines components.
disable-model-invocation
true

02 资源文件 4

03 SKILL.md 正文

Habitat-Lab Extension Patterns

Use this sub-skill when you need to add, inspect, debug, or review custom Habitat-Lab extension classes that are discovered through habitat.registry or baseline_registry. It covers registration and config alignment; it does not run environments or train agents.

Fast Route

  • For exact decorator/getter names, ownership boundaries, and config-name alignment, read references/registry-api.md.
  • For recipes for sensors, measures, actions, tasks, datasets, simulators, policies, trainers, storage, updaters, and observation transforms, read references/extension-patterns.md.
  • For failure diagnosis around missing keys, duplicate registrations, action spaces, measure dependencies, and transform shapes, read references/troubleshooting.md.
  • To safely inspect currently importable registry keys without creating a simulator or Env, run scripts/inspect_registry_extensions.py from a checkout or installed environment.

Working Rules

  1. Define the extension class in an importable Python module and decorate it with the matching registry decorator.
  2. Ensure the module is imported before config construction or before a factory calls registry.get_* or baseline_registry.get_*.
  3. Align every config type value with the registry key, not necessarily with the config dictionary key or emitted observation/action UUID.
  4. Keep runtime tests narrow: registry lookup, config-node creation, observation/action space compatibility, and shape contracts can be checked without full simulator assets.
  5. Escalate full Env execution, dataset loading, and training/evaluation to the task/env and baselines sub-skills.

Common Extension Flow

  1. Pick the registry surface: habitat.registry for lab runtime components, baseline_registry for training/inference components.
  2. Pick a stable registry key with @...register_*(name="...") when the public config name should differ from the class name.
  3. Create or update the matching structured config class so its type field equals that registry key.
  4. Add the config node under the correct collection, such as habitat.task.lab_sensors, habitat.task.measurements, habitat.task.actions, habitat.dataset, habitat.simulator, habitat_baselines.trainer_name, habitat_baselines.rl.policy.name, or habitat_baselines.rl.agent.type.
  5. Import the module containing the decorator before the lookup site runs.
  6. Probe with registry lookups and shape/space checks before attempting asset-backed execution.

Safe Probe

From this sub-skill directory, run the bundled probe to list registered keys:

python scripts/inspect_registry_extensions.py --all

Useful narrow probes:

python scripts/inspect_registry_extensions.py --category sensor --category measure
python scripts/inspect_registry_extensions.py --import-module my_project.habitat_extensions --category task_action --category policy
python scripts/inspect_registry_extensions.py --key measure:MyMeasure --key trainer:ppo

The script imports registries and optional extension modules, then prints mapping keys. It does not instantiate Env, datasets, simulators, trainers, or policies.

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