Problem
When subagents (spawned via the Task tool) run in dontAsk mode, they cannot prompt for permissions. Tools not in the user's allow list are auto-denied silently. This creates a frustrating pattern:
- User asks a question
- Claude spawns a subagent to research it
- Subagent needs a tool (e.g.,
WebFetch) that isn't pre-approved
- Tool is auto-denied, agent works around it or fails partially
- User only discovers this if the agent happens to mention the error
The current model forces users to either:
- Pre-approve a broad set of tools (reducing security)
- Accept that agents may silently fail on useful operations
Proposed Solution
Allow subagents to pause and forward permission requests to the foreground conversation:
- Agent hits a tool requiring permission
- Agent pauses execution
- Permission request surfaces to the main conversation
- User approves/denies
- Agent resumes with the result
Alternative/Complementary Approach
Provide a way to verify and pre-approve necessary permissions before launching an agent:
- Agents could declare what tools they might need
- Claude could check permissions before spawning
- User could be prompted to approve missing permissions upfront
Why This Matters
Launching background agents is efficient and smart. But if those agents are silently doomed by missing permissions, the efficiency gain is lost. Users end up with partial results without understanding why.
This would improve:
- Transparency: Users see when agents need additional permissions
- Flexibility: No need to pre-approve everything "just in case"
- Reliability: Agents don't silently degrade due to permission gaps
Problem
When subagents (spawned via the Task tool) run in
dontAskmode, they cannot prompt for permissions. Tools not in the user's allow list are auto-denied silently. This creates a frustrating pattern:WebFetch) that isn't pre-approvedThe current model forces users to either:
Proposed Solution
Allow subagents to pause and forward permission requests to the foreground conversation:
Alternative/Complementary Approach
Provide a way to verify and pre-approve necessary permissions before launching an agent:
Why This Matters
Launching background agents is efficient and smart. But if those agents are silently doomed by missing permissions, the efficiency gain is lost. Users end up with partial results without understanding why.
This would improve: