| name | discover |
| description | Discover gaps, clusters, and stale findings across the research corpus — dimensions with thin coverage, concepts and entities that co-occur through graph edges, and findings that are aging or carry a weakened/quarantined verdict. Use this skill when the user asks 'where are my gaps', 'what's stale', 'what clusters exist', 'research coverage', 'what am I missing', 'audit the corpus', 'what needs updating', 'coverage matrix', 'stale findings'. |
| version | 0.4.2 |
| argument-hint | [--gaps] [--clusters] [--stale] [--all] |
| allowed-tools | Read, Bash, Grep, Glob |
discover — gaps, clusters, and stale findings
Audit the corpus for actionable structure. All three analyses derive from the
MIF substrate — the projected index and the MIF-native graph — never from
tag-co-occurrence recomputation.
Data sources:
research-index.json — flat per-finding projection (id, title, namespace,
dimension, tags, verdict, citations).knowledge-graph.json — MIF concept/entity nodes and typed relationships[]
mentions edges (all urn:mif: ids).
harness.config.json — the authoritative, config-declared dimensions[].
Arguments
--gaps — dimensions with few or zero findings (against the config taxonomy)--clusters — concepts/entities that co-occur via graph edges--stale — findings that are aging or carry a weakened/quarantined verdict--all — all three (default when no flag is given)
Coverage gaps
The dimension taxonomy is config-declared, never fixed. Read it from
harness.config.json, then count index findings per dimension; a dimension with
no (or very few) findings is a gap.
jq -n \
--slurpfile cfg harness.config.json \
--slurpfile idx reports/_meta/sample-session/research-index.json '
($idx[0].findings | group_by(.dimension)
| map({ (.[0].dimension // "unassigned"): length }) | add) as $counts
| $cfg[0].dimensions
| map({ dimension: .id, findings: ($counts[.id] // 0), description })'
Present as a table; flag any dimension at 0, or far below the corpus mean, and
recommend /start --augment <dimension> --topic <topic> to fill it.
Clusters
A cluster is a set of findings tied together through the graph — they share a
referenced entity (mentions edges to the same urn:mif:entity: id) or sit on
a chain of typed relationship edges. Group by shared entity first:
jq '[ .edges[] | select(.via == "entity")
| { entity: .target, finding: .source } ]
| group_by(.entity)
| map(select(length > 1)
| { entity: .[0].entity, findings: map(.finding), size: length })
| sort_by(-.size)' \
reports/_meta/sample-session/knowledge-graph.json
Name each cluster from its shared entity's node label. Highlight clusters that
bridge findings of different dimensions — those are the cross-cutting themes.
Stale findings
Staleness has two independent signals; report both.
-
Verdict-based (from the index). Any finding whose verdict is weakened,
falsified, or quarantined is epistemically stale regardless of age:
jq '[ .findings[] | select(.verdict != "survived" and .verdict != null) ]' \
reports/_meta/sample-session/research-index.json
-
Age-based (from the MIF findings, not the index). The index carries no
timestamp; created and MIF temporal/decay metadata live in the finding
files. Read the findings dir to flag old or low-temporal-strength findings:
jq -r '[ .["@id"], (.created // "n/a") ] | @tsv' \
reports/_meta/sample-session/findings/*.json | sort -k2
Flag the oldest created dates (and any decayed temporal strength) for
re-validation.
Present each stale finding with its id, title, the reason flagged, and a
recommendation: /update <topic> to refresh, or /falsify to re-test.
Output format
For --all, three headed sections — Coverage gaps, Clusters, Stale findings —
then a one-line summary: "N gaps, M clusters, P stale findings."
Prerequisites
research-index.json and knowledge-graph.json for the corpus, plus the
findings dir. If missing:
bash scripts/build-index.sh <findings-dir> && bash scripts/build-graph.sh <findings-dir>.
Examples
/discover
/discover --gaps
/discover --stale
/discover --clusters
