problem
we make isolated LLM calls to generate PR title/body and commit messages.
these generations lack context - they use only the diff to write content.
this often results in misinformed PR descriptions. see PostHog/posthog#53451 for example:
- one-line change to add
LIMIT 50000 to a query
- the intent was to override the implicit default
LIMIT 100 that is applied to hogql queries
- the agent in posthog code was fully aware of this intent
- BUT the original generated PR description said something along the lines of "adding a limit to prevent unbounded queries" (which was actually the opposite of my goal)
desired solution
some way to pass more context to the content generators. ideally the agent just does it.. but imo it'd be a cleaner UX if those prompts/responses didn't end up in the chat logs, so i'm not 100% sure what the best option is.
maybe some sort of sub-task that forks -> summarizes/compacts -> reads summary + diff to generate?
problem
we make isolated LLM calls to generate PR title/body and commit messages.
these generations lack context - they use only the diff to write content.
this often results in misinformed PR descriptions. see PostHog/posthog#53451 for example:
LIMIT 50000to a queryLIMIT 100that is applied to hogql queriesdesired solution
some way to pass more context to the content generators. ideally the agent just does it.. but imo it'd be a cleaner UX if those prompts/responses didn't end up in the chat logs, so i'm not 100% sure what the best option is.
maybe some sort of sub-task that forks -> summarizes/compacts -> reads summary + diff to generate?