Add Jevonian to Community Builders / 05 Harness

[xinyao27]

Section

Community Builders → 05 Harness, added after AbdelStark/bicameral and before the "Canonical examples" line. One entry, one file.

Entry format compliance

The guide's format for Community Builders is:

- [owner/repo](link): <what it does, one line>. <why it's unique, one line>. [self-reported|independent]

The added line follows it exactly, and is labelled [self-reported] because it is my own project and I have published no independent measurement of it.

What it does

The project serves a local endpoint at http://127.0.0.1:8787/v1 speaking the OpenAI Chat Completions, Anthropic Messages, and OpenAI Responses shapes, so an existing coding agent can point at it. For the virtual model jevonian/auto, one Jev call answers both the model route and the thinking level, from a session state snapshot of recent messages and tool results, consecutive error count, context headroom, quota health, candidate capabilities, and cache-switch penalties.

Why it is unique in this section

Two things distinguish it from the other harness entries here:

1. The decision is economic, not just capability-based. Because the state explicitly includes cache-affinity evidence, the route Jev picks is the cheapest capable model given what is already in context, and providers whose quota window is spent are removed from the candidate list before Jev is consulted at all. Unknown quota is treated as neutral rather than a blocker.
2. The decision stays auditable. Every turn is recorded with the model and provider that actually served it, the reason for the choice, real token usage, cache reads, and an estimated cost, and the same data is returned as response headers. The cost figures are estimates from a price table rather than invoice reconciliation, and cache figures are estimates, which the repo documents rather than glosses.

Deterministic code owns the boundaries: filtering by wire protocol, context window, and thinking-level floor, plus every threshold. A pinned real model ID, an explicit jevonian/plan / /execute / /utility / /chat, or routing.mode: "off" skips Jev entirely, so pass-through never pays for a decision. minConfidence marks a low-confidence route in the ledger instead of accepting it silently.

Jev calls live in src/brain.ts (https://api.typesafe.ai/v1/systemone, typesafe/jev-1.13), filtering in src/routing.ts, and the state contract is documented in docs/brain.md.

Eligibility conditions

• Actually uses Jev. Yes, four interchangeable channels.
• Unique. No other entry in this section routes per turn with Jev optimising cost and cache affinity rather than only tier; the closest neighbours (jev-router, tiershift, JevRouter) appear in other sections of this list and are distinct projects.
• What it does + why unique. Both covered in the two-sentence line above.
• Public source. Yes, AGPL-3.0-only, npm package jevonian.

Affiliation

I maintain the project, disclosed as required. This is a self-submission under Community Builders, not a nomination for the maintainer-managed Top Use Cases section.

[anandi1989]

merged