Hypothesis-driven problem solving for AI agents: probe, falsify, escalate — never blind-retry. MCP server + agent skill built on Newton, Popper, and NASA's P10.
Inquisitor has 7 tools with explicit schemas and descriptions visible in server.py. Naming follows verb_noun convention (search, analyze, trace, scaffold, phase_get, phase_set, verify). However, descriptions are brief (avg ~80 chars) and lack actionable context for LLM selection. Parameter descriptions exist but are generic; many lack constraints (enums, ranges, formats). Output schemas are undocumented, LLMs cannot predict response structure. Error handling is absent from visible code. No tool annotations (readOnlyHint/destructiveHint). Two tools (scaffold, phase_set) are destructive but lack confirmation patterns or dry-run support.
Scan a project directory and return a structured overview.
Get the current phase of the Newton investigation.
Advance to a new phase in the Newton investigation process.
Set up a new project with best practices and minimal boilerplate.
Search the web using DuckDuckGo (free), Brave, or SearXNG.
Trace a function, class, or method through the codebase.
Validate investigation findings against the original problem definitions.
No output schemas documented. LLMs cannot predict response structure or plan downstream tool calls. All tools return untyped strings.
Destructive tools (scaffold, phase_set) lack confirmation or dry-run patterns. Agents can accidentally overwrite projects or advance investigation phases without validation.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Agents cannot distinguish safe reads from destructive writes without parsing descriptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | 2026-07-28+ | v2 |
Parameter descriptions lack constraints. 'engine' accepts 'ddg', 'brave', 'searxng' but no enum declared. 'direction' accepts 'callers', 'callees', 'both' but no enum. LLMs will hallucinate invalid values.
No error handling guidance visible. Tools return unstructured strings with no recovery hints. If a search fails or a project path is invalid, LLMs have no actionable next step.