Claude-Mind: Skill for Claude Code

/aspirations-select — Goal Selection + Metacognitive Assessment

Selects the highest-value goal for execution using algorithmic scoring (via script) plus metacognitive assessment (model judgment on familiarity, value, cost, infrastructure).

Step 0: Load Conventions — Bash: load-conventions.sh with each name from the conventions: front matter.

Inputs (from orchestrator)

• first_action: Pre-scored goal from handoff (first iteration only, or None)
• decisions_locked: Carried forward decisions from previous session

Outputs (to orchestrator)

• goal: Selected goal object (or None if no candidates)
• effort_level: "full", "standard", or "skip"
• batch_mode: Boolean
• batch: Array of batched goals (if batch_mode)
• ranked_goals: Full ranked list from selector
• prefetch_goals: Goals for pre-fetch research agents
• selection_context: Raw parsed output from goal-selector.sh (includes by_reason, blocked_goals, blocked_count when all_blocked)
• selection_reason: Why no goal was returned ("all_blocked", "all_blocked_by_gate", or absent when goal selected)
• source: Queue origin of the selected goal ("world" or "agent") — from goal-selector output. Pass to all downstream aspirations-*.sh calls via --source {source}.
• deviation_code: Scorer Sovereignty Layer B (g-115-2812) — the sanctioned-deviation enum code when the selected goal is NOT the scorer's top pick (ranked_goals[0]), else "". Phase 4's world-goal claim forwards it via aspirations-claim.sh {goal.id} --deviation {deviation_code}. Computed in Phase 2.94.

Phase 2: Select Next Goal

First-Action Override (first iteration only)

IF first_action is set (from handoff):
    Look up goal by first_action.goal_id
    effort_level = first_action.effort_level
    Clear first_action (consumed)
    # Still run Phase 2.5 for focus context

Algorithmic Scoring

# ASSERTION: goal-selector.sh MUST run every iteration. No exceptions.
# After autocompact, memory of blockers is unreliable. The script reads live state.
Bash: goal-selector.sh
parsed_output = parse JSON output

# Blocked-goals detection: script returns object with "all_blocked" when
# goals exist but none are executable (all deferred/blocked/gated)
IF parsed_output is a JSON object with "all_blocked": true:
    Output: "▸ ALL GOALS BLOCKED: {blocked_count} goals — {by_reason summary}"
    FOR EACH goal in blocked_goals: Output: "  {goal_id}: {detail}"
    # parsed_output contains blocked_goals, blocked_count, by_reason — orchestrator needs these
    # g-357-94: an idle agent must still HEAR directives. Phase 2.07 sits BELOW this return,
    # so while all-blocked no directive was acked, marked or passed on for as long as the
    # agent stayed idle (measured 2026-09-03: a user scope directive unheard for 4h). ACK
    # here exactly as 2.07 (dedup read + moot short-circuit); the honor set is empty with
    # zero candidates. Hand the ACTIVE set to the all-blocked handler, which reads it as
    # generation SCOPE (B0/B1/B2) — a directive never creates work by itself (guard-732).
    Bash: new_directives = board-read.sh --channel coordination --type directive --since 24h --unread-only --mark-read --json
    FOR EACH directive in new_directives: ack per Phase 2.07 (skip moot targets; else
        echo "Acknowledged directive {directive.id}" | board-post.sh --channel coordination --type status --reply-to {directive.id} --tags "acknowledged,{AGENT_NAME}")
    Bash: parsed_output.active_directives = board-read.sh --channel coordination --type directive --since 24h --json
    RETURN (goal = None, selection_reason = "all_blocked", selection_context = parsed_output)

ranked_goals = parsed_output  # JSON array of scored candidates
# Each entry: {goal_id, aspiration_id, title, skill, category, recurring, score, breakdown, raw, cross_world_origin, intended_agent, routed_to_me}
# Foreign-goal display hint (g-336-12): cross_world_origin is non-null
# "<identity>@<origin-world>" ONLY on goals injected by a cross-world INFLUENCE
# grant; null for native goals. Whenever a candidate is rendered to a
# human-visible line — the ALL-BLOCKED list above, the Program-alignment probe,
# any ▸ Output that names a goal by title — append a " [foreign: {cross_world_origin}]"
# badge when the field is non-null, so the agent KNOWS it is executing another
# world's intent and applies appropriate scrutiny (guardrails, review gate).
# A native goal (null) renders unchanged — no badge, no false positive.
# Routed-to-me display hint (g-115-2940): routed_to_me is true ONLY on a
# source='cross-agent:<owner>' candidate, which — by collect_cross_agent_candidates'
# strict-match contract (intended_agent==agent_name) — is BY CONSTRUCTION routed to
# THIS agent. When routed_to_me is true, render the goal as
# " [routed to YOU (owner: {source.split(':',1)[1]})]" and treat it as YOUR work,
# NOT another agent's: a 'cross-agent'/not-my-lane abstention on it is ALWAYS wrong
# (Phase 2.55 CROSS-AGENT-ROUTED EXEMPTION). Dropping intended_agent made bravo
# abstain 13x from its own HIGH-routed g-001-339.

# Partner-claim filter: drop any goal the partner is already in_flight on.
# This is the live claim-conflict HINT — it avoids wasting decomposition and
# context-fetch effort on a goal we would lose at the Phase 4 claim-conflict
# gate anyway. The authoritative gate still runs in the orchestrator (digest
# Phase 4) immediately before aspirations-claim.sh, because partner state can
# flip between this filter and the claim attempt.
#
# BOTH SHAPES ARE REQUIRED, AND READING ONLY `in_flight` OPENS THIS FILTER
# COMPLETELY (g-306-276). `in_flight` is REDUCER-OWNED: team-state-in-flight.sh
# stamps it only when this box's running-session-id exists AND equals MIND_SID,
# and SKIPs for every other Body — writing `in_flight_bodies.<sid>` instead.
# So a partner running as a WORKER Body is invisible in `in_flight` (measured
# 2026-08-10, alpha cc-07: null for ALL four live agents while three held live
# body-row claims). Same defect g-306-160 repaired in
# goal-pickup-coordination-check.py ("silently opened"); this is the second of
# three readers, the digest Phase 4 hard gate the third.
#
# Staleness is the REAPER's job, not this filter's: body_row_reaper (via
# stranded-claim-sweep) deletes rows whose carrier is stale for
# DEFAULT_REAP_STALE_MINUTES (180). Do NOT add a second freshness heuristic
# here — it would drift and put two policies on one store. The ~3h bound
# holds ONLY because every row is born with a carrier (g-306-349); a
# carrier-less row is unreapable at ANY age.
Bash: team-state-read.sh --field agent_status.<partner>.in_flight.goal_id --json
Bash: team-state-read.sh --field agent_status.<partner>.in_flight_bodies --json
    # `--field` returns the whole nested map, so no new endpoint/helper is needed.
partner_held_ids = ({first call's value} if it is a non-null string else {})
                 | {b.goal_id for b in second call's values if b is a dict and b.goal_id}
IF partner_held_ids is non-empty:
    ranked_goals = [g for g in ranked_goals if g.goal_id not in partner_held_ids]
    IF len(ranked_goals) == 0:
        Output: "▸ Partner holds the only candidate goal(s) ({partner_held_ids}) — yielding"
        RETURN (goal = None, selection_reason = "all_blocked", selection_context = {by_reason: {partner_held: 1}, blocked_count: 1})