DevAgentRadar → Claude Code

You are helping me adopt a real coding-assistant change. Work only from the facts below. Do not invent features.

Context

Assistant: Claude Code Proposal: Upgrade Claude Code to v2.1.293 and switch Haiku to claude-haiku-5-5 Summary: v2.1.293 makes claude-haiku-5-5 the default Haiku model (1M context, $0.10/$0.50 per Mtok) and stops Claude from retracting finished work after compaction. Upgrade, switch Haiku sessions, and wire agentType plus isDeferred:false into status-line scripts and mods. Primary source: https://github.com/anthropics/claude-code/releases/tag/v2.1.293

Why it matters

This release both cheapens long-context Haiku work and removes several ways Claude Code lost or duplicated work. claude-haiku-5-5 is now the Anthropic API default Haiku: 1M context, $0.10/$0.50 per Mtok, and $0.50/$2.50 when prompts exceed 100K—so pinned older Haiku ids should be replaced. Compaction no longer makes Claude treat its last pre-compaction actions as unfinished and retract or redo them, which is why long sessions should move to v2.1.293 first. Status-line scripts can read the new agentType field on subagentStatusLine to tell custom subagent types apart. Mods that must expose a tool without tool search should pass isDeferred:false on $.tool.register so the schema is in the prompt from the start. Also stop inspecting files with Bash cat/head/tail/sed -n/grep if you rely on path-scoped rules or nested CLAUDE.md (those now load for those Bash views, but Read remains the reliable path), and re-grant bypass-permissions in the claude agents consent prompt instead of only .claude/settings.local.json or --settings, which background sessions previously ignored. Do not infer the three omitted changelog items.

Suggested actions

1. Upgrade this machine's Claude Code CLI to v2.1.293, then in the session run /model claude-haiku-5-5 and keep that id as the Haiku default (1M context; $0.10/$0.50 per Mtok; $0.50/$2.50 for prompts over 100K).
2. In every custom subagentStatusLine script, read the new agentType field on the payload and branch the status line by custom subagent type instead of treating all subagents as one kind.
3. For each mod tool that must be visible without tool search, change $.tool.register to pass isDeferred: false so the tool schema is listed in the prompt from the start.
4. When inspecting a file so path-scoped rules or nested CLAUDE.md apply, use the Read tool rather than a Bash cat, head, tail, sed -n, or grep one-liner.
5. Re-run claude agents and grant bypass-permissions in the new interactive consent prompt; do not rely on consent saved only in .claude/settings.local.json or a --settings file for background or --agent sessions.

Config surfaces this release may change

• subagent definitions (high confidence) — check your repo before applying
• MCP servers (high confidence) — check your repo before applying
• permission rules (high confidence) — check your repo before applying
• settings files (high confidence) — check your repo before applying
• skills — check your repo before applying
• sandbox settings — check your repo before applying
• hooks — check your repo before applying
• agent context files — check your repo before applying

After you finish

Do not report this as applied to DevAgentRadar. You cannot write the visitor's loop.

Tell the human: open https://devagentradar.com/proposals/claude-code-v2-1-293-upgrade-claude-code-to-v2-1-293-and-switch-haiku-to and mark Applied, Skipped, or Failed. Proposal id: d10747ad-8ea3-40e8-a5c6-cc0c54d863f5

Your job

1. Restate the change in one sentence.
2. Propose a minimal plan for my repo (or a throwaway pilot).
3. Implement only what I approve; prefer small diffs and tests.
4. Call out risks (permissions, breaking APIs, cost).

Start by confirming you understood the proposal.