>removedM

Context

When using Claude Code agents for tasks that submit HPC/SLURM jobs, the current workflow is:

1. Agent submits HPC job
2. Agent writes job_id to signal file and exits
3. External daemon monitors HPC job status
4. When job completes, daemon spawns fresh agent to validate outputs

Potential Improvement

Instead of spawning a fresh agent, could we resume the same session that submitted the job? This would preserve:

• Context about what was submitted
• Any partial work/analysis done before submit
• Reasoning about expected outputs

Questions to Explore

1. Can claude --resume <session-id> be used programmatically?
2. How long do sessions persist?
3. Is there an API to check session state before resuming?
4. Would resuming be more cost-effective than fresh context?

Current Workaround

Daemon auto-kills agent after HPC submit and respawns with explicit "HPC job X COMPLETED" prompt. Works but loses session context.

Related

This is for the workaholic autonomous task daemon: https://github.com/[user]/workaholic