fix: Codex preflight summary compatibility (max_output_tokens + responses-lite header)

[ranxianglei]

Fixes #663

Two related fixes so Codex (ChatGPT-login) preflight summaries work end-to-end. Both live in requestSummary (src/preflight.ts) and only affect the independently-built summary call — the original model request is untouched.

1. Learn to drop max_output_tokens when the upstream rejects it

The ChatGPT codex backend rejects the Responses max_output_tokens parameter outright (400 {"detail":"Unsupported parameter: max_output_tokens"}), which killed preflight summaries with a 502 after the #626 stream fix. The proxy now detects that rejection, retries the summarization without the optional parameter, and remembers it per session+upstream+model — so a model/provider that accepts the limit keeps the existing 8192 output cap. Both rejection orders (stream-first, param-first) recover, and the retries are bounded (each capability is learned at most once).

2. Drop the Codex Lite header from the summary call only

requestSummary copied ALL main-request headers into its independently-built summary body. The main Codex request carries x-openai-internal-codex-responses-lite (its body has reasoning.context: all_turns), but the summary body has no reasoning field — so the backend rejected the summary with 400 "…requires reasoning.context to be all_turns". The fix strips that one request-shape-specific header from the summary call only; the original model request keeps the header and its reasoning fields. Auth and routing headers are preserved. The strip is keyed on the Codex header name, so it is a no-op for Anthropic/OpenAI.

Tests

tests/preflight-max-output-tokens.test.ts — 6 e2e tests:

• order A (stream-first) + order B (param-first): both rejection orders learn and recover; second request first-shot compatible
• standard provider: 8192 cap retained, no capability learned
• model scoping: rejection learned for model A keeps model B's cap
• codex lite header: summary drops the Lite header, main request keeps it; auth preserved
• anthropic: header strip is a no-op — summary keeps auth, fold + forward OK

Pre-flight

• npm run typecheck ✓
• npm test ✓ (1252/1252)
• npm run build ✓

Validated against the real backend by the reporter: an over-window Lite-mode request now completes after preflight folds the ranges, preserving the canary.

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-09_preflight-max-output-tokens (128f9d3)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-665

Each push to this PR publishes a new version under the pr-665 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf billion-context-pr665.tgz
npm install -g package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review in progress — verifying #663 repro evidence, checking the diff against the #626 capability-learning pattern, and running typecheck/tests/build on the PR branch. Will report findings here.