meridian-privilege-escalation
Privilege Escalation
Meridian agents run with constrained permissions by default: sandboxed filesystems, restricted tools, harness-specific limitations. When a spawn can't complete its task because of these constraints, you can escalate permissions per-spawn without changing the agent profile.
Escalation Discipline
Prefer the least-privilege escalation that unblocks the task. Try targeted fixes first (--approval auto) before broad overrides (--approval yolo, --sandbox danger-full-access). Broad overrides disable safety checks entirely: if you're reaching for yolo or danger-full-access, surface the situation to the user first and let them approve the escalation. An autonomous agent silently granting itself maximum permissions defeats the purpose of having tiers.
Sandbox Tiers (Codex only)
The --sandbox flag controls Codex's process sandboxing: filesystem, network, and process isolation. Other harnesses have different escalation paths: Claude supports approval escalation (--approval) but not sandbox tier changes. OpenCode permissions vary by configuration. See Approval Modes and Model/Harness Switching below.
Tiers from most to least restrictive:
| Tier | What it allows |
|---|---|
read-only |
Read files only. No writes, no process execution. |
workspace-write |
Read/write within the workspace. No network listeners, no access outside project. |
danger-full-access |
Full filesystem and process access with reduced safety checks. |