Menu
Spawn a loomcycle evaluator agent to score the current Claude Code transcript (or a named run) and report the result. Use when the user wants an independent quality/score judgement of work just done via loomcycle.
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
| name | loomcycle-spawn-evaluator |
| description | Spawn a loomcycle evaluator agent to score the current Claude Code transcript (or a named run) and report the result. Use when the user wants an independent quality/score judgement of work just done via loomcycle. |
Use this when the operator wants an independent evaluation of work — either the current Claude Code transcript or a specific loomcycle run — scored by a loomcycle evaluator agent rather than by you directly.
Steps:
Identify what's being evaluated. Ask the operator (or infer) whether
they want to evaluate (a) the work in the current conversation, or (b) a
specific completed run. If (b), get the run_id.
Choose the evaluator agent. Ask which registered loomcycle agent should
act as evaluator (e.g. evaluator, judge). Do not guess a name — if none
is known, list options via the runs/agents the operator mentions, or ask.
Spawn the evaluator with mcp__loomcycle__spawn_run. Pass the material
to be judged as the prompt, wrapped as segments:
{
"agent": "<evaluator-agent>",
"segments": [
{ "role": "user", "content": [
{ "type": "trusted-text", "text": "Evaluate the following and return a score 0-1 with a one-line rationale:\n\n<material>" }
] }
]
}
Use the active user_id / user_bearer from /loomcycle:connect if set.
Report the evaluator's score + rationale back to the operator clearly.
Offer to record it. If the work corresponds to a real run_id, offer to
submit the score via /loomcycle:eval <run_id> <score>. loomcycle does not
auto-promote on scores — recording is the operator's call.
Keep the loop honest: you are orchestrating an independent judge, so don't pre-bias the evaluator's prompt toward a verdict.
Configure a loomcycle runtime — providers, model tiers, user tiers, fallbacks, per-agent sampling and context-compaction, environment variables, deployment profiles (brew/in-system, containerized, true sandbox, server, multi-tenant, cloud), filesystem Volumes, the Bashbox in-process sandbox, the Path VFS, chunked-graph Documents, inbound webhooks, and third-party MCP servers. Use when the user wants to set up or tune loomcycle.yaml or its env, pick a deployment posture, wire provider routing/cost-cascades, gate plans, tune decoding (temperature/top_p) or compaction, lock down tool/sandbox/auth, enable or choose between Bash and the Bashbox sandbox (incl. its host-command fallback), name resources with Path, author chunked-graph Documents (and the SQL Memory they require), receive webhooks, or connect external MCP tools.
Diff two loomcycle AgentDef versions by def_id and show what changed in system_prompt, allowed_tools, max_tokens, and other fields. Use when the user wants to compare two versions of a self-evolving agent definition.
Guide the user through importing a .claude/ directory into loomcycle yaml using the `loomcycle import claude-code` CLI — dry-run first, review the report, then write. Use when the user wants to move Claude Code agents/skills/MCP servers into a loomcycle config.
Walk through replaying a failed loomcycle run — inspect the failure, adjust the prompt or agent, and re-spawn. Use when a loomcycle run errored or produced a bad result and the user wants to retry with a fix.