v2.1.228

2.1.228

(☝️ Click to see full diff.)

+7,141 tokens

• NEW: Data: Claude Code gateway customer-routed inference protocol — Defines offline validation of short-lived, audience-bound CRI JWTs; operator-credential upstream forwarding without credential relay; response-header and error-body hygiene; stable capability-rejection recovery tokens; policy-block responses; discovery metadata; and fixed Messages API endpoint behavior.
• NEW: Skill: Artifact document — Adds creation guidance for live, editable word-processor-style document artifacts with status and ownership metadata, block-level edits, inline comments, stable updates, and preservation of the template's editor machinery.
• NEW: Skill: Artifact spreadsheet — Adds creation guidance for live, editable spreadsheet artifacts with persistable rows, cell editing, formulas, sorting, comments, status metadata, and preservation of the template's editor machinery.
• REMOVED: Agent Prompt: Bash command prefix detection — Removes the standalone policy prompt that extracted allowlistable command prefixes and flagged suspected command injection.
• Data: Self-hosted runner command help — Documents a non-disableable grace hold that keeps a just-finished background task in flight until the follow-up turn reading its result starts, bounded by SELF_HOSTED_RUNNER_BG_RESULT_GRACE_MS and falling back to the default when zero or unusable.
• Skill: Artifact design, Skill: Prototype, Skill: Whiteboard, and Tool Description: Artifact — Tighten artifact naming around short, distinctive, product-style titles, keep explainers in descriptions rather than dash- or colon-appended title suffixes, and rename topical boards as <topic> whiteboard instead of Whiteboard — <topic>.
• System Prompt: Artifact comment list framing and System Prompt: Artifact comment thread framing — Add optional anchor-path guidance to comment lists and keep thread instructions synchronized with the runtime anchor-path marker while continuing to treat viewer-influenced paths and element snippets as untrusted data.
• Tool Description: ListAgents — Clarifies that Remote Control-connected account listings cover both sessions on other machines and cloud sessions, with each row labeled by kind.