Patch release with early self-hosted runner support, stronger memory and focus-mode behaviour, and tighter managed-agent / plan-mode guidance.
Changelog
[new] Self-hosted runner support: register and manage runner pools, poll spawn hints, and operate sessions against your own hosts, with matching Settings guidance under Self-hosted environments.
[new] CLI flags --cloud, --environment, and --teleport for cloud / remote session workflows.
[new] Operator tools for self-hosted setups — pool health, log tailing, and point-in-time metrics — plus API routes for register/deregister, pools, sessions, and SCM connector tunnels.
[changed] Memory writes are stricter and more actionable: save durable, applicable lessons in the same turn as a correction; respect read-only memory sessions; handle explicit remember/forget; avoid sensitive data and transient environment quirks.
[changed] Focus mode: when enabled, only the final text of each response is shown — skip mid-turn progress chatter and put everything the user needs in the closing message.
[changed] Plan mode blocks artifact-database reads/writes that need a consent surface no one can answer in-session; keep those in the plan and raise them in chat instead of retrying.
[changed] Managed-agent sessions: clearer rules for mid-session tools / mcp_servers updates, create-only vault_ids, non-propagating overrides to roster agents, budget-reached handling, and thin webhook payloads.
[changed] Subagent defaults and isolation notes refined (background vs sync, remote CCR isolation, teammate spawn limits).
[changed] Permission and correction behaviour: treat repeated user affirmation as a decision and proceed; refuse only genuinely harmful or prohibited requests; save applicable corrections before finishing the turn.
[new] Built-in specialist system prompts for research delegation, concurrency review, focused test writing, and source-backed Q&A.
[fix] Git stash safety guidance when multiple sessions/worktrees share a stash stack (tagged stash + apply by SHA, not bare pop).
[changed] Chrome extension decline is session-sticky (no re-prompt); continue with read-only web tools or manual browser steps; /chrome to revisit.
[changed] SDK compatibility window updated (wrapper 0.3.222 added; older 0.3.182 dropped from the tested set).
Models & prompts
Memory guidance now stresses applicability, durability, and legibility; path limits and index/version handling are spelled out more carefully, including when memory is read-only.
Context management and partial-summary prompts distinguish full-session summaries from “recent portion only” summaries after retained earlier context.
Auto-mode classifier review and high-effort code-review prompts remain available with clarified blocking preamble rules (response must start with <block>).
Coordinator / multi-agent prompts clarify roster naming, no shared conversation context across threads, and correct interpretation of thread message direction on child streams.
PR workflow still optionally offers Slack posting when CLAUDE.md mentions it and a Slack tool is discoverable — only after user confirmation.
Under the hood
Large self-hosted orchestrator surface: pool/session queue metrics, spawn-hook timing, warm-hint dispatch, SCM connector status, and related runner environment variables (CLAUDE_RUNNER_*, messaging socket, remote session UUID).
Directory-sync lifecycle events (seed/upload/apply) and ultrareview quota/awareness hooks.
MCP hardening signals around OAuth issuer echo mismatches and tripwires.
Bedrock region prefix and control-plane 5xx handling knobs; optional CCR v2 path (CLAUDE_CODE_USE_CCR_V2).
Debug-only --debug-token-dir for writing live tokens to disk (not for production).
Artifact DB and harbor-kite related internal flags for upcoming workflows.
Upgrade notes
If you use the Claude Code SDK wrappers, prefer a tested version in the current compat list (0.3.183–0.3.222); 0.3.182 is no longer listed.
Self-hosted runner APIs and Settings copy distinguish the new flow from the older “Self-hosted cloud environments” / environment-profile path — use the Self-hosted environments section when configuring runners.
Sessions that pin older agent versions keep that pin; new sessions pick up latest agent config unless you pin explicitly.
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.
