Runway (Nano Banana 2) used a 30s whole-request timeout that includes reading the response body, so downloading a rendered image/video artifact off a slow CDN aborted with 'request canceled (Client.Timeout or context cancellation while reading body)'. Raised to 300s; genCtx still bounds the overall generation.
imagine/krea capped polling at 4min and leonardo at 5min while the caller's genCtx budget is 8min, so slow-but-valid generations were killed early with 'generation timed out'. Now the poll deadline tracks ctx.Deadline() minus 60s download headroom (fallback to the old fixed cap if ctx has no deadline). chatgpt already ties its budget to the deadline (pollBudget) and runway/grok honor ctx, so no change needed there.
pollImage/pollVideo capped themselves at 3min/10min while the caller's genCtx budget is 8min (image) / 12min (video), so slow-but-valid generations were killed early with 'adobe generation timed out'. Now the loop honors ctx.Err() and uses the full budget.
Now only the generate/video submit uses proxy.url; cookie-exchange/login, token refresh, credit-balance and project/session bootstrap all go direct. Also splits adobe GenerateVideo (submit=proxy, poll/download=local). Verified live: adobe gpt 1K cookie-exchange via local IP, submit 23.244.30.102 (proxy) vs local 76.209.9.65, 2.0MB PNG. Build+vet pass.
Reference-frame upload, the mp4 download and OpenAsset (/content) streaming now egress on the local IP; only the generate submit uses proxy.url. Verified live: submit egress 98.97.26.37 (proxy) vs upload/download 76.209.9.65 (local), 3.47MB video generated OK.
Extends the adobe split to leonardo, krea, imagine, runway (image+video) and chatgpt: reference-image upload, polling and result download egress on the local IP; only the generate/task-create submit uses proxy.url. chatgpt keeps its anti-bot-guarded submit+upload phase on the proxy and moves only the existing second (poll/resolve/download) session to local. custom is already direct. Build+vet pass; adobe verified live, the other providers are code-complete but not yet live-tested (no dev accounts).
Reference-image upload, job polling and result download now run on the local IP; only the generate submit egresses via proxy.url. Verified live: submit egress 74.244.159.228 (proxy) vs upload/poll/download 76.209.9.65 (local), 1K adobe gpt image generated OK.
The server session slides its TTL on use but the browser cookie's Max-Age was frozen at login, so it lapsed mid-session and broke cookie-only auth (private image <img>/download loads returned 401) while the SPA still looked logged in via its Bearer token. Re-issue the cookie with a fresh Max-Age whenever an authenticated request carried it.
Derive the 49-byte header seed and the salt F per-session from the grok homepage (browser-free tls-client GET), instead of hardcoding constants that go stale on every grok web rebuild (403 anti-bot). F is reproduced natively in Go (curve keyframe sampling: cubic-bezier easing, color lerp, rotation matrix, JS-exact number->hex). Static constants remain as env-overridable fallback. Adds offline regression test against server-verified ground-truth triples.
chiyi
Steven