fix(agent): constrain image payloads and enforce a request-body budget for the ChatGPT Codex transport

[jessepollak]

Summary

The ChatGPT Codex transport (chatgpt.com/backend-api/codex/responses) hard-drops oversized request bodies: the server reads the whole upload, then closes the connection without sending any response. Measured by bisection (2026-08-02): bodies up to ~1,175,000 bytes return HTTP 200; ~1,200,000 and above are dropped after ~15s. The client can only surface this as a retryable-looking APIConnectionError — and because sessions run store=False, the identical oversized history is re-sent on every call and every retry, permanently wedging the session. One unresized phone photo (11MB JPEG → ~15MB base64 data URL) is enough.

Token-based context compression cannot protect this limit: images and large tool outputs are byte-dense but token-cheap, so a body can be 6x over the wire ceiling while token accounting reports plenty of headroom (observed: 7.05MB body at ~170k tokens, 98% of it replayed function_call_output items).

Two commits:

1. Constrain native image payloads before embedding — inbound images are re-encoded in memory (EXIF-oriented, JPEG/PNG, quality-then-dimension reduction) to fit a configurable data-URL byte cap before entering conversation history; bmp/tiff transcode to PNG; unidentifiable images are skipped with a warning instead of poisoning the session. Config: agent.native_image_max_payload_bytes / agent.native_image_max_dimension.
2. Enforce a total request-body budget at codex preflight — agent/request_budget.py measures the serialized body just before send (including retries) and, when over budget, degrades a deep copy via a ladder: truncate history tool outputs largest-first (the most recent tool exchanges stay verbatim) → re-encode embedded images → placeholder history images → truncate active-tail outputs → last-resort placeholder any remaining oversized image. User text is never modified; under-budget bodies pass through untouched. Progress is tracked with exact per-item serialization deltas (2 full dumps per over-budget call). Gated on the codex issuer classification so other Harmony-speaking backends are never silently degraded. Config: agent.max_request_body_bytes (default 1,000,000 — the measured ceiling minus headroom).

Related issues: #27478 (pre-resize inbound images in gateway), #47339 (413 — context compression doesn't evict image/vision payloads), and the same failure class in OpenAI's own clients (openai/codex#13508, #7682, #2908 — there it surfaces as a clean 413; through this client it's a connection drop).

Verification

• Bisected the live endpoint to locate the ceiling (12 probes: clean pass/fail flip between 1,175,000 and 1,200,000 bytes).
• Replayed a real failed 7.05MB request dump through the budget: constrained to 931KB in 1.4s and accepted live (HTTP 200) by the same endpoint that had silently dropped it.
• 62 unit tests for the new modules + the transport suite (239) green on this branch.
• An independent correctness review (GPT-5.6 agent with repo access) produced 5 findings with concrete reproductions — a size-keyed retry memo that could substitute an unrelated equal-length request (memo removed), aggregate under-cap images bypassing the last resort, untouched function_call.arguments/reasoning.encrypted_content, first-pass truncation floors exhausting on 200-output histories, and a chars-vs-bytes marker label. All five are fixed and covered by regression tests in tests/agent/test_request_budget.py.

Notes for reviewers

• The budget deliberately lives at transport preflight, not the compression layer: with store=False + client retries, enforcement must be per-request. A follow-up could make the shrink durable in stored history (today the degradation is recomputed per call).
• There are now parallel image-fitting implementations (agent/image_payloads.py vs tools/vision_tools.py/conversation_compression.py); consolidating them is worthwhile but was kept out of this PR to keep the diff reviewable.

[aerialdreamxl]

Nearly same issue here on Aliyun Dashscope Token Plan registration, here's error.log, saying "An existing connection was forcibly closed by the remote host" when trying to continue long sessions

errors.log