mastermind-agent-select

Agent Selection from Registry

Use this pattern whenever a mastermind skill or command needs to select specialist agents. Never hardcode a roster: the agent set differs per install, and a subagent_type that is not installed fails at spawn time.

The pick order (always in this order)

1. [PICK] line. When the user prompt carries a hook-injected line [PICK] agent: <name> · skill: <invoke>, use that agent/skill unless it is clearly wrong for the task.
2. MCP tool mcp__monomind__pick (when the monomind MCP server is connected): mcp__monomind__pick({ task: "<task>", kind: "agents", categories: ["engineering"], top: 3 }) → same JSON as monomind pick --json; each agents.ranked[].name is a spawnable subagent_type, and summary is a one-line recap. Use kind: "skills" or "both" for skills.
3. Local CLI monomind pick — the Standard Selection Block below (local binary only, never npx).
4. Registry keyword scorer — the fallback inside the same block, when no usable CLI is installed.
5. Fixed fallback defaults (bottom of this file) — real agent names only.

The shared index behind 2 and 3 ranks the agent registry (.monomind/registry.json, built from .claude/agents/) and every skill — Claude skills plus the Org skill library — in one call.

Standard Selection Block

# AGENT SELECTION — pick best agents for the current task
# Set these before the block:
#   REGISTRY=".monomind/registry.json"
#   PROMPT="<the user's prompt or idea description>"
#   CATEGORIES="marketing specialized"        # space-separated; see Category Map
#   TOP_N=6                                   # how many agents to return

REGISTRY="${REGISTRY:-.monomind/registry.json}"

# Local only — never npx (no registry egress). A candidate binary is accepted
# only when its output is the unified index: every skill entry carries a
# `source` field. Installs without `pick`, and pick builds that predate the
# unified index, never emit it — so the call deliberately ranks skills too
# (no --agents), and that output doubles as the version check.
mmpick() { for c in monomind ./node_modules/.bin/monomind; do
    command -v "$c" >/dev/null 2>&1 || continue
    out=$("$c" pick "$@" --json 2>/dev/null) || continue
    printf '%s' "$out" | jq -e '[.skills.ranked[]? | has("source")] | (length > 0 and all)' \
      >/dev/null 2>&1 && { printf '%s\n' "$out"; return 0; }
  done; return 127; }

selected_agents=$(mmpick -t "$PROMPT" --categories "$CATEGORIES" --top "$TOP_N" \
  | jq -c '[.agents.ranked[] | {name: (.name // .id), slug: .id, category}]' 2>/dev/null)

# Fallback when no usable CLI: score the registry against name + description +
# capabilities + tags + whenToUse (keywords of 4+ letters from the prompt).
if { [ -z "$selected_agents" ] || [ "$selected_agents" = "[]" ]; } && [ -f "$REGISTRY" ]; then
  keywords=$(echo "$PROMPT" | tr '[:upper:]' '[:lower:]' | grep -oE '[a-z]{4,}' | sort -u | tr '\n' ' ')
  selected_agents=$(jq -c \
    --arg cats "$CATEGORIES" \
    --arg kw "$keywords" \
    --argjson n "$TOP_N" \
    '($kw | split(" ") | map(select(length > 0))) as $keywords
     | [ (.agents // [])[]
         | select(.deprecated != true)
         | select(($cats | length) == 0 or (.category as $c | $cats | split(" ") | any(. == $c)))
         | . as $a
         | ([ $a.name, $a.description, ($a.capabilities // [] | join(" ")),
              ($a.tags // [] | join(" ")), ($a.whenToUse // "") ]
            | map(. // "") | join(" ") | ascii_downcase) as $text
         | {name: $a.name, slug: $a.slug, category: $a.category,
            score: ([ $keywords[] | select(. as $k | $text | contains($k)) ] | length)}
       ]
     | unique_by(.slug)
     | map(select(.score > 0))
     | sort_by(-.score)
     | .[0:$n]
     | map({name, slug, category})' \
    "$REGISTRY" 2>/dev/null)
fi

echo "${selected_agents:-[]}"