#!/usr/bin/env bash
#
# opencode-builder.sh
#
# Script-driven orchestrator for the auto-agents system.
# Queries the OpenCode server fleet, calculates worker pool gaps,
# fetches work items directly via list_prs_* / list_issues scripts,
# and launches workers via session_start.ts — no supervisor agents needed.

set -uo pipefail

# ── Debug mode ─────────────────────────────────────────────────────────────
if [[ "${BUILDER_DEBUG:-0}" == "1" ]]; then
  set -x
  log "DEBUG MODE ENABLED"
fi

# ── Configuration (override via environment) ──────────────────────────────
PORT="${OPENCODE_PORT:-4096}"
HOST="${OPENCODE_HOST:-127.0.0.1}"
BASE="http://${HOST}:${PORT}"
HEALTH_TIMEOUT=60
MAX_IDLE_PER_MINUTE=10
SCRIPT_TIMEOUT_SEC=1200  # timeout for list_prs / list_issues scripts (Forgejo is slow — allow 20 min)
SESSION_SCRIPT_TIMEOUT=30  # timeout for session_* scripts

# Worker pool sizing
total_max_workers="${CA_MAX_PARALLEL_WORKERS:-4}"
# Calculate per-type max workers (same logic as auto-agents)
merge_max_workers=$(awk "BEGIN {print int((${total_max_workers} + 7) / 8)}")
imp_max_workers="${total_max_workers}"
rev_max_workers=$(awk "BEGIN {print int((${total_max_workers} + 1) / 2)}")
# Grooming runs cheap metadata-only work via Forgejo API. Each task is fast
# (a few REST calls), so we don't need many parallel slots — ceil(N/4).
groom_max_workers=$(awk "BEGIN {print int((${total_max_workers} + 3) / 4)}")

# Health check intervals (seconds)
DOOM_CHECK_INTERVAL=900   # 15 minutes
DOOM_AGE_THRESHOLD=1800   # 30 minutes
CONTINUE_BURST_WINDOW=60  # seconds
CONTINUE_BURST_LIMIT=5
SLEEP_WHEN_IDLE=30        # seconds
IDLE_THRESHOLD_MS=180000  # 3 minutes — session considered idle if last_active older than this

# Script paths
SCRIPT_DIR="/app/.opencode/skills/auto-agents-system/scripts"

# Make tsx discoverable on PATH. Running `npx --yes tsx` concurrently with
# a cold or partially-warm npm cache causes a race in `~/.npm/_npx/<hash>/`:
# multiple npm processes simultaneously try to rename `node_modules/esbuild`
# to the same atomic-backup name and fail with
# `ENOTEMPTY: rename '.../esbuild' -> '.../esbuild-7kCFO8Lm'`, crashing
# all but one of them and leaving the cache broken for follow-on calls.
# When tsx is on PATH, npx skips its install/cache logic entirely and just
# execs the existing binary — so the race disappears without changing any
# call sites. We install tsx once into ~/.local/node_modules and prepend
# its bin dir to PATH for the rest of this script's process tree.
TSX_PREFIX="${HOME}/.local"
TSX_BIN_DIR="${TSX_PREFIX}/node_modules/.bin"
if ! [[ -x "${TSX_BIN_DIR}/tsx" ]]; then
  printf "[%s] Installing tsx into %s …\n" "$(date +%H:%M:%S)" "$TSX_PREFIX" >&2
  mkdir -p "$TSX_PREFIX"
  npm install --prefix "$TSX_PREFIX" --silent tsx >/dev/null 2>&1 \
    || { printf "ERROR: failed to install tsx into %s\n" "$TSX_PREFIX" >&2; exit 1; }
fi
export PATH="${TSX_BIN_DIR}:${PATH}"

SESSION_LIST="npx --yes tsx ${SCRIPT_DIR}/session_list.ts"
SESSION_FIND_PREFIX="npx --yes tsx ${SCRIPT_DIR}/session_find_by_prefix.ts"
SESSION_START="npx --yes tsx ${SCRIPT_DIR}/session_start.ts"
SESSION_MESSAGES="npx --yes tsx ${SCRIPT_DIR}/session_messages.ts"
SESSION_DELETE="npx --yes tsx ${SCRIPT_DIR}/session_delete.ts"
SESSION_STOP="npx --yes tsx ${SCRIPT_DIR}/session_stop.ts"

# ── Internal state ───────────────────────────────────────────────────────
SERVER_PID=""
CURL_PID=""
STOP_LOOP=false
LAST_SIGINT=0
OWN_SERVER=false
N=0
IDLE_EVENTS_FILE=""

