fix(responses): count input tokens locally for Codex OAuth - #2
Merged
anhtran-ai merged 2 commits intoSep 10, 2026
Merged
Conversation
…t rejection, not a ban
A Cloudflare managed/JS challenge served in front of an upstream provider is the
same class of block as a Cloudflare 1010 — the edge refused the CLIENT's
signature — but it is a different product surface and carries none of the 1010
markers isCloudflareFingerprintRejection() looks for.
It therefore fell through the entire 403 ladder in classifyProviderError() to the
terminal default FORBIDDEN, which chatCore persists via writeTerminalStatus() as
testStatus=banned / isActive=false. That state never auto-recovers, so a single
challenge takes the whole provider offline until an operator reconnects in the
dashboard.
Observed on POST chatgpt.com/backend-api/codex/responses/input_tokens for a
healthy Codex OAuth account: the response carried cf-mitigated: challenge,
server: cloudflare and a ~12KB text/html interstitial with
window._cf_chl_opt = {... cType: 'managed', cZone: 'chatgpt.com' ...}. The same
connection refreshed its OAuth token successfully in the same second and served
normal /responses traffic seconds before and after, so the account was never
banned upstream.
Classify the interstitial as FINGERPRINT_REJECTION, reusing the existing
non-terminal precedent from diegosouzapw#9929: authTerminalStatus already treats that type as
non-terminal, so the request falls through to the next combo target and the
account state stays untouched.
The markers are matched as full, distinctive Cloudflare-internal strings
(_cf_chl_opt, cdn-cgi/challenge-platform, the challenge-error-text span id
including its escaped-quote nested form) and never as the loose word
"challenge", so provider bodies discussing a challenge in prose are unaffected.
Tests cover the full interstitial, the gateway-nested error.message form, each
marker individually, prose false-positive guards, and regression guards proving
a genuine permission 403 and the ChatGPT Web Sentinel/Turnstile 403 (diegosouzapw#8813) both
still classify as FORBIDDEN.
The ChatGPT subscription backend does not serve /backend-api/codex/responses/input_tokens for the affected account. Native requests to that path are intercepted by an OpenAI Cloudflare managed challenge, while the same path over the bundled Chrome transport returns 404 Not Found. Forwarding the client preflight can therefore never return a useful count and, before the companion classifier fix, permanently disabled the healthy Codex connection on the first challenge. Add a static /v1/responses/input_tokens route that shadows the generic Responses passthrough, uses the existing offline o200k_base token counter, and returns the standard response.input_tokens contract without issuing any upstream request. Count instructions, structured input, tool definitions and config; apply a conservative five-percent margin so the failure mode is earlier client compaction rather than a context-window overflow. Preserve the API-key and model-policy boundary from the catch-all Responses path. Tests pin the public schema, prove fetch is never called, cover text, instructions, structured input, tools, non-text parts, server-held context ids, invalid JSON, OPTIONS, and the conservative lower bound.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
POST /v1/responses/input_tokenswas handled by the generic Responses passthrough and forwarded tochatgpt.com/backend-api/codex/responses/input_tokens. For the Codex OAuth account in production, that upstream subpath is not served (404through the Chrome transport) and native transport is intercepted by a Cloudflare managed challenge. The forwarding path can never return a useful count, adds latency/fallback, and previously terminalized the healthy connection.Fix
Add a static
/v1/responses/input_tokensroute that shadows[...path]and counts locally with the existingjs-tiktokencounter (o200k_baseforcx/Codex). It returns the exact OpenAI contract:{"object":"response.input_tokens","input_tokens":123}No upstream request is issued. The route preserves API-key/model policy and process-wide body admission, counts instructions, structured input, tool definitions and Responses config, and applies a conservative overhead model measured against live Codex usage.
The companion classifier safety net is already merged in fork PR #1 and proposed upstream in diegosouzapw#13161.
Validation
Local:
npm run typecheck:core: clean.Production (
llm-gateway/omniroute:3.8.51-113ddf9a-azox1):usage.input_tokens: 8/8 valid, 0 under-counts, median absolute error 0%; five cases exact, structured input +12.2%, one tool +1.5%, two tools +30.7% (safe direction: early compaction).is_active=1,test_status=active, no error fields.Production rollback artifacts:
llm-gateway/omniroute:3.8.50-5458026c.llm-gateway/omniroute:3.8.51-55155b92-azox1.omniroute-data-pre-55155b92-20260910-105915.tgz(mode 0600).ok, 2 provider connections.