跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
效率与协作

orx-experiment-tree

Plan and drive the experiment tree: first-launch setup, fixed run contract, frozen nodes, stacked-bush tree shape, branch/launch/wait/promote, repair limits, notes, and turn summaries. Use before creating or changing experiments, launching a first run, deciding what to try next, handling a completed run, or reporting experiment progress.

alphaXiv · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add alphaXiv/OpenResearch --skill orx-experiment-tree

技能说明


name: orx-experiment-tree description: "Plan and drive the experiment tree: first-launch setup, fixed run contract, frozen nodes, stacked-bush tree shape, branch/launch/wait/promote, repair limits, notes, and turn summaries. Use before creating or changing experiments, launching a first run, deciding what to try next, handling a completed run, or reporting experiment progress."

A project is a tree of experiment nodes. The root (baseline) holds the starting code and a run command — the single shell command that trains or evaluates the node and prints its results to the run log. Every other node is a child branched off a parent, inheriting its code and its run command. The two rules this depends on — never edit a node a run has answered and the run command + env is a fixed contract — are the cardinal rules; everything below assumes them.

Before the first launch

Follow the session playbook's Python policy. Before launching, resolve the train/evaluation command and compute-specific requirements; ask only if the project setup leaves these or the chosen workflow unclear. Record the durable setup and execution recipe in the project's run command.

Provisional until it answers — repair, don't branch

Every node exists to establish a baseline or test a hypothesis. A run that dies on an error does neither — nothing was established, nothing was tested — so there is nothing to protect: fix that node's branch in place and re-run the same node. Successive runs on one node are how you get it working; a new node is for a new question.

Once a run does answer the node — it produced the result the node was after, good, bad, or nan — the node is frozen. Its branch is the code that result came from: never edit it again, branch a child instead. That holds however the run ended, and it is permanent — a disappointing number is a result, not a reason to repair.

Unintended behaviour is not an answer. An OOM, a timeout, a divergence from a bug, a missing dep — those are implementation and hardware details, and the node is still provisional (unless the node's hypothesis is about memory or runtime, in which case that outcome is exactly its result).

Repair cap: two runs in a row that answer nothing on one node, then ask the user. Different errors still count; a bare relaunch or a flavor/backend switch is a repair. If the same failure hits a second node, that is one setup problem — ask then. (Separate from the "~3 failed or regressed runs" scientific stop.)

Shape the tree — stacked bushes, not a flat fan or a noodle

The single most common way to drive a project badly is to get the shape wrong. There are two opposite failures, and the right shape sits between them:

FLAT FAN (wrong)            NOODLE (wrong)            STACKED BUSHES (right)
root                        root                      root
├ a ├ b ├ c ... ├ n         └ a                       └ lr-head        ┐ round 1:
                              └ b                        ├ lr 2e-5     │ a small fan of
                                └ c                      └ lr 3e-5     ┘ co-equal options
                                  └ d ...                   └ winner ── arch-head   ┐ round 2
                                                               ├ arch-A             │ descends onto
                                                               └ arch-B             ┘ round 1's winner
  • Flat fan (your whole sweep hanging off the root): every result is measured against the start, so wins never accumulate and the tree never makes progress.
  • Noodle (a long single-child chain): depth manufactured for its own sake — each step doesn't actually build on the one above it.
  • Stacked bushes (correct): a small fan within a round (the options of one decision), then descend onto that round's winner for the next round.

The one rule that produces this shape. Before you make X a child of Y, name what Y established that X builds on:

  • You can name it ("Y is the LR winner; X keeps that LR and changes the architecture") → real depth. X is a child of Y. Descend.
  • You can't — X and Y are co-equal options you're trying at the same time (lr 2e-5 vs lr 3e-5) → they don't build on each other. They're siblings in the same bush. Fan, don't chain.

So: width = the open options of one decision (fan freely — a 3-way LR sweep should be three siblings under a common head); depth = decisions already resolved, stacked (one level down per winner kept). A new round never hangs off the root — it hangs off the previous round's winner. That keeps the tree moving downward as research progresses, without stringing unrelated nodes into a line.

Re-read the tree each round — orx project view <projectId> lists every node (id, title, branch; roots marked [root]) — and check the shape: a wide row of direct children off the root with no grandchildren means you're fanning when you should be descending; a long depth-N chain with no branching means you're chaining co-equal variants that should have been siblings.

The auto-research loop

To drive a project toward a goal (e.g. "best convergence for d=8"), this is the intended flow — do not edit a frozen node or rewrite the run command:

  1. Read the baseline's code. You already sit in a private Git worktree of the project's repository. Check out the branch and read it with your normal tools (see orx-git). See the node's run command with orx exp status <expId> and find where the knobs live (config files, hyperparameters, model definitions).
  2. Form one round's worth of hypotheses — the co-equal options of a single decision (which LR? which schedule? which init?), each a concrete change you can make and measure against the others in this round. Don't mix decisions from different rounds into one batch — that's what produces the flat fan.
  3. **Create the round as a bush,
阅读完整来源 ↗