abadge permission

`abadge permission create`

Grant one or more capabilities for an agent on an item. Multiple capabilities land atomically — either every row is written or none is. permission create always targets a single item, so it accepts only the legacy capability names: read_ciphertext and reveal_plaintext (mapped to canonical read), and mount_env and mount_file (mapped to canonical use).

# Allow an agent to inject a server-managed secret as an env var
abadge permission create \
  --agent-id <agent-id> \
  --item-id <item-id> \
  --capability mount_env

# Grant several capabilities atomically — repeat the flag or comma-separate
abadge permission create \
  --agent-id <agent-id> \
  --item-id <item-id> \
  --capability mount_env \
  --capability mount_file

abadge permission create \
  --agent-id <agent-id> \
  --item-id <item-id> \
  --capability mount_env,mount_file

# With expiry — applied to every capability in this batch
abadge permission create \
  --agent-id <agent-id> \
  --item-id <item-id> \
  --capability reveal_plaintext \
  --expires-at 2026-12-31T00:00:00Z

Flag	Required	Description
`--agent-id <id>`	yes	Target agent
`--item-id <id>`	yes	Target item
`--capability <cap>`	yes	One of `read_ciphertext`, `reveal_plaintext`, `mount_env`, `mount_file`. Repeat the flag or comma-separate to grant multiple.
`--expires-at <ts>`	no	ISO timestamp expiry; applied to every capability in this submission
`--json`	no	Print raw JSON

The canonical capabilities read / use apply to profile-target grants. Item-target grants — the only kind the CLI creates — reject canonical names, so use the legacy aliases above.

Capability legality

Capability legality depends on the agent’s locality and the item’s storage mode (ZK = zero-knowledge, SM = server-managed):

Legacy capability	Canonical	Local + ZK	Local + SM	Remote + ZK	Remote + SM
`read_ciphertext`	`read`	✓	✗	✗	✗
`reveal_plaintext`	`read`	✗	✓	✗	✓
`mount_env` / `mount_file`	`use`	✓	✓	✗ (no daemon)	✗ (no daemon)

Profile-target grants are not available from the CLI. Use the dashboard’s profile-grant flow — it surfaces the blast-radius confirmation before granting canonical read on a whole profile.