fix(proxy):-restore-compaction-after-ChatGPT-retired-the-responses-compact-endpoint

[WangErgouaaaa]

Summary

ChatGPT retired the dedicated compact endpoint /backend-api/codex/responses/compact on 2026-08-12 18:44 UTC. Every compaction request through codex-lb has 404'd since then, regardless of client version or account — compaction is fully broken on the current release.

This PR restores compaction by serving it through the plain /codex/responses endpoint with the current v2 protocol:

1. fix(proxy): serve compaction via plain responses endpoint — route compact requests to /codex/responses, force stream: true, set store: false explicitly (the plain endpoint rejects requests without it with 400 Store must be set to false), and re-append the terminal compaction_trigger input item that the API layer strips before the compact service.
2. fix(proxy): restore compaction trigger and collect streamed output items — the upstream compaction result is delivered via SSE response.output_item.done events (the response.completed payload has an empty output array). Fold those events back into the legacy response.compact payload, preserving encrypted_content, so the API layer's compaction validation passes (otherwise 502 Compact response did not include a compaction output item).

Protocol notes

• New protocol (verified against chatgpt.com on 2026-08-13): plain POST /codex/responses + stream: true + store: false + terminal {"type":"compaction_trigger"} input item + x-codex-turn-metadata: {"request_kind":"compaction"} header.
• The compaction item arrives as {"id":"cmp_...","type":"compaction","encrypted_content":"..."} inside response.output_item.done events, ordered by output_index.
• Client side (codex CLI) needs no changes — compact_remote_v2 already speaks this protocol.

Testing

• Full-path verification through the proxy: HTTP 200 with a compaction event stream; request logs record compaction success.
• Unit-tested the SSE fold-back against captured real upstream traffic.

[Komzpa]

Status note: the compact-endpoint retirement fix this PR addresses now lives in #1749 (the #1705 line was folded there). Keeping this open until that lands; if #1749 merges first, this can close as covered. Thanks for the report and fix — the shipped implementation follows the same approach.