The Claude Code Changelog

This episode explores how Claude Code’s `ultrareview` command uses exit codes, streamed output, and PR targeting to make AI review behave like a real CI check. It also digs into the multi-agent verification model, cost tradeoffs, and the network constraints teams need to consider before wiring it into GitHub Actions.

We break down Anthropic’s new xhigh effort level for Claude Code on Opus 4.7, why it now defaults for Max subscribers, and how the new slider and config flow make it easier to use.The discussion also covers when deeper reasoning is worth the extra latency and cost, plus how to spot the tasks where xhigh really pays off.

Claude Code now asks your configured ANTHROPIC_BASE_URL for /v1/models, so the model picker reflects the IDs your gateway actually exposes instead of a hardcoded Anthropic-only list. The episode breaks down why that matters for LiteLLM, Bedrock, and other proxies, plus the catch: the backend still has to return truthful model names or the client will pass them through verbatim.

We break down the new claude project purge command, which wipes stale transcripts, tasks, file history, and project config so broken project state can be reset cleanly. The episode also covers the expanded behavior of --dangerously-skip-permissions, the new direct OAuth code paste flow for login failures, and a subtle policy bug fix affecting Claude Code memory.

April 29, 2026 brought a small Claude Code update with big budget implications: the new ANTHROPIC_BEDROCK_SERVICE_TIER setting lets Bedrock users choose default, flex, or priority and sends that choice as the X-Amzn-Bedrock-Service-Tier header. The episode breaks down when each tier makes sense, why latency and throughput guarantees matter, and how to avoid overpaying for interactive and batch workloads.

This episode digs into two major v2.1.121 changes in Claude Code: the alwaysLoad flag for loading selected MCP server tools at session start, and a widened PostToolUse hook that can rewrite outputs from built-in tools like bash and file reads. It also covers the practical tradeoffs around context, trust boundaries, and the release’s memory leak fixes.

We break down how ultrareview moved from an interactive slash command to a real CLI that can run in scripts, Make targets, and GitHub Actions as a potential build gate. The episode also covers JSON output for automation, the trust and latency tradeoffs of model-based PR checks, plus smaller quality-of-life updates like PowerShell fallback on Windows and better GitHub attribution.

This episode breaks down Claude Code v2.1.119’s expanded support for GitLab, Bitbucket, and GitHub Enterprise PRs, plus the value of loading full diff, commit, and description context automatically. It also covers smarter PR link handling, configurable review URL templates, and a small privacy win for demos with hidden working directories.

Claude Code v2.1.108 extends prompt caching from 5 minutes to 1 hour, making long system prompts and big repo contexts stay warm through real work sessions instead of expiring during normal breaks. The episode also covers the `DISABLE_TELEMETRY` bug fix, the optional 5-minute fallback, deprecated Bedrock compatibility, and why any prefix change still invalidates the cache immediately.

The hosts unpack how Claude Code 2.1.119 moved /config changes into ~/.claude/settings.json, making theme, editor mode, and verbose output persist across restarts. They also break down the config hierarchy across user, project, and managed policy scopes, plus what teams should audit after upgrading.

This episode explores how Claude Code 2.1.118 and 2.1.119 turn hooks into first-class automation, letting PostToolUse events call MCP tools like Slack without wrapper scripts or brittle bash glue. It also digs into duration_ms as a clean timing signal for smarter logging, alerts, and workflow routing.

This episode digs into Anthropic’s postmortem on Claude’s silent output corruption, from misrouted traffic to TPU/compiler bugs that caused garbled or truncated responses for a significant share of users. It also covers Opus 4.7’s breaking API changes, including hard failures on familiar generation settings and a new tokenizer that can quietly raise token costs.

This episode breaks down a compaction bug that made Claude Code treat Opus 4.7 like it had a 200K-token limit instead of 1M, causing long coding sessions to summarize far too early. It also covers the new default high effort setting for Pro and Max users on Opus 4.6 and Sonnet 4.6, and what that means for reasoning quality and token usage.

This episode breaks down Claude Code 2.1.116’s biggest changes: a much faster /resume flow for huge sessions, plus a critical safety fix that stops broad auto-allow rules from bypassing dangerous-path checks.We also cover hook behavior in --agent mode, MCP startup improvements, and a new GitHub rate-limit hint that makes automation feel less flaky.

The hosts dig into Claude Code 2.1.111’s new xhigh effort tier for Opus 4.7, explaining how it fills the gap between high and max for tougher coding and reasoning tasks. They also cover the redesigned /effort control, safe fallback behavior in mixed-model workflows, and the removal of the old auto-mode flag.

The hosts unpack a cloud-based multi-agent code review command that runs in the background, checks either your current branch or a GitHub PR, and returns only independently reproduced findings. They also dig into why the feature feels more trustworthy than a single AI pass, plus the practical details of diffstat previews, faster launch checks, and how to avoid wasting review runs.

This episode breaks down a small-looking Claude Code update that swaps in a native binary for faster cold starts and a leaner CLI, especially in CI. It also digs into important Bash rule fixes, including a simple env bypass, dangerous find -exec/find -delete approvals, and tighter handling of macOS private paths.

We break down Anthropic’s Claude Code Routines research preview, where jobs run in the cloud so your laptop can sleep, and explore the three ways to trigger them: schedules, API calls, and GitHub events.We also dig into the safety model behind the feature, from claude/ branches and repo scoping to the shift from autonomous execution to human review.

This episode digs into the latest Claude Code 2.1.110 improvements, including the new fullscreen TUI that stops terminal flicker without restarting your session and the cleaner split between /focus and Ctrl+O.It also covers practical upgrades like disabled auto-scroll, editor context injection, and important hardening changes that make long-running workflows and file handling more reliable.

This episode breaks down Claude Code 2.1.108’s new prompt-caching options, including the tradeoff between longer-lived session continuity and short-lived cache turnover for testing. It also covers the new /recap workflow and background session summaries that make it easier to jump back into long-running coding work.

Get a plain-English take on the latest Claude Code updates, with a focus on what changed, who it helps, and whether it actually matters. We separate the useful workflow tweaks from the noise so you can decide what to try, test, or ignore.