Repository navigation
Replies: 4 comments
|
The same need comes up when the orchestrator is not a T3 thread at all. I drive threads from a controller process over What I see on t3 0.1.9, macOS, Claude provider:
From reading the code, any process that runs as the machine's owner and can read T3's data directory can do the same, an agent in a For this case, the API the idea proposes would need two properties:
Acceptance for my case: launch a thread in Related: #15464 added approval responses over MCP, and #22152 reverted it for missing maintainer sign-off on a permission change. |
|
I use T3 Code every day and run into this constantly, so I'd like to add a second real-world case. My setup: one supervising thread (a Claude planner) launches and steers several long-running Codex executor threads (top-level threads started with So every time an executor needs approval, all work in that thread stops until I click Approve by hand, even when the supervisor already reviewed and requested that exact command. In my case these are mostly guarded deletes of the executor's own scratch copies, which still prompt in Full access (see #37298). What would help:
Today the only alternatives are clicking Approve every few minutes or giving unrestricted access, which is exactly the gap this idea describes. |
|
A concrete pointer, from reading the source at 7b44788, in case it helps whoever picks this up:
So a small upstream step could be: let |
|
Same need here, from outside T3 entirely. I launch top-level threads over The smallest version that would unblock me is decline-only:
I filed this as #39166 by mistake (it's a feature, not a bug), so I'm moving it here. For whoever picks it up: #15464 added MCP approval responses, and #22152 reverted it pending review. |
Uh oh!
There was an error while loading. Please reload this page.
Problem
When delegating tasks to another agent, command execution approvals can become a bottleneck.
Currently, the available permission modes do not fully address a workflow where the parent agent should be allowed to review and approve commands requested by a delegated child agent, without granting the child agent unrestricted execution privileges.
Using
Full accessremoves too many safeguards, while requiring human approval for every routine command defeats much of the purpose of delegation.Real-World Example: Claude Code → OpenCode
Consider this workflow:
Claude Code delegates the task to OpenCode so it can focus on higher-level work.
OpenCode may need to execute commands such as:
These are generally low-risk, read-only operations.
However, I do not want to enable
Full accessfor OpenCode simply to avoid repeated approval prompts.Instead, I would like the following behavior:
This would allow the parent agent to act as a lightweight permission reviewer while preserving the separation between orchestration and execution.
Proposed Feature
Delegated Command Approval by Parent Agent
Introduce an optional permission mode or policy for delegated agents, for example:
Delegated approvals: Parent agent reviews / Human escalationUnder this mode:
Full access.Approval decisions should be scoped to individual requests rather than granting broad, persistent permissions.
Ideally, the parent should receive enough context to make an informed decision, including the command, working directory, affected paths, and relevant execution metadata.
Expected Workflow
Safety Considerations
This should be an explicitly enabled feature, not the default behavior.
Suggested safeguards:
Full access.Why This Matters
The parent agent already understands the delegated task's objective and context.
For simple but tedious tasks, requiring the human to repeatedly approve routine commands adds unnecessary friction.
At the same time, granting
Full accessto a delegated agent is often excessive.The goal is to delegate limited approval authority, not unrestricted execution authority.
This would make mixed-provider workflows such as Claude Code → OpenCode significantly more practical while retaining meaningful permission boundaries.
Related Issues
This request goes one step further than notifying the parent agent about pending approvals: it proposes allowing the parent agent to explicitly resolve those requests within a constrained permission policy.
All reactions