[danipl]

What

Closes #9. routePayload now returns the silent passthrough before askJev when there is no candidate function tool, or when the lone candidate is the only entry in payload.tools — every Jev outcome (trim / bypass) hands the big model the identical request, so the round-trip is pure added latency (up to timeoutMs, default 2 s).

Correction to the issue's proposed fix

Issue #9 proposed guarding on names.length < 2. Verified against the trim implementation (src/index.ts): the filter keeps only entries matching the chosen name, so a Responses-API payload like [{type:web_search}, {type:function, function:{name:shell}}] has exactly one candidate but a successful trim drops the built-in — a real, tested behavior (test 7). The literal guard would have regressed it. Shipped guard instead fires only when the trim is provably a no-op:

if (names.length === 0 || (names.length === 1 && (payload.tools as unknown[]).length === 1)) {
  return false;
}```

## Acceptance criteria

- [x] Black-box test 20: single function tool → byte-identical passthrough and `calls.length === 0` (no fetch pattern)
- [x] DEVELOPMENT.md §5 silent-paths list updated
- [x] `typecheck` / `test` (22/22) / `build` green
- [x] Invariants: no throw, rebuilt passthrough, `tool_choice` untouched, latch untouched

## Skipped (known ceilings)

- Duplicate-name payloads (`tools: [read, read]`) still take the round-trip; trim there is also a no-op but OpenCode never emits duplicate tool names — tighten only if real.