Map T3 Permissions onto Droid's autonomy levels

[factory-ain3sh]

Droid threads in T3 Code show two permission controls, T3's Permissions and Droid's own Autonomy Level, and both apply. T3 already handles this for Codex, Claude and Grok: Permissions sets the harness's own permission mode, and the harness's picker stays hidden. Would you take a PR that does the same for Droid in the factory-droid registry flavor?

Today

sequenceDiagram
    actor U as User
    participant T as T3 Code
    participant D as Droid
    U->>T: Permissions: Full access (T3's default)
    U->>T: Autonomy Level: Auto (Medium)
    T->>D: session/set_config_option autonomy_level = auto-medium
    Note over D: runs what Auto (Medium) allows
    D->>T: session/request_permission for rm -rf ~
    Note over T: Full access allows every tool kind
    T-->>D: allow_once
    Note over U,D: Droid's deny list requires a person to approve this. Nobody saw the ask.

Loading

Auto works the same way, since T3 answers asks by tool kind and allows every kind there. And Plan looks for a mode named plan or architect, so it never puts Droid in its read-only spec mode.

Proposal

sequenceDiagram
    actor U as User
    participant T as T3 Code
    participant D as Droid
    U->>T: Permissions: Full access
    T->>D: session/set_mode auto-high
    Note over D: runs what Auto (High) allows
    D->>T: session/request_permission for rm -rf ~
    T->>U: approval prompt
    U-->>T: approve or decline
    T-->>D: allow_once or reject_once

Loading

T3 Permissions	Droid level	Mode id
Supervised	Auto (Off)	normal
Auto-accept edits	Auto (Low), which also runs low-risk commands	auto-low
Auto	Auto (Medium)	auto-medium
Full access	Auto (High)	auto-high
Plan	Spec, and Build restores the level	spec

When an org caps Droid's autonomy, the mapping stops at the highest level the org allows.

The flavor change is small:

• sessionModeForPolicy returns the mode id from the table, as Antigravity's does.
• permissionDisposition returns "ask", as Grok's does in Auto. Droid's deny list asks at every level, even Auto (High), so under the mapped level every ask Droid sends is one its own rules left to a person.
• approvalOptions drops Droid's "Allow & auto-run" answers, since they would raise Droid's level behind Permissions' back.
• Two small generic hooks: one lets a flavor name its plan mode (Droid's is spec), and one lets a registry flavor hide an option it owns, so the Autonomy Level picker goes away. Model and reasoning options stay.

I work on Droid at Factory, and I'm happy to send the PR, scoped to this flavor and those two hooks, with tests and a live run against Droid. Thanks for taking a look!