ADR-041: Member Tool Permission Model
Date: 2026-11-16
Status: Accepted
Context
A member runs with a prompt someone else wrote. If that prompt can reach any tool the runtime has, a
misleading skill is a privilege escalation: it can push branches, open PRs, or write files the
reviewer never approved. allowed-tools narrows a skill's declared surface (ADR-028), but the
declaration alone is not enough: an omitted declaration must not mean "everything", and a new tool
must not be granted by default.
Three profiles existed (restricted, standard, trusted) whose difference was implicit in code.
Decision
A member's effective tools are the intersection of its allowed-tools declaration and its profile
grant. Both sides must allow a tool; either side can only narrow.
The profiles are cumulative:
| Profile | Read-only tools | Write tools | Trusted tools | Delegation |
|---|---|---|---|---|
restricted | yes | no | no | no |
standard | yes | yes | no | read-only roles only |
trusted | yes | yes | yes | read-only roles only |
Classification is a fixed allowlist in MemberAgent:
- READ-ONLY:
ask_human,file.read,file.search,code.explain,generate_image,speech_to_text,generate_video - WRITE:
file.write,code.write,code.refactor,text_to_speech(it writes a file) - TRUSTED:
pr_sync
Anything not classified is denied for every profile. That is the load-bearing rule: a new tool is
unusable until someone classifies it, so a tool cannot leak into a restricted member by being added to
the vocabulary. ask_human is read-only by definition; it asks, it does not mutate.
delegate_task is separate: opt-in per member spec (canDelegate), and limited to the read-only
analysis roles in DELEGATION_ALLOWED_ROLES. A write-capable agent is never a delegation target.
Alternatives Considered
| Option | Pros | Cons | Why Rejected |
|---|---|---|---|
| Profile only | one place to reason | a third-party skill could not narrow further than its profile | Rejected: least privilege is per skill |
| Declaration overrides profile | maximal skill control | a SKILL.md edit could grant pr_sync to a restricted member | Rejected |
| Deny-by-default with explicit grants per member | tightest | every member carries a full tool list that drifts from the vocabulary | Rejected: one classification table is easier to audit |
| Default-allow for unknown tools | never blocks a new tool | silently widens every member when a tool is added | Rejected: the failure is invisible |
Consequences
- Adding a state-touching tool requires classifying it, and the classification is a reviewable line in the diff.
- A third-party member starts read-only: its profile is chosen by the reviewer, not by its author.
- The enforcement point is the reasoning loop, before the tool executes: a denied call is a tool error the model sees and can recover from, not a crash.
- Hosts that build an
ExecutionContextby hand must wire the skill-scope hook; omitting it skips enforcement (ADR-032).