[BUG] Regression: claude.ai/code web sessions back to random <adjective>-<surname>-<hash> branch names instead of task-derived names

What's Wrong?

For some period before late May 2026, branches created for claude.ai/code web sessions were semantically named — a short slug derived from the initial prompt plus a random suffix, e.g. claude/add-stripe-webhook-EhDOB. Over the last few days the naming has reverted to Docker-style claude/<adjective>-<surname>-<hash>:

• claude/relaxed-brown-EhDOB
• claude/lucid-hamilton-MigdG
• claude/peaceful-noether-2MaZ1

These names carry no information about what the session worked on, which hurts searchability and history navigation across both git log / git branch --list and the GitHub PR list. The random suffix alone was already sufficient for uniqueness; the value of the previous behavior was the prefix.

This is distinct from existing reports:

• #54653 asks for mid-session renaming of local worktree branches and doesn't reference the web product.
• #50752 / #58095 are about word-list cosmetics (avoid slurs / avoid surnames), not whether names should be task-derived at all.
• #33744 / #25902 / #39009 (all closed) requested configurability but predate the period when semantic naming was actually shipping in the web product.

What Should Happen?

Restore the task-derived prefix for claude.ai/code web sessions: generate the branch name from the initial prompt (same Haiku-based summarizer that powers session auto-naming per #58870 / the ANTHROPIC_DEFAULT_HAIKU_MODEL background path) and keep the existing random suffix for uniqueness. Concretely:

• claude/<slug-from-prompt>-<5-char-hash> (current suffix length preserved)
• Fall back to the current <adjective>-<surname>-<hash> only when the slug can't be generated (no Haiku model configured / summarizer failure).

Is this a regression?

Yes.

Platform

claude.ai/code (web sessions)