2026-05-12-partial-cycle-claude-code-2026-05-07_2026-05-12-frontier-v0

The most significant change this window is a new surface, not a new feature. claude agents (v2.1.139, Research Preview) opens a full-screen list of every Claude Code session by state: working, waiting on you, done, or failed. Background sessions run under a persistent supervisor process -- they keep going when you close the terminal, isolate their file edits to separate git worktrees automatically, and show one-line summaries without requiring you to open the transcript. You can reply to a blocked session from a peek panel without attaching.

The design is deliberate: claude agents is a management surface, not just a session list. You can dispatch new sessions from the prompt, background an active session with one keystroke, filter by state or directory, and pin sessions that need attention. The combination of persistent background execution, automatic worktree isolation, and lightweight peek/reply closes the gap between "running five Claude sessions in separate windows" and having a proper multi-agent workspace.

Alongside it, the /goal command sets a completion condition that Claude works toward across turns until it is met. It is the first explicit goal-completion primitive in Claude Code. You can step away while the work continues, and the live overlay tracks elapsed time, turns, and tokens consumed.

The governance additions are quieter but load-bearing. settings.autoMode.hard_deny completes the auto-mode policy model: hard blocks that no allow rule can override. The new continueOnBlock option for PostToolUse hooks feeds the rejection reason back to Claude so it can adapt rather than just stop. Together they move auto mode from "classifier decides, full stop" to "classifier decides, explains, and tries again with the constraint in context."

One behavioral change to note: worktree.baseRef reverts to "fresh" (branch from origin/<default>) as the default. Operators who relied on new worktrees carrying unpushed local commits need to set worktree.baseRef: "head" explicitly.

What To Try

• Dispatch background sessions with claude --bg "<prompt>" and use claude agents to monitor them. Peek and reply from the list to handle blocked sessions without leaving the supervisor view.
• Set a /goal on a multi-step refactoring or investigation and check back after several minutes. Inspect the turn/token overlay to calibrate how complex your goals should be for a given context budget.
• Add continueOnBlock: true to a PostToolUse hook that currently rejects certain file writes. Include a reason that explains the constraint. Watch whether Claude routes around the rejected path or clarifies its intent.

What Remains Uncertain

• What artifact does /ultrareview produce, and how should a CI pipeline ingest or route the verdicts? The research preview returns findings to CLI/Desktop but the schema is not yet documented.
• Whether /goal state survives context compaction. The compaction prompt now asks the model to preserve sensitive user instructions; whether a goal counts as sensitive is not specified.
• How parentSettingsBehavior (the admin-tier SDK managedSettings merge key, v2.1.133) interacts with enterprise policy deployments at scale. The changelog documents the key but not the policy-merge semantics in detail.