OmniRoute/@omniroute/opencode-plugin-v2

@omniroute/opencode-plugin-v2

OpenCode v2 plugin (define({ id, setup }), Promise API) that publishes the live OmniRoute catalog — models from /v1/models, combos from /api/combos (least-common-denominator join), enrichment (names + pricing), and usable-provider filtering — into the v2 catalog.transform, with key + env auth via integration.transform.

Companion to @omniroute/opencode-plugin (OpenCode v1, same repo). The two packages are independent: this one carries its own catalog-mapping logic and the v1 plugin is left untouched.

Install

npm install @omniroute/opencode-plugin-v2

opencode.json:

{
  "plugins": [
    {
      "package": "@omniroute/opencode-plugin-v2",
      "options": {
        "providerId": "omniroute",
        "baseURL": "http://localhost:20128"
      }
    }
  ]
}

Local file:// install

OpenCode resolves a local plugin directory by probing the subpaths server.* / index.* (then tui, rpc) at the package root — it never reads package.json main/exports. A folder exposing only dist/ is therefore silently skipped (no loading plugin, no error).

This package ships a root server.js re-exporting ./dist/index.js for exactly that probe, so pointing OpenCode at a local checkout works:

{
  "plugins": [
    {
      "package": "file:///path/to/OmniRoute/@omniroute/opencode-plugin-v2",
      "options": {
        "providerId": "omniroute",
        "baseURL": "http://localhost:20128"
      }
    }
  ]
}

Prerequisites when targeting a folder: run npm run build first (the root server.js re-exports ./dist/index.js), and keep the folder's root server.js — dist/ alone is not resolvable by the host.

Credentials

The plugin needs a gateway key to read the catalog, and looks for one in this order:

1. The credential you connected in OpenCode. The plugin registers an integration, so opencode auth (or the Connect action in the model picker) can store a key for it. Nothing is written to opencode.json — this is the recommended route.
2. apiKey in the plugin options, when you want a per-project override. Remember that this puts the key in a config file you may be committing.
3. OMNIROUTE_API_KEY in the environment.

If none of the three yields a key, the catalog is empty and the plugin says so once at startup rather than leaving you with a silent empty model list.

The management token is a different key

Combos, provider health and enrichment (display names, pricing, free-tier budgets) come from the gateway's /api/* endpoints, which most deployments gate behind a management token rather than the inference key. Set it explicitly:

"options": {
  "baseURL": "http://localhost:20128",
  "managementReadToken": "<management read token>"
}

The token can also come from the OMNIROUTE_MANAGEMENT_API_KEY environment variable (the option wins when both are set). Resolution order: managementReadToken option, then OMNIROUTE_MANAGEMENT_API_KEY, then the apiKey fallback.

Left unset, managementReadToken falls back to apiKey for backwards compatibility, and the plugin warns once at startup that the fallback is active. When a gateway rejects that fallback, the catalog still publishes — but with raw model ids instead of display names, no canonical alias dedupe, no pricing and no combos. The plugin warns once per endpoint when this happens, naming the endpoint and the consequence, so the degraded catalog is never a mystery. Environment variables must reach the opencode server process: when the host runs as a service, export the key where the server reads it, not only in the shell that runs opencode models.

Options

Key	Default	Notes
providerId	"omniroute"	Provider id and integration id; models publish under <providerId>/…
baseURL	required	OmniRoute gateway root (no /v1 suffix needed)
apiKey	connected credential, then OMNIROUTE_API_KEY	Chat key for /v1/* — see Credentials
managementReadToken	option, then OMNIROUTE_MANAGEMENT_API_KEY, then apiKey	Management key for /api/* (combos, providers, enrichment) — usually not the same key
displayName	"OmniRoute"	Provider display name
timeoutMs	10000	Per-endpoint fetch timeout
modelCacheTtlMs	300000	Catalog cache TTL; disk snapshot warms cold starts
timeouts	per-endpoint override	{ models, combos, enrichment } in ms; falls back to timeoutMs (autoCombos still parses but is ignored — the virtual-entries route is no longer requested)
enrichment	true	Fetch names + pricing (/api/pricing*, /api/free-tier/summary)
providerTag	true	Prefix a display name with the upstream provider it routes to
geminiSanitization	true	Strip $schema/additionalProperties from tool schemas sent to Gemini models ($ref tools are forwarded untouched)
usableOnly	false	Filter to healthy provisioned providers (/api/providers)
combosSource	"api"	Combo source. "api" reads /api/combos and publishes a combo only when every member passes the filters; "models" skips that request and publishes the combo rows /v1/models serves. Any preset, usableOnly, providersAllow or visible/hidden list keeps "api"
freeOnly / toolsOnly / visionOnly	false / false / false	Filter to free-tier / tool-calling / image-input models; combos with a filtered member are dropped, never partial. Presets filter the pool before the default view ranks it, and a combo publishes only when every resolved member passes (dropped when a member fails or none resolves)
showcasePerOwner	10	Default-view entries kept per provider
freshPerOwner	10	Default-view fresh entries kept per provider
freshWindowDays	90	Freshness window in days for the fresh branch
usageMemory	true	Restore statically dropped entries named by 30-day usage analytics (needs a management token; inert without one)
visibleModels / hiddenModels	[]	Exact-or-suffix allowlists, deny wins
providersAllow	[]	Provider allowlist by id prefix (empty = full catalog); unknown names warn once, empty result publishes empty + warns. Unknown catalog prefixes always pass, and an allowlist or vocabulary that names nothing known keeps the full catalog + warns
apiFormat.allowAnthropic	false	Route allowlisted ids to the Anthropic API block
apiFormat.anthropicModels	[]	Full model ids routed to Anthropic
apiFormat.anthropicPrefixes	v1 defaults	Deprecated, warns once — prefer anthropicModels
logLevel / startupDebug	warn / false	Logger verbosity

toolsOnly is off by default, so the full catalog is published. Set toolsOnly: true to hide models without tool calling from the host picker — useful for agent sessions that call tools on almost every turn. It does not cap the catalog size, and a very large catalog can still slow the host picker (see opencode#47615).

Usage memory is on by default. Without a management token it stays inert (a startup notice is logged) and nothing is restored — the picker stays filtered.

Tool calling on Gemini models

Gemini answers 400 INVALID_ARGUMENT — for the whole request, not just the offending tool — when a tool declaration carries $schema or additionalProperties. Anything that emits standard JSON Schema therefore breaks tool calling as soon as the chain routes to Gemini.

The plugin strips those keywords from tool schemas bound for a Gemini model of this provider, and leaves every other request untouched. A tool carrying a $ref is forwarded untouched instead of stripped: removing the reference would widen the schema to "accept anything". Set "geminiSanitization": false to turn it off.

Migrating from the v1 plugin

The v2 plugin publishes provider id X bare. The v1 plugin published opencode-X (native-adapter gate). Sessions pinned to opencode-X/... must re-select the model under X/....

License

MIT