Reduce unsolicited replies and repair the build dependency audit

[EugeneDraitsev]

The bot could answer unaddressed group questions and replies intended for other people, while remembered preferences could override its instruction to ignore praise. The reply-gate prompt now requires a clear request to the bot, preserves genuine follow-ups, corrections and aliases, and treats chat memory as background data. It also ignores clearly unauthorized cyber requests while allowing defensive security and account recovery. No runtime heuristics, additional model calls, or agent routing changes are introduced; registered commands such as /q and /o still bypass the gate.

CI was also blocked by GHSA-vfj7-8cjw-p6xm in the build dependency braces@3.0.3, which has no fixed npm release. A native Bun dependency patch caps parser nesting at 256 levels and treats deeper input as literal text, following the upstream fix proposal. Bun reapplies the patch on install. bun run audit first runs security regressions against the installed package, then excludes only this locally fixed advisory because registry metadata still reports version 3.0.3. All other advisories remain blocking. The patch and exception are documented for removal after an upstream release.

Validation:

• Security regressions fail without the patch and pass with it: deeply nested braces/parentheses have a bounded AST, compile/expand/stringify safely, and normal build glob expansion stays unchanged.
• bun run audit passed after the security checks.
• All 687 unit tests passed, including the four new security regressions.
• bun run tsc, bun run biome, and git diff --check passed.
• bun run build successfully packaged the service with CI placeholder credentials.
• Live Luna evaluation: 73/73 labeled cases passed, including eight selected unsolicited replies and 41 requested replies. One API timeout passed on retry.

The seven-day investigation covered 6,075 gate decisions, including 160 engage decisions. Incoming text was retained for 49 engage decisions, so this is not a full-week false-positive measurement. Live evaluations used current chat memory.

[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 2026-10-06T19:59:56.402988Z	d08539e	Manual request

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.

[EugeneDraitsev]

@codex review

Please review the updated head d08539e. The CI audit failure is addressed with a native Bun patch for braces nesting depth, installed-package security regressions that run before the narrowly scoped advisory exception, and documentation. Audit, all 687 tests, types, lint, and service packaging pass locally. The original reply-gate changes remain prompt-only.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. More of your lovely PRs please.

Reviewed commit: d08539ee91

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".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".