inquisitor

inquisitor

Hypothesis-driven problem solving for AI agents Probe · Falsify · Escalate — never overcomplicate, never blind-retry

Overview

inquisitor makes AI agents solve problems the way a chess engine plays chess: it cannot explore every branch, so it estimates complexity first, prunes paths that add no information, and spends its search budget only where the problem actually is.

It ships as two coordinated layers:

• MCP server (inquisitor-mcp) — the engine. Web search, project analysis, code tracing, project scaffolding, and a persistent investigation state machine. Works with any MCP-compatible agent: OpenCode, Claude Code, Claude Desktop, Cursor.
• Agent skill (skills/inquisitor/SKILL.md) — the behavioral layer. Injects the probe-and-gate loop, the pruning rules, and the full methodology into the agent's reasoning.

The method is not invented here — it is assembled from primary sources: Newton's Analysis→Synthesis skeleton, hardened with the scientific method's own defenses against wrong assumptions (falsification, strong inference, competing-hypotheses analysis, anti-fixation reframes, pre-mortems) and constrained by engineering discipline (NASA/JPL's Power of Ten, surgical-change guidelines, the minimalism ladder). Every source is cited in Foundations.

Web search, codebase scans, and code tracing are tools invoked when local evidence is insufficient — never mandatory rituals.

---

How it decides

Depth is an output, not a label. An LLM's up-front difficulty guess is its least reliable signal — poorly calibrated, and biased to under-rate exactly the hard problems that matter. So inquisitor never classifies a problem before understanding it: it runs a cheap probe (the single cheapest action that could confirm or kill the best current hypothesis), and the probe's result — never the prediction — sets the depth:

flowchart TD
    P([Problem]) --> F["FRAME<br/>done = ? · must not break = ?"]
    F --> D{"delegate?<br/>a specialist skill<br/>owns this"}
    D -->|yes| SK["hand off — /tdd,<br/>/code-review, ..."]
    D -->|no| PR["PROBE<br/>cheapest action that could confirm<br/>or kill the best hypothesis"]
    PR -->|"obvious, local"| SH["<b>Shallow</b><br/>fix → verify<br/><i>no ceremony</i>"]
    PR -->|"clear single-<br/>component cause"| ST["<b>Standard</b><br/>frame → minimal evidence<br/>→ fix → verify"]
    PR -->|"root cause unknown,<br/>multi-component"| DE["<b>Deep</b> — Newton 7-phase:<br/>DEFINE → AXIOMS → ANALYSIS →<br/>EXPERIMENT → SYNTHESIS →<br/>VALIDATE → QUERY<br/><i>+ session tracking as memory</i>"]

    SH -.->|"gate trigger /<br/>failed fix / low confidence"| ST
    ST -.->|"gate only raises depth —<br/>never lowers"| DE

Escalation is enforced, not just allowed. The probe is only a starting point: objective gate triggers (touching infra/deploy/routing/config, auth/security, data migrations, multi-file fixes, prod-only symptoms) force a minimum depth regardless of how "clear" the problem feels, and a 3-question confidence check (read the runtime path? can name the runtime signal? verified the platform assumption?) bumps the depth up per unanswered question. Downgrades need cited evidence, never a feeling. Inflated ceremony is not allowed either — a 7-phase investigation of a typo is as wrong as a blind guess at a race condition.

Retry is never blind. "Loop until it passes" agents (the Ralph-loop pattern) have persistence but no memory: on failure they revert, flush, and re-roll — the same wrong idea, retried with fresh confidence. Inquisitor keeps a failure ledger instead: every dead hypothesis is recorded with the evidence that killed it, every retry must name what is different and why that changes the outcome, and two dead hypotheses from the same family force a reframe — re-audit an assumption, invert the question, widen the system boundary — never a third attempt at the same idea. Persistence with memory, creativity on evidence.

---

Foundations

Each rule in the method traces to a primary source. The left column is the citation; the right column is the exact mechanism inquisitor takes from it.

The scientific method

Source	Mechanism adopted
Isaac Newton, Opticks, Query 31 (1704)	The investigation skeleton: Analysis before Synthesis — define, decompose, experiment, only then reconstruct — and closing with open Queries instead of false certainty. Hypotheses non fingo.
T.C. Chamberlin, The Method of Multiple Working Hypotheses, Science (1890)	Hold at least two rival explanations at all times; a single hypothesis turns every subsequent observation into confirmation.
Abraham Luchins, Mechanization in Problem Solving (1942) — the Einstellung effect	The failure ledger's trigger: repeating a familiar approach after it stopped working is a measurable fixation, and the cure is a forced reframe, not another attempt.
Karl Popper, The Logic of Scientific Discovery (1959)	Falsify first: for the leading hypothesis, name the observation that would disprove it and hunt for that observation before anything else.
John Platt, Strong Inference, Science (1964)	Design the experiment that excludes a hypothesis, not the one that corroborates the favorite — discriminating tests over confirming tests.
Richards J. Heuer, Psychology of Intelligence Analysis, CIA (1999)	Analysis of Competing Hypotheses: rank hypotheses by the evidence inconsistent with each — confirming evidence is cheap and usually fits several at once.
Gary Klein, Performing a Project Premortem, Harvard Business Review (2007)	Before shipping: assume the fix is live and the problem still happens — name the likeliest reason and probe it now.

