favorite-harness

Look up the user's preferred coding-agent harness, model, effort, and autonomous launch command. Use when choosing or launching a coding agent or when another routing skill needs defaults; do not use

SKILL.md

GitHub

Favorite harness

This skill selects defaults; it does not orchestrate or launch by itself. Match rows top-to-bottom, and preserve each table value the user did not explicitly override. When a caller launches, use the complete autonomous command shown.

Job Harness Model Effort
Orchestration Claude Opus 5.5 xhigh
Final review, security review Claude Opus 5.5 high; Fable 5.1 xhigh for the hardest cross-cutting reasoning or a high-stakes security audit
Implementation Codex GPT-6 Sol high; stuck → xhigh; still stuck → GPT-6 Astra high
Tests, retrieval, checking, creating issues Cursor Agent Grok 4.6 high; medium if mechanical
Observation (an auto-improve observer, polling a transcript for hours) Claude Opus 5.5 medium; Cursor Agent Grok 4.6 high once its spend runs high
Anything that is not tool/repo work Kimi Code CLI K3 max
# Claude — Opus 5.5 xhigh (orchestration); high (review); medium (observers)
claude --model opus --effort xhigh --dangerously-skip-permissions
claude --model opus --effort high --dangerously-skip-permissions
claude --model opus --effort medium --dangerously-skip-permissions

# Claude — Fable 5.1 xhigh (hardest reasoning, high-stakes security audit)
claude --model fable --effort xhigh --dangerously-skip-permissions

# Codex — GPT-6 Sol high (stuck → xhigh); Astra high as the last step
codex -m gpt-6-sol --approve-for-me -c 'model_reasoning_effort="high"'
codex -m gpt-6-sol --approve-for-me -c 'model_reasoning_effort="xhigh"'
codex -m gpt-6-astra --approve-for-me -c 'model_reasoning_effort="high"'

# Cursor Agent — Grok 4.6 high (mechanical → medium); never a *-fast variant
agent --model cursor-grok-4.6-high --yolo --sandbox disabled
agent --model cursor-grok-4.6-medium --yolo --sandbox disabled

# Kimi Code CLI — K3 max. No launch-time effort flag; send /effort max once idle.
kimi -m k3 --yolo

In Herdr, --kind is claude / codex / cursor / kimi. Pass everything after the executable after --. See herdr-orchestration for the refused-launch fallback.

Opus 5.5 thinks more per turn than Fable at the same effort name, so carry an effort level across models only through this table. When a Claude session reports "Switched to" an older model (a safety flag), report it; the work did not run on the model this table picked.

Grok 4.7 (grok-4.7-*, no cursor- prefix) is opt-in: per task it costs about twice as much as 4.6, and Cursor subagents set to it silently run on Auto.

Never pin max or ultra on any model, or Grok xhigh. Do not put Kimi on a tool loop.