initial idea
This commit is contained in:
commit
536b308096
5 files changed
+1092
No files matched your search
@@ -0,0 +1,20 @@
|
||||
Gral auto. It's gral but better, keke.
|
||||
|
||||
gral is an existing chat, with advanced features for terminal based interactions and constrained context. I don't want to kill it, it's really good at what it does.
|
||||
gral-auto is a single-run utility that runs a preprogrammed script, spawning agents with configurable tools and resources for concrete runs with defined deliverables.
|
||||
it must provide an api to a '.gral/auto.js' script in the working directory and execute it. This api allows spawning AI agents on preprogrammed tasks.
|
||||
for now just a prompt. we'll add more utilities later. it would be good that the design were extensible.
|
||||
|
||||
actually, we will begin with this feature set also:
|
||||
- each run has these cofigurable options:
|
||||
- agent name, or agent id?
|
||||
- write/read permission: file or directory list where writing and reading is enabled, as separate lists or in the data structures
|
||||
- file context: list of initially open files
|
||||
- verification callback: runs after the agent did the task, can either succeed, fail, or retry with modifications or feedback, and it's optional
|
||||
- feedback callback?
|
||||
- functions
|
||||
- agents only have basic file operations, restricted by the default permission system
|
||||
- functions can be defined in the auto script, which are exposed to the agent besides the default functions
|
||||
|
||||
technical details: functions are implemented as remotes. a remote is created and connected for each task, and the agent is connected to the remote via the chat.
|
||||
also, agents are managed via chat interactions. A single temporary chat is created for the auto, and each run creates a thread in the chat to interact with the agent.
|
||||
Reference in new issue
Block a user