Improve context by providing a repo map; Add Vertex AI integration; Add Greedy action parser for more robust code block action syntax

[Henri-Laiho]

I've been experimenting with the SWE-Agent and made some improvements on the go that I thought would be worth considering adding to the project

What does this implement?

• Repo map implementation from Aider
• Google Vertex AI integration
• A greedy thought-action parser to handle files that contain triple backticks (```)

Repo map

When using the agent on an existing codebase I encountered the issue where the agent starts to reimplement existing code, to give the LLM context for avoiding this I have added the repo map code from Aider
The Repo map feature adds a snippet like this to the first user message

Here is a map of your working directory's latest state, showing files and signatures in code files:

.gitignore

readme.md

requirements.txt

sonoff_diy_api/smart_socket_mock.py:
│class SmartSocketMock:
│    def __init__(self):
│        # Initialize device state and attributes
│        self.switch_state = "off"  # Initial state of the switch
│        self.startup_state = "stay"  # Power on state
│        self.signal_strength = -70  # WiFi signal strength
│        self.pulse_state = "off"  # Inching state
│        self.pulse_width = 500  # Inching pulse width
│        self.ssid = "sonoffDiy"  # WiFi SSID
│        self.password = "20170618sn"  # WiFi password
⋮...
│    def switch(self, switch_state):
⋮...
│    def startup(self, startup_state):
⋮...
│    def signal_strength(self):
⋮...
│    def pulse(self, pulse_state, pulse_width=None):
⋮...
│    def wifi(self, ssid, password):
⋮...
│    def ota_unlock(self):
⋮...
│    def ota_flash(self, download_url, sha256sum):
⋮...
│    def info(self):
⋮...

Google Vertex AI integration

Added a Vertex AI integration for access to 1M context window models

Greedy parser

When the agent tried to write documentation in markdown files I encountered the issue of the action parser failing to read file with triple backtick (```) blocks. To fix this I added the ThoughtActionGreedyParser class that parses the string between the outer backticks as the only action, as multiple action outputs seem to be unsupported. I know that a workaround would be using an xml-agent, but that might be too complex syntax for smaller LLMs

Any other comments?

I hope any of these features would be of use. Let me know if you would like me to cherry-pick and polish any of them to a separate PR

I have not had the time to write tests and make sure these features smoothly integrate with everything else

🧡 Thanks for contributing!

[ofirpress]

Hi!
Thanks for making a PR. Have you run any of these 3 new features on SWE-bench to see if they improve the accuracy?

[Henri-Laiho]

Hi! Thanks for making a PR. Have you run any of these 3 new features on SWE-bench to see if they improve the accuracy?

Hi!
No, I haven't run on SWE-bench yet. I opened the PR to share the code I added during my testing, and could be helpful to the project.
I will run on SWE-bench once I have more time

[klieret]

Repo map was since independently added. Closing this, because we didn't end up moving in this direction.