v2.1.236

2.1.236

(☝️ Click to see full diff.)

+26,676 tokens

• NEW: Agent Prompt: Security monitor host-context line guidance and Data: Hook classifier context field — Distinguish live host-attached user statements, which may satisfy a soft-block consent bar, from restored or mixed host context, which remains unverified; define call binding, trust, size, timing, and rewrite-integrity rules, and integrate the guidance into the autonomous-action monitor.
• NEW: Data: SDK register device hooks request schema — Documents how cloud device clients register forwarded hooks and vetted worker-side templates, including the size limit, worker-epoch handling, and retryable versus terminal registration errors.
• NEW: Skill: Anthropic Python SDK 0.x to 1.x upgrade — Adds an executable migration workflow for the Python 3.10 floor, httpx2, removed deprecated APIs and parameters, raw-response and streaming changes, Bedrock region handling, verification, and user-owned migration decisions, with /claude-api upgrade routing and an authoritative migration-guide source.
• NEW: System Reminder: Artifact auto-replies resumed — Reports that a requested Artifact comment auto-reply resume is re-arming the live watch, explains which comments from the stopped period will be handled based on the stop cause, and warns that the stop persists until reconnection succeeds.
• NEW: System Reminder: Artifact type page untrusted content warning — Treats HTML supplied by an owned Artifact's type publisher as untrusted data that cannot grant instructions or permission escalation.
• NEW: System Reminder: Background task notification with concurrent user input — Separates an automated background-task event from genuine user input delivered in the same turn and prevents the notification from being treated as approval or consent.
• REMOVED: System Reminder: Ultrareview launch acknowledgement — Removes the standalone acknowledgement prompt for already-visible cloud review launches and remembered --fix intent.
• Agent Prompt: Managed Agents onboarding flow; Data: Managed Agents self-hosted sandboxes, memory stores reference, and related Managed Agents/AWS references — Expand self-hosted SDK-worker guidance with single-item handling, safe shutdown and file confinement, synchronized memory stores and recovery, resource and deployment constraints, and Claude Platform on AWS authentication, session-duration, and memory-store limitations.
• Data: Managed Agents tools and skills and related Managed Agents references — Add per-tool web_search and web_fetch domain filters, search location and fetch-size settings, validation and runtime behavior, typed-SDK configuration changes, multiagent restriction layering, and guidance that these tools run server-side rather than under environment networking rules.
• Data: Managed Agents events and steering and Data: Managed Agents overview — Document the Console session viewer's searchable transcript, timeline, raw events, tool/resource/thread inspection, cost views, JSON export, and event deep links.
• Data: Plugin eval and skill-doctor reference — Requires every eval case to provide an execution prompt, treating it as the resumed session's next user turn when a history file is also supplied.
• Skill: Artifact document — Clarifies that publishing assigns comment-anchor block IDs, the editor assigns IDs to user-added blocks, and existing IDs must still be preserved rather than copied or hand-authored unnecessarily.
• System Prompt: Artifact comment reply composer — Makes change-request reply guidance conditional instead of always prescribing a generic work-in-progress acknowledgement, while retaining brief plain-text answers and prohibitions on claiming completed edits.
• Tool Description: Artifact database guidance — Adds file-backed database reads via out_dir and writes via a local JSON file_path so large or numerous documents need not be returned or retyped inline.
• Tool Description: Artifact publishing and update guidance and Tool Parameter: Artifact watch actions guidance — Add terminal and web paths for reopening artifacts, distinguish watch arming from an established connection, limit comment wakes to armed auto-replies, restore at most one watch after resume or continue, and tighten status-based claims about active watches.
• Tool Description: Edit, Tool Description: Edit single replacement, and Tool Description: Write — Make read-before-edit and read-before-overwrite guidance path-sensitive, explicitly retaining the prior-Read requirement for files outside the working directory.
• Tool Description: SendMessage cross-session guidance — Adds one-shot notify_when_idle subscriptions for local sessions, including pure subscriptions, approval-held notices, expiry behavior, and guidance to use them instead of polling or status-chasing messages.