[lydakis]

Requested by George · project thread

Role-file text only: no code, no tool, no daemon change. Each edit sharpens a principle that already exists instead of adding cases.

What changes

• Care paragraph (lead.md and task.md, so every bot Sonny starts): "tell them once, plainly and kindly" becomes "tell them as soon as you see it, plainly and without softening it, with what to do instead". It now covers what the person missed or said they would do. A repeat is the news: name the pattern and coach the habit. Bots save how the person works, and which nudges they found worth hearing, to the person's memory.
• Lead principle (lead.md): what the sources it hears report is weighed like task news, and each thing is weighed against the rest of the person's work. Things that look unrelated can answer, block or improve each other.
• Quiet default (lead.md): a lead also speaks when it has something the person would want and cannot see from where they sit.
• superlead.md: the outer lead names every source in its subscription.
  • Fixes its example, which passed "updates": true, a field the subscriptions tool refuses since Sonny: the lookout forwards everything and judges nothing #310.
  • Fixes "add it again with the new name list", which the tool refuses for a name in use. It now says to remove the subscription first.
• sonny/README.md describes both changes.

Margin nudges are deliberately not here: the margin text belongs to the conversation-noise work.

Grounding

Behavior	Existing pattern
Proactive cross-source nudges with a feedback loop	ChatGPT Pulse synthesizes memory, chat history and connected mail and calendar, and "remembers your feedback for next time" (OpenAI)
Straight feedback, no sycophancy, no lying by omission	OpenAI Model Spec, "firm sounding board" (spec)
Coaching a habit across conversations	No shipped assistant found. This part is our own design.

Eval

A scripted day: a lead driven headless through sonny acp, with fake chat and mail sources, real tasks and four messages from the person. It plants four links between the news and the work, and three unread merges. Model chatgpt/gpt-6.1-sol on the ChatGPT login. A = main, B = this branch, 3 runs each. Every step settled in every run. Scored blind, before reading the key.

Arm	Links raised (P1-P4)	Habit named (H1-H3)	Habit saved to memory	Noise lines	Softened lines	Input tokens, mean
main	12/12	0/9	0/3	0	0	2.10M
this PR	12/12	2/9	0/3	1	0	1.91M

• Links don't separate the arms: today's text already catches all four.
• Only this branch named the habit. One run, at the third unread merge: "You've now skipped review on three changes. Delegating the diff review is fine; skipping the behavior decision isn't."
• Both arms still sometimes reassured the person ("you don't need to read the diff") and covered for the habit with their own review.
• No run saved anything about how the person works.
• n=3 is a direction, not a measurement.

The harness is kept outside the repo.

---

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 2026-10-30T00:57:06.088771Z	2536cfb	PR opened
🔒 Security Review	✅ Completed 2026-10-30T00:59:02.833397Z	2536cfb	PR opened

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.