fix(providers): stop silently dropping reasoning effort on opencode families (#10788)

[oyi77]

Summary

Addresses the code facets of #10788 (part 1). Ollama Cloud tier declarations already landed upstream in #11307; this covers what is still missing on release/v3.8.51 for opencode / opencode-go / opencode-zen and documents nvidia.

Dispatch — the silent drop

OpencodeExecutor.transformRequest stripped the effort-suffixed alias (e.g. glm-5.2-high) down to the base id and injected a flat reasoning_effort body field. But opencode-go's ChatCompletionRequest has no such field for non-DeepSeek families (its reasoning is a structured object), so the tier never reached the upstream — every request ran at default effort with no error.

Fix:

• DeepSeek V4 keeps the base-rewrite + reasoning_effort injection — that IS its native contract via opencode-go (feat(opencode): add go deepseek reasoning variants #4647, sanitizer whitelist)
• Every other family now forwards the aliased model id verbatim. The suffix is their only native effort mechanism (exactly the ids opencode models opencode-go --verbose lists), so the tier finally reaches the upstream

Registry — tiers invisible in the catalog

Declared supportedThinkingEfforts on base rows that had aliases but no tier vocabulary:

Registry	Models declared
opencode-go	glm-5.2 [high,max], mimo-v2.5 [high,max], grok-4.5 [low,medium,high], hy3 [none,low,high], kimi-k3 [max], qwen3.7-max/plus [high,max]
shared zen/go row	qwen3.6-plus [high,max]
opencode-zen	deepseek-v4-pro/flash [none,low,high,max], glm-5.2 [high,max], kimi-k3 [max]

This lets catalog variant synthesis (#9485) and sanitizeReasoningEffortForProvider clamping work from one source of truth matching the executor's EFFORT_TIERS.

nvidia — honest empty declaration

nvidia/z-ai/glm-5.2 is now marked reasoning-capable with an empty tier list: its upstream only exposes a binary chat_template_kwargs.enable_thinking switch (mapNvidiaGlm52ReasoningParams), so there are no honest -low/-high/-max tiers to advertise. Synthesizing fake variants would trade one bug for another; the empty list suppresses them while keeping the reasoning flag accurate.

Tests changed/added

• opencode-go-effort-aliases-8353.test.ts: transform samples updated — non-DeepSeek aliases must reach the wire untouched with NO injected flat field; new registry-parity test asserting every base row's vocabulary equals EFFORT_TIERS; nvidia binary-switch declaration test
• opencode-go-effort-aliases-6922.test.ts: glm/mimo transform expectations updated to alias-forwarding
• DeepSeek paths (feat(opencode): add go deepseek reasoning variants #4647 tests, clobber guard) unchanged and still green

Verification

opencode/nvidia/deepseek effort suites (8 files): 191/191 pass
tsc --noEmit: no errors in changed files

Base: release/v3.8.51 (3192eb88d). Release-engineering facet of #10788 (npm publish flow) is out of scope here.

[oyi77]

CI triage

Check	Verdict	Evidence
Fast Quality Gates / No new ESLint warnings	Pre-existing on the base branch	run-eslint-json --max-warnings 0 exits 2 with "suppressions left that do not occur anymore" — reproduced verbatim on a stashed clean release/v3.8.51 checkout (22 baseline problems). The currently-green PRs all target release/v3.8.50, whose suppression state differs. None of the flagged surface is touched by this PR (registry declarations + one executor transform).
Unit Tests fast-path (4/4)	Under investigation → likely flake	Same class of single-test flakes as documented on #11408 (tmp-dir rmSync ENOTEMPTY race and jitter-timing assertions in files this PR does not touch). Both pass locally on this branch.
dast-smoke	Does not reproduce from this diff	The fuzz 500s target DELETE /api/keys/{id}, OPTIONS /api/keys, OPTIONS /api/auth/logout — probed locally against a server built from the lazy-executor sibling branch: 204/401/204, all documented statuses.

No code changes required; will act on any maintainer findings.

[diegosouzapw]

Thanks — the DeepSeek-vs-rest split and the registry-parity test are exactly right, and we reproduced the silent-drop bug on the current tip. One blocker before merge: your new supportedThinkingEfforts declarations activate the nearest-tier clamp from #11295/#11305, so a Codex request asking effort=low on opencode-go/glm-5.2 is now rewritten to high — 'Codex Responses routing keeps reasoning effort while dropping GPT-only verbosity' fails on this branch but passes on the base tip (reproduced locally; it is also why Unit Tests 4/4 is red). Please either update that guard expectation deliberately (with a note on what the upstream actually accepts) or scope the declarations so lower requested efforts survive. The ESLint/quality-gate reds are inherited base-red and not your problem.

[oyi77]

Nearest-tier clamp fix pushed (63ea2ed1f): the Codex-reroute characterization now locks the #10788 behavior — glm-5.2 declares {high, max} on opencode-go, so out-of-vocab low clamps up to high while GPT-only verbosity is still stripped. The fixed test passes in Unit Tests fast-path (3/4).

Heads-up on the current red batch (Docs Gates, all 4 unit shards, Vitest): these reproduce at pure upstream tip 6435f618f with none of this branch's content — e.g. tests/unit/agent-card-route.test.ts fails 0/5 and tests/unit/providers-constants-split.test.ts (231 APIKEY_PROVIDERS entries) fails at the tip itself, and check:docs-counts reports README/AGENTS/llm.txt still saying "159 migrations" while code has 160 after today's quota-share merge (#11408). They hit any PR merging against the current tip.