[BUG] Dispatch conductor session grows unbounded with no rotation/compaction, forcing premium 1M-context billing

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?

[BUG] Dispatch conductor session grows unbounded with no rotation/compaction, forcing premium 1M-context billing

What's Wrong?

The Dispatch conductor — the single persistent orchestrator conversation in Cowork — accumulates context monotonically across every task it manages, with no client-side way to rotate, reset, or compact it. On a long-lived thread it eventually crosses the 200K standard context window, at which point it requires a 1M-context model and begins billing as premium usage. The user has no control to prevent or recover from this.

In my case the conductor reached ~357K tokens (peak 357,449) over ~2,227 turns, starting from ~29,884 tokens. Once past 200K, attempting to continue produces:

API Error: Usage credits required for 1M context · turn on usage credits at claude.ai/settings/usage, or use --model to switch to standard context

The only in-the-moment option to let in-progress work finish was to enable premium usage credits. Anthropic support has confirmed they cannot manually reset the session or reverse the billing, so there is currently no recovery path once a conductor reaches this state.

Steps to Reproduce

1. Use the Dispatch conductor over an extended period, assigning many tasks from the single persistent conductor conversation.
2. Observe the conductor's context grow turn over turn (it will report its own token count if asked).
3. As the conductor approaches and crosses ~200K tokens, it requires a 1M-context model.
4. Attempting to continue triggers the "Usage credits required for 1M context" error and premium billing.

Expected Behavior

A long-lived orchestrator should not silently grow until it tips into premium billing with no user control. Any of the following would resolve it:

• A user-triggered "start a new conductor / archive current conductor" control.
• Automatic conductor compaction at a threshold below 200K, so the conductor never crosses into 1M-context territory unintentionally.
• At minimum, a visible warning as the conductor approaches the 200K boundary, so the user can wind down deliberately rather than being ambushed.

Actual Behavior

• The conductor grows without bound; there is no client-side rotation, reset, or compaction.
• Disk-based handoff/breadcrumb files only reduce what a successor session reloads at startup — they do nothing about in-session accumulation, because the conductor still holds the full running transcript.
• Restarting the desktop app reconnects to the same server-side session rather than creating a fresh one. (The local bridge-state.json only holds a pointer — a cse_ remoteSessionId — to the server-side session; clearing it severs the pointer and risks orphaning pending message acks rather than producing a fresh conductor.)
• Support cannot manually reset the session or reverse the resulting premium charges.

Billing Question (Max plan)

I am on the Max plan. My understanding is that 1M context is included on Max for at least some models without extra usage charges, so it is unclear why this session required premium usage credits rather than being covered under the Max entitlement. Clarification on whether this billing behavior is correct for Max would be helpful, as it may indicate a separate billing bug.

Environment

• Claude Desktop version: 1.9659.2
• Operating system: macOS
• Feature: Cowork Dispatch
• Plan: Max
• Exact error: API Error: Usage credits required for 1M context · turn on usage credits at claude.ai/settings/usage, or use --model to switch to standard context

What Should Happen?

A long-lived orchestrator should not silently grow until it tips into premium billing with no user control. Any of the following would resolve it:

• A user-triggered "start a new conductor / archive current conductor" control.
• Automatic conductor compaction at a threshold below 200K, so the conductor never crosses into 1M-context territory unintentionally.
• At minimum, a visible warning as the conductor approaches the 200K boundary, so the user can wind down deliberately rather than being ambushed.

Additionally one might ask, does Dispatch really need to be a single long lived orchestrator?

Error Messages/Logs

Exact error: `API Error: Usage credits required for 1M context · turn on usage credits at claude.ai/settings/usage, or use --model to switch to standard context`

Steps to Reproduce

1. Use the Dispatch conductor over an extended period, assigning many tasks from the single persistent conductor conversation.
2. Observe the conductor's context grow turn over turn (it will report its own token count if asked).
3. As the conductor approaches and crosses ~200K tokens, it requires a 1M-context model.
4. Attempting to continue triggers the "Usage credits required for 1M context" error and premium billing.

Claude Model

Other

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

1.9659.2

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

This is bout the use of Claude CoWork Dispatch, an extremely valuable tool. I repeated asked it how it would clear its context and deal with turns and it assured me that was all handled. It isn't.