Two point releases in two days is a normal Claude Code cadence, but v2.1.145 (May 19) and v2.1.146 (May 21) together signal something worth tracking: Anthropic is treating code review and multi-agent observability as core product surface, not afterthoughts.
Here’s what changed and why it matters.
/code-review Replaces /simplify#
The clearest rename in the release: /simplify is now /code-review. The command now accepts an optional effort level:
/code-review # default effort
/code-review high # deeper analysis passThis isn’t just cosmetic. /simplify framed the command as “make this code less complex.” /code-review frames it as a first-class review workflow — the same positioning Anthropic uses for Code Review GA (the $15-25/PR multi-agent product launched at Code with Claude SF on May 6).
The inline version is lighter than Code Review GA: it runs in the current session rather than spawning dedicated cloud reviewers, and it doesn’t post GitHub PR comments automatically. Think of it as the in-terminal counterpart to the paid PR review product. Use /code-review for fast, local review during development; use Code Review GA for pre-merge gate review at the PR level.
The effort flag is the important addition. High-effort review engages the same xhigh reasoning level introduced in April — extended thinking budget, more passes, better coverage of subtle issues like race conditions, invariant violations, and security edge cases.
Agent Observability: OTel Spans Get Agent IDs#
Multi-agent tracing just got a critical piece. v2.1.145 adds agent_id and parent_agent_id attributes to OpenTelemetry spans emitted by Claude Code sessions.
What this enables: if you’re running a multi-agent workflow — an orchestrator spawning subagents, Agent Teams running parallel tasks, or Managed Agents coordinating across sessions — you can now trace the full execution tree in your observability stack. Each subagent’s spans are tagged with its own agent_id and the parent_agent_id of whatever spawned it.
Before this change, OTel spans from a Claude Code multi-agent session were a flat stream of events with no way to reconstruct the agent hierarchy. Tool calls, token usage, latency spikes — all of it appeared as a single-level trace. You could see that the orchestrator used 50K tokens; you couldn’t see which subagent consumed 40K of them.
With agent ID propagation, the Datadog/Grafana/Honeycomb view of a multi-agent run becomes a proper trace tree. This is the kind of observability data that makes the difference between “we vibe-checked it” and “we can defend the cost and latency of this agent system to engineering leadership.”
Pair this with the invocation_trigger OTel attribute added in v2.1.126 (which tags whether a session was started by CLI, API, Routine, or GitHub event) and you have a fairly complete telemetry story for production multi-agent workflows.
claude agents –json#
The claude agents command (introduced alongside Agent View in v2.1.139) now accepts a --json flag:
claude agents --jsonOutput is a machine-readable JSON array of active sessions with their status, session IDs, current task descriptions, and whether they’re awaiting input. This makes it scriptable.
Practical use: CI pipelines that spin up multiple Claude Code agents via the API can now poll claude agents --json to check completion status without screen-scraping the terminal UI. Build tooling that blocks the deploy step until all active agents report done. Alert when an agent enters awaiting_input state because it hit an ambiguous file conflict.
The companion feature: claude agents also sets the terminal tab title to include the count of sessions currently awaiting input. If you run multiple agent windows in tmux, you can see at a glance which tabs need attention.
Plugin Discovery Before Installation#
The /plugin command’s Discover and Browse screens now show full metadata before you install anything: available commands, agents, skills, hooks, MCP servers, and LSP servers. You get cost estimates and permission requirements upfront.
This is a direct response to the plugin security concern highlighted when CVE-2026-21852 (CLAUDE.md supply-chain attack) was disclosed. The pre-install disclosure model means you can audit what a plugin will add to your environment — and what it will cost in tokens per invocation — before committing.
The per-session token cost estimate (introduced in v2.1.139’s plugin details screen) now extends to the Browse screen, so comparison shopping between plugins is possible without installing each one.
Status Line Gets Repository Context#
The status line JSON format (consumed by Starship, Oh My Posh, and custom prompts) now includes GitHub repository name and PR number when Claude Code detects them. This means your shell prompt can show something like:
[claude: in-progress | repo: myorg/myrepo | PR #847]Minor ergonomic win for teams that run Claude Code inside CI/CD environments or developer VMs where multiple repo contexts are common.
Reading the Cadence#
v2.1.139 on May 20 shipped Agent View and /goal — the architecture features. v2.1.145-146 ships the operational features: observability hooks for ops teams, scripting support for tooling engineers, and the /code-review rename for developers who want a faster path to the review workflow.
Anthropic’s release pace has stayed high throughout May (four point releases in three weeks). The pattern: one or two architectural features that get the press coverage, then a rapid follow-on that wires up the observability, scripting, and UX surface around them.
The OTel agent ID story is the one worth watching. As multi-agent Claude Code deployments move from “interesting demo” to “production infra,” the teams running them will need exactly this kind of trace data to debug, optimize, and justify the cost. v2.1.145 quietly gave them the foundation.
Sources: Claude Code releases — GitHub · Releasebot changelog — Anthropic Claude Code May 2026 · Claude Code v2.1.139 announcement · Claude Code Review GA