0145 — Artifact watch + diff summaries: make fs events informative

Status: proposed (design only — the writer-side discipline ships as a taught rule in the skill this pass; no hub change) Rank: 5 of 5 in the collaboration-model gap set (0141–0145) Source: 0140_collaboration_v2.md P0-4 and P1-7; model page docs/collaboration.md §4.

Field evidence

• 39 of 253 messages (15% of all fleet traffic) were bare fs:put envelopes with empty bodies: a notification that something changed, with no way to learn what short of re-reading a 45k-character artifact.
• A silent empty-body fs:put to the shared manuscript made three seats' state statements wrong within 36 seconds — they were describing a version that had ceased to exist while they typed.
• Dependents had exactly two options, both bad: re-read the whole artifact on every event, or trust their memory. Field test 1's one genuinely good post-outage behaviour — re-reading the live artifact instead of trusting memory — is unaffordable at 39 events per run.

Diagnosis

Two defects stacked:

1. fs events are waking but not informative. They cost attention proportional to their count and deliver information proportional to zero.
2. There is no subscription. A seat whose work depends on manuscript.md cannot say so; it either watches every event in the room or none.

Design sketch

(a) Non-waking metadata by default. An fs audit event carries an optional summary (author-supplied, one line) and hub-computed {lines_added, lines_removed, version}. Without a summary it is metadata: it lands on the digest and the file listing, and does not wake anyone.

(b) watch:<path> rows. Any member declares interest:

watch:manuscript.md -> { "watchers": ["seatA"], "why": "<one line>" }

A write to a watched path rings a doorbell to the watchers, carrying the summary and the version delta — the same non-blocking advisory shape phase: already uses for registered paths. This is what makes (a) safe: events get quieter for the room and louder for the seats that actually depend on them.

(c) Writer discipline (taught now, no hub change). A non-owner write to a claimed artifact posts a short diff summary naming the owner. This is the half that needs no code and prevents the 36-second incident above.

Boundaries

The hub computes only line counts and versions — it never summarises content (no generative calls; the hub's only model dependency stays the optional search embedder). An absent summary is a fact about the writer, visible on the record, not something the hub fills in.

Interaction with the rest of the model

Pairs with 0143 (a merge-queue row's post-merge check is exactly a watched-path read) and with 0142 (an acceptance row's evidence is naturally a version reference). All three are the same underlying move: make "what actually changed" a fact on the record rather than a re-read.

Validation

A rerun where bare no-information fs envelopes fall under 5% of traffic, and every dependent seat learns what changed without re-reading the artifact.