# In-memory tracking for health checks
declare -A QUESTION_COUNT
declare -A CONTINUE_TIMES
LAST_DOOM_CHECK=0

# ── Helpers ───────────────────────────────────────────────────────────────
die() { printf "ERROR: %s\n" "$*" >&2; exit 1; }
log() { printf "[%s] %s\n" "$(date +%H:%M:%S)" "$*"; }
log_warn() { printf "[%s] WARN: %s\n" "$(date +%H:%M:%S)" "$*" >&2; }
log_error() { printf "[%s] ERROR: %s\n" "$(date +%H:%M:%S)" "$*" >&2; }

is_timeout() {
  local ec="${1:-0}"
  [[ $ec -eq 124 || $ec -eq 137 || $ec -eq 143 ]]
}

# run_timed OUT_VAR TIMEOUT_SEC CMD [ARG...]
#
# Runs CMD under timeout(1), capturing stdout into OUT_VAR and exit code into
# RUN_TIMED_EXITCODE.  The command's stderr is captured separately (fd 3) and
# logged via log_warn so it never pollutes the output that callers parse with jq.
# A warning is always logged: on timeout (exit 124/137/143), non-zero failure,
# or zero-with-stderr.  Empty output on timeout/failure is intentional — callers
# that pipe into jq must handle it gracefully.
#
# Usage:
#   run_timed result 30 npx --yes tsx script.ts --arg1 val1
#   if is_timeout "${RUN_TIMED_EXITCODE:-0}"; then
#     echo "Timed out after 30s"
#   fi
run_timed() {
  local out_var="$1"
  local timeout_sec="$2"
  shift 2
  local cmd_str="$*"
  local tmpout
  local tmperr
  tmpout=$(mktemp)
  tmperr=$(mktemp)
  local exitcode=0

  timeout "$timeout_sec" "$@" > "$tmpout" 2> "$tmperr"
  exitcode=$?

  if is_timeout $exitcode; then
    log_warn "TIMEOUT after ${timeout_sec}s: ${cmd_str}"
    printf -v "$out_var" '%s' ""
  elif [[ $exitcode -ne 0 ]]; then
    log_warn "FAILED (exit ${exitcode}) after ${timeout_sec}s: ${cmd_str}"
    local err_snippet
    err_snippet=$(head -n 3 "$tmperr" 2>/dev/null | tr '\n' '|' | sed 's/|$//')
    if [[ -n "$err_snippet" ]]; then
      log_warn "  stderr: ${err_snippet}"
    fi
    printf -v "$out_var" '%s' ""
  else
    local output
    output=$(cat "$tmpout")
    local err_output
    err_output=$(cat "$tmperr")
    if [[ -n "$err_output" ]]; then
      local err_first_line
      err_first_line=$(echo "$err_output" | head -n 1)
      if [[ "$err_first_line" == ERROR:* || "$err_first_line" == WARNING:* || "$err_first_line" == FAILED:* ]]; then
        log_warn "Command produced stderr: ${cmd_str}"
        local err_snippet
        err_snippet=$(echo "$err_output" | head -n 2 | tr '\n' '|' | sed 's/|$//')
        log_warn "  stderr snippet: ${err_snippet}"
      fi
    fi
    printf -v "$out_var" '%s' "$output"
  fi

  RUN_TIMED_EXITCODE=$exitcode
  rm -f "$tmpout" "$tmperr"
}

check_opencode_running() {
  if curl -sf "${BASE}/global/health" >/dev/null 2>&1; then
    return 0
  else
    return 1
  fi
}

cleanup() {
  if [[ -n "$IDLE_EVENTS_FILE" ]]; then
    rm -f "$IDLE_EVENTS_FILE" 2>/dev/null || true
  fi
  if $OWN_SERVER && [[ -n "$SERVER_PID" ]] && kill -0 "$SERVER_PID" 2>/dev/null; then
    log "Stopping opencode server (PID $SERVER_PID)..."
    kill "$SERVER_PID" 2>/dev/null
    wait "$SERVER_PID" 2>/dev/null || true
  elif ! $OWN_SERVER; then
    log "Leaving existing OpenCode server running (not managed by this script)"
  fi
}

handle_sigint() {
  local now
  now=$(date +%s)

  if $STOP_LOOP && (( now - LAST_SIGINT <= 2 )); then
    printf "\nForce quit!\n"
    [[ -n "$CURL_PID" ]] && kill "$CURL_PID" 2>/dev/null || true
    cleanup
    exit 130
  fi

  STOP_LOOP=true
  LAST_SIGINT=$now
  printf "\nWill stop after the current request completes. Ctrl+C again within 2s to force-quit.\n"
}

trap handle_sigint INT
trap cleanup EXIT

# ── Idle-event tracking ───────────────────────────────────────────────────
init_idle_tracking() {
  IDLE_EVENTS_FILE="/tmp/opencode-builder-idle-events-$$.txt"
  : > "$IDLE_EVENTS_FILE" 2>/dev/null || true
}

record_idle_event() {
  local now
  now=$(date +%s)
  echo "$now" >> "$IDLE_EVENTS_FILE"
}

count_idle_events_last_minute() {
  local now cutoff
  now=$(date +%s)
  cutoff=$((now - 60))
  if [[ -f "$IDLE_EVENTS_FILE" ]]; then
    awk -v c="$cutoff" '$1 > c {count++} END {print count+0}' "$IDLE_EVENTS_FILE" 2>/dev/null
  else
    echo 0
  fi
}

