Pipeline Pause Control Research

Question

How should Signet let an operator temporarily stop extraction work to free local resources without dropping queued memory capture?

Current behavior

Signet already has several pipeline-adjacent switches, but none match the operator need cleanly:

1. memory.pipelineV2.enabled = false

   • fully disables pipeline enqueue/start behavior
   • reduces work, but new memories stop entering the extraction queue
   • not a good fit for “pause now, catch up later”

2. memory.pipelineV2.shadowMode = true

   • still runs extraction and decision stages
   • preserves observability, but does not reduce model/runtime pressure

3. memory.pipelineV2.mutationsFrozen = true

   • blocks writes and destructive actions
   • extraction still runs, so local compute and provider pressure remain

The missing mode is operational pause:

• keep raw memory capture working
• allow pending jobs to accumulate safely
• stop worker activity so the machine gets relief
• resume later and drain backlog with existing pipeline behavior

Constraints from the current codebase

Daemon startup

packages/daemon/src/daemon.ts decides whether to start the extraction pipeline at startup. If the pipeline does not start, Signet already falls back to retention-only background work.

Enqueue behavior

Remember/session-end paths check pipelineV2.enabled before enqueueing jobs. That means a pause mechanism must remain distinct from enabled if we want backlog preservation.

Worker lifecycle

packages/daemon/src/pipeline/index.ts already exposes a clean startPipeline() / stopPipeline() split. This is a strong fit for a paused state because the worker graph is already centralized.

Recommended design

Add a dedicated persisted flag:

memory:
  pipelineV2:
    paused: true

Semantics:

1. enabled = true, paused = false

   • normal operation

2. enabled = true, paused = true

   • keep enqueue behavior
   • do not start extraction workers
   • surface pipeline mode as paused

3. enabled = false

   • fully disabled, regardless of paused flag

This gives users a real “pause/resume” control instead of overloading enabled or mutationsFrozen.

Recommendation for CLI behavior

Expose operator-first commands:

• signet daemon pause
• signet daemon resume

The CLI should:

1. persist the paused flag in the active config file
2. restart the daemon so the running process picks up the new state
3. report whether queued work will resume later

Restart-on-toggle is acceptable for the first implementation because the current daemon does not hot-reload pipeline runtime state from agent.yaml.

Non-goals

• no scheduler or timed auto-resume
• no per-worker selective pause
• no change to memory retention semantics
• no dropping or reclassifying existing queued jobs