Engineering discipline

Source	Mechanism adopted
Gerard J. Holzmann (NASA/JPL), The Power of Ten: Rules for Developing Safety-Critical Code (2006)	The P10 template: a rule set small enough to remember and strict enough to check mechanically.
Andrej Karpathy's LLM coding guidelines (2025)	Think before coding · simplicity first · surgical changes · goal-driven execution.
Ponytail decision ladder	YAGNI → reuse → stdlib → native platform → installed dependency → one line → minimum code that works.

---

Installation

Requires uv and Python 3.12+.

Claude Code — plugin install (recommended)

inquisitor ships as a Claude Code plugin that installs both the skill and the MCP server — no cloning, no editing absolute paths, no manual symlink.

From within Claude Code, first add the marketplace:

/plugin marketplace add iamalisson/inquisitor

Then install the plugin:

/plugin install inquisitor@inquisitor

That's it. The plugin bundles the inquisitor-mcp server (registered automatically via ${CLAUDE_PLUGIN_ROOT}) and the inquisitor skill. uv syncs the server's dependencies on first launch. Update later with /plugin marketplace update inquisitor.

Prefer to point at a local checkout instead of GitHub? /plugin marketplace add /path/to/inquisitor works too.

For OpenCode and Claude Desktop (which don't use Claude Code plugins), or for a manual Claude Code setup, use the steps below.

Step 1 — Get the server

Option A — no clone (recommended). Once published to PyPI, uvx fetches and runs it on demand — no clone, no absolute paths:

uvx inquisitor-mcp   # prints a ready message and waits for a client — Ctrl+C to exit

You'll reference uvx inquisitor-mcp directly in the config below.

Option B — from a checkout (for local development, or before the PyPI release):

git clone https://github.com/iamalisson/inquisitor.git ~/tools/inquisitor
cd ~/tools/inquisitor
uv sync

The clone path is up to you — just use the same absolute path in the config below. ~ does not expand inside JSON config files, so write the full path (e.g. /home/you/tools/inquisitor).

You do not run the server manually. It's a stdio MCP server: your agent spawns and manages it automatically. (If you run it by hand it prints a ready message on stderr and waits silently — that's normal.)

Step 2 — Register the MCP server with your agent

OpenCode — add to ~/.config/opencode/opencode.json (global) or ./opencode.json (per-project):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "inquisitor": {
      "type": "local",
      "command": ["uvx", "inquisitor-mcp"],
      "enabled": true
    }
  }
}

Claude Code — add to .mcp.json in your project, or ~/.claude.json for all projects:

{
  "mcpServers": {
    "inquisitor": {
      "command": "uvx",
      "args": ["inquisitor-mcp"]
    }
  }
}

Claude Desktop — same mcpServers block in claude_desktop_config.json (Settings → Developer → Edit Config).

Using a checkout instead of uvx? Swap the command for uv with args ["run", "--directory", "/home/you/tools/inquisitor", "inquisitor-mcp"].

Step 3 — Install the skill (the behavioral layer)

Symlink it so it stays up to date with the repo:

# OpenCode
mkdir -p ~/.config/opencode/skills
ln -s /home/you/tools/inquisitor/skills/inquisitor ~/.config/opencode/skills/inquisitor

# Claude Code
mkdir -p ~/.claude/skills
ln -s /home/you/tools/inquisitor/skills/inquisitor ~/.claude/skills/inquisitor

(Copying the folder works too — you'll just need to re-copy after updates.)

Or, no clone — install just the skill across agents with the skills CLI:

npx skills add iamalisson/inquisitor

(This installs the skill only. For the inquisitor_* tools you still need the MCP server from Steps 1-2 — or use the Claude Code plugin above, which bundles both.)

Step 4 — Restart your agent

Config is loaded at startup. Quit and reopen OpenCode / Claude Code, then verify: the inquisitor_* tools appear in the tool list, and the inquisitor skill is available.

---

Tools

Tool	Purpose	When
inquisitor_search	Multi-backend web search (DuckDuckGo free/keyless, Brave, SearXNG) with content extraction (HTML + PDF)	Local evidence insufficient: unknown errors, unfamiliar libraries, current best practices
inquisitor_analyze	Project overview: languages, frameworks, tests, deps, git history	Entering an unfamiliar codebase
inquisitor_trace	Symbol tracing: definition, callers, callees with file:line refs	Bug spans multiple functions/files