AGENT SKILL

kit-builder

Use when a selected config must be assembled into a runnable kogitsune collection or launch command, when materializing a new named kit in kits.yaml, or after kit-selector produces a decision that needs to become an actual kit to run.

License 许可
MIT
最近更新
2026-11-07
Stars
★ 0
Forks
0

01 前言 Front Matter

name
kit-builder
version
0.1.0
description
Use when a selected config must be assembled into a runnable kogitsune collection or launch command, when materializing a new named kit in kits.yaml, or after kit-selector produces a decision that needs to become an actual kit to run.

03 SKILL.md 正文

kit-builder

Overview

Takes a kit-selector decision and assembles the collection: either a launch command for an existing kit or a new named kit entry (model + MCP + skills) in the catalog. It then records whether the build worked, so future builds learn which compositions hold up. Builds what the selector selected.

When to Use

  • A selector decision needs to become a real launch (kit <name> -- "<task>")
  • A recurring composition deserves promotion to a named kit in kits.yaml
  • Capturing build success/failure to feed the learning loop

Resolve the datastore first

This skill is published to ~/.claude/skills/ and runs from whatever directory the session is in, so decider is not on PATH and no repo-relative path is safe. kit is on PATH and points into the repo — resolve through it:

KOG_ROOT="$(cd "$(dirname "$(readlink "$(command -v kit)" 2>/dev/null || command -v kit)")/.." && pwd)"
DECIDER="$KOG_ROOT/skills/lib/decider.sh"

If kit is not on PATH, say so and stop rather than guessing a location.

How it works

  1. Read the decision (from decisions.jsonl or passed inline).
  2. If it maps to an existing kit → emit the launch command. If it's a novel composition → assemble a kits.yaml entry (model, mcp, skills) and validate it resolves (kit show <name> / dry-run).
  3. Record the outcome: "$DECIDER" append-build '<json>' with from_decision, built, success, notes (see decisions/SCHEMA.md).

Improve with experience

builds.jsonl records which built configs actually launched/worked. Recurring successful compositions are promotion candidates (make them named kits); repeated failures are a signal to the selector to stop choosing them. Over time the builder favors proven builds.

Versioning

  • Skill: frontmatter version:.
  • Outcomes: append-only decisions/builds.jsonl (ids B001…); promoted kits are versioned by git history of kits.yaml; current counts in decisions/VERSIONS.md.

Common Mistakes

  • Building without validating the config resolves — verify before recording success.
  • Not recording failures — the loop only learns if losses are logged too.
  • Silently inventing a kit not grounded in the selector's decision — build what was selected.

04 同仓库更多 Skills