gptme
gptme copied to clipboard
feat(wip): started work on subagent and tools refactor
Related issues:
- https://github.com/ErikBjare/gptme/issues/61
- https://github.com/ErikBjare/gptme/issues/75
Questions
- Should these subagents be ran in the background? Waking the main agent when done?
From my notes
I should add a function that serves as a wrapper to chat, making a prompt that makes the agent give a JSON return value.
- Then give the LLM a tool to spawn subagents and check on their status.
- Give it a sleep/await tool to wait for one or more tasks to complete
- Feels like I could just prompt it to do this via the reply by providing examples
- Could also give the sub-LLM a tool to update the status.
- Low value, but would be nice UX with an indicator
Codecov Report
Attention: Patch coverage is 75.00000%
with 49 lines
in your changes are missing coverage. Please review.
Project coverage is 79.71%. Comparing base (
e256fc9
) to head (c7121c2
).
:exclamation: Current head c7121c2 differs from pull request most recent head cd7943b
Please upload reports for the commit cd7943b to get more accurate results.
Files | Patch % | Lines |
---|---|---|
gptme/tools/subagent.py | 46.29% | 29 Missing :warning: |
gptme/tools/__init__.py | 81.33% | 14 Missing :warning: |
gptme/prompts.py | 25.00% | 6 Missing :warning: |
Additional details and impacted files
@@ Coverage Diff @@
## master #79 +/- ##
==========================================
- Coverage 81.38% 79.71% -1.67%
==========================================
Files 27 29 +2
Lines 1896 2066 +170
==========================================
+ Hits 1543 1647 +104
- Misses 353 419 +66
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
Ended up refactoring how tools are defined and registered, including how their prompts are constructed from their tool specs. Tools are extremely modular now.