URL: https://mkdocs.justinsforge.com/memory/general/reference_gpt_sol_in_claude_code/

GPT-5.6 Sol inside Claude Code¶

Tested end-to-end 2026-07-13. Both paths run on the ChatGPT subscription, no OpenAI API key (~/.codex/auth.json has auth_mode set and OPENAI_API_KEY: null).

Models that actually exist¶

gpt-5.6-sol, gpt-5.6-luna, gpt-5.6-terra, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark. Confirmed from ~/.codex/models_cache.json and the proxy's /v1/models. ~/.codex/config.toml already defaults to gpt-5.6-sol at high effort.

Option 1: openai/codex-plugin-cc (INSTALLED, USE THIS)¶

Plugin codex@openai-codex v1.0.6. Shells out to the local codex CLI, reuses its existing ChatGPT OAuth. Seven commands:

Command	Purpose
`/codex:review`	Read-only review; `--base <ref>` for branch review. Not steerable.
`/codex:adversarial-review`	Steerable challenge review. Pressure-tests design and tradeoffs. Best pre-`/ship` use.
`/codex:rescue <task>`	Real delegation, not just review. Takes `--model` / `--effort`.
`/codex:transfer`	Hand the current session to Codex.
`/codex:status` / `/codex:result` / `/codex:cancel`	Background job management.

Plus a codex-rescue subagent and a stop-review-gate hook. All long commands take --background. Resume any job natively with codex resume <session-id>.

GOTCHA: the plugin sweeps untracked files into the review payload and dies at Input exceeds the maximum length of 1048576 characters. This is how the 337MB of sites/sipnservesociety.com/node_modules blew it up. Fixed by gitignoring that nested repo. If a review 1MB-fails again, check untracked bytes first, not the tracked diff.

Option 2: CLIProxyAPI, Sol as MAIN model (WORKS, BUT TERMINAL-ONLY)¶

CLIProxyAPI v7.2.71 at ~/.local/opt/cliproxyapi, config.yaml binds 127.0.0.1:8317, isolated auth-dir at ~/.cli-proxy-api (does NOT read ~/.codex/auth.json, separate OAuth grant). Login is ./cli-proxy-api -config config.yaml -codex-device-login -no-browser; it buffers output, so run it detached to a log file or you will see nothing.

Claude Code with ANTHROPIC_BASE_URL=http://127.0.0.1:8317 + ANTHROPIC_AUTH_TOKEN=<local key> + --model gpt-5.6-sol works. Verified reply: "CLAUDEX OK. I am GPT-5.6 Sol."

THE HARD BLOCK (the reason not to bother)¶

Remote Control is only available when using Claude via api.anthropic.com.

Setting a custom ANTHROPIC_BASE_URL disables Remote Control and claude.ai connectors. So a proxied Sol session can NEVER bridge to claude.ai/code. That kills, for Sol: /spawn, the warm pool, /rebridge, the remote-bridge Telegram bot, every mobile surface.

Consequence: claudex is a local-terminal-only toy. Given Justin's mobile-first workflow, its value is low, and it costs a persistent daemon holding an OpenAI OAuth refresh token for a project whose README sponsors are ChatGPT-account resellers (PackyCode, AICodeMirror, BmoPlus). Recommendation: use the plugin, skip the proxy.

Status bar showing "API Usage Billing" under the proxy is cosmetic; Claude Code says that for any custom base URL. Real usage bills the ChatGPT subscription.

Teardown (if the proxy is ever installed again)¶

pkill -f 'cli-proxy-api -config'
rm -rf ~/.cli-proxy-api ~/.local/opt/cliproxyapi
# then revoke the grant at chatgpt.com account settings

[Claude Code]