moleworks-subagent-orchestrator

name	moleworks-subagent-orchestrator
description	Orchestrate sub-agents for long-horizon Moleworks code, RL, and ROS tasks. Use when the task needs many code/doc searches, Slurm or W&B inspection, benchmark comparisons, or ROS process inspection without polluting the parent context.

Moleworks Sub-Agent Orchestrator

Use this skill when the task is too large or too noisy for one parent context window.

Goal

Keep the parent thread in the smart zone. The parent agent should orchestrate, decide, and integrate. Sub-agents should do context-heavy exploration and return compact results.

When To Use

Use sub-agents when the task involves one or more of these:

• repeated rg or file-discovery passes across the Newton or ROS repos
• Slurm log inspection, squeue, sacct, and run-dir archaeology
• W&B run lookup or metric comparison
• benchmark JSON comparison across multiple artifacts
• ROS/tmux/process inspection with lots of noisy output
• tracing a code path through several packages before patching

Do not spawn sub-agents just because the task is hard. Spawn them when there is clear context-isolation value.

Recommended Delegation Pattern

Parent thread responsibilities:

• decide scope
• choose which repo or branch matters
• assign a narrow question
• integrate the result
• make final code changes unless write ownership is explicitly delegated

Good sub-agent task shapes:

• "Find the exact branch-local entrypoints and docs for shared-turn benchmark + terrain-bank replay."
• "Inspect synced Slurm log X and summarize failure mode, first traceback, and likely owning script."
• "Compare benchmark JSON A vs B and report success/full/close/timeout deltas."
• "Inspect tmux panes and process list for the parity stack; return only stale PIDs, missing topics, and TF failures."

Role Suggestions In Codex

• Use explorer for read-only codebase or doc questions.
• Use worker only for bounded code changes with a disjoint write set.
• Reuse the same agent for follow-up questions on the same narrow topic instead of respawning.

Output Contract For Sub-Agents

Require compact answers:

• answer first
• then evidence as file paths, line refs, commands, or artifact names
• no raw log dumps unless the parent explicitly asks
• no speculative architecture essays

Good result shape:

1. direct answer
2. 3-8 bullets of evidence
3. open risks or unknowns

Moleworks-Specific Delegation Patterns

• Newton repo locator: docs, scripts, task config, benchmark entrypoints
• Cluster/log digester: cluster/*.sh, synced Slurm logs, docs/experiments/*
• Benchmark diff analyst: outputs/..., logs/rsl_rl/..., benchmark JSON, W&B identifiers
• ROS parity inspector: tmux panes, process lists, TF/topic health, duplicate publishers

Fallback If Sub-Agents Are Unavailable

If the harness cannot spawn sub-agents:

• keep exploration tightly scoped
• prefer rg over broad file reads
• inspect one subsystem at a time
• summarize findings before moving to the next subsystem
• avoid pasting large logs or command output into the main thread