Skip to content

daemon.md

Latest commit

 

History

History
54 lines (41 loc) · 2.2 KB

daemon.md

File metadata and controls

54 lines (41 loc) · 2.2 KB
Raw

Background daemon and operations

stakecli daemon is a detached local process that owns long-running work. Uploads and reviewer polling therefore continue when Codex, Claude, a terminal, or the MCP stdio process exits.

Lifecycle

stakecli daemon start
stakecli daemon status
stakecli daemon status --json
stakecli daemon stop

start_upload automatically starts the daemon when needed. Only one daemon may hold the runtime lock. Its heartbeat, log, and status use private files in the StakeCLI state directory.

The daemon reads authentication from STAKE_SID or the active OS-keyring account. Upload queue records contain the target, local path, plan ID, compliance context, non-secret account ID, progress, and desired control state. They never contain the SID, authorization headers, tokens, or presigned URLs. Credentials supplied to stakecli mcp --sid/--key are inherited only by the detached child process. A daemon claims jobs only for its authenticated account; StakeCLI refuses to reuse a live daemon from another account.

Upload execution

start_upload still requires the exact plan_id returned by plan_upload. Before executing a queued operation, the daemon rescans local files and remote scratch. Any change invalidates the plan and fails safely before a mutation.

Available controls:

  • pause_upload: stop at the next safe transfer boundary
  • resume_upload: continue and ensure the daemon is running
  • cancel_upload: cancel queued or active work with confirm=true
  • retry_upload: create a new attempt for failed/cancelled work with confirm=true; the original record remains available

Control and progress updates use per-operation file locks plus atomic private writes. Finished records are retained for one hour.

Reviewer polling

The daemon polls all accessible teams every 15 seconds by default. Configure a different interval with STAKE_DAEMON_REVIEW_INTERVAL=30s or run the foreground worker with stakecli daemon run --review-interval 30s.

The first poll establishes a baseline and does not replay historical events. Later message and approval event IDs are delivered exactly once to each successful hook. Failed deliveries are retried without replaying successful sibling hooks.