feat(opencode): opencode v2 plugin publishing the OmniRoute catalog

[maxmad64bis]

⚠️ base-red inherited: #12732

Summary

OpenCode v2 has no route to an OmniRoute gateway. The existing @omniroute/opencode-plugin targets v1, whose loader expects plugin factories; v2 expects a default define({id, setup}) with catalog and integration domains. This adds a second, self-contained package. The v1 one is untouched — no move, no migration, no breaking version — and they share no code and no release.

It publishes models, combos and auto-combos into the host catalog, refreshes them behind a 300s TTL, and falls back to a disk snapshot when the gateway is down. Models and combos go out as soon as they're known; the optional sources fold in when they land, so one slow endpoint can't hold the picker hostage.

Four things it does that the v1 plugin can't. The key comes from OpenCode's own credential store when the integration is connected, so no secret has to sit in opencode.json. Failures are named instead of swallowed: three empty catch blocks used to turn a refused management token into a picker full of raw ids with no explanation. Display names now carry what the gateway knows — the upstream provider, a free marker, the budget — so two connections selling the same model stay distinguishable. And tool schemas bound for Gemini are stripped of the keywords it rejects with 400 INVALID_ARGUMENT, at the model level rather than by rewriting HTTP bodies.

Four v1 behaviours are deliberately dropped, because v2 owns them or no longer needs them: the plugin-side debug log, the compression suffix on combo names, the MCP auto-emit, and the omni-sync command with its timer (v2 commands are prompt templates, not callbacks — the TTL and a fingerprint drive catalog.reload()).

Related Issues

• Closes #
• Related to #

Validation

• Change type: other (new npm package + docs; no change to src/, open-sse/, electron/ or bin/)
• Focused tests and category gates from the golden path — 189 tests in the new package (npm test inside @omniroute/opencode-plugin-v2), plus tsc --noEmit, Prettier and the production build
• npm run lint
• Reconciled with the current active release base; focused checks rerun afterward
• Production-code changes include a new or updated automated test in this PR

npm run lint doesn't start on a clean checkout of the base either: string.prototype.repeat@1.0.0 needs es-abstract/2019/RequireObjectCoercible, which es-abstract@1.24.1 (pinned by the lockfile) no longer ships. Installing es-abstract@1.23.9 out of the lockfile makes it run, and it then exits 0 with only the pre-existing violations frozen in config/quality/eslint-suppressions.json.

Three red checks on this PR come from the base, not from this branch — listing them so nobody chases them here:

Check	Failure	Why it isn't this branch
Docs Gates (fast-path)	169 migrations in README.md, AGENTS.md, llm.txt vs 170 migration files	This branch adds no migration and edits none of those files
Merge integrity (changelog)	changelog.d/fixes/reset-aware-model-family.md doesn't start with -	Added by 2a6eff0ae (#12637); not in this diff
API Route Typecheck	open-sse/executors/glm.ts TS2554	This branch touches no file under open-sse/

Tests Added Or Updated

All under @omniroute/opencode-plugin-v2/tests/ (202 tests, new package, nothing pre-existing changed):

• credentials.test.ts, enrichment-report.test.ts, shared-enrichment-source-errors.test.ts — credential resolution order, OAuth refusal, old hosts, and every degraded-source path, on the plugin and library entry points alike
• enrichment-render.test.ts — provider tag, free marker and budget suffix reaching the published name
• parity.test.ts + fixtures/v1-parity.json — the v1 plugin's recorded output for the shared fixture, so the parity claim is reviewable and runs without building the v1 package
• staged-refresh.test.ts — a source that never answers must not hold the publish; a late one must reach the catalog
• gemini-language.test.ts, shared-gemini.test.ts — schema stripping, and that it applies to this provider's Gemini models only
• catalog.test.ts, host-contract.test.ts, api-package.test.ts, parity.test.ts, nested-combos.test.ts, auto-combos.test.ts, usable-only.test.ts, cache-ttl-snapshot.test.ts, refresh-failopen.test.ts, publish-guard.test.ts, snapshot-stale-entries.test.ts, management-token.test.ts, timeouts-logger.test.ts, options.test.ts, compat.test.ts, enrichment.test.ts, smoke-types.test.ts, and the shared-*.test.ts files covering the mapping layer

Coverage Notes

No file under src/, open-sse/, electron/ or bin/ changes, so the repo coverage gate is unaffected. The new package ships its own suite and runs it in CI on Node 22 and 24 before publishing.

Reviewer Notes

The catalog contract is still moving, so the plugin reads the shape the host seeds into the draft rather than keying off a version list, and publishes the api block, the pre-api fields, or both. host-contract.test.ts pins it; the SDK stays pinned, so a contract move surfaces as a failing test on the bump. Three capability probes (integration.connection, aisdk, catalog.reload) keep the plugin loadable on older hosts.

Migration note: v1 published provider id opencode-<id>, this one publishes <id> bare, so a pinned session has to re-select its model. The plugin id itself is fixed (omniroute-v2) — the host reads it before any option exists.

Three things I left alone on purpose. The publish job copies the v1 one, bump included (npm version patch never lands back in the repo) — worth fixing, but that's a repo-wide convention, not mine to change here. Two v1 mapper behaviours stay as they are because fixtures/v1-parity.json freezes exactly that contract: a combo whose members are combos routes openai-compatible even when every leaf is Anthropic, and two aliases sharing a bare model id can overwrite each other's pricing. Both predate this package.

The Gemini sanitiser is unit-tested only; the 400 itself is inherited from v1 rather than re-measured against a live route.