codeitall — Rust API behavioral oracle

Version-pinned behavioral oracle for Rust package APIs, built for AI coding agents. Five deterministic tools backed by a Postgres substrate of 66,549,370 pattern_records and 107 successful probe runs across 7 task families (datetime, duration, base64, JSON strict-vs-permissive, JWT verify, AEAD encrypt/decrypt, constant-time compare).

Hosted, free, no signup: https://api.codeitall.dev

This repository is codeitall's public surface — the API reference, OpenAPI spec, the "For AI Agents" page, and the per-family pages. The ingest/indexer and server source are not redistributed; you consume codeitall by querying the hosted API, not by running it yourself.

What it answers

The differentiator from a docs RAG: the substrate captures runtime behavior (what a function actually does on a canonical input set, observed empirically), not just documentation (what it claims to do). It targets the documentation-vs-runtime gap — yank status, RustSec advisories, deprecation renames, and cross-implementation behavior tables.

Tools

Tool	Use it for
`answer_api_question`	Open-ended Rust API questions; routes deterministically (rule-based, no LLM) and falls back to bge-m3 cosine retrieval.
`signature_search`	You know the signature shape (`fn(&str) -> Result<Vec<u8>, _>`); get ranked candidate functions + crate, downloads, yank status.
`behavior_lookup`	A specific `(crate, fn_name)`; get the probe-observation table verbatim — each row `(input, outcome=ok\|err\|panic, value_or_error)`.
`compare_implementations`	A task category (e.g. "base64 crates"); get the side-by-side cross-implementation behavior table.
`find_modern_equivalent`	A suspected-deprecated call; get the current canonical alternative + yank/advisory rows.

Determinism contract

No generative LLM sits between a request and its response. Every response is SQL + deterministic templating; the only model in the hot path is bge-m3, a deterministic text→vector embedder used for the cosine fallback (same input → same vector → same hits). This is a regression-pinned invariant.

Quickstart

bash

# Mint a free anonymous key (no email, no signup):
curl -sX POST https://api.codeitall.dev/v1/keys \
  -H 'content-type: application/json' -d '{}'

# Ask a question (HTTP REST):
curl -sX POST https://api.codeitall.dev/v1/answer_api_question \
  -H 'content-type: application/json' \
  -d '{"question":"how do I parse a duration string in rust"}'

GET aliases exist for every tool (browse-only clients); signature_search also accepts a shape_b64 (base64url) parameter so URL re-encoding can't mangle type-signature characters.

Access modes

MCP — SSE/HTTP at https://api.codeitall.dev/mcp (Bearer key), or stdio for self-hosters.
HTTP REST — POST/GET /v1/<tool>.
OpenAPI 3.1 — openapi.json (live: https://api.codeitall.dev/openapi.json).

Docs in this repo

for-ai-agents.html — positioning, tool list, key snippet (live: https://api.codeitall.dev/for-ai-agents.html).
api-reference.html / api-reference.md — generated tool reference + scope/coverage.
agents/ — one page per probe family.

Cost & limits

Free. Per-IP key issuance cap (10/day) and per-key quota (10,000/day). No paid tier, no email capture, no signup.