hosted agents: declare workspace upload size and negotiate before sending

[sdharavath]

Companion to cthulhu #169365 (MARSOHS-626), which makes OHS shed saturated workspace transfers with a 503 and refuse over-cap uploads from their headers. Both of those verdicts need a client that declares a size and, ideally, asks before sending.

Summary

Declare Content-Length. UploadWorkspace handed its io.Reader straight to http.NewRequest, which derives a length only for a few in-memory types. The *os.File callers pass falls through, so uploads went out chunked with no declared size.

OHS caps uploads at 50 MiB and limits concurrent transfers, but with no length to inspect it cannot apply either verdict until it has read the body. An over-cap upload was therefore only refused mid-stream, after the whole payload had crossed the network — measured on staging at 176s for a 51 MiB file that was doomed from its first byte.

uploadBodyLength measures a regular *os.File from its current offset, or any reader exposing Len(). A new optional ContentLength field covers readers we cannot measure.

Negotiate above 1 MiB. With Expect: 100-continue the payload stays on the client until OHS approves, so a rejection at headers costs one round trip instead of a full transfer. The threshold keeps the extra round trip off small uploads, and an intermediary that ignores the handshake only delays a transfer — the transport falls back to sending the body after ExpectContinueTimeout.

Both changes are additive. doctl needs no source change: it already passes an *os.File rewound to the start after hashing.

Testing

Tests count bytes on the socket rather than trusting headers. A 4 MiB upload refused with 503 puts under 32 KiB on the wire, and a paired control confirms an accepted upload still delivers every byte. Covered over HTTP/1.1, over HTTP/2, and through the exact stack doctl builds (oauth2 wrapped in retryablehttp), where the payload is withheld across both the initial attempt and its retry, and the retry waits out Retry-After rather than its much shorter configured backoff.

Two things worth knowing for review, both easy to get wrong:

• Honouring the handshake depends on a non-zero ExpectContinueTimeout on the underlying transport. httptest leaves it at zero, which silently sends the whole payload and proves nothing, so the tests set it to match production (http.DefaultTransport and cleanhttp both use 1s).
• An h2 handler cannot assert on the Expect header: net/http lifts it into an internal flag before the handler runs. Only the byte count shows the negotiation happened.

go build ./..., go vet ./..., and the full suite pass; the new tests were also run 10x under -race.

No CHANGELOG entry, matching the other hosted-agents commits on this branch.

Made with Cursor