feat(openai): GPT-6 models + per-model reasoning effort on the ChatGPT subscription

[ericleepi314]

Summary

• Subscription login tested live. The browser "Sign in with ChatGPT" flow ran end to end through clawcodex login in an isolated CLAWCODEX_CONFIG_DIR. Credentials were saved, the catalog was discovered, and a Read-tool agent turn succeeded on gpt-6-astra, gpt-6-sol and gpt-6-luna at --effort max.

• GPT-6 now reaches subscription users. The Codex catalog hides any model whose minimal_client_version is above our pinned client_version. gpt-6-astra needs 0.153.0 and gpt-6-sol/luna need 0.155.0, but we were pinned at 0.149.1. The pin is now 0.155.1.

• Subscription reasoning effort is clamped per model. Levels come from the login's cached catalog, with a static fallback. Live probes on 2026-09-24:

  • gpt-6 and gpt-5.6 accept up to max.
  • gpt-5.5 accepts up to xhigh.
  • ultra is advertised in the catalog but returns 400 on every model, so it is filtered out.
  • The old blanket rule sent high for every model, which capped GPT-6 two notches low.

• Public API: max passes through for GPT-6 and still degrades to xhigh for everything else.

• /model picker: the effort step uses the same per-model ladder and knows whether the session is on the subscription.

• Model registration:

  • gpt-6-sol and gpt-6-luna added to the OpenAI model list; gpt-6-astra was already there.
  • Context rows set to 872K, the smaller of the 922K API input limit and the 872K subscription limit. OpenAI's overflow error does not trigger reactive compaction, so an over-estimate is a hard failure.
  • List pricing added, with the 272K long-context tier.
  • Verbosity support added.
  • Harness effort docs updated.

• GPT-5.6 context window fixed too. All five gpt-5.6 rows drop from 1.05M to 872K for the same reason: the API caps input at 922K and the subscription at 872K.

Follow-ups (not in this PR)

• OpenRouter has no openai/gpt-6-* rows.
• The advisor_cost context-tier note is out of date.
• Tests could be isolated from the real config dir suite-wide.

Test plan

• Targeted suites pass: providers, picker, openai routing/subscription, pricing, effort, compact.
• Full suite: 1 failure, test_system_prompt_full::...damped_wording.... It fails the same way on origin/main.
• Live against the ChatGPT backend: catalog at several client_versions, an effort ladder per model, and -p agent turns with a tool call on all three GPT-6 models.
• A fresh browser login saves gpt-6-astra as the default model, and a flag-less session runs on it.
• Critic review: REQUEST_CHANGES on the context window, fixed, then APPROVE.

🤖 Generated with Claude Code

[github-actions]

Test Results

     5 files   1 021 suites   19m 33s ⏱️
16 054 tests 16 032 ✅ 22 💤 0 ❌
32 078 runs  32 007 ✅ 71 💤 0 ❌

Results for commit 4bf9762.

♻️ This comment has been updated with latest results.