Repository navigation
fix(api): resolve auto-combo candidate capabilities once in GET /api/combos/auto - #15353
Merged
diegosouzapw merged 1 commit intoOct 6, 2026
Conversation
maxmad64bis
force-pushed
the
fix/combos-auto-resolved-capabilities
branch
from
October 2, 2026 14:29
de22d79 to
a0c4adb
Compare
Merged
4 of 5 tasks
…combos/auto Pool prepared once with resolved capabilities; variants read the snapshot.
maxmad64bis
force-pushed
the
fix/combos-auto-resolved-capabilities
branch
from
October 3, 2026 01:06
a0c4adb to
2fb4498
Compare
4 of 5 tasks
maxmad64bis
marked this pull request as ready for review
October 3, 2026 01:41
diegosouzapw
merged commit Oct 6, 2026
98593c2
into
diegosouzapw:release/v3.8.52
119 of 200 checks passed
diegosouzapw
added a commit
that referenced
this pull request
Oct 8, 2026
* test(translator): expect boolean is_error on tool_result blocks (#15754) #15754 made the OpenAI-to-Messages translator always emit a boolean is_error on every tool_result block (strict upstreams such as the Zed hosted proxy reject requests without it). The older multimodal translation test still pinned the block shape without the field. * test(images): include Codex GPT Image models in the image catalog listing (#14976) #14976 registered gpt-image-2.5-flare and gpt-image-2 under the codex image provider (routed to the dedicated Codex Images API) and updated the registry test to the five-id list, but the image catalog route test still pinned the three GPT-5.6 hosted-tool ids. * docs(flags): keep RATE_LIMIT_AUTO_ENABLE catalog default in sync with its definition #15736 changed the catalog Default cell to _(unset)_, but the catalog documents FEATURE_FLAG_DEFINITIONS 1:1 and the definition (what the feature-flags API/UI resolves when nothing is set) is still "false". Restore the code value and keep the corrected runtime explanation in the description: rateLimitManager reads only the env var and falls back to the dashboard setting (default on) when it is unset. * test(combos): pin on-demand per-candidate reads on the tables #15378 left unmemoized #15353 added a contrast test asserting the on-demand capability path re-reads the synced catalog per candidate (>50 reads). #15378, merged just before it, memoized the synced vision verdict behind the catalog version, so that path now issues 7 synced reads and the contrast went red on the tip. Keep the contrast on the capability and override tables (still thousands of per-candidate reads on demand vs <=2 / bounded on the snapshot route) and pin the memo itself: synced reads stay within the per-provider bound. With the pre-#15378 module the new assertion fails (336 synced reads). * test(chatcore): end the prompt-cache fixture on a user turn (#15830) #15830 re-landed #13572: the first-party Messages provider now strips a trailing text-only assistant turn (upstream rejects assistant prefill). The prompt-cache metadata fixture ended on an assistant text block carrying a cache_control breakpoint, so the strip dropped it and the recorded totalBreakpoints fell from 3 to 2. Append a user turn so the fixture still exercises system, user and assistant breakpoints; the assertions are unchanged. * fix(providers): keep discovered Codex effort variants in the exclusive dashboard listing #13224 lists one <model>-<tier> row per reasoning level the Codex account advertises (appendSyncedEffortVariants, non-exclusive branch). #15132 then made Codex an exclusive synced-listing provider, and the exclusive branch returns before that step, so the provider dashboard / Test All listing lost the discovered tier rows while /v1/models still advertised them (cx/<model>-max). Apply the same variant step on the exclusive branch for codex. The variants come from the live inventory, so #15132's contract (no static aliases the account does not advertise) still holds; its own tests stay green. The one assertion in codex-discovered-reasoning that expected a static registry-only row (gpt-5.6-sol-ultra) next to the live inventory is flipped to #15132's contract. * test(providers): keep the custom gpt-6 max_tokens case on an unregistered id (#15164) The #14869 custom-model case used gpt-6-luna because it was not in the openai registry, so it stayed on Chat Completions and had to be renamed to max_completion_tokens. #15164 registered gpt-6-luna with targetFormat openai-responses, so the request now goes through the Responses translator (max_output_tokens) and the Chat Completions assertion read undefined. Use an unregistered gpt-6 id for the custom-model case and add a case for the registered gpt-6-luna Responses path (max_output_tokens, no max_tokens). * chore(changelog): add fragment for the v3.8.52 tip unit-red drain
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
GET /api/combos/autoprepares the candidate pool once since #14925, but without resolved capabilities, so every listed variant re-resolves context length, output limit, vision and reasoning for every candidate, re-reading and re-parsing the provider's synced model rows each time while nothing on that path yields. The route now asks for the pool with capabilities resolved, the option the catalog and the duplicate route already use: one bulk read per request plus the existing cooperative yield. The pre-resolved vision check gets the vision-bridge exclusion the on-demand check already had, so the response is unchanged; the built-in catalog path shares that filter and is covered by its suite.Related Issues
Validation
npm run lint— ESLint on the touched files is clean; the full run is red on the base (🔴 Release branch not green: release/v3.8.52 #15306).Tests Added Or Updated
tests/unit/14889-combos-auto-prepare-once.test.ts: bounded capability reads, set-equivalence across real parsed suffix and family variants, pre-resolved parity mirror, resilience on unknown models.tests/unit/autoCombo/vision-filter-excludes-forced.test.ts: pre-resolved branch excludes bridge-forced models, unchanged on-demand cases.auto/chaosjudge model already vary between processes on the base. Each new case fails without the change.Coverage Notes
src/app/api/combos/auto/route.ts: covered bytests/unit/14889-combos-auto-prepare-once.test.ts.open-sse/services/autoCombo/suffixComposition.ts: covered bytests/unit/autoCombo/vision-filter-excludes-forced.test.tsandtests/unit/14889-combos-auto-prepare-once.test.ts.Reviewer Notes
catch {}blocks of the same route are handled by fix(api): log the auto-combo variants GET /api/combos/auto skips instead of failing silently #15362, which touches the same file: either order works, the second one needs a rebase on the release tip before landing.auto/*request path keeps on-demand resolution, to be measured separately.