Chat completions api

Chat Completions Compatibility Reference

Chat Completions is the lingua franca of LLM serving — and because everyone has implemented and extended it longest, it carries the MOST undocumented divergence of the three open protocols. OpenAI keeps it fully supported but second-choice ("we recommend trying Responses") — and since 2026-07-29 OpenAI's own publications call it "our legacy Chat Completions API" (ARC-AGI-3 post; still no deprecation or sunset date); xAI, Groq, and Azure declare it legacy too. Legacy /v1/completions loses its last first-party OpenAI models 2026-09-28 and survives as a local/third-party surface.

Fleet property: stateless. Full history resent every turn, like Messages; no server-side session state to break load-balanced fleets (contrast responses-api's previous_response_id).

Last refreshed: 2026-07-19 (source-examined at commits of 2026-07-16..18; provenance in references/sources.md).

Critical Gotchas

• The reasoning-field schism: reasoning_content (SGLang, llama.cpp, mistral.rs, DeepSeek, xAI) vs reasoning (vLLM, Ollama, Together, Groq, OpenRouter) vs inline <think> — and vLLM silently RENAMES incoming reasoning_content→reasoning. Servers should emit both; clients should read both. Full table: references/backend-implementations.md.
• Reasoning must be passed BACK in tool loops on DeepSeek v4 (with tool calls) and OpenRouter (exact block sequence) — clients that strip reasoning break agentic loops. references/cloud-compat.md.
• Silent-drop vs hard-400 split: Ollama silently drops tool_choice (forced tool calls no-op!) and max_completion_tokens (unbounded generation); Groq and xAI-reasoning hard-400 on specific params; LiteLLM raises UnsupportedParamsError as HTTP 500. Identify the target's failure mode before debugging.
• finish_reason is not a closed enum: abort (vLLM, SGLang), repetition, canceled/generated_image (mistral.rs); llama.cpp DEFAULTS to length; named tool_choice returns stop on vLLM/OpenAI but tool_calls on SGLang/mistral.rs. Parse tolerantly, key loops on tool_calls.
• seed is a lottery: honored (vLLM, llama.cpp), silent no-op unless server flag (SGLang --enable-deterministic-inference), ignored (mistral.rs) — and OpenAI has formally deprecated seed AND system_fingerprint.
• json_object ≠ json mode everywhere: SGLang and mistral.rs implement it as schema {"type":"object"} — top-level arrays forbidden. json_schema.strict is ignored by ALL local servers (always fully enforced anyway).
• Tool-call streaming split: whole-blob single delta (Ollama, mistral.rs) vs incremental argument diffs (vLLM, SGLang, llama.cpp). First delta must carry id+function.name or AI-SDK clients kill the stream; loose parsers finalize args the moment they parse as JSON.
• Cached-token reporting is flag-gated: vLLM --enable-prompt-tokens-details, SGLang --enable-cache-report; llama.cpp reports always; Ollama never. OpenAI CC now has explicit caching (prompt_cache_options + per-block breakpoints, cache_write_tokens billed 1.25×): references/spec.md.
• data: [DONE] is not universal — Llama Stack/OGX never sends it; AI SDK clients ignore it, official-SDK clients expect it. Send it; don't require it.

Quick Reference

When invoked with a topic argument (spec, backends, gateways, cloud, clients), load that reference file first and answer from it. Without an argument, pick by question shape: official params/deprecations/what-CC-gets → spec; which-local-server-does-what → backends; proxy/routing behavior → gateways; hosted-provider compat → cloud; what-clients-send / parser tolerance → clients.

• Spec: references/spec.md — full request/response/chunk schema highlights, explicit prompt caching, stored completions, legacy /v1/completions + shutdown timeline, deprecation archaeology, Responses-only feature delta
• Backends: references/backend-implementations.md — divergence matrix
  • per-server sections for the 7 local servers
• Gateways: references/gateways.md — LiteLLM param cascade and prefix routing, Bifrost, Superagent (CC outbound only), gateway-tax table
• Cloud: references/cloud-compat.md — 10 providers' CC-compat endpoints and the cross-provider gotcha matrix
• Clients: references/clients.md — opencode's three CC paths, AI SDK parser tolerance, server tolerance checklist
• Sources: references/sources.md — dated per-URL index with commits examined and the live-verification log

Translation between protocols is homed elsewhere: CC↔Messages mapping in the messages-api skill, CC↔Responses mapping in the responses-api skill.

Procedures

Pointing an OpenAI-SDK client at a local backend

1. Base URL http://host:port/v1 (SDKs append /chat/completions); auth Bearer anything unless the server enforces keys.
2. Check the backend's row in the divergence matrix FIRST — especially Ollama (tool_choice and max_completion_tokens silently dropped) and reasoning field naming.
3. For agent clients (opencode etc.): use the openai-COMPATIBLE provider path, not the openai provider — the openai path model-id-sniffs ("o3-…" gets temperature stripped), drops reasoning fields, and its factory default is the Responses API (references/clients.md).
4. Reasoning models: confirm how thinking is toggled (reasoning_effort mapping vs chat_template_kwargs.enable_thinking vs vendor thinking fields) and which field the CoT comes back in.
5. Sanity curl:

   curl -sS http://host:port/v1/chat/completions -H "Content-Type: application/json" \
     -d '{"model":"<name>","max_tokens":32,"messages":[{"role":"user","content":"Say OK."}]}'
   

   Expect object:"chat.completion", non-null choices[0].message.content, finish_reason of stop OR length (llama.cpp defaults to length). An error body here → check the server's error-envelope row in the matrix before parsing (mistral.rs sends {"message"}, not {"error":{...}}).

Serving CC through LiteLLM to a fleet

1. Prefix decides fidelity: openai/<model> = full surface passthrough; hosted_vllm/<model> = tool schemas silently edited (strict + additionalProperties stripped). Pick deliberately.
2. Set drop_params: true (or per-model additional_drop_params) — otherwise unsupported params surface as HTTP 500 UnsupportedParamsError.
3. Statelessness makes any replica valid — no affinity needed.

Debugging a CC streaming issue

1. Capture raw SSE:

   curl -sN http://host:port/v1/chat/completions -H "Content-Type: application/json" \
     -d '{"model":"<name>","max_tokens":64,"stream":true,"stream_options":{"include_usage":true},"messages":[{"role":"user","content":"hi"}]}' | head -40
   

2. Expect data-only SSE: first chunk delta.role, content/tool deltas keyed by index, finish chunk, optional choices:[] usage chunk, [DONE].
3. Check the backend's streaming row in the matrix for known deviations (always-present null keys on SGLang, vendor timings on llama.cpp, role-on-every-chunk + whole-blob tools on Ollama, raw non-JSON error lines on mistral.rs, missing [DONE] on OGX).
4. If tool calls vanish client-side: check whether the first delta carried id+function.name, and whether the client requires tool_calls[].index (references/clients.md).