Permissions - Plurai Simulation

The plugin’s flows run under Claude Code’s auto mode with no configuration. This page explains the one capability auto mode removes, how to restore it, and how to allow a plugin script if your environment blocks one. It also explains which prompts are your consent.

Auto mode

Auto mode lets Claude Code’s permission classifier approve routine actions for you. The plugin is built for it:

Every step that manages the flow’s progress is a single call to one plugin script.
The flow keeps its own state in ~/.plurai/flow/, outside ~/.claude/, so nothing the flow does looks like the agent changing its own configuration.

The dropped-turn heartbeat

Outside auto mode, the plugin schedules a heartbeat that picks a flow back up when a turn drops, for example on an API error. Auto mode denies creating scheduled jobs, so under auto mode the plugin does not try. The flow prints a one-line notice instead, makes no call and triggers no denial. The flow runs normally, without that rescue. To restore the heartbeat under auto mode:

Force the heartbeat

Export the variable in the shell that launches Claude Code:

export PLURAI_FLOW_HEARTBEAT=force

Allow scheduled jobs

Add CronCreate to permissions.allow in ~/.claude/settings.json:

~/.claude/settings.json

{
  "permissions": {
    "allow": ["CronCreate"]
  }
}

Restart Claude Code

Claude Code reads both at session start.

If a turn drops without the heartbeat, re-run the same command. It resumes where it stopped.

Copilot CLI and Codex CLI have no scheduled-job tool, so the heartbeat never runs there. Re-run the command after a dropped turn.

If the classifier blocks a plugin script

If auto mode ever blocks a call to a plugin script in your environment, add an autoMode.allow (or autoMode.environment) entry to ~/.claude/settings.json. Keep the literal "$defaults" entry in any list you set, or you lose the built-in rules:

~/.claude/settings.json

{
  "autoMode": {
    "allow": [
      "$defaults",
      "Running scripts under the plurai-simulation plugin directory (flow-window.sh, poll_job.sh, wait.sh) is routine simulation-flow work"
    ]
  }
}

Restart Claude Code after you change the file. Uploads of your data run as ordinary shell commands (plurai ingest-*). In your coding agent’s default permission mode, its approval prompt shows you the exact command before it runs, so approving it is your consent to that upload. Under /simulate or auto mode, uploads run without that prompt. Before anything leaves your machine, the SDK also prints a disclosure of what will be sent and records the upload in .plurai/uploads/audit.log. Some decisions are asked as explicit questions, not as command approvals:

The mock-or-real map. The scan shows which tools are mocked and which run for real, and writes no file until you approve it.
Fixes to the Harness. When the suite’s quality loop finds a fix that belongs in simulate.py, it asks you first. For unattended runs, see approve_harness_edits.
Agent fixes. /simulation-fixes never edits your agent’s source.

See Data and privacy for what each upload contains.

Next steps

Data and privacy

What leaves your machine, and the audit record.

Environment variables

Every variable the plugin reads.

​Auto mode

​The dropped-turn heartbeat

​If the classifier blocks a plugin script

​Approval prompts are your consent

​Next steps

Data and privacy

Environment variables

Auto mode

The dropped-turn heartbeat

If the classifier blocks a plugin script

Approval prompts are your consent

Next steps