[lydakis]

Requested by George · project thread

This is the first half of the memory-first prompt audit (write-up). George's principle: what agents can learn into memory beats behavior written into role files and skills. He approved all three recommendations. This PR changes prompt text only and ships in 0.7.0.

What changes

• Memory overrides the role files' defaults. lead.md and task.md now end with one paragraph saying so. What the person's memory says about how they want work done overrides the defaults in the file, and in a task's role. The person's latest words override both. The defaults stay for someone whose memory is empty. Claude Code works the same way: memory files override default behavior.

• Save what worked, not only corrections. lead.md had five separate "save X to memory" sentences:

  • decisions
  • traps
  • work under way elsewhere
  • how far to go unasked (Sonny: a lead starts what the person said they will do when nothing owns it #399)
  • how the person works (Sonny: leads and tasks give straight feedback, coach repeated habits and weigh what sources report #394)

  They are folded into that one paragraph, which also asks agents to save what the person welcomed as well as what they corrected. The Sonny: leads and tasks give straight feedback, coach repeated habits and weigh what sources report #394 eval showed that save prose scattered through the turn never fired. The nightly reflection (Sonny: the nightly cleanup reflects on habits, from what the person wrote and sent #403) is the reliable writer.

• Memory skill and memory tool now describe feedback as both what the person corrected and what they confirmed worked. Memory that holds only corrections steers agents away from what already works. Guardrail: save the approach, never that the person liked being agreed with. That keeps it from wearing down Sonny: leads and tasks give straight feedback, coach repeated habits and weigh what sources report #394's straight-feedback rule. OpenAI's 2025 sycophancy rollback traced back to thumbs-up data and memory.

• Pinned picks left to the model and memory, as Sonny lead: the lead picks the emoji for a pickup in a task #340 did for lead.md's emoji:

  • The literal reply label "Needs you:" becomes "lead a reply with what needs the person, pointing to the ask or margin note that holds each". That fits Sonny: a task's question reaches its lead, and a decision stands out #395, which moved decisions to asks and open items to margin notes.
  • "at most three at a time" open decisions becomes "few, the one that matters most first".
  • The react tool's emoji meanings and "React rarely" are replaced by what a reaction is for.
  • The margin tool's "draw it in Mermaid" becomes "such as in Mermaid".

• Stale line dropped: "Triggers need Sonny on a Mac". Linux has had triggers since Sonny: triggers on Linux, kept by systemd's user manager #344, and the triggers tool refuses with triggers_unsupported where there is no scheduler.

• README, docs/SONNY.md and FEATURES.md rows are updated.

Not in this PR

PR B comes after #403 merges and the coach thread finishes its two-day eval of #403, so each eval measures one change:

• Reactions reach the nightly said, each with the reply it answered. Today lookout::said_by_person drops them.
• The habit step also saves approaches the person confirmed, and patterns of ask picks that go against the recommendation.
• Each feedback fact gets confirm and push-back counts, shown in the index line (ExpeL / ACE style).

Checks

• cargo test -p sonny on main 1ef4fb7: 516 passed. a_task_is_told_who_it_works_for now expects task.md's new last sentence.
• cargo fmt --check and cargo clippy -p sonny --all-targets -D warnings pass.
• No daemon, protocol or schema change.

🤖 Generated with Claude Code

https://claude.ai/code/session_013jkdGGVwTdQTtcHeKUqRa6

[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 2026-10-30T04:22:44.485213Z	9ab64d4	New commits
🔒 Security Review	✅ Completed 2026-10-30T04:26:15.348662Z	9ab64d4	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.