Guides delegation of code location, editing, and review tasks to compressed-output subagents.
npx skills add https://github.com/th-efool/SKILLS --skill cavecrew-th-efool
Please help me install this Agent Skill. Skill: cavecrew Source: https://github.com/th-efool/SKILLS/tree/main/cavecrew Command: npx skills add https://github.com/th-efool/SKILLS --skill cavecrew-th-efool
cavecrew-investigator (locate code), cavecrew-builder (1-2 file edits), or cavecrew-reviewer (diff review) versus using vanilla agents or working inline. - Compressed Output Contracts: Each subagent returns structured, caveman-compressed results roughly 60% smaller than vanilla prose, with grep-friendly formats like path:line — symbol — note. - Chaining Patterns: Supports locate → fix → verify pipelines, parallel scouting with multiple investigators, and single-shot edits when the site is already known. - Use Case: During a large refactor, spawn two investigators in parallel to find definitions and callers, hand the chosen sites to the builder, then have the reviewer audit the diff — all while keeping main-context usage low. ## Quick Start Ask the assistant to delegate a code search to cavecrew, for example: use cavecrew to find where the authentication middleware is defined and then fix the token expiry bug there.High-intent search queries and answers about installing and using this skill.
Spawn cavecrew-investigator for read-only code location tasks like finding definitions, callers, or usages. It returns a compressed list of path:line entries with backticked symbols, which costs far less main-context budget than a vanilla Explore agent.
Use cavecrew-builder for surgical edits of one to two files where the scope is already obvious. For three or more files or cross-cutting refactors, it returns too-big and you should use the main thread or a code architect agent instead.
cavecrew-reviewer returns one-line findings with severity emoji and totals, sorted by file and line. The vanilla Code Reviewer provides deep reviews with rationale and alternatives, so choose it when you want prose rather than compressed findings.
Yes, set environment variables before launching Claude Code: CAVECREW_REVIEWER_MODEL, CAVECREW_BUILDER_MODEL, or CAVECREW_INVESTIGATOR_MODEL with values like haiku, sonnet, or opus. The override patches only the model line in the installed agent frontmatter.
Cavecrew subagents emit caveman-compressed structured output to shrink tool results injected into main context by roughly 60 percent. They automatically switch to normal English for security warnings and irreversible-action confirmations.
Avoid cavecrew-builder when you do not already know the target file, and avoid chaining investigator to builder for five-file refactors. Also skip cavecrew-reviewer for general feedback since it returns findings only, not architecture opinions.
dwaynimayDelegate locate, edit, and diff review tasks to compressed caveman-style subagents.
bmad-code-orgImplements work items into reviewed, verified code through a five-step spec-driven workflow.
MINT-IADispatch parallel agents to investigate independent test failures and tasks.
TrayMachiGuides delegation to compressed-output subagents for code location, editing, and review tasks.