LLM policy

[Imre Samu]

Greg Troxel <gdt at lexort.com> ezt írta (időpont: 2026. jún. 9., K, 12:16):

> Regardless of how the consensus ends up



Since consensus on a main-repository AI/LLM policy seems difficult right
now, I propose an experimental compromise:
- create *a separate "postgis-agentic-lab" repository* !
It would allow AI/agentic experimentation without adding noise or review
pressure to the main postgis repo.
It would create no obligation for maintainers, but useful results could be
manually extracted by a human postgis maintainer and submitted upstream
through the normal process.

So some postgis users may be able to contribute not only bug reports, but
also AI-assisted candidate solutions. ( Fable/Mythos welcome )
Even if these are not accepted upstream, they may still be useful to other
postgis users with similar issues.

Imre / wearing my postgis power user hat /


=====================

*Proposal: ~ `postgis-agentic-lab`*
( brainstormed with Chatgpt-5.5-High )

The goal is to ensure that AI-assisted work demonstrably reduces, rather
than increases, the maintenance burden on PostGIS maintainers.

*Main PostGIS repository*
- Only human contributors may submit upstream issues, pull requests, review
comments, or mailing list messages.
- *Existing trusted postgis commiters/contributors may use AI assistance f*or
small, reviewable changes, preferably under a soft limit such as ~50-100
changed lines. ( new contributors with AI --> `postgis-agentic-lab` )
-  Substantial AI assistance must be disclosed, including the tool/model
used where relevant.
-  The human contributor remains fully responsible for understanding,
testing, licensing, and maintaining the contribution.
-  Any `AGENTS.md`, skills, prompts, or agentic instructions in the main
repository should clearly say *that agentic experiments belong in
`postgis-agentic-lab`,* not in the main repository. ( including postgis
end-users agentic code suggestions , fix )
- Autonomous agents must not act directly in official PostGIS project
spaces.

*Postgis-Agentic-lab  repository:*
-  Create a new fork* `github.com/postgis/postgis-agentic-lab
<http://github.com/postgis/postgis-agentic-lab>*` as* an experimental AI
staging/lab repository for postgis*
-  The lab is for AI-assisted reproduction, testing, analysis,
benchmarking, patch prototyping, prompt sharing, and agentic workflow
experiments.
-  The lab is continuously synchronized from the main PostGIS repository,
but nothing flows back automatically.
-  AI agents and human+agent workflows may create issues and pull requests
there under special rules.
*-  Work in the lab creates no obligation for PostGIS maintainers to
review, respond, accept, or merge anything.*
-  Stale lab issues/PRs without a human owner or clear progress may be
*automatically
closed after 30 days.*
*-  Agentic tools such as openclaw, hermes, or similar systems should work
in this lab repository, not directly against the main PostGIS repository.*

*Supported use cases in the lab*
- Generate a minimal, reproducible, human-friendly bug report from a
real-world problem ( by the postgis-end users or agents  )
- Sanitize and reduce private or company data into a safe public reproducer.
- Create a failing new PostGIS regression test as a separate commit.
- Analyze the root cause and document competing hypotheses.
- Create a detailed fix prompt for a general AI coding agent.
- Produce a candidate patch as a separate commit.
- Document test results, benchmark methodology, and performance impact.
- Share useful prompts, agent instructions, negative results, and failed
approaches.
- ....

*Contribution categories in the agentic-lab:*
- "Supported": reproducible bug reports, minimized test cases, benchmarks,
root-cause analysis, documentation of methodology.
- "Accepted for human upstreaming":  small, tested, human-owned patches
with clear provenance and AI disclosure.
- "Tolerated / prototype" : speculative fixes, large refactors,
experimental prompts, agent-generated ideas, incomplete prototypes.
- "Forbidden": autonomous / generated spam, hallucinated test results,
undisclosed private data, unclear license provenance, or any change the
human owner cannot explain.

*Upstream promotion rule*
- The lab is not an official upstream contribution queue.
- Only a human trusted postgis committer may manually prepare and submit a
final PR to `postgis/postgis`.
- The upstream PR must be small, reviewable, tested, disclose substantial
AI assistance, and link back to the relevant lab artifact if useful.
- Ideally, the upstream PR should contain the prompts and:
     - one commit for the failing test,
     - one commit for the fix,
     - optional documentation or benchmark notes.

etc.
=====
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/postgis-devel/attachments/20260616/d5210e4c/attachment.htm>