Claude Code v2.1.293: Haiku 5.5 Default, 1M Context, $0.10/Mtok

Anthropic released Claude Code v2.1.293, which swaps in Claude Haiku 5.5 (claude-haiku-5-5) as the default Haiku model on the Anthropic API. The new model ships with a 1M context window and pricing of $0.10/$0.50 per Mtok for standard prompts, rising to $0.50/$2.50 per Mtok once a prompt crosses 100K tokens.

New APIs and payload changes

• agentType is now included in the subagentStatusLine payload, so scripts can distinguish custom subagent types instead of getting a generic label.
• $.tool.register for mods accepts an isDeferred flag. Setting isDeferred: false lists the tool's schema in the prompt from the start rather than hiding it behind tool search.
• claude plugin test now supports mock.session so tests for mods that call $.session.append can read the appended rows back.

Fixes that touched real workflows

• Context compaction: Claude no longer treats its own pre-compaction actions as done afterward and retracts or redoes finished work.
• MCP memory leak: an HTTP MCP connection used to retain every request it had sent until the connection closed. No longer.
• Backgrounding with ←: messages sent while Claude was working could be lost when ← moved a session to the background. If a queued message can't move, ← now stays put and says so.
• Permission prompts: Esc or "No without feedback" now actually stops the turn when ← was just pressed to background the session.
• ToolList handling: subagents and --agent sessions were told a built-in tool was disabled for the whole session when only their own tool list left it out. Fixed.
• SendMessage: Claude was incorrectly told to continue or message subagents with SendMessage in sessions where a host, permission rule, or --tools list had removed it.
• Path-scoped rules and nested CLAUDE.md: these weren't loading when Claude viewed a file with a single-file cat, head, tail, sed -n, or grep in the Bash tool rather than the Read tool.
• Credential/session bugs: claude logs, stop, kill, rm, and claude daemon status, stop, uninstall could sign you out when your login had expired or was about to. PushNotification no longer reports "Remote Control inactive" in claude remote-control sessions.
• /model effort wrapping: ←/→ used to wrap past the highest or lowest level and could silently save Low as a model's default effort. Fixed.
• Docker Desktop GitHub credential store: claude plugin eval no longer refuses every Bash-granting run on Macs where ~/.docker/bin holds a link; the refusal names which part of the credential store held it.
• Agent naming: a custom agent named worker no longer shows as "Agent" on start, and the agent detail dialog keeps the agent type once the agent finishes.

Other notable items

• Very long Remote Control and cloud sessions now stream replies instead of dumping them a block at a time.
• Remote Control no longer re-uploads a session's starting history after every credential recovery.
• claude agents now asks for consent before offering bypass permissions when consent was saved only in .claude/settings.local.json or a --settings file, and shows a persistent notice if a session ignores bypass.
• The footer's agents count no longer disappears after a momentary failure to read the sessions folder (e.g. too many open files).
• The Artifact tool's transcript row no longer briefly reads Artifact("(unprintable path)") while a publish call is streaming in.

If you use Claude Code with subagents, mods, or Remote Control, this release is mostly a cleanup pass on the sharp edges around tool lists, backgrounded sessions, and MCP lifecycles. Upgrading is low-risk and worth it for the Haiku 5.5 default alone if you lean on the cheap tier.