SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsadvise-max

advise-max

Agents & models

View repository

Updated 10-16 03:15

About this skill

advise-max — flagship models, maximum effort, side by side

Purpose

advise, but every member is the strongest model its harness offers, at the highest reasoning tier the CLI accepts. Slower and much heavier on subscription quota than plain /advise — reserve it for questions that deserve it.

Workflow

  1. Run from the researchkit repo root (or --directory a checkout):
uv run researchkit advise "QUESTION" --harnesses \
  claude:claude-fable-5@max \
  codex:gpt-6-astra@ultra \
  agy:gemini-3.8-flash-high \
  grokcli:grok-4.6@xhigh \
  kimicli:kimi-code/k3

--context-file notes.md appends context, same as /advise.

  1. Relay each ## <harness-spec> answer side by side, attributed — no merging (run /council with these same --harnesses for synthesis).

Model notes (Codex re-verified live 2026-09-07; the rest 2026-08-21): gpt-6-astra is the current Codex flagship, and it supersedes gpt-5.6-sol. Effort is a REASONING TIER, never part of the model id — pass -m gpt-6-astra with model_reasoning_effort, since an id like gpt-6-astra-ultra does not exist. Astra accepts both ultra and xhigh (both answered a probe cleanly); ultra is what this skill uses and xhigh is what codex's own config defaults to. Antigravity now lists real ids (agy models prints <id>\t<display name>) and bakes the thinking level into the id, so pass agy:gemini-3.8-flash-high — no quoting, and a @high suffix does nothing. The Grok CLI's effort ceiling rose to xhigh (max is still rejected: "use one of: xhigh, high, medium, low"); grok models offers grok-4.6 (default) and grok-4.5. Kimi Code has no effort flag — kimicli:kimi-code/k3 runs at the CLI config's [thinking] effort. Effort tiers and model ids drift — uv run researchkit doctor --preset harness catches dead pins in seconds, and when a member fails with an "unknown effort/model" error, trust the CLI's error text over these notes.

Failure modes

  • Same isolation as /advise: a member out of quota ("usage limit", "Individual quota reached") or not logged in fails alone, reason inline.
  • Ultra/max tiers multiply token burn per answer — expect ~2–15 min and meaningful subscription usage per run. A long, multi-question prompt can push a single member past 10 minutes; harness calls get a 30-minute budget for exactly that reason (RESEARCHKIT_CLI_TIMEOUT=<seconds> raises or lowers it). Do not shorten it to "speed the run up" — that is what used to turn codex into a *failed: timed out* line.