claude plugin marketplace add cdeust/zetetic-team-subagents
claude plugin install zetetic-team-subagentsThe plugin's setup.sh runs automatically and copies agents, skills, hooks, and tools into ~/.claude/.
git clone https://github.com/cdeust/zetetic-team-subagents.git
cd zetetic-team-subagents
bash scripts/setup.sh installsetup.sh sub-commands:
| Command | What it does |
|---|---|
install |
Install or re-install — copies tracked files to ~/.claude/ |
update |
Re-apply agent models, pull new assets, prune orphans |
configure |
Create ~/.claude/zetetic-agent-models.json with calibrated defaults |
uninstall |
Remove tracked files; keeps user-modified copies |
--dry-run |
Preview any of the above |
cp -r zetetic-team-subagents/skills/ ~/.claude/skills/~/.claude/zetetic-agent-models.json — yours to keep, plugin updates never overwrite. Run setup.sh configure to seed with calibrated defaults.
{
"patterns": [
{ "glob": "genius/*", "model": "sonnet" }
],
"agents": {
"refactorer": { "model": "haiku", "effort": "low" },
"engineer": { "model": "sonnet", "effort": "medium" },
"architect": { "model": "opus", "effort": "high" },
"code-reviewer": "sonnet",
"test-engineer": { "effort": "medium" }
}
}- String value (
"sonnet"): model only. Effort kept from frontmatter. - Object value (
{ model, effort }): set either or both. - Precedence: per-call parameter > this file > frontmatter default.
| Agents | Model | Effort | Why |
|---|---|---|---|
refactorer, latex-engineer |
haiku | low | Mechanical (Fowler catalog / template) |
engineer, code-reviewer, test-engineer, frontend-engineer, dba, devops-engineer, data-scientist, experiment-runner, mlops, ux-designer, professor |
sonnet | medium | Procedural with decision points |
orchestrator |
opus | medium | Routing, not deep reasoning |
architect, security-auditor, research-scientist, paper-writer, reviewer-academic |
opus | high | Deep structural / evidence reasoning |
| 97 genius agents (default) | sonnet | medium-or-high (from frontmatter) | Procedural patterns at medium; deep-reasoning patterns (Feynman, Dijkstra, Lamport, Gödel, Einstein) at high |
Estimated token savings vs. all-opus default: ~2.5–3× on typical sessions (measured against opus-only baseline; report your own measurements at Discussions if they differ).
Re-run setup.sh update after editing.
Five team agents — engineer, architect, code-reviewer, security-auditor, research-scientist — adjust reasoning depth dynamically:
- Low-stakes change → reason one level below baseline (skip exploratory alternatives)
- Medium-stakes → baseline
- High-stakes → reason one level above baseline (enumerate alternatives, full verification loop)
Stakes classification is objective (see rules/coding-standards.md §10). No config — built into the agent's Moves.
ZETETIC_PROFILE=standard # strict | standard | permissive
ZETETIC_CHECK_DATA_FORMATS=false # true to scan .json/.yaml/.toml/.sql/.csvProfiles:
strict— UNSOURCED, MAGIC_NUMBER, TODO_NO_REF all block commitsstandard(default) — UNSOURCED blocks; MAGIC_NUMBER / TODO_NO_REF warnpermissive— informational; never blocks (use during transition)
You cannot disable rules via .zetetic.conf. Directives like DISABLE_UNSOURCED=true, SKIP_RULE, EXCLUDE_RULE cause the checker to refuse to load the config. You can override the profile and file/path exclusions; you cannot silence the checks.
Optional Bash script sourced by zetetic-checker.sh if present. Can add project-specific checks that increment ERRORS or WARNINGS. Cannot disable built-in checks.