💡 Claude Code v2.1.137: Sub-Agent Progress Summaries Now Include Prompt Cache — Cutting cache_creation by ~3×
Claude Code v2.1.137 fixes a cost-amplifying bug in sub-agent workflows. When a sub-agent generated a progress summary to pass back to the orchestrating session, those summaries were not including the prompt cache state — causing the receiving session to re-process the full context from scratch rather than reusing cached tokens. For long-running multi-agent sessions, this could inflate cache_creation token counts by approximately 3×. The fix aligns sub-agent progress summaries with the same cache-aware format used by the main session.
How to verify you were affected
# Check your session's cache usage after upgrading:
# In your usage logs, look at cache_creation vs cache_read tokens
# Before v2.1.137 — high cache_creation, low cache_read on sub-agent turns
# After v2.1.137 — cache_read should dominate on subsequent sub-agent turns
# In API response headers (if using the API directly):
# anthropic-cache-creation-input-tokens: should decrease
# anthropic-cache-read-input-tokens: should increase
Cache economics in multi-agent workflows
Prompt cache read tokens cost significantly less than cache creation tokens (roughly 10× cheaper for input). In a multi-agent session that processes the same large codebase context across multiple sub-agent turns, the difference between a cached and uncached session can be substantial. With v2.1.137, sub-agent handoffs should now benefit from the same cache hit rates as the main agent — which, for a 100k-token context, reduces per-turn cost dramatically on turns after the first.
Claude Code
prompt cache
sub-agents
v2.1.137
cost optimisation
💡 Claude Code v2.1.138 & v2.1.129 Recap: VS Code Windows Fix and --plugin-url for Session-Specific Plugins
v2.1.138 resolves a VS Code extension activation failure on Windows caused by a hardcoded build path in the bundled SDK — a regression that had broken Claude Code for VS Code users on Windows who installed from certain build channels. It also fixes MCP stdio servers receiving corrupted arguments when CLAUDE_CODE_SHELL_PREFIX is set and an argument contains spaces or shell metacharacters, which had caused silent tool failures in shell-wrapped environments.
v2.1.129 — --plugin-url for per-session plugin loading
Released earlier in May, v2.1.129 added a --plugin-url flag that fetches a plugin .zip archive from a URL and loads it for the current session only. This enables ephemeral plugin deployments: share a plugin URL with a colleague, and they can try it in a single session without installing it globally or modifying their settings.
# Load a plugin from a URL for this session only:
claude --plugin-url https://internal.example.com/plugins/my-reviewer.zip
# The plugin is active for this session, not persisted to settings.json
# Useful for: testing plugins, sharing experimental tools with teammates,
# or CI environments where session-specific tooling is preferable
Hosting your own plugin archives
If your team maintains internal Claude Code plugins, serve them from an authenticated internal endpoint (S3 presigned URL, Artifactory, or an internal nginx path). The --plugin-url flag respects HTTP authentication headers, so you can restrict access to authorised team members without publishing plugins to the public Claude Plugin directory.
Claude Code
VS Code
plugin-url
v2.1.138
v2.1.129
Windows
