Claude Code v2.1.225~v2.1.226 Major Updates - SendMessage Cross-Machine Origin Support and Authentication-Related Fixes

This is Ishikawa from the Cloud Business Division. I'd like to summarize the Claude Code updates for v2.1.225 ~ v2.1.226 (2026-08-07 ~ 2026-08-08). While there aren't many new features, there are fixes related to authentication and session handling, but no breaking changes. Today I tried out "SendMessage to other machines."

The previous update article is here.

https://dev.classmethod.jp/articles/20260807-cc-updates-v2-1-224/

Update Summary

This covers 2 versions (v2.1.225 ~ v2.1.226, 2026-08-07 ~ 2026-08-08), with 15 changes listed in the CHANGELOG. The breakdown is 2 new features, 4 security/authentication-related changes, 1 improvement, 8 fixes, with no breaking changes or deprecations. The majority of changes are concentrated in v2.1.225, centered on stabilizing Remote Control and authentication.

Additionally, on the Claude Platform / Managed Agents side, it was announced that "Auto mode (automatic approval mode)" will become the default starting August 14, 2026, for Pro, Max, and Team plans.

https://claude.com/blog/auto-mode-default-in-claude-code

Highlighted Updates

SendMessage Can Now Initiate Conversations to Sessions on Other Machines (v2.1.225)

Continuing from the recent release (v2.1.225), SendMessage can now initiate conversations by specifying a name to Remote Control sessions running on other machines. In ListAgents, they are displayed as name [ref]. Previously, you could only reply after receiving a message from the other party.

Since you can now reach out first to sessions on other machines from your local machine, those who divide work across multiple machines will find this useful.

Usage Warnings Now Support Gateway Spend Limits (v2.1.225)

Usage warnings now support gateway spend limits, and when the limit is reached, the message now displays the limit value, reset time, and a message from the operator. The gateway side also needs to be v2.1.225.

Since the limit-reached display now shows the limit value, reset time, and operator message together, those who operate a gateway internally should find it reduces the effort needed to handle inquiries.

Added Workspace Trust Prompt to claude agents (v2.1.225)

When launching claude agents for an untrusted directory, a workspace trust confirmation prompt is now displayed, making it behave the same as claude.

The more you run background agents in directories of separate projects, the more significant it is to have the same confirmation as claude intervene.

Fixed Bug Where CLAUDE_CODE_OAUTH_TOKEN Was Replaced with Short-Lived Token (v2.1.225)

A bug was fixed where a temporary 401 error would cause a long-lived CLAUDE_CODE_OAUTH_TOKEN to be replaced by a short-lived token from a saved login, breaking headless sessions until restart.

For those using CLAUDE_CODE_OAUTH_TOKEN in CI or batch processes, since this was the kind of failure that wouldn't recover until a restart, it's worth prioritizing an upgrade.

Fixed Bug Where MCP OAuth Server on macOS Would Repeatedly Return 401 (v2.1.225)

A bug was fixed where, after a keychain read timeout on macOS, the MCP OAuth server would repeatedly return 401 errors as if never authenticated, causing intermittent failures.

Those who regularly use OAuth MCP servers on macOS and occasionally saw authentication appearing to drop out may find this fix explains what was happening.

Fixed Bug Where Conversation History Was Corrupted on Remote Control Session Resume (v2.1.225)

A bug was fixed where resuming a Remote Control session after a very large conversation was compacted would corrupt the conversation history.

For those who use Remote Control while continuing long conversations, it's significant to avoid the situation of only discovering history corruption after resuming.

