fix(cli): gate OpenAI-only reasoning effort tiers to gpt-5 family - #13563
fix(cli): gate OpenAI-only reasoning effort tiers to gpt-5 family#13563maphew wants to merge 5 commits into
Conversation
…t-variants-non-openai Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
Code Review SummaryStatus: No Issues Found | Recommendation: Merge Files Reviewed (3 files)
Previous Review Summaries (2 snapshots, latest commit 5f2740e)Current summary above is authoritative. Previous snapshots are kept for context only. Previous review (commit 5f2740e)Status: No Issues Found | Recommendation: Merge Files Reviewed (4 files)
Previous review (commit ba4a8dc)Status: No Issues Found | Recommendation: Merge Files Reviewed (3 files)
Reviewed by grok-4.6 · Input: 182.7K · Output: 11K · Cached: 279.9K Review guidance: REVIEW.md from base branch |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ba4a8dc686
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| // kilocode_change start - like GPT5_FAMILY_RE, but also accepts a "." prefix so | ||
| // provider-prefixed ids like Bedrock Mantle's "openai.gpt-5.5" count as gpt-5 | ||
| // family. Still excludes lookalikes such as "gpt-50" and "gpt-5o". | ||
| const GPT5_FAMILY_PREFIX_RE = /(?:^|\/|\.)gpt-5(?:[.-]|$)/ |
There was a problem hiding this comment.
Move this Kilo-specific helper and behavior gate out of the shared upstream src/provider/transform.ts, with its regression tests moved from test/provider/transform.test.ts to the corresponding test/kilocode/ mirror. The package rule requires Kilo-specific logic to live under src/kilocode/ and leave only a single marked call in the upstream file, whereas this commit adds two inline blocks and 53 lines of tests to shared files, increasing recurring upstream merge conflicts.
AGENTS.md reference: packages/opencode/AGENTS.md:L76-L78
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Addressed in 5f2740e. The gate now lives in src/kilocode/provider/transform.ts (gpt5FamilyReasoningTiers) and the shared src/provider/transform.ts keeps only a single kilocode_change-marked call; the regression tests were moved to test/kilocode/provider/transform.test.ts.
| // models routed through @ai-sdk/openai on a compatible base URL (e.g. Grok | ||
| // via api.x.ai) reject them, which surfaced as offering an unusable `xhigh` | ||
| // variant (issue #13342). | ||
| if (!GPT5_FAMILY_PREFIX_RE.test(id)) return WIDELY_SUPPORTED_EFFORTS |
There was a problem hiding this comment.
Extend this family check to the provider-prefixed ID forms already used by the SAP path. For example, azure-openai--gpt-5.4 reaches openaiReasoningEfforts() through the SAP adapter's id.includes("gpt") branch, but this regex rejects it because gpt-5 follows - rather than the start, /, or .; the new early return consequently removes its supported none and xhigh variants. The existing SAP tests demonstrate the same azure-openai--<model> naming form with azure-openai--o3-mini.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Addressed in 5f2740e. The family regex now accepts hyphen-prefixed deployment ids ((?:^|\/|\.|-)gpt-5(?:[.-]|$)), so azure-openai--gpt-5.4 through the SAP adapter keeps its none/xhigh variants. Regression coverage added for the SAP and Bedrock Mantle forms.
|
Addressed both Codex review points in 5f2740e:
Verification: |
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Merge-readiness check complete:
Ready for human review. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5f2740eddc
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| @@ -0,0 +1,23 @@ | |||
| // kilocode_change - new file | |||
There was a problem hiding this comment.
This file is already under src/kilocode/, but its first line marks it as shared upstream code. The repository rule explicitly prohibits kilocode_change markers in directories whose names contain kilo; retaining it misclassifies this Kilo-owned mirror for future upstream-conflict work.
AGENTS.md reference: AGENTS.md:L189-L190
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Addressed in b93b74d: the // kilocode_change - new file marker was removed from src/kilocode/provider/transform.ts since the file lives in a kilo-owned directory and needs no marker.
| // kilocode_change start - `none`/`xhigh` are OpenAI rollout tiers only gpt-5-family | ||
| // models accept; non-OpenAI models on an OpenAI-compatible base URL (e.g. Grok | ||
| // via api.x.ai) reject them (issue #13342). See gpt5FamilyReasoningTiers. | ||
| if (!gpt5FamilyReasoningTiers(id)) return WIDELY_SUPPORTED_EFFORTS |
There was a problem hiding this comment.
