Agent automation
Agents may coordinate Voidscape non-interactively, but automation does not weaken consent gates.
Canonical sources: workflow, CLI reference, and the reader protocol decision
Guided coordination#
Use the guided CLI when an agent can present approvals to a user:
voidscape inspect "clip.mp4" --json
voidscape preview "clip.mp4" --tier both --backend captions
voidscape read "clip.mp4" --tier both --backend captions --workdir evidenceThe agent records the preview decision, stops when required, and adds only the approval flags the user granted for that invocation.
Raw coordination#
Use focused readers for integrations that need stable envelopes:
python skill/scripts/video.py manifest --envelope --compact
python skill/scripts/video.py probe "clip.mp4" --envelope --compact
python skill/scripts/video.py estimate "clip.mp4" --tier both --envelope --compact
python skill/scripts/video.py run "clip.mp4" --tier both --workdir evidence --envelope --compactDispatch order in the guided CLI is image, then article, then video/audio. Integrators should use reader manifests rather than duplicating command flags from memory.
Safe state machine#
new -> inspected -> previewed -> approval_required
| |
| +-> wait for current user decision
+-> ready -> read -> evidence_ready -> answer_with_citationsBind the decision to the input, scope, tier, backend, and preview. If those change, preview again. Never auto-retry exit 4 with approval flags.
Failure handling#
- Exit 2 or 3: correct invocation/input before retrying.
- Exit 4: stop for the named approval.
- Exit 5: explain the missing dependency; installation is a separate action.
- Exit 6: surface the sanitized operation error and whether a narrower/public/local path exists.
- Exit 1: preserve the error and investigate; do not assume the run is safe to repeat.
Voidscape ships no scheduler, background worker, or production MCP host. A calling system remains responsible for queue durability, concurrency, cancellation, and human approval presentation.