inquisitor AGENTS.md

AGENTS.md — Inquisitor (always-on)

Disciplined problem solving for AI agents: probe, prune, and spend your budget where the problem actually is. You cannot explore every branch — match the depth of the method to the depth of the problem. Overcomplication is a failure mode, exactly like a wrong answer.

This file is the always-on core. The full 7-phase method lives in the inquisitor skill (skills/inquisitor/SKILL.md), mandatory for Deep problems. The inquisitor-mcp server adds web search, project analysis, code tracing, and persistent phase tracking — tools used when local evidence is insufficient, never as ritual.

Don't classify — probe (depth is an output)

An up-front difficulty guess is an LLM's least reliable signal — biased to under-rate the hard problems. So never declare depth; let it emerge. Every problem runs one loop: frame (done = ? / must not break) → delegate? → probe (cheapest action that confirms/kills your best hypothesis) → gate → act (min change) → verify (name the runtime signal). The probe's result sets the depth, not your prediction.

Depth	Probe found	Do
Shallow	Obvious, local, one-liner	Fix → verify → done. No ceremony.
Standard	Clear single-component cause	Frame → minimal evidence → fix → verify. In your head.
Deep	Unknown root cause, multi-component, or 2 failed fixes	Load inquisitor skill: full 7-phase method + session tracking

Delegate before you dig (router first)

If a purpose-built skill squarely owns the task (/tdd, /code-review, /security-review, …), invoke it via /skill prose and let it drive — the Newton method is the fallback for problems no specialist owns, not the first resort. Also check docs/companion-skills.md if present, unprompted, every pass — inquisitor's own curated pairing list, installed or not. Re-deriving a skill's discipline inline is slop, like re-writing a helper two files over. Finding or installing a NEW skill is a trust-boundary action: surface it to the user, never automatic.