Target Versions and Period

New Features

• SendMessage can now initiate conversations by specifying a name to Remote Control sessions on other machines (v2.1.225)
• Usage warnings now support gateway spend limits, displaying the limit value, reset time, and operator message (gateway side also requires v2.1.225) (v2.1.225)

Security / Authentication

• A workspace trust confirmation prompt for untrusted directories has been added to claude agents (v2.1.225)
• A bug was fixed where a temporary 401 would cause a long-lived CLAUDE_CODE_OAUTH_TOKEN to be replaced by a short-lived token from a saved login (v2.1.225)
• A bug was fixed where, after a keychain read timeout on macOS, the MCP OAuth server would repeatedly return 401 errors (v2.1.225)
• A bug was fixed in auto mode where the model's own permission check being rejected by safety filters was counted toward consecutive block counts. The operation is still rejected, but the model is now told to move on rather than retry (v2.1.225)

Improvements

• In Remote Control, photos attached from the Claude app are now passed directly to Claude instead of being read from disk via a separate tool call (v2.1.225)

Fixes (Major Items)

Highlighting fixes related to stability and usability.

• Fixed bug where conversation history was corrupted on Remote Control resume: The issue where conversation history would be corrupted when resuming a session after a very large conversation was compacted has been resolved (v2.1.225)
• Fixed inter-session message accumulation: The issue where inter-session messages would remain pending without notification or expiration in headless sessions and at startup has been resolved (v2.1.225)
• Fixed handling of self-hosted-runner startup failure: When claude self-hosted-runner could not create or write to --base-dir, instead of completing registration while all sessions kept failing, it now outputs a clear error at startup and exits (v2.1.225)
• Fixed false detection of Claude Code on the web sessions: The issue where sessions were incorrectly reported as stopped and resent a growing backlog of events on every reconnect has been resolved (v2.1.225)
• Fixed collapsing in VSCode Focus view: The issue where the latest TODO list, pending question context, and confirmed answers were being collapsed has been resolved; thought-only collapsing now displays as "Thought for Ns" and collapses again when that turn completes (v2.1.225)
• Fixed SendMessage destination mix-up: Once a Remote Control destination is confirmed, it will no longer be replaced by a local session of the same name even if the list cannot be re-checked (v2.1.225)
• A quietly appreciated fix: A bug was fixed where simply hovering over a session from a different project in the agent list would change the starting directory for the next agent to be launched (v2.1.225). Since a hover alone changing the working location of the next agent is the kind of accident that's hard to notice, this feels like a fix that quietly makes a difference.
• Additionally, v2.1.226 includes bug fixes and reliability improvements with no individual item descriptions.

Trying Out "SendMessage to Other Machines"

I'll verify within read-only scope the point that ListAgents in v2.1.225 now displays sessions in name [ref] format.

In advance, I ran claude in a separate terminal and used /rename session-a to rename the session for easier identification.

Now, since actually sending via SendMessage would affect the other session, I won't execute that, and will only look at the list display.

# Run under a temporary directory (only allowing read-only ListAgents)
% echo "Please execute the ListAgents tool." | claude -p --max-turns 1 \
  --allowedTools ListAgents --output-format stream-json --verbose > listagents.jsonl

# Extract only the tool execution results from the output JSON Lines
% jq -r 'select(.type=="user") | .message.content[]? | select(.type=="tool_result") | .content' listagents.jsonl
Peer sessions (1):
  session-a [845f99]  ·  interactive  ·  idle  ·  started 25s ago

You can confirm that a reference like [845f99] is appended after the session name, listed in name [ref] format. Note that claude's exit code is 1, but this is because --max-turns 1 was reached, and the ListAgents execution itself succeeded.

Previously we had to wait for the other party to initiate, but being able to assign work by specifying a name from our side feels like it increases flexibility when working across multiple machines. Since I didn't have Remote Control connected locally, I couldn't confirm rows for other machines, but the display format was consistent as name [ref].

Closing

These 2 versions feel more focused on solidifying the foundation around authentication and sessions rather than new features. The more you rely on headless execution or MCP, the more value there is in upgrading.

If there are any changes that interest you, try updating and checking them out.

References

https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

https://code.claude.com/docs/en/changelog

https://dev.classmethod.jp/articles/20260807-cc-updates-v2-1-224/