SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsadvise

advise

Agents & models

View repository

Updated 10-16 07:48

About this skill

advise — every harness answers, side by side

Purpose

One question goes to each subscription CLI harness in parallel; every answer comes back verbatim and labeled. The value is the un-synthesized comparison — you see where four different vendors' flagship models agree and diverge. Runs entirely on logged-in CLI subscriptions; no API keys are read.

When to use / when NOT to use

Use for opinion questions, design calls, tradeoff checks, "sanity-check this idea against several models". NOT for: a single decisive answer (that is /council), live web research or citations (that is /explore), or anything requiring repo file access — members answer from model knowledge only.

Workflow

  1. Run from the researchkit repo root (or point --directory at a checkout):
uv run researchkit advise "QUESTION"
uv run researchkit advise "QUESTION" --context-file notes.md   # append context
uv run --directory /path/to/researchkit researchkit advise "QUESTION"

  1. Read the output: one ## <harness-spec> section per member. A failed member prints *failed: <reason>* and never blocks the others; the stderr tail says N/5 harnesses answered.

  2. Relay the answers side by side, attributed. Do not merge them into one answer — if the user wants synthesis, run /council instead.

Default members come from the harness preset in models.yaml (claude:claude-opus-5@xhigh, codex:gpt-6-astra@xhigh, agy:gemini-3.8-flash-high, grokcli:grok-4.6, kimicli:kimi-code/k3). Override per run: --harnesses codex:gpt-6-astra grokcli:grok-4.6 (a @<effort> suffix sets reasoning effort where the CLI supports it). uv run researchkit doctor --preset harness verifies every pin before spending anything.

Failure modes

  • A harness that is not installed, not logged in, or out of subscription quota ("usage limit", "Individual quota reached") fails that member only — the reason prints inline; exit code is 1 only when ALL members fail.
  • Requires the CLIs to be signed in (claude, codex, agy, grok login, kimi login) — there is no API-key fallback by design.
  • Expect ~1–5 min wall clock (parallel; slowest member dominates). Each harness call gets a 30-minute budget — a dense prompt at a high effort tier really can take >10 minutes, and a tighter clock just turns the slowest (usually codex) into a *failed: timed out* line. Override with RESEARCHKIT_CLI_TIMEOUT=<seconds> if you need a hard cap.