uv tool install codexcomp      # install
codexcomp                      # run (127.0.0.1:8787)
# then append to ~/.codex/config.toml:  openai_base_url = "http://127.0.0.1:8787/v1"

It overrides the built-in provider's base URL in place — model_provider is unchanged, so session grouping, remote compaction, and remote-control keep working.

Credits. The detect-and-continue mechanism originates from neteroster/CodexCont (MIT); this is an independent, from-scratch implementation that keeps the built-in provider intact.

The problem

gpt-5.5's reasoning is intermittently truncated at reasoning_tokens == 518·n − 2 (516, 1034, 1552, …): the turn stops mid-reasoning and answers from an incomplete thought, degrading quality sharply. Aggregate telemetry in the upstream report shows ~44 % of gpt-5.5 responses that reach 516 reasoning tokens end at exactly that boundary — an upstream defect with no official fix (openai/codex#30364).

codexcomp sits on 127.0.0.1 between Codex and the upstream Responses API. On a 518n−2 truncation it drives the model to keep reasoning and folds the extra rounds into a single downstream response — Codex sees one complete, untruncated answer.

Features

• Detect → continue → fold — spots the 518n−2 fingerprint, replays the round's reasoning with a continue nudge, and folds all rounds into one response.
• Zero-footprint wiring — one official top-level openai_base_url key; no [model_providers] entry, no provider id change, no session re-bucketing.
• WebSocket-first transport — native responses_websockets protocol (envelope frames, serial connection reuse, prewarm); no "Falling back" noise in Codex logs.
• Resilient SSE fallback — the POST path transparently decompresses zstd/gzip upstream responses.
• Full /v1/* passthrough — including GET /v1/models (model catalog refresh).
• Protocol-faithful headers & compaction — upstream response headers reach Codex on the POST path (x-codex-* rate-limit snapshots, x-models-etag, request ids), the x-codex-turn-state sticky-routing token is replayed across fold rounds when the upstream issues one, and remote-compaction requests are never folded.
• Live streaming — reasoning streams in real time even mid-fold; only the final clean round's output is released downstream.
• Honest accounting — the true cumulative cost of folded rounds is reported under metadata.proxy_billed_usage.
• Loopback-only, auth passthrough — forwards Codex's Authorization header; never reads, persists, or logs a credential.
• Opt-in autostart — installation registers nothing; one command sets up a systemd user unit (Linux/WSL) or LaunchAgent (macOS).

Quick start

Requires uv and the Codex CLI (ChatGPT OAuth; tested on 0.142.x).

uv tool install codexcomp                                  # from PyPI
# uv tool install git+https://github.com/dzshzx/codexcomp  # or from source
codexcomp                                                  # foreground, 127.0.0.1:8787

Point Codex at the proxy with one top-level config key:

# ~/.codex/config.toml  (top level, before the first [table])
openai_base_url = "http://127.0.0.1:8787/v1"

That's it. Disable by removing that line and stopping the proxy; upgrade / uninstall with uv tool upgrade codexcomp / uv tool uninstall codexcomp.

How it works

A state machine (codexcomp/fold.py) runs per round:

1. Detect — reasoning_tokens == 518n − 2 (any tier by default; cuts up to n=21 observed — see --max-n / --max-continue) marks the round as truncated.
2. Continue — discard the tentative output and replay the round's reasoning items (incl. encrypted_content) plus one phase:"commentary" "Continue thinking..." message as the next input.
3. Fold — stream reasoning live, flush only the final clean round, and rebuild the terminal event as one response (reasoning summed, true cost under metadata.proxy_billed_usage).

CLI reference

Autostart (optional, off by default)

codexcomp install-service     # register + start (current platform)
codexcomp uninstall-service   # remove

• Linux / WSL — systemd user unit; loginctl enable-linger starts it at boot without login.
• macOS — launchd LaunchAgent in ~/Library/LaunchAgents/.
• Windows — prints manual steps only: point a Startup shortcut (Win+R → shell:startup) at the windowless codexcompw (where.exe codexcompw). Delete it to disable.

