library-push

slug	library-push
name	Library Push
description	On every new TM or Hallway entry, search prior memoranda for relevance and push the top-K to the relevant researcher's queue as a citation suggestion. Not search; push.
version	0.1.0
metadata	{"sources":[{"mode":"port-original","author":"Jannes Stubbemann","added_in":"0.1.0"}]}

Library Push

Port-original skill. Hand-authored for this Agent Company; not from upstream. Owned by the Librarian role.

When to fire

On every new TM, Hallway entry, or colloquium section posted by any researcher. The Librarian watches the artifact stream; no explicit request is needed. This is a push service, not a reference desk — the Librarian initiates, researchers do not summon.

Inputs

• The new artifact: the TM, Hallway entry, or colloquium section that just arrived. Full text.
• The prior corpus: all TMs under memos/ and all Hallway entries under hallway/. Scan the full corpus each run; recency and overlap both affect scoring.
• Manifest metadata: manifest.yaml metadata.library_push_k field sets K. Default K=3 if the field is absent.

Outputs

Push notes appended to relevant researchers' agents/<role>/library-suggestions.md. Each push note contains:

• The artifact being cited (title, path, date).
• One sentence explaining why this prior artifact is relevant to the recipient's current work.
• Posted to researchers other than the artifact's author.

Up to K suggestions per push event (hard cap). Quality determines which K are selected; do not pad to K if fewer strong matches exist.

Procedure

1. Extract key terms from the new artifact. Identify the 5–10 most distinctive concepts, named objects, or research questions in the artifact. Use the artifact's own language; do not normalize to a taxonomy.

2. Search the prior corpus for relevance. Match key terms against TM abstracts, Hallway entry bodies, and colloquium sections. Cast wide on the first pass; prune in the scoring step.

3. Score candidates by recency × topical overlap. Recency: artifacts from the last 90 days score higher than older ones. Topical overlap: number of matching key terms. Combine multiplicatively. Ties broken by recency.

4. Select the top-K. K is the value from manifest.yaml metadata.library_push_k; default 3. If fewer than K candidates score above zero, push only those that do. Never pad with weak matches.

5. Write push notes to relevant researchers' library-suggestions.md files. A "relevant researcher" is any researcher currently working on a thread that overlaps with the artifact's key terms — identified by their recent Hallway entries or queue contents. The author of the new artifact is excluded. Each note is a short block: artifact title + path + date, then the one-line relevance justification.

Invariants

• Push, not search. The Librarian initiates the push on every new artifact. Researchers do not file requests for their own citation queue; the service is proactive. A Librarian that waits to be asked has misunderstood its role.
• K is a hard cap. The default is 3; manifest.yaml can raise or lower it. Never exceed K, regardless of how many plausible matches exist.
• Push to other researchers, never the author. If the artifact's author is the only plausible recipient, post nothing. The push enriches the network, not the author's own awareness of their own work.
• Every push note includes a one-line "why this is relevant." The relevance sentence is not optional. A bare citation is noise; context is what makes the push useful.

Anti-patterns

• "Push everything you find." K is a cap because quality matters. Flooding a researcher's suggestion file with ten marginal citations degrades the signal. Pick the best K; discard the rest.

• "Push to the agent who wrote the artifact." The author already knows their own work. Push flows to neighbors in the research graph, not back to the source. If the only match is the author, skip the push.

• "Require researchers to request suggestions." The Librarian is a push service. The entire value of this role comes from surfacing connections researchers did not know to ask for. A passive Librarian that waits for requests is a search engine, not a routing layer.

• "Skip the one-line 'why this is relevant'." Relevance without justification is noise. The researcher receiving the suggestion needs to understand immediately whether it is worth reading. A bare title and path forces them to read the artifact to judge its relevance — that is the Librarian's job, not the researcher's.