Implement agentskills for OpenDevin to helpfully improve edit AND including more useful tools/skills

[xingyaoww]

co-authored by OpenDevin

OpenDevin Skill Sets

This PR implements a skill/tool set agentskills for OpenDevin.

It is intended to be used by the agent inside sandbox.
The skill set will be exposed as a pip package that can be installed as a plugin inside the sandbox.

The skill set can contains a bunch of wrapped tools for agent (many examples here), for example:

• Audio/Video to text (these are temporarily solution, and we should switch to multimodal models when they are sufficiently cheap)
• PDF to text
• etc.

Inclusion Criteria

We are walking a fine line here.
We DON't want to wrap every possible python packages and re-teach agent their usage (e.g., LLM already knows pandas pretty well, so we don't really need create a skill that reads csv - it can just use pandas).

We ONLY want to add a new skill, when:

• Such skill is not easily achivable for LLM to write code directly (e.g., edit code and replace certain line)
• It involves calling an external model (e.g., you need to call a speech to text model, editor model for speculative editing)

Intend functionality

• Tool/skill usage (through IPythonRunAction)

# In[1]
from agentskills import open_file, edit_file
open_file("/workspace/a.txt")
# Out[1]
[SWE-agent open output]

# In[2]
edit_file(
    "/workspace/a.txt",
    start=1, end=3,
    content=(
        ("REPLACE TEXT")
))
# Out[1]
[SWE-agent edit output]

• Tool/skill retrieval (through IPythonRunAction) -- TODO for future PRs

# In[1]
from agentskills import help_me

help_me("I want to solve a tasks that involves reading a bunch of PDFs and reason about it")

# Out[1]
"Here's the top skills that maybe helpful to you:
- `pdf_to_text`: [documentation about the tools]
...
"

---

For now, I've migrated all the swe-agent's bash tools into Python (i.e., the agentskills) and wrote extensive unit tests to ensure those tools' robustness.

I've tested on the first 10 SWE-Bench dataset for now, it did not change the performance (3/10) compared to previous version (but i reduced the max turns to 30 - which saves costs). I plan to run on the full swe-bench lite to measure the changes.

[mczhuge]

I think it is a very important thing for agent self-improving, and I am very interesting on it.

[xingyaoww]

This is a dependency for this: #1995

[Shimada666]

What a great refactoring! The maintenance cost of plugins will be much lower in the future, and adding plugins will be much easier! I love it!

[rbren]

This LGTM overall!

Have we been able to run a mini-eval to see if there are any regressions on SWE-bench lite?

[xingyaoww]

@rbren Yep! I ran on the first 54 instances - This is 13/54, compared to the previous version which is 12/54. At least this is not make things worse :)

[xingyaoww]

and these improvements happens when we reduce the max iter from 50 to 30 -> cheaper to run yet maintains similar performance :)