vulyk: Command for Claude Code

Run the evolution cycle. Use the insight-harvester and skill-gardener skills - they define the method. "$ARGUMENTS"

Nobody has to remember this command: the SessionStart brief says maintenance due: ... evolve (...) when it never ran, or ran 7+ days ago with a council round since, and the Queen runs it after the owner's task, on the default branch with a clean tree.

1. Ledger (memory/stats/evolve.jsonl, owned by scripts/evolve-ledger.py). Run bash scripts/py.sh scripts/evolve-ledger.py . resolve - it records the owner's verdict on earlier changesets from git (merged = accepted, branch deleted unmerged = rejected; add --reason-for <branch>=<text> when the owner gave one) - then bash scripts/py.sh scripts/evolve-ledger.py . window and read it before anything else. A hypothesis it shows rejected is not proposed again unless this run's evidence is newer than the rejection. If bash scripts/py.sh scripts/evolve-ledger.py . pending names a branch, stop: tell the owner in one line that <branch> waits for review (merge to accept, delete the branch to reject) and propose nothing new.

2. Harvest. Read memory/learnings/ (raw + CONSOLIDATED), memory/stats/skills.json, and - if the user can paste it - the output of the built-in /insights command (ask once; proceed without it if unavailable). Then the week's spend: bash scripts/py.sh scripts/token-report.py . --since "$(date -u -d '-7 days' +%Y-%m-%d 2>/dev/null || date -u -v-7d +%Y-%m-%d)", printed as-is - one line per spec with raw and weighted tokens, dispatches and rounds. That report is the spend; the totalTokens a Workflow run prints is a sum of final contexts and never counts as spend. A spec whose dispatches or rounds stand out is evidence for step 3's friction list. Then the owner's corrections of the week and how many reached docs/defects/. The lexicon is the intake hook's own (--lexicon), and the reader is litopys 0.4.0+ corrections. Run the block as written. It prints one line and blocks nothing:

   X="$(mktemp)"; bash .claude/hooks/defect-intake.sh --lexicon > "$X"; L=""; OUT=""; seen=0
   while IFS= read -r c; do   # PATH's copy first, then cached copies newest first: the first with the verb wins
     [ -f "$c" ] || continue; seen=1
     if OUT="$(bash "$c" corrections --since "$(date -u -d '-7 days' +%Y-%m-%d 2>/dev/null || date -u -v-7d +%Y-%m-%d)" --lexicon "$X" 2>/dev/null)"; then L="$c"; break; fi
   done < <(command -v litopys; ls -d "$HOME"/.claude/plugins/cache/litopys/litopys/*/bin/litopys 2>/dev/null | sort -rV)
   if [ -n "$L" ]; then   # a quote opening with ">" is a relayed side-agent note: model text, not the owner's
     r=$(printf '%s\n' "$OUT" | grep -c ' · «> '); OUT="$(printf '%s\n' "$OUT" | grep -v ' · «> ')"
     n=$(printf '%s\n' "$OUT" | grep -c ' · lexicon · '); m=$(printf '%s\n' "$OUT" | grep -c ' · record · '); u=0
     while IFS= read -r l; do q="${l#*«}"; q="${q%»}"; grep -rqF -- "$q" docs/defects 2>/dev/null || u=$((u+1)); done \
       < <(printf '%s\n' "$OUT" | grep -E ' · (lexicon|record) · ')
     echo "corrections (7d): $n by lexicon · $m in records · $u not in docs/defects · $r relayed notes left out"
   elif [ "$seen" = 1 ]; then echo "corrections: no installed litopys has the corrections verb - update it to 0.4.0+"
   else echo "corrections: litopys not installed - no count"; fi
   rm -f "$X"
   

   A line counts as filed when docs/defects/ holds it verbatim. The lexicon also catches remarks that are not corrections, so treat the three numbers as a lead with its n, never as a rate on their own. A remark that is not filed is evidence for step 3: file its quote into the matching card (Law 6), not a new rule in prose.

3. Council check-in (read-only; nothing here writes anything). If memory/stats/council.jsonl does not exist, print council: no rounds yet and skip to step 3. Otherwise run the same awk pass /vulyk-status step 2 uses, restricted to the last 7 days:

   SINCE="$(date -u -d '-7 days' +%Y-%m-%dT%H:%M:%SZ 2>/dev/null || date -u -v-7d +%Y-%m-%dT%H:%M:%SZ)"
   awk -v since="$SINCE" '
     { match($0, /"ts":"[^"]*"/);      ts=substr($0, RSTART+6, RLENGTH-7) }
     ts < since { next }
     { match($0, /"spec":"[^"]*"/);    spec=substr($0, RSTART+8, RLENGTH-9) }
     { match($0, /"round":[0-9]+/);    round=substr($0, RSTART+8, RLENGTH-8) }
     { match($0, /"verdict":"[^"]*"/); verdict=substr($0, RSTART+11, RLENGTH-12) }
     {
       specs[spec]=1
       if (verdict=="GREEN" && (!(spec in firstgreen) || round+0 < firstgreen[spec]+0)) firstgreen[spec]=round
       if (verdict=="ESCALATE") escalations++
     }
     END {
       n=0; for (s in specs) n++
       g=0; for (s in firstgreen) { g++; rounds[g]=firstgreen[s]+0 }
       for (i=1;i<=g;i++) for (j=i+1;j<=g;j++) if (rounds[j]<rounds[i]) { t=rounds[i]; rounds[i]=rounds[j]; rounds[j]=t }
       med=0; if (g>0) { if (g%2==1) med=rounds[(g+1)/2]; else med=(rounds[g/2]+rounds[g/2+1])/2 }
       printf "specs:%d median:%s escalations:%d\n", n, med, escalations+0
     }
   ' memory/stats/council.jsonl
   

   plus the same escaped-defect grep as /vulyk-status step 2, restricted to briefs whose file mtime falls in the last 7 days (date -r <brief.md> +%s vs. date -u -d '-7 days' +%s 2>/dev/null || date -u -v-7d +%s). Print council (7d): <specs> specs · median <n> rounds to green · <n> escalations · <n> escaped defects. Then the comparison the grill names - human.jsonl rows with "verdict":"REJECTED" in the same window:

   awk -v since="$SINCE" '
     { match($0, /"ts":"[^"]*"/); ts=substr($0, RSTART+6, RLENGTH-7) }
     ts < since { next }
     $0 ~ /"verdict":"REJECTED"/ { c++ }
     END { print c+0 }
   ' memory/stats/human.jsonl
   

   Print human.jsonl REJECTED (7d): <n>. Every number in this step is printed with its n, and none of them proves a trend on its own: a week holds a handful of specs, so a count of 0-2 against 0-2 is noise (report rrsi-self-improvement §1). Nothing in this step applies anything; it feeds step 3's diagnosis the way skills.json already does.