SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsexplore

explore

Writing & research

View repository

Updated 10-16 07:48

About this skill

explore — boosted research on subscriptions only

Purpose

The full researchkit boost pipeline with every step on subscription CLI harnesses: the council (5 harnesses) refines the topic and decomposes it into sub-investigations; each sub-project researches with the five CLI providers (web search included); summaries and the cross-cutting super-summary run through CLIs too. No API key is read at any step.

When to use / when NOT to use

Use for deep, citation-backed research when API spend is unwanted or keys are unavailable. NOT for: quick questions (a run is 15–30 minutes of real subscription usage), opinion-only questions (/advise, /council), or maximum source breadth — API providers harvest more citations faster; the harness path trades breadth for depth (agents that actually read pages).

Workflow

  1. Pick the output directory first. If the project you are working in has a gitignored .local/ directory, put the run there — research output is bulky and must never reach a commit:

    git check-ignore -q .local && export RESEARCHKIT_PROJECTS_DIR="$PWD/.local/research"

    Check it is actually ignored (git check-ignore -q .local), not merely present: an un-ignored .local/ would stage the whole pack. With no .local/, leave the variable unset and output goes to projects/.

  2. Run from the researchkit repo root (or --directory a checkout). Quote the topic; add flags as needed:

    uv run researchkit explore "TOPIC" --days 7
uv run researchkit explore "TOPIC" --materials --ingest ../brainkit/brain

  3. It is alive, not hung: [boost] heartbeat lines on stderr track the council, each sub-project's providers, and the super-summary.

  4. Outputs land under <projects-dir>/<timestamp>_<topic>/ — .local/research when step 1 set it, else projects/ (the absolute path is printed as wrote: …): sub-reports in subprojects/*/report.md, the super-summary as the parent report.md. --materials archives cited pages per sub-project; --ingest <brain> hands the run to brainkit.

  5. Relay the super-summary; point at the sub-reports for depth. If stderr says "super-summary unavailable", the sub-reports still stand.

Defaults come from the harness preset in models.yaml (providers codex:gpt-6-astra / agy:gemini-3.8-flash-high / grokcli:grok-4.6 / kimicli:kimi-code/k3 / claude:claude-opus-5; site research off — its connectors are API-key paths). --preset <name> swaps the whole model set. uv run researchkit doctor --preset harness verifies every pin first — a 20-minute run is a bad place to discover a dead model id.

Failure modes

  • Providers degrade independently: a harness that fails or returns nothing (observed live: codex refusing heavy research prompts) costs its sources, not the run — Claude Code and the Grok CLI carry the depth today. Each harness call gets a 30-minute budget (RESEARCHKIT_CLI_TIMEOUT=<seconds> overrides), so slow agents finish rather than time out.
  • All five CLIs must be signed in; there is no API fallback by design.
  • Budget note: a boosted run is many harness invocations against your subscriptions — don't loop it casually.