Claude-Mind: Skill for Claude Code

/aspirations-strategic-scan -- Strategic Environmental Scan

Periodically "step back and look at the world." Unlike sparks (reactive to one goal) or evolution (watching agent learning curves), this skill reads the ENVIRONMENT -- recurring goal outputs, knowledge freshness, portfolio balance, and unexplored territory -- and generates work from what it observes.

Design principle: Sparks ask "what did I just learn?" Evolution asks "how am I growing?" Strategic scan asks "what does the world need?" This is the intrinsic motivation engine.

Inputs

• scan_trigger: Why the scan was triggered ("goal_cadence", "recurring_settling", "time_cadence")
• source: Source identifier for goal/aspiration creation

Step 0: Load Conventions

Bash: load-conventions.sh with each name from the conventions: front matter. Read only the paths returned (files not yet in context). If output is empty, proceed.

Phase S1: Recurring Goal Output Review

Read recent execution history for each recurring goal. Recurring goals are the agent's "sensors" -- they periodically observe the world and produce data. This phase reads that data and looks for signals that demand new work.

# ⛔ ALREADY OWNED — DO NOT RE-FILE. The header used to read "KNOWN-INERT"; that
# is FALSIFIED as of 2026-08-16 — the `achievedCount` gate is LIVE (echo, cc-03,
# both files in one call: summary 36 of 220 carry the key and 34 clear `>= 2`, so
# S1 selects ~34 sensors/scan; the full compact carries 92/74). The prior
# 0-of-2437 reading (zeta, cc-02, 2026-08-12) was real and is SUPERSEDED; when and
# why the field appeared is unmeasured — do not assert a cause. A recurrence now
# announces itself via the zero-guard below rather than printing as health.
#
# ⚠ THE GATE GOING LIVE MADE THIS PHASE MORE DANGEROUS, NOT LESS — a silent
# no-op became a confidently-stale detector, and the second owner below is now
# the binding defect. S1 reads only the BOUND AGENT's experience file, and world
# sensors are run by whichever agent picks them up. Measured the same run on
# `g-335-09` (the live customer-spend monitor — a revenue sensor): this box held
# **7 of 30** records fleet-wide (23%), newest local **2026-08-02** against fleet
# newest **2026-08-16** on zeta. Read locally, S1 would have reported "30th
# consecutive zero-live run" as a current finding while the fleet was at run 48+,
# and zeta's two newest records are titled `refusing-a-zero-from-the-wrong-channel`
# and `dark-envs-keep-billing` — i.e. that zero had already been refused.
# BEFORE REPORTING ANY S1 TREND: count the sensor's records across ALL agent
# stores (`/opt/ayoai-mind/agents/*/experience*.jsonl`) and compare the newest
# fleet-wide timestamp to your newest local one. A local-only read of a world
# sensor is a claim about this box, never about the sensor.
# ⚠ THAT COMPARISON CANNOT RUN ON THE WORST CASE: `mine == 0` yields no local
# timestamp, and `len(entries) < 2 -> continue` drops the sensor BEFORE any
# detector — no signal, no warning, no count, so "healthy" and "invisible to this
# box" print identically (guard-1715). Measured 2026-08-19 (alpha, cc-04,
# 6.8.0-137-generic), top-10 sensors / 6 stores: 10/10 cross-agent, 9/10 local <
# fleet; `g-353-03` read **mine 0 of 12** (bravo holds all 12), `g-115-151`
# (production health, ach=140) mine 1/5 with local newest 26d behind fleet. The
# only mine==fleet row was alpha-private by construction. SO: report the
# `mine/fleet` census per sensor and name `mine < 2` rows DROPPED. Owned by
# g-115-3215 — file nothing.
#
# OPEN OWNERS (re-verify before acting): **g-115-3215** — the cross-agent
# blindness, now the LIVE one; g-115-5318 (8 of 10 recurring goals have <2
# experience records). g-115-3246 CLOSED 2026-08-21: its S1 premise expired with
# the FALSIFIED line above, and it shipped the zero-guard below.
#
# This marker exists because the ritual honestly recomputes S1 every scan, so with
# nothing here saying the finding is known, each pass re-derives it as new — the
# guard-1984 shape (a guardrail cannot outvote the instrument it guards), so the
# note belongs in the INSTRUMENT, not in another goal (rb-7613). A suppression
# marker with no re-examination path makes the first suppression the last one
# forever (tree node `detector-dedup-lease-without-release`); the FALSIFIED line
# above is that release — RE-MEASURE, do not inherit.
#
# The old `lastAchievedAt is not None` substitute is NO LONGER NEEDED (it was a
# workaround for the absent field). Run the gate as written. Do NOT file a goal
# about the gate itself.
Bash: load-aspirations-compact.sh -> IF path returned: Read it
recurring_goals = [g for asp in compact for g in asp.goals
                   if g.get("recurring", False) and g.get("achievedCount", 0) >= 2]
# REGRESSION GUARD (g-115-3246), symmetric with S2's below: the fleet HAS recurring
# sensors, so a 0 here means the projection lacks `achievedCount`, NOT that the
# fleet is quiet. A detector whose broken state renders identically to its healthy
# state is invisible -- that asymmetry is why S2a was caught and S1 was not.
recurring_total = sum(1 for asp in compact for g in asp.goals if g.get("recurring", False))
IF len(recurring_goals) == 0:
    Output: ">> WARN strategic-scan S1: 0 sensors from {recurring_total} recurring goals -- projection likely lacks `achievedCount` (g-115-3246). Check WHICH file load-aspirations-compact.sh returned before trusting 'no recurring-sensor signals': the full compact carries the field, the summary does not."

signals = []
FOR EACH rg in recurring_goals (cap at 10 most-recently-achieved):
    # Read last 3 experience entries for this recurring goal
    Bash: experience-read.sh --goal {rg.id}
    entries = parse result

    IF entries is empty or len(entries) < 2: continue  # need 2+ for trend

    # S1a: Regression detection
    # Are metrics, health indicators, or quality measures getting WORSE
    # across recent entries? The LLM interprets the experience text --
    # this is domain-agnostic because any recurring goal's output works.
    # Look for: error counts increasing, success rates decreasing,
    # response times increasing, data quality declining, scores dropping.
    IF entries show worsening trend across 2+ consecutive executions:
        signals.append({
            type: "regression",
            source_goal: rg.id,
            aspiration: rg.parent_asp_id,
            description: "Recurring goal '{rg.title}' shows worsening trend: {what_is_declining}",
            severity: "HIGH",
            evidence: [concise entry summaries]
        })

    # S1b: Anomaly detection
    # Did the most recent execution produce results significantly
    # different from the prior pattern? Not necessarily worse -- just different.
    # Anomalies are worth investigating because they signal change.
    IF latest entry is significantly different from prior entries:
        signals.append({
            type: "anomaly",
            source_goal: rg.id,
            aspiration: rg.parent_asp_id,
            description: "Anomaly in '{rg.title}': {what_changed}",
            severity: "MEDIUM",
            evidence: [concise entry summaries]
        })

    # S1c: Stagnation detection
    # Has this recurring goal produced identical/near-identical results
    # for 3+ consecutive executions? If so, the monitoring may need to
    # look at something different, or the thing being monitored is stuck.
    IF all entries are semantically identical for 3+ executions:
        signals.append({
            type: "stagnation",
            source_goal: rg.id,
            aspiration: rg.parent_asp_id,
            description: "Recurring goal '{rg.title}' producing identical results for {N} executions -- monitoring may need to evolve or the subject is stuck",
            severity: "LOW",
            evidence: [latest entry summary]
        })