Summary
A read of the Claude Code release notes for v2.1.257 and v2.1.277 (with the official changelog covering v2.1.263–269), taken 2026-09-26. Prior scans of this vault stop at v2.1.239, so this window is new ground.
Four items in the window each name a gap in the author's own agent system: a trust boundary around subagent output, a precedence model for instruction files, MCP connect-failure disclosure with HTTP status, and a productised counterfactual eval harness. One of them — the subagent-output provenance boundary — is judged the most transferable. Remaining items are logged but not acted on.
Key Points
- v2.1.277 — Subagent-output provenance boundary. Subagent results now reach the main agent under a header marking them as subagent output, with the result indented, so subagent text cannot pass as the session's own instructions. This is a prompt-injection boundary between trust levels, not a feature. It is the same defect family as the author's #902: untrusted text arriving where trusted instruction lives, on a different channel.
- v2.1.277 — AGENTS.md as a first-class instruction file. With no CLAUDE.md present, Claude Code reads AGENTS.md instead; the choice is changed under Project instructions in
/config. Contrast with the author's system, which reads AGENTS.md as one of six unconditionally concatenatedPROJECT_CONTEXT_FILES— no precedence, no visibility. - MCP connect-failure disclosure.
claude mcp listand/mcpnow show HTTP status and error text when a server fails to connect, plus a warning for MCP config values with hidden leading/trailing whitespace. The author has the model-facing half (#843, Day 181) but not the/mcp-surface status text or the whitespace warning. - v2.1.263–269 —
claude plugin eval. Runs a plugin against a suite of test cases, scores the results, and by default re-runs each case without the plugin to show what it contributes. This is the author'sscripts/counterfactual_green.pyidea generalised to arbitrary plugins and productised with a delta column.