bab-ilu: Skill for Claude Code

/gap — Lens-aware gap discovery

Synopsis

/gap                             # tier 1; active lens; write candidates
/gap --tier=1                    # explicit tier 1 (default)
/gap --tier=2                    # activate tier 2 clustering (gated)
/gap --lens=<id>                 # override active lens
/gap --dry-run                   # print plan to stdout, do not write
/gap --vault=<path>              # non-default vault root

What /gap produces

Per .agent/spec/gap-algorithm.md (D2 landed):

1. .agent/todos/<tier_1_cluster>-candidates.md — one block per qualifying community emerging from tier-1 clustering. Format per D2 §3.3:

   ## Candidate YYYY-MM-DD-HHMM
   
   Atoms: [[a1]], [[a2]], [[a3]]
   Tier-0 spanning: [[x1]] (1657), [[x2]] (1900), [[x3]] (2019)
   Density ratio: 4.2
   Suggested name: "quiet-interior-light" (LLM suggestion — user may rename)
   Status: [ ] accept [ ] refine [ ] reject
   

   Under aesthetic-warburg: pathosformel-candidates.md. Under engineering-alexander: pattern-language-candidates.md. Under general-zettelkasten: concept-cluster-candidates.md.

2. .agent/todos/<tier_2_cluster>-candidates.md — tier-2 output, only when --tier=2 AND |tier_1_cluster| ≥ lens.thresholds.tier2_activation_min (default 15, or override warning when below).

3. wiki/questions/<slug>.md — one stub per unbridged gap per D2 §5. Bridge candidates are atom pairs in different communities with zero shared tier-0s and LLM semantic similarity > the lens threshold.

4. wiki/_insights.md — human-readable overview (inherited from graph_analyzer). Contains cluster summary, bridge candidates, gap list, bias signals.

5. Planned (v2.3): wiki/_index/insights.json — machine-readable companion output. Not emitted in v2.2.

How it works (one screen)

Parse args → resolve lens (honor --lens; else .agent/lenses/active/)
           ↓
Delegate to tools/gap_runner.py which:
  1. Imports graph_analyzer.{load_wiki, build_graph, compute_clusters,
                              compute_bridges, detect_gaps}
  2. Runs the pipeline under the lens (D1 lens-parameterized)
  3. Qualifies clusters against D2 §3.3 criteria:
     - |atoms in C| ≥ max{3, ⌈log₂(|atoms_total|)⌉}
     - each atom has ≥ max{3, ⌈log₂(|tier_0_total|)⌉} tier_0 exemplifiers
     - internal_density > lens.thresholds.density_multiplier × E[null_density]
  4. For each qualifying cluster, writes the Candidate block to
     .agent/todos/<tier_1_cluster>-candidates.md
  5. For each unbridged pair (community-i-atom, community-j-atom) with
     weight=0 and LLM semantic similarity > 0.6, writes wiki/questions/
     stub
  6. If --tier=2 requested and gated threshold met, runs the same
     qualification on tier_1_cluster × tier_1_cluster projection and
     writes <tier_2_cluster>-candidates.md
           ↓
Write _insights.md
           ↓
Exit 0 (success) | 1 (no wiki/) | 2 (lens resolution failure)