Ar Experiment Runner

AR Experiment Runner

End-to-end controller for ar-runtime skills experiments. The user should be able to say only:

我需要做 <experiment_goal> 实验

and this skill drives everything. Two sibling skills are always in force and MUST be respected even mid-flow:

• ar-gpu-preflight — GPU need / sizing / RED-YELLOW-GREEN
• ar-workspace-safety — filesystem safety + correct conda/pip usage

If either is not loaded in this session, follow the rules below as if they were.

Default Paths

DATA_DISK=.
WORKSPACE=$DATA_DISK/workspace

If $DATA_DISK does not exist, inspect pwd / df -h and ask once before continuing.

Phase 0 — Resolve Profile From Minimal Input

For "我需要做 实验" style input:

1. Restate the inferred goal in one sentence.

2. Build a slug (lowercase, hyphens, ≤ 64 chars). Examples:

   • "GPU smoke test" → gpu-smoke
   • "CPU matmul smoke" → cpu-matmul-smoke
   • "train mnist classifier" → train-mnist-classifier

3. Derive paths:

   profile   = <slug>
   workspace = $DATA_DISK/workspace/projects/<slug>
   env       = $DATA_DISK/workspace/envs/<slug>
   artifacts = $DATA_DISK/workspace/artifacts/<slug>/<run_id>
   run_id    = $(date +%Y%m%dT%H%M)-<slug>
   

4. If config/experiment-profiles*.json defines <slug>, use that file's paths/deps/run instead.

5. If the goal is too vague to infer GPU need or code shape, ask AT MOST 1–3 focused questions. Never ask the user to paste a long checklist.

If the user gave an existing script path under workspace/projects/<name>/, infer profile = <name> and read the script before doing anything else.

Phase 1 — Plan

Print a short plan and wait for go only if the operation is destructive, expensive, or long. For routine smoke tests, proceed without confirmation.

Profile:
Workspace:
Env:
Artifacts:
GPU need: required / optional / none / unknown
Planned code files:
Planned deps:

Phase 2 — Workspace

mkdir -p "$DATA_DISK/workspace"/{projects,artifacts,scratch,envs}
mkdir -p "$DATA_DISK/workspace/projects/<slug>"
mkdir -p "$DATA_DISK/workspace/artifacts/<slug>/<run_id>"

Never write outside $DATA_DISK/workspace/. The only allowed exception is /tmp/ for installer downloads that you delete afterwards. (Full rules: ar-workspace-safety.)

Phase 3 — Code

If you are writing the experiment:

• Put code under $DATA_DISK/workspace/projects/<slug>/.
• Use a clear entrypoint: run.py, train.py, matmul.py, gpu.py, etc.
• Make output paths configurable; default them under the artifacts dir.
• Set a deterministic seed when there is randomness.
• No hardcoded paths outside the workspace.

If the user provided an existing script:

• Read it first.
• Do not run until env + preflight are ready.

Phase 4 — Env (delegate to ar-workspace-safety)

Use the workspace-local Miniconda only:

$DATA_DISK/workspace/envs/.miniconda/

Create or reuse the profile env by path (never -n <name>):

"$DATA_DISK/workspace/envs/.miniconda/bin/conda" create \
  -p "$DATA_DISK/workspace/envs/<slug>" python=3.11 -y

Always install and run via the activated profile env:

source "$DATA_DISK/workspace/envs/.miniconda/etc/profile.d/conda.sh" && \
conda activate "$DATA_DISK/workspace/envs/<slug>" && \
pip install <deps>

HARD RULES (also enforced by ar-workspace-safety):

• NEVER python3 <script> — always activate the profile env first, or call <env>/bin/python directly.
• NEVER naked pip install <pkg> to fix ModuleNotFoundError.
• NEVER use conda base, system Python, or conda init / ~/.bashrc edits.

Phase 5 — GPU + Preflight (delegate to ar-gpu-preflight)

Before any run, output the preflight table from ar-gpu-preflight. Minimum fields:

Profile / Workspace / Env / Artifacts:
GPU need:
Estimated cards:           Estimated memory:
Available cards:           Chosen CUDA_VISIBLE_DEVICES:
A1 env activates / A2 imports resolve
B0 GPU need / B1 CUDA reachable / B2 GPU free / B3 card plan
C1 CPU/RAM / C2 disk
D1 output paths / D2 destructive ops / D3 seed
E1 runtime command
VERDICT: GREEN / YELLOW / RED — <reason>

Verdict rules: RED stops; YELLOW asks; GREEN runs. For GPU code, missing nvidia-smi check, missing memory estimate, or missing CUDA_VISIBLE_DEVICES is automatically RED.

Phase 6 — Run

Short foreground run:

source "$DATA_DISK/workspace/envs/.miniconda/etc/profile.d/conda.sh" && \
conda activate "$DATA_DISK/workspace/envs/<slug>" && \
cd "$DATA_DISK/workspace/projects/<slug>" && \
CUDA_VISIBLE_DEVICES=<ids_if_gpu> python <entrypoint> \
  > "$DATA_DISK/workspace/artifacts/<slug>/<run_id>/run.log" 2>&1

Long job (>~5 min or training): use nohup or tmux, capture PID/session, log to artifacts:

mkdir -p "$DATA_DISK/workspace/artifacts/<slug>/<run_id>"
nohup bash -lc '
source "$DATA_DISK/workspace/envs/.miniconda/etc/profile.d/conda.sh" &&
conda activate "$DATA_DISK/workspace/envs/<slug>" &&
cd "$DATA_DISK/workspace/projects/<slug>" &&
CUDA_VISIBLE_DEVICES=<ids> python <entrypoint>
' > "$DATA_DISK/workspace/artifacts/<slug>/<run_id>/run.log" 2>&1 &
echo $!

Phase 7 — Report

Profile:
Workspace:
Env:
Artifacts:
GPU allocation:
Command:
Exit code / PID:
Key output or log path:

Tail the log for short runs; for background jobs, give the user the PID and log path.

Hard Rules (Defense In Depth)

1. NEVER run user experiment code with system python3 or conda base.
2. NEVER naked pip install. Install only into the activated profile env.
3. NEVER touch GPU code paths without showing the ar-gpu-preflight table first.
4. NEVER write outside $DATA_DISK/workspace/ (except /tmp/ for transient downloads).
5. NEVER rm -rf, git reset --hard, or bulk-delete without listing targets and asking.
6. NEVER ask the user to paste a long ar-runtime skills checklist when minimal input is enough.
7. NEVER let a GPU job grab every card by default — CUDA_VISIBLE_DEVICES is mandatory.

Anti-Patterns

• Running first, deciding env/GPU afterwards.
• Fixing ModuleNotFoundError with pip install <pkg> outside the profile env.
• "It worked because base happened to be active."
• Silently rewriting a GPU task to CPU because torch.cuda.is_available() returned False.
• Skipping preflight on a "small" run that turns out to allocate a lot of GPU memory.