Canonical surface

The endpoints the prober walks under the default openai profile, plus the /v1/videos HT-compat extension as a representative ours row. The catalog is sourced from endpoints.py; the full HT-compat ours surface (rerank, segmentation, encoder rows, omni chat, 3D, image decomposition, audio segmentation) lives in HT-compat profile.

You can produce a current copy of this table with:

aioc spec

Models / discovery

Path	Method	Kind	Notes
/v1/models	GET	core	Required for any compat surface; powers model sniffing
/v1/models/{model}	GET	core	Many OSS impls return 404; OpenAI returns the object

Chat & completions

Path	Method	Kind	Notes
/v1/chat/completions	POST	core	The headline endpoint; supports tools, JSON mode, streams
/v1/chat/completions (stream)	POST	core	Separate row so a missing-stream regression is visible
/v1/chat/completions (logprobs)	POST	ext	Separate row so a missing-logprobs regression is visible
/v1/completions	POST	ext	Legacy text completion; many newer servers omit it
/v1/responses	POST	ext	Newer Responses API; few OSS servers implement
/v1/responses/compact	POST	ext	Compact-output variant of the Responses API
/v1/realtime	GET	ext	WebSocket upgrade for low-latency conversational sessions

Audio

Path	Method	Kind	Notes
/v1/audio/speech	POST	core	TTS; returns audio bytes (mp3/opus/wav)
/v1/audio/transcriptions	POST	core	STT; multipart upload, returns text JSON
/v1/audio/translations	POST	ext	STT to English; many servers fold into transcriptions
/v1/audio/voices	GET	ext	List installed TTS voice ids

Images

Path	Method	Kind	Notes
/v1/images/generations	POST	core	Returns {data: [{url}|{b64_json}]}
/v1/images/edits	POST	core	OpenAI requires multipart; some OSS take JSON (WARN)

/v1/images/variations was pruned from the catalog in v0.3.1 — see the Retired section of the matrix for context.

Videos (ours extension)

Path	Method	Kind	Notes
/v1/videos	POST	ours	Async job creation; mirrors OpenAI's Sora job shape

The polling routes /v1/videos/{id} and /v1/videos/{id}/content are part of the HT-compat protocol but aren't probed directly — they need a live job id, which would couple a probe to a multi- second render.

Embeddings

Path	Method	Kind	Notes
/v1/embeddings	POST	core	Returns {data: [{embedding: [...]}]}

Files / batches / fine-tuning (typically not implemented)

Path	Method	Kind	Notes
/v1/files	GET	ext	Uploads listing; most OSS servers omit
/v1/batches	GET	ext	OpenAI Batch API; almost no OSS server implements
/v1/fine_tuning/jobs	GET	ext	OSS servers don't fine-tune via API
/v1/uploads	POST	ext	Multi-part upload-session creation (OpenAI Uploads API)

The three GET routes (/v1/files, /v1/batches, /v1/fine_tuning/jobs) carry phase_b_skip=True in the catalog — Phase A existence is the meaningful signal; Phase B against an unauth server would just 401.

Moderation / safety

Path	Method	Kind	Notes
/v1/moderations	POST	ext	Few OSS servers implement

Reading the deviations

A WARN row in aioc probe output is not a failure — it's an endpoint that exists and broadly works but deviates from the canonical shape in a way the catalog has flagged as "common but non-spec". Examples:

• /v1/images/edits taking JSON instead of multipart (vLLM fork).
• /v1/audio/speech returning application/octet-stream instead of audio/mpeg (some llama.cpp shims).
• /v1/chat/completions omitting the usage object on streamed responses (llama.cpp's default).

These are documented per-implementation in Implementations so a WARN doesn't surprise you twice.