[velizarofff]

Since 2026-10-14, standalone requests to OpenCode free-tier models (the ones isGatedFreeTierRequest gates) fail with 403 FreeTierError (#16328), while the official OpenCode CLI keeps working.

Two things changed upstream at once. Both were measured by capturing the official CLI 1.19.0 request and replaying it with exactly one element changed at a time (2026-10-15, anonymous lane, 5–10 s between requests).

1. The free tier moved off /zen/v1

The client now takes its base URL from the models.dev catalog (https://opencode.ai/inference/v1). The captured request on /inference/v1/chat/completions answers 200; byte-for-byte the same request on /zen/v1/chat/completions answers 403 FreeTierError.

2. The tools check now inspects parameter schemas

Replay (single change vs the captured baseline)	Result
captured CLI request unchanged	200
same request on /zen/v1	403
no tools	403
only read with the client's real schema	200
only grep; only bash,edit,glob; any subset without read	403
read with any non-empty properties (filePath, file_path, x), with or without required	200
empty parameters on every tool	403
read + bash with properties: {}	403
read + bash{command}	200
read + webfetch{} / read + noop_tool{}	200
only foo with a non-empty schema	403
empty descriptions on all tools	200
one-line system prompt	200
router-like headers (UA 1.18.31, client: desktop, no session id)	200

So: a property-bearing read is required, an empty bash breaks the request, and empty schemas on names outside the quartet still pass. Descriptions, system prompt and headers do not matter.

What OmniRoute does today (release/v3.8.52)

buildUrl resolves this.config.baseUrl for every request format — the zen registry base — and the request contract injects placeholder tools (applyFreeTierRequestContract, mergeClientToolsWithObserved) with parameters: { type: "object", properties: {} }. The surface machinery added in 3.8.52 (surfaceFromBaseUrl) already tells the gated zen surface apart, but nothing uses it to change the base URL, so a standalone request violates both conditions of the new gate.

Fix

• add FREE_TIER_INFERENCE_BASE_URL (https://opencode.ai/inference/v1) to opencodeFreeTierContract.ts, next to the surface constants;
• buildUrl sends a gated free-tier request (zen surface, non-premium model) to that base for every request format. Paid models and opencode-go keep the registry base, which still serves them;
• add placeholderToolParameters(name): injected quartet members get minimal schemas mirroring the client's argument names — bash{command}, glob{pattern}, grep{pattern}, read{filePath}, each with required — while names outside the quartet keep the empty schema. Used by both applyFreeTierRequestContract and mergeClientToolsWithObserved.

With an operator configuration like OPENCODE_FREE_TIER_PLACEHOLDER_TOOLS=bash,glob,grep,read, a standalone request carrying no tools of its own now satisfies the tools half of the 2026-10-14 gate. The session half is still open: measured 2026-10-16, the same request with a session id in the pre-#16727 shape (12 arbitrary hex digits) is refused 403 on /inference/v1 as well, so a standalone client that relies on OmniRoute's synthesized x-opencode-session still needs the encoding half of #16727. A client that supplies its own valid session id (the official CLI does) is unaffected. I have that half ready on a second branch of the same fork and will keep it out of this PR unless maintainers want both here.

The same two changes (endpoint + non-empty placeholder schemas) are what unblocked a live self-hosted instance on 2026-10-15, in a build that also carries #15322; this PR is the minimal subset against the current default branch.

Out of scope

• Client-supplied tools still pass through unchanged. The name check is still active on /inference/v1: measured 2026-10-16, a request whose only tools are capitalized Read/Grep/Glob with non-empty schemas is refused 403. Covering such clients is the concealment/rename machinery of fix(opencode): declare the free-tier tool fingerprint the upstream requires #15322, not touched here.
• The session-id encoding half of fix(opencode): make the free tier work standalone (tool fingerprint + session value) #16727 is not touched, and that check is still active on /inference/v1 too.
• fix(opencode): follow the published endpoint and the client's own headers #16608 proposes reading the base from the catalog api field, and fix(opencode): route each key-based model to the endpoint the published catalog declares #16798 routes each key-based model to the endpoint the published catalog declares. The constant keeps this patch minimal and self-contained; happy to rebase onto the catalog field if maintainers prefer that direction.

Tests

• executor: free models now assert /inference/v1/...; paid (gpt-5-nano), unknown models and the Go tier keep their bases;
• request contract: injected quartet members assert the minimal schemas; new unit tests for placeholderToolParameters and for an appended quartet member in mergeClientToolsWithObserved;
• full tests/unit/opencode-*.test.ts suite (92 files, node 26) against release/v3.8.52 + this patch: 855 tests, 0 failures. The four baseline failures of v3.8.51 (accounts-per-request, geo-block-rotation, responses-first-byte-stall, transient-retry-delay) no longer reproduce on clean v3.8.52.

Addresses the endpoint and placeholder-schema halves of #16328.