[BUG] Cross-session message written to target transcript but excluded from model's context

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

A cross-session send_message is written into the target session's transcript
(confirmed via list_events) but never enters the assistant's context.

DIRECT PROOF: the target session enumerated the four human-typed turns
surrounding the injected message ("test 3", "open", "run", "?") and skipped
the injected turn sitting directly between "run" and "?" in its own
transcript. Across ~12 attempts it never once echoed an injected code word.

Symptoms:

• send_message returns "Message sent" (no error)
• target wakes: isRunning becomes true
• target then sits with ZERO token consumption (never reaches the model)
• after ~5 minutes a generic "Error" badge appears, no error text on hover
• the same session responds instantly when the user types in it directly

This is a REGRESSION. It worked reliably in the same project days earlier -
sub-chats received injected instructions and returned substantive
multi-paragraph reports. It then stopped with no configuration change.

What Should Happen?

The target session should receive the injected message into its context and
respond to it, as it did days earlier in the same setup.

Error Messages/Logs

=== 1. The tool itself reports SUCCESS. There is no error to paste. ===

Message sent to session local_361fa42c-92d2-46e7-9d81-8396b2bdc81f ("test 4").

Message queued for session local_0468f13d-0faf-4160-becf-1bf7a33e0b8e
("iPro BI"); it will be processed after the in-flight turn finishes if that
session stays healthy.

=== 2. The UI shows a bare "Error" badge with NO text on hover. ===
No stack trace, no message, no code. That is the entire error surface.


=== 3. list_events on the target - the injected turns ARE in the
        transcript, with no assistant turn after them ===

Session "test 3" (idle) - showing 16 of 30 messages

[user] test 3
[assistant] Got it - test 3. I'm here and ready. What would you like?
[user] open
[assistant] Not clear what you mean by "open" - a file, a page, a chat?
[user] run
[assistant] I don't have a concrete task to run - the requests "test 3",
            "open" and "run" don't specify what exactly.
[user] [user] ?
[assistant] It looks like the last messages ("test 3", "open", "run", "?")
            are input tests rather than a clear request.

^^^ THE KEY EVIDENCE: the assistant enumerated the four human turns and
    skipped the cross-session turn sitting directly between "run" and "?"
    in its own transcript. It was written to disk but not put in context.


=== 4. Also seen, but probably UNRELATED (terminal CLI only; the desktop
        app was authenticated and answering typed input normally) ===

20:31:03 [ERROR] API error (attempt 1/11): ... not valid; run /login to
                 re-authenticate
20:31:03 [ERROR] API auth_error: OAuth refresh token is no longer valid
20:31:03 [ERROR] [engine] turn ended in error
         -> failed instantly (api=0ms, 0 tokens), no retry

Note: /login fixed the terminal. Injection still failed afterwards, from a
session created after the fix.

Steps to Reproduce

1. Open two Claude Code sessions in the same project directory.
2. From session A, call mcp__ccd_session_mgmt__send_message targeting

session B's sessionId with any short message containing a unique code word.

1. Tool returns "Message sent to session <id>".
2. get_session on B shows isRunning: true, then idle. Zero tokens consumed.
3. list_events on B shows the message present in its transcript.
4. Type anything into B manually - it answers immediately, but never

references the injected message or its code word.

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

2.1.227

Claude Code Version

2.1.229 (Claude Code)

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

IMPORTANT TIMING DETAIL - the break does NOT line up with a version update:

• 2.1.227 installed 2026-08-12 11:07
• Injection WORKED on 2026-08-12 (a hub session injected into a sub-chat

and the sub-chat acted on it, confirmed in its transcript)

• Injection BROKE on 2026-08-13 evening - still on 2.1.227
• 2.1.229 installed 2026-08-14 09:29 - AFTER the break

So it regressed WITHIN the same client version. That argues against a
client-side release regression and points to something server-side, or to
accumulated session/project state.
Ruled out by measurement (please don't re-suggest these):

• context size: fails on brand-new empty chats
• burst/load: fails on a single injection
• effort level: fails on both high and medium
• MCP plugins: finance:bigquery / finance:slack disabled - still fails
• app version: full quit + relaunch, and a version update - still fails
• machine resources: 31GB RAM, 13GB free; no hooks defined
• sender session: two different sender sessions, identical failure
• stale session ids in config: corrected - no change
• crossSessionInbound: "accept" in ~/.claude/settings.json, AND a target

session created after that change - still fails

• expired auth: was real but only in the terminal CLI; the desktop app is

authenticated and its sessions answer typed input normally

Additional observation: in older sessions the injected message is not
written to the transcript at all (it vanishes), while in newly created
sessions it IS written but never read. Two distinct behaviours.

Impact: this is the backbone of a 28-chat coordination setup. With it
broken, every instruction must be copy-pasted by hand.

[feedback\[1\].json](https://github.com/user-attachments/files/31139775/feedback.1.json)