With WSL2 networkingMode=mirrored, Windows and WSL share 127.0.0.1: run one proxy in WSL and just add the openai_base_url line on the Windows side — no second proxy needed.

Verify

curl -sS http://127.0.0.1:8787/healthz            # {"ok":true,...}
journalctl --user -u codexcomp -f | grep -E 'round|done'   # Linux/WSL

A live fold — two consecutive 516s folded, answer correct:

round 1: in=21550 out=664 reason=516 total=22214 | n=1 buffered=['function_call'] -> continue
round 2: in=22078 out=652 reason=516 total=22730 | n=1 buffered=['function_call'] -> continue
round 3: in=22606 out=566 reason=291 total=23172 | n=None buffered=[...] -> clean
done: 3 round(s) | ... | status=completed stop=natural

Round verdicts: continue (cut detected → will continue), clean (natural end), tier_out_of_window / max_continue / no_encrypted_content (cut detected but released as-is), compaction_request (remote-compaction request — never folded), upstream_eof (stream ended without a terminal event). done: lines end with stop=natural or the release reason. A fold torn down by a client disconnect logs fold aborted downstream after N round(s) instead of a done: line.

Evals

evals/candy_eval.py measures the fix end-to-end: it runs a model × effort × proxy-on/off grid of codex exec calls on the candy pigeonhole puzzle from haowang02/codex-candy-eval (answer: 21, independently re-verified by brute force) and reports boundary-cut rate, reasoning tokens and accuracy per condition. Both modes wire openai_base_url explicitly, so the ambient config doesn't matter; interrupted grids resume by re-running the same command.

codexcomp &                                    # proxy must be running for `on`
python evals/candy_eval.py -m gpt-5.5 -r xhigh -n 5   # small grid
python evals/candy_eval.py                            # full 80-run grid

An 80-run grid of this eval (2026-07-06) found every unmitigated gpt-5.5 run cut exactly on a 518n−2 boundary, 15% vs 90% accuracy off/on — details in openai/codex#30364.

FAQ

Does it touch normal turns? No. Clean rounds pass through byte-for-byte; the fold path only engages on a detected 518n−2 truncation.

What does a fold cost? Continuation rounds spend extra real tokens, bounded by the continuation cap (--max-continue, default 3). The true cumulative usage is reported under metadata.proxy_billed_usage.

What happens when OpenAI fixes this upstream? Nothing breaks — the detector simply stops firing and the proxy becomes a transparent passthrough. Unwire it by deleting the openai_base_url line whenever you like.

Why not a separate [model_providers] entry? That changes the provider id, which re-buckets session history and drops remote compaction and remote-control. openai_base_url is the official in-place override of the built-in openai provider.

Is my credential safe? The proxy forwards the Authorization header untouched and binds to loopback only; it never reads, persists, or logs a credential.

Security & disclaimer

• Auth passthrough only — forwards Codex's Authorization header; never reads, persists, or logs a credential.
• Loopback only — do not expose it on a non-loopback interface.
• Unofficial — it relies on non-contract upstream behavior; an OpenAI-side change may break it. Use at your own risk.
• Continuation spends extra real tokens (metadata.proxy_billed_usage), bounded by the --max-continue cap (default 3 rounds).

Development

git clone https://github.com/dzshzx/codexcomp && cd codexcomp
uv sync
uv run python test_fold.py        # fold state-machine self-test → ALL PASS
uv run python test_ws.py          # transport self-test (WS protocol, headers) → ALL PASS
uv run codexcomp                  # run locally

Releases go out via PyPI Trusted Publishing (OIDC, no stored token): push a v* tag to build and publish. Version history: CHANGELOG.md.

Contributing

Bug reports, fold-log excerpts, and reproduction details are the most valuable contributions — please file them on GitHub Issues. For code changes, run uv run python test_fold.py and uv run python test_ws.py before opening a PR and keep changes focused.

Community

Built for and shared with the LINUX DO community, where the gpt-5.5 "516" degradation was diagnosed. Feedback and issues welcome there and on GitHub Issues.

License

MIT — mechanism credit to neteroster/CodexCont (MIT), whose 518n−2 detect-and-continue idea this reuses with an independent, from-scratch implementation.