Repository navigation
OpenAI/Responses requests now upload chunked without Content-Length (on-demand body from #854) #3847
Description
Background
Since #854, OpenAI chat and Responses HTTP request bodies are produced on demand from disk segments as the upload proceeds. The bytes match what main sent. Because the length is not known before the body is produced, requests now go out with chunked transfer encoding and no Content-Length header. Main sent Content-Length.
The Kimi cache transport test (packages/providers/src/openai/OpenAIProvider.kimiCacheTransport.test.ts) now asserts that Content-Length is absent and the wire body is non-empty.
Risk
Some proxies, gateways and self-hosted inference servers reject chunked request bodies or buffer them differently. None has been reported. The providers tested so far accept chunked uploads: Codex/Responses, Anthropic, DeepSeek through the OpenAI chat path, and the local test servers.
Options if it becomes a problem
Computing the exact length up front needs a full pass over the content before upload. The disk route JSON-escapes UTF-16LE segments chunk by chunk, and the progressive route converts rows while it uploads. Either way the content is read twice. A counting pre-pass that streams the body once to measure it, without keeping it, would do that. It doubles read I/O but adds no memory. It could be opt-in per provider or base URL.
Acceptance (if implemented)
- A test proves the declared length equals the bytes sent for text, media and tool-heavy requests.
- The body is never fully materialized in memory to compute the length.
Follow-up from #854. Only worth doing if a real endpoint rejects chunked uploads.
Activity
github-actions commentedon Oct 10, 2026
LLxprt Issue Planner — infrastructure failure
The autonomous planner failed with exit code 1. Please inspect the workflow logs (Run planner agent section) and re-run once resolved.
fix(#854): A3 Content-Length on the Responses disk text route (closes #…
acoliver commentedon Oct 14, 2026
Fixed in PR #3852: every OpenAI chat, OpenAI Responses and Anthropic request route now sends
Content-Lengthequal to the bytes it uploads. None of them goes out chunked any more.At the start of the fix only the
openai-responsesprovider sent chunked uploads. The chat, openai-provider Responses and Anthropic routes already had a length because they built a bounded array body. When those routes moved to request-scoped disk bodies, each of them kept the header by declaring the sealed size of the spool or segment files before the first byte.Routes covered, with the commit that put the header on each:
openai-responsesprovider, row-backed calls (disk text route): 798fe12 (A3).openai-responsesprovider, calls with no request rows (the old lazy and progressive body): 253a7d5 (C2), which stages the rows on scratch disk and sends them through the same route. The lazy body type is deleted in 6106d45 (C3).openaiprovider Responses route, rows and one-shot streams: c25f528 (C1) and 253a7d5 (C2).How it is checked:
wireBodyGolden.test.ts, 315 cases) recordsContent-Lengthfor 7 routes, each fed from a row selection and from a one-shot stream. All 14 combinations arepresent, and the recorder throws if the value differs from the body byte count. The fixtures were recorded before this work began, and the request bytes are unchanged for every case. The only fixture change in the series is the header flipping from absent to present onopenai-responses.disk-text-content-length.test.tscovers text, inline media, media references, a 120-pair tool-heavy history, and UTF-16 instructions with lone surrogates. A body that serves more or fewer bytes than declared fails the request.sendPeak.*run asserts that Content-Length equals the bytes sent. In the final run this held for chat (8,596,228 and 11,774,659 bytes), openai-provider Responses (8,596,212 and 11,582,131), openai-responses (8,596,212 and 11,582,131) and Anthropic (8,596,378 and 11,541,325).Not covered: the Codex WebSocket path sends frames and has no HTTP length header. Vercel and the Gemini plugin are separate (#3869).
The Kimi cache transport test now asserts that
Content-Lengthequals the body bytes for both the row-backed and the contents-only Responses request. This issue closes when #3852 merges.