Retarget delegation models from measured coding-agent cost/task

[kelchm]

Summary

• retarget the model ranking and delegation pins to Grok 4.6, Fable 5.1, GPT-6 Astra, and Opus 5
• rank cost and intelligence from measured coding-agent cost/task, not list price or the old Grok deal: DeepSWE and the Artificial Analysis Coding Agent Index
• keep taste as the remaining judgment axis
• add a start effort column instead of a model×effort matrix: scores assume that operating point
• escalate by raising one effort step on the same model, then switching models
• pin those start efforts on the CLI invocations the skills actually run (grok --effort medium, Codex model_reasoning_effort=xhigh, claude --effort high)
• preserve the live @RTK.md include in CLAUDE.md so chezmoi apply does not drop it

Why these start efforts

• grok-4.6 / medium — DeepSWE peak (67.5% at $3.45). xhigh costs more and scores worse.
• gpt-6-astra / xhigh — DeepSWE peak (74.1% at $6.52). max is ~2x the cost with no gain.
• fable-5.1 / opus-5 / high — Claude's product default. Fable 5.1 max is the expensive last step (~1.7x Fable 5 tokens on the Intelligence Index).
• gpt-5.6-luna / max — the exception: lower efforts collapse on DeepSWE (high 44%, medium 11%).

Sonnet-5 stays in the table with honest scores (cost 2, intelligence 4). Measured DeepSWE cost/task at max is $26.40 for 54%; it is not a cheap coding agent.

Out of scope

• interactive Codex default in ~/.codex/config.toml is still gpt-5.6-sol / high (not chezmoi-managed)
• historical grok-4.5 canary in docs/cross-harness-delegation.md

Validation

• git diff --check passed
• chezmoi applied the rendered CLAUDE.md, Codex AGENTS.md, and affected skills locally
• live six-cell matrix not re-run in this PR (model/effort pins only; guards unchanged)

Review notes

The start column is the bet that effort is a cheaper lever than switching models, without turning the ranking into a 2D grid. If a score should move (especially Astra taste 6, Sonnet remaining in the table, or Luna starting at max), that is the discussion this draft is for.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}