Add AgentRejectAction across multiple modules #1615
Conversation
This commit introduces the AgentRejectAction class and integrates it across various modules and actions. It includes updates to READMEs, action definitions, and agent controllers to handle the new 'reject' action. This functionality will allow agents to properly signal task rejection.
Codecov ReportAttention: Patch coverage is
❗ Your organization needs to install the Codecov GitHub app to enable full functionality. Additional details and impacted files@@ Coverage Diff @@
## main #1615 +/- ##
=======================================
Coverage ? 39.05%
=======================================
Files ? 92
Lines ? 3731
Branches ? 0
=======================================
Hits ? 1457
Misses ? 2274
Partials ? 0 ☔ View full report in Codecov by Sentry. |
| @@ -105,6 +105,17 @@ def message(self) -> str: | |||
| return "All done! What's next on the agenda?" | |||
|
|
|||
|
|
|||
| @dataclass | |||
| class AgentRejectAction(Action): | |||
| outputs: Dict = field(default_factory=dict) | |||
There was a problem hiding this comment.
I don't think this should have an arbitrary dict of outputs, the way finish does. All we want is a reason
There was a problem hiding this comment.
I see where you are coming from, but I have a few counter-arguments:
- An agent can choose to have more than just a
reasonattribute. For example, it can suggest how to fix the task in addition to a reason. - Both AgentRejectAction and AgentFinishAction are termination actions, so it's better for them to have the same interface IMO.
- By design, a micro-agent has a schema for
inputsandoutputs. We don't really enforce it at the moment (i.e. a micro-agent can finish with whateveroutputstype), but I feel like we should do it sooner or later. IfAgentRejectActionandAgentFinishActioncould have different schema for outputs, things will become trickier.
There was a problem hiding this comment.
Will create a follow-up to enforce the schema for micro-agents. We could discuss if we want to continue this approach there.
|
Very cool! Agree this will be useful |
| { | ||
| 'action': AgentRejectAction(), | ||
| 'observations': [], | ||
| }, |
There was a problem hiding this comment.
The agent_controller will stop after the previous action AgentFinishAction. If this action is placed before, the same will happen. So, is this action needed?
There was a problem hiding this comment.
Probably not needed
This commit introduces the AgentRejectAction class and integrates it across various modules and actions. It includes updates to READMEs, action definitions, and agent controllers to handle the new 'reject' action. This functionality will allow agents to properly signal task rejection.
===========BELOW COMMIT MESSAGE IS GENERATED BY OPENDEVIN=============
Yes, again I am too lazy to write a commit message by myself. So I asked OpenDevin to do me the favour:
Why this PR is needed?
CommitWriterAgentnever finishes when running in a non-git folderThere is no diff in the staging area; no changes to commit.output fromCommitWriterAgentreally means there's no change to commit, OR the output itself is literally the compiled commit message.TODO in this PR:
UpdateThis should be a separate issue.backend_architecture.pumlFuture work: