zetetic-team-subagents

Claude Code agents whose commits are blocked when constants lack source citations. 97 genius reasoning agents (plus 19 team-role agents = 116 total) each citing their primary paper and documenting their refusal conditions, paired with a pre-commit hook that blocks any floating-point constant with 3+ significant digits that lacks a source: annotation. Not a prompt library. A methodology with commit-time enforcement.

---

The system enforcing its own standard

$ git commit -m "tune retry backoff"

UNSOURCED   (error)    retry.py:1: # It always works
MAGIC_NUMBER (error)    retry.py:2: DELAY = 2.741592

Profile: strict  (staged mode)
Errors:   2  (blocking)
Warnings: 0  (informational — promoted to errors when profile=strict)
FAILED: 2 blocking violation(s).

BLOCKED: Zetetic violations in staged files.

Composite output: lines 1–2 are verbatim from tools/zetetic-checker.sh --staged; the closing BLOCKED: line is the wrapper from hooks/pre-commit-zetetic.sh that returns exit 2 to git. Reproduce on your machine: echo "DELAY = 2.741592" > /tmp/x.py && cd /tmp && git init -q && git add x.py && ZETETIC_PROFILE=strict bash <repo>/tools/zetetic-checker.sh --staged.

The commit re-runs once each flagged line carries a # source: comment, a benchmark reference, or a measured-on note.

---

Why this exists

Every AI agent system ships a role prompt. "You are a senior engineer." The agent sounds confident. It invents numbers, cites papers it hasn't read, and ships code with conviction inversely proportional to its correctness.

Zetetic Agents are different in one specific way: they can say "I don't know."

97 reasoning patterns drawn from primary sources — Dijkstra's correctness discipline, Curie's residual-with-a-carrier method, Hamilton's fault-tolerance protocol, Cochrane's evidence synthesis — routed automatically to your problem by shape, not by field. Every output is sourced. Every commit is checked. The standard is not a prompt. It is a gate.

---

What you type → what happens

/paper-vs-code-audit arxiv:2401.12345 ./src/
→ Extracts every claim → finds corresponding code → flags mismatches → traceability matrix

/autoresearch-loop "optimize beam search for abstention"
→ Hypothesis → implement → commit → benchmark → keep/revert → iterate until diminishing returns

/deep-research "transformer attention alternatives 2024-2026"
→ Plans search → parallel researchers → synthesizes → verifies citations → writes cited brief + provenance sidecar

/incident-investigation
→ Forensic timeline → three-timescale decomposition → common vs special cause → structural root cause → remediation

/genius route "p99 latency exceeds the sum of profiled components"
→ Routes to the reasoning procedure that fits the problem shape

These aren't prompts dressed up as commands. Each is a multi-step pipeline that names the procedure used, surfaces blind spots in its output, and refuses to ship if a step fails. See docs/EXAMPLES.md for full session transcripts.

---

Install

claude plugin marketplace add cdeust/zetetic-team-subagents
claude plugin install zetetic-team-subagents

That's the whole install. The plugin's installer copies agents, skills, hooks, and tools into ~/.claude/. Manual install + advanced config: docs/INSTALL.md.

---

What you actually get