Patch release 2.1.223: richer feedback reports (structured details, failure mode, and task category), clearer permission and multi-session git guidance, and assorted agent behaviour polish.
Changelog
[changed] Feedback reports use a structured details format: labeled bullets for what happened, what the user said, repro steps, and evidence, instead of a free-form narrative.
[new] Feedback can tag a model-behaviour failure_mode (for example instruction following, looping, overconfidence, unwanted scope) and a task_category (code edit, debug, plan, shell, review, and similar).
[changed] Permission denials are reported more clearly as rejected tool use, with optional user wording when provided, so the agent can adjust instead of stalling on vague “user doesn’t want this” language.
[changed] Stronger guidance when the git stash stack is shared across worktrees and concurrent Claude sessions: prefer a temporary WIP commit, or tag stashes and apply by SHA rather than bare stash / stash pop.
[changed] Corrections handling is tighter — a follow-up question alone is not treated as proof of error; accurate prior statements are left alone unless the user points at a real mistake.
[changed] Focus mode behaviour is spelled out: only the final text of each response is user-visible, so progress chatter between tool calls is skipped and essentials go in the last message.
[changed] Long-session context management: when history is compacted, the summary plus remaining context is what continues the work — no need to wrap up early mid-task.
[changed] Memory path rules are explicit (no dot-leading folder segments, allowed file extensions, no control characters or backslashes).
[changed] After opening or updating a PR, if CLAUDE.md mentions Slack, the agent may offer to post the PR URL there only after confirmation (and stays quiet if no Slack tool is available).
[changed] Subagent lifecycle guidance clarified (background by default, no fabricated completion results; remote isolation always background where applicable).
Models & prompts
Agent prompts reinforce fair handling of repeated user requests after a concern is raised: proceed when the user reaffirms ordinary work; reserve hard refusals for genuinely harmful or prohibited cases.
Plan mode still blocks edits and non-readonly tools except for the plan file itself.
Summary prompts distinguish full-conversation digests from “recent portion only” digests when earlier context is retained.
Computer-use MCP availability is called out when those tools are present (screenshots and desktop control).
SendFeedback can write a local draft under ~/.claude/feedback/drafts/ so a connected host UI can show a review card without polling; nothing is submitted without explicit user action.
Under the hood
New environment knobs for plan artifacts, tasks, child sessions, session kind, custom headers, organization id, Foundry API key, bash working-directory retention, nonessential traffic control, unknown-model window enforcement, token-reminder budget, voice interim forwarding, and related internal toggles.
Feature gates added for bridge inline image attachments, org memory connected mode, OAuth token refresh edge cases, post-compact survey, remote notification routing, transcript id scan fallback, and several experimental paths (including parchment/fern and harbor kite limits).
Older experimental flags and gates were removed or renamed (including brief upload, observer agents, investigate-first, and assorted scroll/session-name toggles).
Upgrade notes
If you integrate the feedback tool or parse its payload, expect the richer details shape and optional failure_mode / task_category fields.
Multi-session or multi-worktree workflows should avoid bare git stash pop; follow the tagged stash / WIP-commit guidance above.
content/
Browse captured content/
Same tree as claude-code/latest/content for this version. Open the full browser for sidebar navigation between docs and other files.
