Token-efficient subagent delegation for long AI coding sessions.
Subagent tool results inject back into the main agent context verbatim. Vanilla explore/review agents often return 1,500–2,000 tokens of prose per delegation. cavecrew presets emit structured, compressed output (~60% smaller) while preserving path:line locators and actionable fixes.
Across 20 subagent delegations in one session:
| Mode | Avg tokens/delegation | 20 delegations |
|---|---|---|
| Vanilla Explore | ~2,000 | ~40,000 |
| cavecrew-investigator | ~700 | ~14,000 |
The difference is often context exhaustion vs finishing the task.
| Preset | Job | Use when |
|---|---|---|
cavecrew-investigator |
Locate code (read-only) | "Where is X defined / what calls Y" |
cavecrew-builder |
Surgical edit, ≤2 files | Scope obvious, single-function fix |
cavecrew-reviewer |
Diff/file review | One-line findings with severity |
Use vanilla agents when you want prose, architecture commentary, or 3+ file refactors.
- Copy
SKILL.mdandagents/*.mdinto your agent skills folder (e.g..cursor/skills/cavecrew/or.agents/skills/cavecrew/). - Configure subagent types that load the agent prompts from
agents/. - Trigger with: "use cavecrew", "spawn investigator", "delegate to subagent".
See SKILL.md for the full decision matrix and docs/OUTPUT_CONTRACTS.md for parseable output formats.
investigator → main picks sites → builder → reviewer
Parallel scout: spawn 2–3 investigators (defs / callers / tests) in one turn; aggregate in main thread.
- caveman — ultra-compressed communication mode
- Agent Skills (SKILL.md) — Anthropic / Cursor skill format
- Deterministic tool-result compression: see skill-runtime-adapter companion repo
MIT — see LICENSE.