[lydakis]

lead.md goes from 20,344 to 16,356 bytes (−20%), and its budget in BUDGETS drops from 20 KiB to 16 KiB. That's the CLAUDE.md "under 200 lines" size we set as the goal in #440. This is the first hill-climb step with bench/role_eval.py.

What changed

Every paragraph says the same rules in fewer words. Nothing on the behaviour checklist was dropped: delegation, holding the work, briefs, the machine view, the same folder, tasks, Task updates, replies, triggers, the next look, care, the margin, memory, and pause. Command lines stay word for word. The care, memory and pause paragraphs are unchanged.

Five sentences went as repeats or rationale:

• "Start a new task only for separable work." The opening paragraph already says this.
• "The person makes a project whenever they open Sonny in a subfolder." The subfolder rule stays.
• "Later messages to a task need no --workspace."
• "Every later agent reads it, so it sees a pattern no single conversation shows." This was the rationale for saving memory; the rule stays.
• Why a one-off next look is wrong. "never at or in" stays.

#449's next-look changes are carried over in full: every or cron, the ready if command, the web.lead.next example, and "setting it again with replace".

Eval (scored blind on the owner's Mac, 10-30)

The arms were main at 759c0db and this lead.md at 2b2b525. 2b2b525 is this text before the rebase onto #449, which changed only the next-look paragraph (the eval doesn't exercise it), plus about 30 bytes of small cuts. There were 3 runs per arm per model.

model	main	16 KiB
anthropic/claude-opus-5-5	8.67 / 11	8.67 / 11
chatgpt/gpt-6.1-sol	7.33 / 11	7.33 / 11

The two asks split into tasks, the quick question answered without a task, the planted ticket, finance ask and freeze date all scored full marks in all 12 runs. Each arm also missed once where the other didn't (H2/H3, M, Noise), which is within noise. On sol, all three runs with the 16 KiB text finished the day's merges, while two main runs stalled on their own open questions.

Checks

• cargo test -p sonny passes, including shipped_roles_fit_their_budgets at the new 16 KiB.
• The FEATURES.md roles row is updated.

---

Generated by Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-30T07:07:18.974215Z	1365ab1	New commits
🔒 Security Review	✅ Completed 2026-10-30T07:08:26.260104Z	1365ab1	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.