Agent + git diff + 4 other enricher sources.
Fetch: git diff, files, websites, connectors, agent input (text transcript)
Process:
- We first combine git diff and agent transcripts to make a first pass.
- Then we analyze and query the AI model if it would add additional value if we read files, crawl the web or add info from GitHub.
- If the branch we work on has a PR, we look for issues mentioning this PR, and if these issues might have info we need, we can mention this issue in the commit if it's relevant.
Human source is not used in CLI commit.
We do the same as CLI commit with the only difference that we don't use agent input and instead use human input.
With human input we can add prompt instruction via text input when pressing the preview button with options. This allows the user to emphasize how the commit description and title should be written.
There is also a way to touch-up the commit description after the preview has been generated, by selecting text and pressing the right mouse button to show the context menu, then pressing "Edit with AI". This will present the user with a prompt text input.
We can make PRs by selecting source and target branches. The commit diff between these two branches constitutes the body of the PR.
Firstly we try to analyze what the commits accomplished. For smaller models like SFM and others we heavily truncate and interpolate. Using larger models yields better results that provide more nuances etc. Then we ask the AI if running passes with external sources can improve the PR description. We also look at the agent desktop transcript and combine this with the branch diff/commit text.
The app does the same as the CLI PR routine except accepting agent input. In the app we can also provide user input via the prompt input on the preview button. And once the text has been generated we can then use the AI touch-up feature to tweak things and pull in external knowledge from connectors we point to: add design tokens from Notion for instance, or reference similar PRs or relevant issues, or add API docs from a GitHub repo via the web crawler.
To generate changelog, select the next tag and all commits from previous to current tag. Using the app to generate changelog is a great way to make releases.
For CI/CD we can use the release feature in CI. To learn more about how that works please visit this link.