* test(infra): retry recursive temp-dir removal instead of failing a shard on ENOTEMPTY (#11966)
Two shards on release/v3.8.51 went red in one day with the same signature —
"ENOTEMPTY, Directory not empty: /tmp/omniroute-<test>-XXXXXX" — from
combo-same-provider-cascade (Unit Tests fast-path 4/4, on a PR that touches only
.github/) and auth-policy-embeddings-webfetch-7785 (the 20k-test TIA step). Both pass
alone and on re-run: the cleanup races something still writing into the directory
(SQLite WAL/-shm checkpoint, a worker, the backup) and under a loaded hosted runner
the window opens. 1154 test files do their own cleanup with
fs.rmSync(dir, { recursive: true, force: true }); 57 already asked for retries.
One-shot codemod (scripts/ad-hoc/codemod-rm-maxretries.mjs, kept for the record):
every rm / rmSync / rmdirSync option object with `recursive: true` and no
`maxRetries` gains `maxRetries: 5, retryDelay: 100` — Node itself then retries
ENOTEMPTY/EBUSY/EPERM for up to ~0.5 s before giving up. 2243 call sites in 1292
files under tests/, the shared tests/_setup/isolateDataDir.ts exit hook included.
Only the option object changes: no call site, assertion or import is touched.
Validation: prettier and ESLint (with the frozen suppressions) clean on all 1292
files; a random 20-file sample runs green (quota-redis-store hangs identically on
the untouched tree — it needs a Redis on localhost, an environment matter). The
four unit shards on this PR are the full run.
* fix(quality): let check-forgotten-sibling-tests read a 1,000-file diff
The gate shells out to `git diff` through execFileSync with Node's default 1 MB
maxBuffer; the 1,292-file codemod in this PR is the first diff large enough to
overflow it, and the gate died with `spawnSync git ENOBUFS` before comparing
anything. 64 MB is far above any real PR and costs nothing when unused.
Merged via /merge-batch (v3.8.51 provenance sweep). Boarded and validated together with the batch's other provenance/retirement PRs in a combined worktree — full gate suite green, including the audio/speech-combo regression suites. Thank you.
* fix(providers): resolve combo names on /v1/audio/speech and /v1/videos/generations
`GET /v1/models` advertises combos with `owned_by: combo`, and chat, embeddings,
transcriptions (#9134) and images (#8986, #9239) all resolve those names. Speech
and video did not: both rejected a combo name at model validation, before any
resolution could happen.
POST /v1/audio/speech {"model":"my-combo","input":"hi"}
-> 400 Invalid speech model: my-combo. Use format: provider/model
POST /v1/videos/generations {"model":"my-combo","prompt":"a cube"}
-> 400 Invalid video model: my-combo. Use format: provider/model
A client picking a model out of /v1/models therefore could not tell which
entries the catalogue would actually accept, and callers ended up hardcoding
vendor ids for these two routes while using combo names everywhere else.
Both routes now mirror the images route: detect a combo name before the
provider lookup and divert to a strategy executor. The two new executors follow
imageCombo — expand targets with resolveComboTargets(), filter to targets the
route can actually serve, walk them in priority order, and return the first
success or the last failure, with 400/401/403 treated as terminal.
Two details differ from the image strategy:
Speech filters at model level rather than provider level. parseSpeechModel()
resolves a provider prefix without checking that the model behind it can speak,
so `openai/gpt-4o` would otherwise be accepted as a target and fail only once
dispatched. The filter now checks the provider's own model list, and keeps
targets from dynamic provider nodes that do not enumerate models.
Speech also returns the handler's Response untouched instead of building a JSON
body, because that route streams audio; only the ADD-only meta headers are
attached, exactly as the direct path does. The failure branch is the only place
the body is read.
successfulMediaGenerationResponse() gains optional `strategy` and
`fallbackAttempts` so the video strategy can report them the way imageCombo
does, rather than duplicating the cost calculation. Both are omitted on the
direct single-model path, where neither is meaningful.
Tests mirror tests/unit/combo/image-combo.test.ts for both routes: combo not
found, no capable targets, empty combo, and targets present with no provider
connection. 16/16 pass across the three combo test files.
* fix(providers): preserve local overrides, custom models and per-target prompt rules through video combo dispatch
executeVideoCombo() diverged from the direct /v1/videos/generations route in
three ways: it dropped the ComfyUI-style local-override credential lookup for
authType:"none" targets, its capability filter only matched the built-in
video registry (skipping custom OpenAI-compatible provider nodes tagged with
the "videos" endpoint), and the route validated the prompt against the
unresolved combo name before combo targets were expanded — rejecting
prompt-optional I2V targets that never got the chance to opt out.
Extracts the shared resolution rules (resolveVideoModelTarget,
isVideoPromptOptional, resolveLocalOverrideCredentials) into
src/app/api/v1/_shared/videoModelResolution.ts so the direct route and the
combo executor apply identical rules, moves the combo-name diversion ahead of
the prompt-required check so validation runs against the real resolved
target, and adds per-target prompt validation inside the combo loop so a
missing prompt only rules out that target instead of the whole combo.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
