feat: add GPT-6 Astra to Codex Frontier dispatch

[tkstang]

feat: add GPT-6 Astra to Codex Frontier dispatch

Summary

New Codex Frontier recommendations now order GPT-6 Sol xhigh, GPT-6 Astra
high, and GPT-6 Astra xhigh. GPT-6 Sol max remains supported for existing
configurations, and adopting the updated recommendation preserves populated
user-owned cells. Public benchmarks show task-dependent Astra gains and
tradeoffs, while matched OAT workload evidence has not qualified a routine
dispatch change. Astra's Frontier placement is user-directed.

Scope

• Admit GPT-6 Astra at low, medium, high, xhigh, and max for exact
  phase-implementer and reviewer roles. ultra remains outside the supported
  contract.
• Regenerate the ten Codex role variants, synchronize bundled assets, update
  guidance and documentation, and bump the five public packages to 0.3.3.
• Compare all twelve provider/tier routes with the dated accepted model
  guidance and newer review-pending notes. Claude and Cursor ladder changes
  remain separate follow-up work.

This quick-mode project uses discovery and plan artifacts; separate spec and
design artifacts were not required.

Verification

• Fresh uncached workspace run: 10/10 Turbo tasks and 7,532 CLI tests passed.
• Focused provider-sync regression suite: 242 tests passed.
• Repository check, type-check, test, build, skill-bump, release-version,
  release-validation, docs-build, lint, and format gates passed.
• Project sync dry-run planned zero changes.

Reviews

The OAT phase and manual final reviews found no Critical or High issues. Their
Medium documentation and lifecycle-artifact findings were fixed. The
configured independent cross-runtime gate passed with zero Critical, High, or
Medium findings; its two Low wording findings were addressed without a repeat
gate. Review artifacts are archived locally and indexed in the project plan.

References

• Project record
• Official GPT-6 Astra model details

[cursor]

PR Summary

Medium Risk
Frontier dispatch and new Codex agent roles change which model runs phase work (cost and behavior), but routine routing defaults stay on Sol/Luna and the change is configuration and documentation rather than auth or data handling.

Overview
Adds GPT-6 Astra to the supported Codex catalogue and reshapes the bundled Frontier recommendation to Sol xhigh → Astra high → Astra xhigh, dropping Sol max from that ladder while keeping Sol max available for explicit configs.

New .codex/agents phase-implementer roles cover Astra at low through max effort (same OAT implement/fix/recover contract as existing Sol/Luna workers). Plan-writing, subagent-orchestration, and dispatch-matrix-recommendation guidance are aligned: Astra is documented as a user-directed Frontier option, with Sol/Luna task-class defaults unchanged until matched OAT workload evidence says otherwise. Skill metadata and oatVersion / public packages bump to 0.3.3.

^{Reviewed by Cursor Bugbot for commit cfb26ba. Bugbot is set up for automated code reviews on this repo. Configure here.}