Menu
Protocol for requesting human intervention when an agent encounters a blocker it cannot resolve autonomously.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Strategies for fitting a paper to a venue's page limit without destroying content quality. Applies to any phase that modifies paper content.
Rules for maintaining figure and table data integrity. All visual data must trace to result files — never invented or modified for convenience.
How to submit SLURM jobs on the SANDS Lab ROCS testbed cluster (KAUST). Covers GPU GRES selection (v100/a100 — p100 is off-policy), QoS caps, conda env activation, multi-node DeepSpeed, preemptible spot jobs, and the watchdog that kills jobs with <15% GPU utilization. Use for any project whose compute_backend is slurm and whose cluster nodes are named `mcnode*`.
Containerized isolation via Apptainer for reproducible experiments. Use whenever the project executes third-party code (agent frameworks, browser agents, codegen tools, benchmark harnesses), runs GPU workloads via Slurm on a shared cluster, needs a reproducible runtime environment for paper claims, or isolates untrusted/adversarial inputs from the host. Broadly applicable to measurement, benchmark, and systems papers — not only to malicious-code experiments.
Rules for maintaining research integrity. No simulation, no fabrication, no shortcuts.
Best practices for project-isolated environment setup. All installations must be local to the project, never global.
| name | human-intervention |
| description | Protocol for requesting human intervention when an agent encounters a blocker it cannot resolve autonomously. |
| tags | ["system","intervention","credentials","blocking"] |
Request human help when you encounter ANY of these situations:
Write a JSON file to results/needs_human.json with this format:
{
"urgency": "blocker",
"summary": "One-line description of what is blocked and why",
"stage": "Phase label / pipeline step where this surfaced",
"what_failed": "Exact command or condition that reproduces the blocker",
"evidence": {
"tested_commands": [
{"cmd": "pip install foo", "exit_code": 1, "output": "…"}
],
"error_output": "<truncated stderr if useful>"
},
"options": [
{"id": "1",
"title": "Provide API key ANTHROPIC_API_KEY",
"consequence": "Unblocks exp2/exp3; user pastes the key"},
{"id": "2",
"title": "Defer these experiments to next iteration",
"consequence": "Pipeline continues without them; paper claims weakened"}
],
"default_option": "2",
"timeout_minutes": 60
}
Each option is an action the user can take. The framework renders a numbered Telegram menu; the user's reply is mapped back to the option (or treated as free-text guidance if they write prose).
auto_research/state/hitl_decisions.yaml.
Read this file before retrying any previously-blocked experiment;
if the decision was "deferred", mark it deferred and do not retry."status": "blocked" — not "completed_fallback" or "degraded".needs_human.json first."urgency" field should almost always be "blocking". Use of "degraded" is reserved for truly optional enhancements, not for core experiments.