When a native Azure GPT-5.2+ custom model uses an opaque deployment name such as production and has a post-cutoff release_date, this check sees only model.api.id (the deployment name) and now returns low/medium/high, removing supported none and xhigh variants that the previous release-date logic exposed. This is a supported configuration shape: packages/kilo-docs/pages/code-with-ai/agents/custom-models.md documents that the config model key identifies GPT-5 while id is the independently named Azure deployment, and provider.ts stores that deployment in api.id; the gate therefore needs access to the model key/family rather than relying exclusively on the transport ID.
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Addressed in b93b74d: openaiReasoningEfforts now takes isFirstPartyOpenAI, set for the native @ai-sdk/azure path, which serves only first-party OpenAI models. Opaque deployment/model keys (e.g. production) there keep their none/xhigh rollout tiers while non-OpenAI models on OpenAI-compatible base URLs remain gated. Regression tests added for both the opaque-deployment and the documented model-key/deployment-id shapes.
|
Follow-up: the Codex re-review findings were addressed in b93b74d (removed stray marker from the kilo-owned file; native Azure now passes |
Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>
1 participant
Fixes #13342
What
openaiReasoningEfforts()inpackages/opencode/src/provider/transform.tshanded out OpenAI's rollout-gated reasoning tiers —none(rolled out 2025-11-13) andxhigh(rolled out 2025-12-04) — to any model whose release date is new enough, regardless of whether it is actually a gpt-5-family model. Non-OpenAI models routed through@ai-sdk/openaion an OpenAI-compatible base URL (e.g. Grok viaapi.x.ai, the config shape the maintainer workaround in #13082 recommends) were therefore offered anxhighvariant they cannot honor, and selecting it failed withinvalid xai provider options(the AI SDK's zod validation rejectingreasoning_effort: "xhigh").Change
The gpt-5 family gate lives in a Kilo mirror (
src/kilocode/provider/transform.ts) per the fork-merge policy; the shared upstream file keeps a single markedgpt5FamilyReasoningTiers(id)call and a marked import. The family check accepts bare (gpt-5.5), slash-prefixed (openai/gpt-5.6), dot-prefixed (openai.gpt-5.5, Bedrock Mantle), and hyphen-prefixed deployment ids (azure-openai--gpt-5.4, SAP AI Core) while still excluding lookalikes likegpt-50/gpt-5o. Non-gpt-5 ids get only the widely-supportedlow/medium/highset — matching what the bundled@ai-sdk/xaiprovider already exposes for Grok.All existing gpt-5 / o-series / azure / mantle / SAP variant expectations are preserved (the shared
transform.test.tssuite passes, including theopenai.gpt-5.5Mantle case that still receivesnone/xhigh).Why this scope
Scanned the issue space before picking this: #13342 is unaddressed (no PR), and a search for related reasoning-effort PRs found no intersections. The same class of mis-offering exists in
openaiCompatibleReasoningEfforts()for@ai-sdk/gateway, but that path is deliberately tested to expose the full OpenAI set (Cloudflare's gateway translatesreasoning_effortupstream), so it is intentionally left unchanged to avoid a behavioral regression on the gateway's contract.Verification
test/kilocode/provider/transform.test.ts(Kilo-owned path), failing before the fix and passing after:grok-4.6via@ai-sdk/openaiwith a post-cutoff release date → onlylow/medium/high.low/medium/high.gpt-50/gpt-5oexcluded.openai.gpt-5.5(Mantle) andazure-openai--gpt-5.4(SAP) still receivenone/xhigh.bun test ./test/kilocode/provider/transform.test.ts→ 5 pass;bun test ./test/provider/transform.test.ts→ 422 pass.bun run typecheck(opencode) → clean.bun run script/check-opencode-annotations.ts --worktree→ clean.Note on CI
The JetBrains typecheck in the pre-push hook could not run in the sandbox (no Java 21 / SDKMAN available); it is unrelated to these opencode-only changes.
test/provider/provider.test.tsshows 3 pre-existing failures (defaultModel/opencode loaderresolve to the Kilo gatewaykilo-auto/efficientmodel instead of a direct provider) that are unrelated to this diff.