fix(providers): correct opencode-zen muse-spark context length and Responses auth header (#12681, #12633)

[diegosouzapw]

Summary

Two related, small OpenCode Zen registry/executor bugs fixed together (both files: open-sse/config/providers/registry/opencode/zen/index.ts / open-sse/config/providers/registry/opencode/index.ts for #12681, open-sse/executors/opencode.ts for #12633):

Closes #12681
Closes #12633

#12681 — wrong context_length for Muse Spark 1.2 (200000 vs real ~1M)

Root cause: muse-spark-1.2 and muse-spark-1.2-contributor-free were declared in the opencode and opencode-zen registry entries without their own contextLength — unlike sibling models in the same list. With no per-model window, contextManager.resolveTokenLimit() fell back to the provider-wide defaultContextLength: 200000.

Fix: declared the real window (contextLength: 1048576, maxOutputTokens: 131072) on both Muse Spark 1.2 entries in both files. This value is not invented — it mirrors the same model family already declared correctly on opencode-go's muse-spark-1.2-contributor* entries (open-sse/config/providers/registry/opencode/go/index.ts), which are exercised by existing tests (tests/unit/opencode-go-catalog-alignment.test.ts).

Regression test: tests/unit/issue-12681-opencode-muse-spark-context.test.ts

• RED (before fix): muse-spark-1.2/-contributor-free had contextLength === undefined in both registries; getTokenLimit("opencode", "muse-spark-1.2-contributor-free") resolved to 200000.
• GREEN (after fix): both entries declare contextLength: 1048576; getTokenLimit() resolves to 1048576 on both opencode and opencode-zen.

Note: the plan-file also flagged a second, architecturally larger bug — vendor-prefixed passthrough ids reached through nous-research/kilocode (e.g. poolside/laguna-s-2.1) not cross-referencing the vendor's own registry entry in contextManager.resolveTokenLimit(). That is a separate, larger design change (a new cross-reference resolution step) that is out of scope for this small, TDD-scoped fix and is not addressed here.

#12633 — OpenCode Zen Responses models reject the OmniRoute auth header

Root cause: OpencodeExecutor.buildHeaders() only special-cased x-api-key for _requestFormat === "claude"; every other format — including "openai-responses", used by Muse Spark Contributor models routed to /v1/responses — fell through to Authorization: Bearer <key>. OpenCode Zen's /v1/responses endpoint actually requires x-api-key, matching the reporter's live upstream 401.

Fix: added a usesZenApiKeyAuth() helper that also emits x-api-key when _requestFormat === "openai-responses" and the executor's baseUrl is the main Zen host (https://opencode.ai/zen/v1, shared by the opencode and opencode-zen registry entries). Scoped by baseUrl rather than provider id/alias so opencode-go — a separate upstream (https://opencode.ai/zen/go/v1) that already works with Bearer for its own Responses-routed models (grok-4.5, deepseek-v4-pro, muse-spark-1.2-contributor) — is unaffected.

Regression test: tests/unit/issue-12633-opencode-zen-responses-auth.test.ts

• RED (before fix): openai-responses format on opencode/opencode-zen sent Authorization: Bearer, no x-api-key.
• GREEN (after fix): both send x-api-key, no Authorization; opencode-go (negative control, different upstream) still correctly sends Authorization: Bearer; claude format behavior unchanged.

Scope note: the plan-file's open "Scope Question" (is x-api-key required only for Muse Contributor, or every Zen Responses model?) is answered by scoping to the whole Zen host rather than guessing per-model — any current or future model routed to openai-responses on the main Zen host now gets the correct header, without touching opencode-go. The plan-file's secondary checkbox (adding muse-spark-1.3/muse-spark-1.3-contributor-free registry entries) is not included here — that model's real capability metadata is unconfirmed and is already the subject of a separate open contributor PR (#12675, opencode-go's 1.3 registration), so adding it here as well would risk a conflicting/duplicate registration.

Gates run

• node --import tsx/esm --test tests/unit/issue-12681-opencode-muse-spark-context.test.ts tests/unit/issue-12633-opencode-zen-responses-auth.test.ts — 7/7 pass
• Existing area tests (buildHeaders, target-format, muse-spark routing/output, opencode-go catalog alignment, account rotation, ambient proxy, stream payload collector, etc. — ~289 tests across 24 files) — all pass, unchanged
• node scripts/check/check-file-size.mjs — OK
• node scripts/check/check-complexity.mjs — OK (2798 vs baseline 3218)
• node scripts/check/check-cognitive-complexity.mjs — OK (1265 vs baseline 1437)
• npm run typecheck:core — exit 0
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> — exit 0
• node scripts/check/check-changelog-integrity.mjs — OK

⚠️ base-red inherited: #12732 — unit #12058, integration codex-cache, package-artifact, tarball-smoke, agent-skills-sync