Agent fails to follow established patterns and verify outbound action targets

Resolved 💬 3 comments Opened Mar 12, 2026 by rosscopaddison Closed Mar 16, 2026

Summary

During a client delivery workflow, Claude Code made two compounding errors that resulted in a broken deliverable being sent to a client, and a follow-up apology email being sent to the wrong recipients.

Environment

  • Claude Code CLI (Opus)
  • macOS
  • Using custom agent brains (shipper, PA) with detailed CLAUDE.md instructions

What happened

1. Packaging regression — Shipper brain ignored established pattern

A "shipper" agent was used to package v2.2.0 of a client project. Previous versions (v2.1.1, v2.1.2) had established a clear pattern of including full project contents (src/ + dist/ + configs — 53-55 files). The SESH.md handoff document explicitly stated "Created v2.1.1.zip (full project: src + dist + configs)".

The shipper agent regressed to dist-only packaging (9 files), despite:

  • The previous two versions establishing the full-project pattern
  • The handoff document explicitly describing what was included
  • The shipper brain's own packaging rules listing source code as includable

The broken zip was shipped to the client.

2. Reply sent to wrong recipient — PA brain didn't verify reply target

When using a reply_to_message() function to send a follow-up apology, the agent replied to an outbound email (sent BY the user TO the client). The reply function resolved the recipient as the original sender (the user themselves) rather than the original recipients (the clients).

The agent did not:

  • Consider that replying to your own sent email replies to yourself
  • Verify the actual recipient of the reply before sending
  • Flag that the resolved recipient didn't match the intended targets

Expected behavior

  1. When a clear packaging pattern is established across multiple versions and documented in handoff notes, subsequent versions should follow that pattern unless explicitly told otherwise.
  1. Before executing any outbound communication, the agent should verify the resolved recipient matches the intended target — especially when replying to sent mail where the "From" is the user themselves.

Impact

Client received a broken deliverable. The follow-up apology never reached them. User had to manually intervene and send the correction, resulting in a poor client experience.

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