fix(cloudflare): sync reasoning controls from model schemas

[opencode-agent]

Summary

Fetch each Cloudflare Workers AI model's input schema and sync the reasoning controls it declares.

The format=openrouter catalog advertises reasoning as a feature but currently contains no per-model reasoning objects. Simply retaining the shared translator's computed options is not enough: the sync needs the separate model-schema endpoint to discover the controls.

• Fetch every listed model's schema in batches of four, with a 30-second timeout per schema request. Fail the sync on failed or malformed schema responses rather than writing a partial catalog.
• For catalog-tagged reasoning models, read explicit reasoning_effort / reasoning.effort enums and the chat_template_kwargs.enable_thinking boolean through request-schema anyOf / oneOf alternatives.
• Preserve curated fallback options when the schema has no usable controls. Do not infer reasoning capability from generic schema fields, invent effort levels from null, or treat completion-token limits as reasoning budgets.
• Keep explicit catalog reasoning metadata when supplied instead of unconditionally overwriting it with stale authored options.
• Refresh schema-derived reasoning comments, including toggle wire paths, while preserving unrelated notes and source links. Header-only updates are idempotent, and other providers retain their existing header behavior.

No catalog TOMLs or credentials are included in this change. Schema declarations describe the provider's advertised API surface; this does not claim inference-tested behavioral differences between effort levels.

Live verification

Read-only API checks on 2026-08-26:

• The marketplace catalog returned 26 models, 15 tagged as reasoning, and no reasoning objects.
• All 26 model-schema requests succeeded through the updated fetch/parse/translate path.
• GLM-5.3-Flash's schema declares reasoning_effort = low|medium|high and chat_template_kwargs.enable_thinking = true|false. GPT-OSS declares nested reasoning.effort without that toggle.
• The full dry run reports 1 create, 11 updates (including comment-only refreshes), 0 deletions, and 14 unchanged entries. No inference calls or catalog writes were performed.

Sources: Cloudflare model search and model-schema endpoint.

Validation

• Regression tests reproduced discarded metadata, missing per-model schema fetches, and missing schema-derived controls before their fixes.
• bun test packages/core/test/cloudflare-workers-ai-sync.test.ts: 24 pass, including fetch failures, schema extraction, curated fallback, serialized header preservation, and repeat-sync idempotence.
• bun test packages/core/test/sync.test.ts --test-name-pattern 'OpenRouter|reasoning options|reasoning efforts|reasoning metadata|reasoning controls|header': 32 pass.
• bun models:sync cloudflare-workers-ai --dry-run --no-issues: passes against the live API.
• bun validate and git diff --check: pass.
• The broader sync suite still fails DeepInfra preserves live modalities for new base models; this was reproduced without the Cloudflare fix and is outside this change.
• Package-wide typecheck remains blocked by existing diagnostics, including the shared runner's SyncedMetadata.reasoning_options access. No diagnostics in the changed Cloudflare provider or its tests.

Requested by: @rekram1-node (Aiden via Slack)

[github-actions]

No actionable findings.

[github-actions]

Action items

• [high] [violation] packages/core/test/cloudflare-workers-ai-sync.test.ts:66 - Check: Temp dirs in tests must be created under an existing parent (repo convention: path.join(tmpdir(), …)). Why: mkdtemp("/tmp/opencode/workers-schema-") requires /tmp/opencode to already exist; on a clean runner that path is missing, so the header-refresh/idempotence test throws ENOENT instead of validating sync behavior. Action: Create the temp dir with path.join(tmpdir(), "workers-schema-") (and import tmpdir from node:os), matching the other sync tests.

[superhighfives]

Got an MR to fix this, @rekram1-node: #7920