[RFC] Auto Roadmap Generation

[tzhouam]

let the agent be able to auto dive the relationship of feature/rfcs/prs to build interactive roadmap service like https://world-model.43.155.186.30.nip.io/roadmap for better project management

[tzhouam]

Enriched into a full RFC in #157 (doc/RFC-auto-roadmap.md).

The measurements moved the design. Over 300 merged PRs in vllm-project/vllm-omni (2026-08-17 → 09-11, ~12/day), deriving feature↔PR relationships from traffic does not work: 2 of the 25 PRs curated onto #7074 cite that roadmap at all (8% recall), title keywords give 15% precision, the diffusion label 6%. A human picked 8 of 300. The taxonomy is judgment, and it was never written into the artifacts.

What is derivable without a model is status — and that is where all the cost is: #7074 took 88 body edits in six days, and #6494 tracks 129 numbers, was last edited 2026-09-05, and has had five tracked PRs merge since. Of 19 roadmap/tracking issues here from 15 authors, 8 are untouched for two weeks or more.

So the agent maintains and extracts rather than inventing: enrollment by a marker the issue author places, bootstrap as quote-gated extraction from the prose 19/19 roadmaps already contain, status refreshed every poll cycle with no model call, and attachment left as a proposal a maintainer accepts.

One thing worth flagging for review: the recurring lane writes a comment the bot solely authors, not the issue body. GitHub has no conditional update for an issue body, so a read-check-write against a body edited 24×/day can destroy a maintainer's edit — and could restore an enrollment marker someone had just deleted, defeating revocation.