/radar — Ambient Intelligence Digest¶

Surfaces recent feed entries, synthesizes them into a grouped digest, optionally suggests new sources, and optionally appends a structural view of the knowledge graph (--structure).

Usage¶

/radar                                  # Default: last 7 days, up to 35 entries
/radar --days 3                         # Last 3 days
/radar --limit 10                       # Limit to 10 entries (overrides feeds.digest.candidate_limit)
/radar --topic "build hermeticity"      # Use an explicit topic instead of mining tags
/radar --topic agentic-eval --days 14   # Topic + custom window
/radar --topic build --topic wheels     # Multiple topics; one search query each
/radar --structure                      # Append orphans + bridges + communities
/radar --suggest                        # Include source suggestions
/radar --store                          # Store the digest as an entry (default: display-only)
/radar --include-evergreen              # Surface older / first-poll backfill items

Trigger phrases: "what's new", "show my digest", "ambient digest", "what have I missed", "feed digest"

Options¶

Option	Description	Default
`--days <n>`	Look back period in days (overrides `feeds.digest.window_days`)	`feeds.digest.window_days` (7 if unset)
`--limit <n>`	Maximum entries to include (overrides `feeds.digest.candidate_limit`)	`feeds.digest.candidate_limit` (35 if unset)
`--topic <query>`	Use the literal string as the semantic-search query instead of mining tags. Repeatable.	Off (auto-mine)
`--structure`	Append a Knowledge Structure section (orphans + bridges + communities)	Off
`--suggest`	Include new source suggestions	Off
`--store`	Store the digest as an entry	Off (display-only)
`--include-evergreen`	Include items older than the window or flagged as first-poll backfill	Off

The look-back window is bounded by metadata.published_at (the feed item's publication time), not the ingest timestamp. First-poll backfill batches — items pulled when a feed source is first registered — are flagged metadata.backfill = true and excluded from the default candidate set so they don't surface as "new intelligence". Use --include-evergreen to override.

Query selection¶

Two paths feed the candidate-search step:

Path 1 — --topic override. Each --topic <query> flag becomes one literal distillery_search call. Topics are deduplicated case-insensitively, preserving order. The 5-query namespace cap does not apply — every distinct user-supplied topic is honored, up to --limit. Tag mining is skipped entirely.

Path 2 — Namespace-diverse interest profile (default). The skill mines curated entries (session, reference, bookmark, idea, note, minutes) for a tag profile, then picks 5 namespace-diverse tags rather than the raw top-5 by count. A tag's namespace is its hierarchical path with the leaf removed, capped at two segments (e.g., domain/build/hermeticity → domain/build; tech/duckdb → tech). One leader per namespace prevents a dominant cluster from crowding out distinct topics. The reference implementation lives in distillery.feeds.radar_selection.select_namespace_diverse_tags.

Limit semantics and the 5-query default¶

The candidate budget is distributed across queries so the sum of per-query limits never exceeds --limit:

Compute Q (the query count): up to 5 for Path 2, up to min(distinct_topics, --limit) for Path 1.
Let base = limit // Q and rem = limit % Q. The first rem queries get base + 1, the rest get base.

With the default --limit 35 and Q=5 (Path 2), this yields 5 queries × 7 results each → ~30 unique candidates after dedup. For small overrides like --limit 3, Q is reduced to 3 so the override is honored exactly (no zero-budget queries are issued). If Q == 0 (e.g., --limit 0 or both paths yielded an empty set), the search step short-circuits and falls back to a recent-listing query.

Configure the system-wide default via feeds.digest.candidate_limit in distillery.yaml:

feeds:
  digest:
    window_days: 7
    candidate_limit: 35

Both YAML loading and the MCP distillery_configure tool validate this knob the same way (positive integer); --limit overrides it per-invocation.

`--structure` flag — Knowledge Structure section¶

When --structure is set, an additional section is appended to the digest with three subsections:

Orphans. Up to 10 entries with no incoming or outgoing relations, surfaced via distillery_list(structural=["orphans"]). These are knowledge fragments worth reviewing for connection opportunities. Empty list renders as No orphan entries — every entry has at least one relation.

Bridges. Top-5 entries by betweenness centrality in the entry-relations graph, surfaced via distillery_relations(action="metrics", metric="bridges"). Scores are rendered to 3 decimals (e.g., 0.412). High-centrality entries are knowledge "joints" — losing or contradicting them disconnects the graph the most.

Communities. Detected clusters in the entry-relations graph, surfaced via distillery_relations(action="metrics", metric="communities"). Communities are sorted by total_members (largest first). For each community, the line shows the first three member ids: Community 1 (12 entries): <id1>, <id2>, <id3>. If every member has updated_at < now - 60 days, the community is tagged [stale].

Bridges and communities require the [graph] extra (NetworkX). If the MCP server returns INTERNAL with "NetworkX not installed", /radar emits a single one-line note — Run \pip install distillery-mcp[graph]` to enable bridges/communities.` — and continues without the structural metrics. The orphans subsection does not require NetworkX.

Output¶

# Radar Digest — 2026-05-07

12 feed entries from the last 7 days.

## DuckDB
DuckDB 1.2 shipped with improved HNSW index performance and a new
JSON extension for nested document queries.

- v1.2 release includes 3x faster vector search indexing
- New JSON extension supports path-based queries

## Claude Code
Two new features landed: hook-based automation and MCP server improvements.

- Session hooks enable custom automation on start/end events
- MCP server now supports parallel tool execution

## Overall Summary
This week saw performance improvements in DuckDB's vector search (directly
relevant to Distillery's storage layer) and new automation capabilities
in Claude Code.

## Knowledge Structure

### Orphans
- a1b2c3d4 — One-off reference doc on token rotation
- e5f6g7h8 — Standalone bookmark, no follow-up captured

### Bridges (top by betweenness centrality)
1. m1n2o3p4 — OAuth design session (score: 0.412)
2. q1r2s3t4 — DuckDB migration plan (score: 0.387)

### Communities (ordered by total member count; top 3 members shown per community)
- Community 1 (12 entries): m1n2o3p4, q1r2s3t4, u1v2w3x4
- Community 2 (8 entries) [stale]: y1z2a3b4, c1d2e3f4, g1h2i3j4

Tips¶

Digests are display-only by default; pass --store to persist with tags digest/radar/ambient for future retrieval
--topic is the right flag when you want to follow a specific thread and you don't want the digest diluted by your dominant tag clusters
--structure is most useful as an occasional pulse-check, not every run — bridges and communities don't move much day-to-day
Source suggestions (--suggest) are based on the same query set used in Step 3a
Use /watch add to subscribe to suggested sources