context-folding

To prevent context bleed during multi-agent tasks, you create isolated computation branches with dedicated token budgets using branch_create. This isolates sub-task execution, performs the work, and returns structured data to the parent session via branch_return.

Context folding is a task-orchestration technique that creates token-budgeted branches to isolate complex sub-tasks like multi-file codebase analysis. It prevents context bloat by automatically cleaning up branches upon return, preserving the main session state.

You manage token budgets for parallel agent coordination by assigning a specific budget and timeout to each isolated branch during creation. You can then monitor real-time execution status using branch_status before returning the final results.

Yes, you can execute dependent sub-tasks while protecting the main session state. The context folding mechanism supports both parallel and dependent workflows, ensuring each branch operates independently with clear return values and memory records.

No, context folding does not require external dependencies to isolate branches. It implements branch creation, status checks, and returns natively, with optional contextd integration available for advanced multi-agent orchestration workflows.

Limitations of using token-budgeted branches include strict execution ceilings defined by the assigned budget and timeout. If exploratory tasks exceed these limits, the branch cannot complete, requiring careful budget allocation for complex analyses.