codex-fast-proxy

name	codex-fast-proxy
description	Codex App Fast proxy and auth-split for third-party OpenAI-compatible APIs. Supports Sign in with ChatGPT, priority service_tier, Responses API benchmark, enable/check/update/uninstall. Use when this capability is needed.
metadata	{"author":"gaoguobin"}

Use this skill when the user wants Codex to manage the local auth-split and Fast proxy for Codex App.

Trigger patterns

• Enable requests such as "enable Codex Fast proxy".
• App Fast requests such as "make Codex App use Fast".
• Provider-specific requests such as "enable Fast for PackyAPI".
• ChatGPT login compatibility requests such as "plugins work but model requests return 401".
• Benchmark requests such as "run the Fast proxy benchmark" or "check whether my provider supports Fast".
• Upstream URL changes such as "set the Codex Fast proxy upstream to https://api.example.com/v1".
• Maintenance requests such as "show status", "check updates", "stop", or "uninstall".

How to execute

Run the manager as the source of truth:

python -m codex_fast_proxy doctor
python -m codex_fast_proxy install --start
python -m codex_fast_proxy install --start --use-provider-auth-file
python -m codex_fast_proxy prepare-chatgpt-login
python -m codex_fast_proxy prepare-chatgpt-login --apply
python -m codex_fast_proxy verify-upstream --upstream-base https://api.example.com/v1
python -m codex_fast_proxy set-upstream --upstream-base https://api.example.com/v1
python -m codex_fast_proxy set-upstream --use-provider-auth-file
python -m codex_fast_proxy set-upstream --clear-upstream-auth
python -m codex_fast_proxy set-upstream --service-tier-policy auto
python -m codex_fast_proxy set-upstream --service-tier-policy inject_missing
python -m codex_fast_proxy status
python -m codex_fast_proxy check-update
python -m codex_fast_proxy benchmark
python -m codex_fast_proxy autostart --quiet
python -m codex_fast_proxy stop --force
python -m codex_fast_proxy uninstall --defer-stop
python -m codex_fast_proxy uninstall

Safety model

• Installing the repo or skill must not change Codex provider config.
• Enable with install --start; it starts the local proxy before switching Codex config.
• Enable also installs one user-level Codex SessionStart hook in ~/.codex/hooks.json and enables the Codex hooks feature flag. Newer Codex builds use features.hooks = true; older docs/builds may refer to features.codex_hooks. During the CLI/App transition, write both keys and treat either key as enabled. The hook starts a missing proxy on future Codex sessions only when the recorded provider still points to the local proxy. It must not restart an already healthy proxy just because runtime code is stale. Current Codex builds may also require a trusted hook state entry, so treat startup_hook: true as installed, enabled, and trusted; if startup_hook_trust reports modified or untrusted, rerun enable/update instead of relying on ~/.codex/hooks.json alone.
• After an enabled update, install --start compares the running proxy runtime with the installed code and restarts stale proxy runtime before returning when config still points to the local proxy. Codex may fire SessionStart for each new or resumed session; autostart --quiet does not log normal no-op checks and does not refresh stale runtime implicitly.
• Do not run plain install to enable the proxy; the manager rejects config switching without --start.
• Default service tier policy is auto: ChatGPT-login or unclear states preserve Codex App/CLI Fast choices, while API-key mode can inject priority when Codex omits service_tier because the App Fast UI may not be available. Use --service-tier-policy inject_missing only when the user explicitly asks for global Fast injection and accepts that Codex App's Fast UI toggle will no longer control missing tiers. Use --service-tier-policy preserve only when the user explicitly wants no proxy-side Fast injection.
• Before first enable or model-path setting changes, install --start verifies the candidate upstream and auth source with one side-path Codex-style POST /v1/responses request using stream=true. If verification fails, do not pass --no-verify unless the user explicitly accepts that future Codex model requests may fail.
• Existing enabled installs that do not yet have service_tier_policy in settings are legacy global-Fast installs when they do not also have split upstream auth; preserve that behavior as unless the user explicitly asks for App-controlled Fast. Missing policy plus split upstream auth belongs to the ChatGPT-login auth split path and should be treated as App-controlled .

Sandbox and approval discipline

• Operations that clone from GitHub, install with pip, create ~/.agents skill links, write ~/.codex/config.toml, write ~/.codex/hooks.json, start a background proxy, or remove installed files may need user approval or elevated sandbox permissions.
• If the harness supports escalation, request approval for the intended command instead of trying alternate paths.
• If a command fails because of network, permissions, sandbox write limits, skill link creation, or background-process restrictions, stop and rerun the same intended action with approval. Do not invent workarounds that bypass the user's sandbox policy.
• Do not print API keys, request bodies, prompts, or Codex history. Do not edit auth.json unless the user explicitly asks for a recovery action; prefer copying the current working provider key to the proxy-managed provider auth file with prepare-chatgpt-login --apply.

