Claude Code: Review, Recap, Plugins, and Permission Surfaces Are Converging

What Changed

Claude Code's Week 17 notes introduced /ultrareview, a public research preview that runs a cloud fleet of bug-hunting agents against a branch or pull request and returns findings to Claude Code and Desktop.

The same week added automatic one-line session recaps when returning to the terminal, named themes, JSON theme files under ~/.claude/themes/, plugin theme support, and a redesigned claude.ai/code surface with sessions, layout, and routines.

The broader changelog in the sprint window included plugin zip/session support, MCP tool-count visibility, MCP retry and auth fixes, hook output replacement, subagent fork behavior, AI_AGENT subprocess attribution, OTel metrics for MCP-created PRs and MRs, and permission-prompt tooling.

Operator Consequence

Claude Code is moving verification and review up into provider-native capability while also improving session continuity, plugin distribution, telemetry attribution, hooks, and MCP governance.

The important operator question is no longer "can Claude review this?" It is: what did the provider-native review inspect, what evidence did it produce, and how should that verdict be composed with local tests, BitterQA, and other worker judgments?

Bitter Implication

Bitter should treat /ultrareview as a worker-native verification engine that can produce evidence, not as a feature to reimplement.

Adapter work should focus on:

• capturing review target, provider, model, and result format
• preserving session recaps as worker-native context receipts
• recording plugin, MCP, hook, and permission-prompt state
• distinguishing Claude's verdict from Bitter's final verification judgment
• feeding review findings into wake packets only after settlement

Signal

Verification is becoming a provider-native service. Bitter's role is to make those verdicts comparable, source-attributed, replayable, and subordinate to operator-owned evidence standards.