AnswerDotAI / AnswerDotAI/llms-txt
Proposal: Extend llms.txt to installed packages to support local filesystem resolution
- Dominant language
- Jupyter Notebook
- Stars
- 2.6k
- Forks
- 163
- Avg merge
- 1m
- Merged PRs (30d)
- 1
Description
llms.txt already targets the developer use case; the spec explicitly mentions "coding library documentation" and "development environments" as key scenarios. The convention works well when an agent is browsing the web for information or docs.
However, when a coding agent is working locally with an installed dependency, the package is already on disk. The agent already resolved the import to a directory. A web fetch for information is less efficient than co-location with the code. It is slower, can introduce version mismatch, and unhelpful for private or internal-only packages.
**Proposed Extension**
Packages should ship a llms.txt at their root, and agents should read it when they encounter that package.
Resolution Rule:
import foo -> resolves foo's install path -> read /llms.txt if present
This mirrors the existing web convention: agents just resolve from /llms.txt rather than /llms.txt. The solution works for any ecosystem where an import resolves locally.
Package authors ship llms.txt alongside their source. Agent runtimes update to run a single stat prior to source exploration for an unfamiliar import. Private and internal packages have a non-domain dependent support path.
Contributor guide
No contributing guide indexed for this repository
Research direction
Start by reviewing the existing llms.txt specification and the documented package-author and agent use cases. Define how local package resolution, package versions, and agent behavior should be specified; done means the proposal has an agreed resolution rule and clear documentation for package authors and agent runtimes.
Written by the indexing model from the issue text.
Assessment
- Domain
- ai, developer-experience, documentation
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100