User handoff messages

• After .codex/INSTALL.md or .codex/UPDATE.md changes skill files, explicitly tell the user to restart Codex App and return to the conversation, or open a new CLI process, so Codex can rescan ~/.agents/skills; then ask Codex to enable Codex Fast proxy.
• After .codex/UNINSTALL.md, explicitly tell the user to restart Codex App, or open a new CLI process, so Codex removes codex-fast-proxy from the skill list.
• After a successful install --start, explicitly tell the user that Fast proxy is enabled, but the current Codex process will not hot-switch; they should restart Codex App and return to the conversation, or open a new CLI process.
• After a successful install --start, always append this optional ChatGPT-login UI note even if the status summary is already long: the user may keep API-key mode for third-party API plus global Fast. If they want richer Codex App UI such as plugin marketplace, GitHub/Apps/connectors, manual Fast controls, status hints, and voice input, they should run prepare-chatgpt-login before switching Codex App to ChatGPT login; switching directly may cause 401.
• After uninstall --defer-stop returns status="confirmation_required", explicitly tell the user no uninstall changes were applied because ChatGPT login appears active and direct upstream may return 401. Do not tell the user to restart yet.
• After uninstall --defer-stop returns status="uninstalled", explicitly tell the user that Codex config has been restored to direct upstream, and the proxy was left running temporarily to avoid interrupting the current process. They should restart Codex App and return to the conversation, or open a new CLI process, then run uninstall again to finish cleanup.
• If direct_upstream_auth_warning is present after a confirmed uninstall, first warn that ChatGPT login appears active. After direct upstream restore, requests no longer pass through the proxy upstream auth override. Keeping ChatGPT login may send ChatGPT auth to the third-party provider and return 401. The user should switch back to API-key/third-party provider auth before restarting, or keep the proxy enabled for ChatGPT-login UI with a third-party provider.

Use --provider <name> only when the user names a provider or when doctor reports that no active provider can be selected.

Use --upstream-base <url> only when Codex config does not contain a usable provider base_url or the user explicitly wants a different upstream. Use --upstream-api-key-env <ENV_NAME> only as an advanced compatibility path with an environment variable name, never a literal key value. Use --clear-upstream-auth when the user wants to stop overriding upstream Authorization and return to Codex's original provider auth behavior.

For upstream URL changes after enable, prefer set-upstream --upstream-base <url> over rerunning install --start --upstream-base <url>.

Result handling

• Treat the JSON output as the source of truth.
• Report provider, base_url, upstream_base, service_tier_policy, upstream_auth, running, diagnosis, provider_auth_preparation, chatgpt_login_hint, next_user_action, chatgpt_login_windows_troubleshooting when present, runtime_matches, needs_restart, and backup or restore status.
• Use status.runtime when diagnosing stale runtime, wrong Python executable, or a startup hook that points at a different checkout. Do not infer hook readiness from ~/.codex/hooks.json alone.
• For App-specific traffic checks, use recent /v1/responses events as model-generation evidence. Treat GET /v1/models as provider metadata checks; do not report isolated /v1/models failures as model-generation failure unless /v1/responses also fails.
• Do not print API keys, auth.json, request bodies, prompts, or Codex history.
• For benchmark results, report profile, medians, observed speedup, priority_accepted, observed_priority_effective, provider-confirmed priority metadata when present, sample counts, service_tier_control.valid, and errors. Prioritize full-response total latency and first-output latency over first-event/TTFB. Treat priority_accepted=true as proof that the wire parameter is accepted, and observed_priority_effective=true as proof that this measured workload benefited. Report benchmark_mode and do not present Codex CLI/app-server benchmark results as an App-specific guarantee. For App-specific verification, use recent dashboard/proxy traffic after the user sends an App message. priority_accepted=true means at least one priority sample succeeded; always report the displayed ok/count sample counts with it. Do not claim a guaranteed speedup from a single run.
• If install or update changed the skill files, tell the user to restart Codex.

Expected behavior

• install --start backs up ~/.codex/config.toml.
• The selected provider's original base_url becomes upstream_base.
• The selected provider's base_url becomes http://127.0.0.1:8787/v1.
• Before switching config on first enable, install --start verifies a streaming /v1/responses request against the candidate upstream/auth route unless the user explicitly accepted --no-verify.
• verify-upstream reports the same candidate route validation without changing persistent state.
• set-upstream updates the saved upstream_base, service tier policy, upstream auth source, and uninstall recovery baseline without changing model, reasoning, tools, input, or literal API key values. Before writing settings, it sends one side-path Codex-style POST /v1/responses request with stream=true to the candidate upstream/auth source. This is real provider traffic; if it fails, do not add --no-verify unless the user explicitly accepts that future Codex requests may break. It applies immediately only when the proxy is not running or the user explicitly accepted --restart; otherwise it defers restarting a running proxy to avoid cutting off the current response.
• A SessionStart hook calls the current Python executable with -m codex_fast_proxy autostart --quiet on future Codex sessions.
• By default auto preserves Codex App/CLI service_tier choices in ChatGPT-login or unclear states, and can inject service_tier="priority" in API-key mode when that field is absent. Only explicit inject_missing forces global Fast regardless of login mode.
• Optional upstream auth split applies to proxied provider API requests, not to ChatGPT plugin, GitHub, Apps, connector, cookie, or token traffic; override mode replaces Authorization and drops unexpected Cookie headers before forwarding upstream.
• benchmark compares synthetic Codex-style requests with no service_tier against service_tier="priority". The default codex-cli mode is intended to measure real Codex acceleration; --profile smoke is only for low-cost connectivity checks, and --mode direct is a less representative fallback when Codex CLI is unavailable. It stores only redacted metrics in . The local dashboard shows the latest saved benchmark summary and never starts benchmark runs.

---

Source: gaoguobin/codex-fast-proxy — distributed by TomeVault.