1. Why Do You NEED This Feature?
Subagent progress is lost, wasting tokens and time, if I close my laptop mid-think during a subagent tool call. Likewise, if my auth token expires mid-session, resulting in an error coming back from the LLM API, the tool call aborts.
2. What Is NOT Possible Right Now?
Resuming a subagent session. Calling agent always re-prompts the subagent from scratch, duplicating effort.
3. What WILL Be Possible With This Feature?
I, and other users, can safely put their computer to sleep and resume where they left off later even if Forge is doing stuff without losing work.
Proposed Solution (User Experience)
Forge should offer the user of resuming a session from where it left off in a subagent tool operation, asking them to use /retry if they want to. It'd be nice if the agent itself also had the ability to optionally choose to resume such a subagent operation on next user input, even if the user doesn't use /retry.
Alternatives Considered
Switching to the more extensible Pi agent harness where I could use an arbitrary subagent plugin, possibly written by me, that doesn't have this limitation; using Claude Code, which has options for resumption of subagent tool calls under most circumstances.
Feature Category
Multi-Agent Workflows
Priority/Impact
High - Would significantly improve my workflow
Examples from Other Tools
Claude Code allows asynchronous subagent calls, in which case the calling agent receives a conversation_id immediately which can be used to resume or restart the subagent call. In the case of synchronous subagent tool invocation it also receives a subagent handle if the subagent tool call fails (but seemingly not if the subagent tool is manually cancelled by the user).
Additional Context
I'm building a complex multi-agent process with Forge which involves evaluating Excel and/or CSV files, running SQL-based analysis, and using subagents to review at multiple steps (resulting in potentially 3- and 4-level deep subagent invocation trees). Multiple times I've lost tens of thousands of tokens worth of subagent work due to network issues, API errors/auth failures, or having to go home for the day.
Pre-submission Checklist
1. Why Do You NEED This Feature?
Subagent progress is lost, wasting tokens and time, if I close my laptop mid-think during a subagent tool call. Likewise, if my auth token expires mid-session, resulting in an error coming back from the LLM API, the tool call aborts.
2. What Is NOT Possible Right Now?
Resuming a subagent session. Calling agent always re-prompts the subagent from scratch, duplicating effort.
3. What WILL Be Possible With This Feature?
I, and other users, can safely put their computer to sleep and resume where they left off later even if Forge is doing stuff without losing work.
Proposed Solution (User Experience)
Forge should offer the user of resuming a session from where it left off in a subagent tool operation, asking them to use /retry if they want to. It'd be nice if the agent itself also had the ability to optionally choose to resume such a subagent operation on next user input, even if the user doesn't use
/retry.Alternatives Considered
Switching to the more extensible Pi agent harness where I could use an arbitrary subagent plugin, possibly written by me, that doesn't have this limitation; using Claude Code, which has options for resumption of subagent tool calls under most circumstances.
Feature Category
Multi-Agent Workflows
Priority/Impact
High - Would significantly improve my workflow
Examples from Other Tools
Claude Code allows asynchronous subagent calls, in which case the calling agent receives a
conversation_idimmediately which can be used to resume or restart the subagent call. In the case of synchronous subagent tool invocation it also receives a subagent handle if the subagent tool call fails (but seemingly not if the subagent tool is manually cancelled by the user).Additional Context
I'm building a complex multi-agent process with Forge which involves evaluating Excel and/or CSV files, running SQL-based analysis, and using subagents to review at multiple steps (resulting in potentially 3- and 4-level deep subagent invocation trees). Multiple times I've lost tens of thousands of tokens worth of subagent work due to network issues, API errors/auth failures, or having to go home for the day.
Pre-submission Checklist