CtrlK
BlogDocsLog inGet started
Tessl Logo

harness

Delegate a hard decision or conundrum to the Kyotee harness engine — a cross-vendor Council that debates to consensus, a Two-Brain divergent/ convergent debate, or a forced tool-checking solo run. Use when a decision is high-stakes, genuinely contested, or benefits from multiple model perspectives, and you want a synthesised answer (with dissent noted) back in this session. Requires a running kyotee engine (see setup).

71

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Harness Skill

When to use

  • A design/architecture decision with real trade-offs and no obvious winner → --strategy council.
  • A creative-but-constrained problem needing option generation then rigorous pruning → --strategy twobrain.
  • A question hinging on current/present-state facts you must not answer from memory → --strategy solo --thinking slow (forces the tool-check pre-pass).

When NOT to use

  • Routine coding you can do directly. This spins up external models and costs money; reserve it for genuinely hard sub-decisions.

Limitation (read this)

Claude Code is already an agent loop. This skill does not turn this session into a council — it delegates to the external Kyotee engine and brings the synthesised result back. Think "consult a panel," not "become the panel."

Setup

  • Ensure the engine is running (kyotee serve) and KYOTEE_URL points at it (default http://127.0.0.1:8484; HARNESS_URL is honoured too).
  • Ensure provider API keys are configured in the engine's config (~/.kyotee/config.yaml; check with kyotee providers).

Usage

Run the CLI and capture the synthesised answer (progress streams to stderr; stdout carries only the answer):

kyotee ask "Should we adopt event sourcing for the billing service?" \
    --strategy council --consensus judge --budget 3.00 --wait

Then incorporate the printed answer (and any noted dissent) into your work.

For machine-readable output add --json, which emits a stable contract:

{
  "task_id": "…",
  "answer": "…",
  "strategy": "council",
  "consensus": { "reached": true, "method": "judge", "rounds_used": 2 },
  "dissent": ["gemini-3-pro favored a simpler queue-based approach"],
  "total_cost_usd": 1.87,
  "total_tokens": 41230
}

dissent is populated when the council deadlocked with noted disagreement or a judge flagged holdouts — surface it rather than presenting false confidence.

Other subcommands: kyotee resume <task_id> --wait, kyotee status <task_id>, kyotee providers, kyotee config validate <file>.

The CLI exits non-zero on engine errors or when the budget was exhausted before any answer was produced; without a running engine it fails fast with a "start kyotee serve / check KYOTEE_URL" message.

Repository
stukennedy/kyotee
Last updated
First committed

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.