Ping about workspaces only when they need you, with real names

[elitecoder]

Summary

Follow-up to #35. A review of the channel's history (2,500 posts in 14 days) showed that after #35, workspace pings were the remaining noise. This makes them rare and readable.

• Idle alerts ping only when the workspace needs you. cmux sends every Claude alert as one bare Notification event with no message (verified in a live capture). Most are Claude's idle alert about a minute after every turn. A ping now goes out only when a permission or plan prompt is up (a tool call with no result yet), the session stopped on an API error, or the agent's last few lines ask you something. By that rule, 25% of 1,496 recent turns would ping.
• Patterns match the agent's own words. Turn-end patterns ("CI red", "PR opened") now run on the agent's message from the current turn, not 50 lines of screen. The screen is only the fallback when the transcript can't be read.
• Three noisy word lists are muted. stranded, awaiting-review, and emit-card sent 168 pings in two weeks, mostly on status prose. A muted pattern no longer hides a real one in the same message.
• Workspaces get real names. When cmux's title is the default "Terminal", the ping uses Claude's own session title, else the folder and branch.

Validation

• Full suite: 2,950 passed, 1 skipped. The changed-code coverage gate passes. Mutation probes on the prompt check, the API-error check, the current-turn scoping, and screen-versus-message matching all fail a test.
• Replayed a real captured Notification through the new code: it named the session "assembler-aspect-ratio-reframe test session" and pinged, because the agent's last line asked a question.
• Two independent reviews (missed signals; regressions and coverage).

Known limits

• Patterns that only ever matched terminal output, like committed ([main abc1234]), no longer fire.
• lesson-extractor.py still proposes new patterns from screen evidence, which the watcher now checks against messages instead.

🤖 Generated with Claude Code