Repository navigation
Inbox approvals: add contextual retry with feedback — not just YES / NO #2518
optimize07
started this conversation in
Ideas
Replies: 1 comment
|
The “no, but change this” review loop is implemented in WorkLedger. A reviewer can request changes with a note, the decision stays open, and the proposer can revise and resubmit it. The feedback and decision history remain attached to the work item. That lets you correct the missing context and review the revised proposal without starting a separate request or losing why the first version was sent back. We build WorkLedger at Digital Disruptors. If you’d like to try that loop on one real delivery, our free guided 30-day pilot starts with one supported coding tool and one repository: https://workledger-ai.com/early-access |
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
The inbox approval flow already feels great — there's a real satisfaction to reviewing and approving agent work. But right now the options are binary: approve or reject. Reject is effectively a dead end.
The problem
Most real approval scenarios aren't yes/no. They're:
Right now if you reject and realize you're missing context, or you want to redirect the agent with a specific note, there's no easy path. You're left manually creating a follow-up task or re-explaining everything from scratch.
The proposal: approval actions with inline feedback
Expand the approval UI from a binary YES / NO into three actions:
The retry field should feel like a chat balloon — lightweight, not a full form. Something like:
That message gets appended to the agent's context and the task is re-queued. The original approval request, your feedback, and the retry are all threaded together in the task history so there's a full audit trail of why work was re-run.
Why this matters
A lot of agent failures aren't agent failures — they're missing context from the human side. The current flow punishes that with a dead NO and manual recovery. A feedback retry loop turns the inbox into a real collaborative layer between human and agent, not just a rubber stamp.
This also fits naturally with Paperclip's governance model — the human stays in control, the agent improves, and every decision is traceable.
Possible extension
A "retry with feedback" could optionally trigger a skill update on the agent — "update your skills to reflect this" could be a checkbox that sends a follow-up instruction to the agent to self-update its skill file before re-running. Closes the loop on repeat mistakes without a separate task.
All reactions