README.md

Third-party agent skills

Vendored and submodule copies of upstream agent-efficiency projects used by this repo.

Upstream	License	What we use
DietrichGebert/ponytail	MIT	YAGNI code ladder — Cursor rule + Temporal edit prompts
JuliusBrussee/caveman	MIT (skill)	Terse prose — Cursor rule + Temporal plan prompts

Not vendored: Caveman Engine / Proxy (BSL-1.1). Use upstream CLI if you want input-token compression locally.

Layout

third_party/
  ponytail/           # git submodule (optional full upstream)
  caveman-skill/      # MIT skill only (SKILL.md + NOTICE.md)
.cursor/rules/
  ponytail.mdc        # synced from ponytail submodule or upstream raw
  caveman.mdc         # condensed Cursor rule; full skill in caveman-skill/
temporal/config/agent_skills.py   # prompt injection for worker agents

Install / update

./scripts/install-agent-skills.sh

Init submodule, refresh Cursor rules from ponytail, verify caveman skill copy.

Temporal worker toggles

export AGENT_PONYTAIL=full   # default — inject YAGNI ladder into edit prompts
export AGENT_CAVEMAN=lite    # default — terse plan step strings
export AGENT_PONYTAIL=off    # disable
export AGENT_CAVEMAN=off

Multi-agent roles (planned)

Role	Skill	Purpose
Planner / coordinator	caveman	Short task descriptions on mission board
Editor / builder	ponytail	Minimal diffs, reuse stdlib
Reviewer	ponytail-review pattern	Delete-list before merge

See docs/agent-skills.md.