Add Jevonian to agent tools and MCP servers

[xinyao27]

What this adds

Jevonian — xinyao27/jevonian, one bullet in Agent tools and MCP servers, alphabetically near the other jev* entries. One file changed, one line added.

Exactly how Jev is used

The project serves an OpenAI / Anthropic / Responses-compatible endpoint locally at http://127.0.0.1:8787/v1. For the virtual model jevonian/auto, one Jev call answers both which candidate model serves the turn and how deeply that model should think, against a state snapshot built from the session: last user message, recent messages and tool results, consecutive error count, context headroom, quota health, candidate capabilities, and cache-switch penalties.

https://api.typesafe.ai/v1/systemone with typesafe/jev-1.13 is the primary channel; OpenRouter, Vercel AI Gateway, and OpenCode Zen are the others, all configured in src/brain.ts. Channel setup, the state payload, and the confidence behaviour are documented in docs/brain.md.

Where the boundaries are

• Code first, always. Candidates are filtered by wire protocol, context window, and thinking-level floor, and providers whose quota window is spent are dropped before Jev is asked (unknown quota is neutral, never a blocker). All thresholds live in src/routing.ts, not in prompts.
• Jev is skipped when it should be. A pinned real model ID, an explicit jevonian/plan / /execute / /utility / /chat, or routing.mode: "off" bypasses the brain entirely, so pass-through never pays for a decision.
• Low confidence is surfaced. minConfidence marks a low-confidence route in the ledger and the response headers instead of accepting it silently, and a secondary Jev channel is only tried when the primary channel fails, not when confidence is low.
• The decision is auditable. Each turn is recorded with the model and provider that actually served it, the reason, real token usage, cache reads, and an estimated cost — and returned as x-jevonian-model, x-jevonian-provider, x-jevonian-phase, x-jevonian-reason, x-jevonian-skipped.

Limits are stated in the entry's own repo README: costs are estimated from a price table rather than a vendor invoice, cache figures are estimates (prefixMatch stays unknown because the on-wire prefix is not yet compared), a pinned model is not necessarily a pinned provider, and a route is session affinity rather than a task-level lock.

Evidence

Primary source is the repository itself — vp test covers the routing decisions and pnpm smoke runs end-to-end assertions against a mock upstream with no API keys. The entry claims no speed, cost, or accuracy figure. The two links inside the bullet point directly at the files that call Jev and that own the filtering.

Disclosure

I maintain Jevonian. v0, TypeScript, AGPL-3.0-only, published to npm as jevonian. Not affiliated with TypeSafe AI. One project, one PR, one bullet; no existing entries were reformatted.