# ── Session lifecycle helpers ─────────────────────────────────────────────
query_all_sessions() {
  local result
  run_timed result "$SESSION_SCRIPT_TIMEOUT" \
    $SESSION_LIST --server "${BASE}"
  echo "$result"
}

find_sessions_by_prefix() {
  local prefix="$1"
  local status_filter="${2:-all}"
  local result
  run_timed result "$SESSION_SCRIPT_TIMEOUT" \
    $SESSION_FIND_PREFIX --server "${BASE}" --prefix "$prefix" --exclude-supervisor --status "$status_filter"
  echo "$result"
}

get_busy_worker_tags() {
  local prefix="$1"
  local now_ms
  now_ms=$(date +%s000)
  local cutoff_ms=$((now_ms - 120000))
  local result
  result=$(find_sessions_by_prefix "$prefix" "all" | jq -r \
    --arg cutoff "$cutoff_ms" \
    '.[] | select(.last_active >= ($cutoff | tonumber)) | .tag // empty' 2>/dev/null)
  if [[ -z "$result" ]]; then
    log_warn "[get_busy_worker_tags:${prefix}] no busy tags found (or find_sessions_by_prefix timed out)"
  fi
  echo "$result"
}

create_session() {
  local title="$1"
  curl -sf -X POST "${BASE}/session" \
    -H "Content-Type: application/json" \
    -d "{\"title\":\"${title}\"}" 2>/dev/null
}

delete_session() {
  local sid="$1"
  $SESSION_DELETE --server "${BASE}" --session-id "$sid" --force >/dev/null 2>&1 || true
}

stop_session() {
  local sid="$1"
  $SESSION_STOP --server "${BASE}" --session-id "$sid" >/dev/null 2>&1 || true
}

send_message_capture() {
  local sid="$1"
  local body="$2"
  local out_file="$3"

  # Refuse to send empty bodies — the server logs them as
  # "No user message found in stream. This should never happen."
  if [[ -z "$body" ]]; then
    log_warn "send_message_capture: refusing to send empty body to ${sid}"
    : > "$out_file" 2>/dev/null || true
    return 1
  fi

  # Use prompt_async — fire-and-forget; server returns 204 immediately.
  # The synchronous /message endpoint would block until the agent finishes
  # its full response (potentially many minutes for a busy worker).
  ( trap '' INT
    exec curl -sf -X POST "${BASE}/session/${sid}/prompt_async" \
      -H "Content-Type: application/json" \
      -d "$body" \
      -w "\n"
  ) > "$out_file" 2>/dev/null &
  CURL_PID=$!

  while kill -0 "$CURL_PID" 2>/dev/null; do
    wait "$CURL_PID" 2>/dev/null || true
  done
  CURL_PID=""
}

get_messages_poll() {
  local sid="$1"
  local limit="${2:-10}"
  local timeout_secs="${3:-60}"
  local deadline
  deadline=$(($(date +%s) + timeout_secs))

  while true; do
    local msgs
    msgs=$($SESSION_MESSAGES --server "${BASE}" --session-id "$sid" --limit "$limit" 2>/dev/null)
    if [[ -n "$msgs" ]]; then
      local has_assistant
      has_assistant=$(echo "$msgs" | jq -r 'map(select(.info.role == "assistant")) | length')
      if [[ "$has_assistant" -gt 0 ]]; then
        echo "$msgs"
        return 0
      fi
    fi
    if [[ $(date +%s) -ge $deadline ]]; then
      echo "[]"
      return 1
    fi
    sleep 2
  done
}

extract_last_assistant_text() {
  local msgs="$1"
  echo "$msgs" | jq -r '[.[] | select(.info.role == "assistant")] | last | if . == null then "" else (.parts // [] | map(select(.type == "text")) | .[-1].text // "") end'
}

extract_json_array() {
  local text="$1"
  echo "$text" | grep -oP '\[\s*[^\]]*\]' | tail -n1
}

# filter_json_array — strips any non-JSON prefix (e.g. progress messages from
# list scripts that write to stderr which ends up in stdout) and returns clean
# JSON.  Returns "[]" if no JSON array is found.  Use this after any command
# whose stderr might have leaked into stdout.
filter_json_array() {
  local text="$1"
  local first_char
  first_char=$(echo "$text" | head -c1)
  if [[ "$first_char" == "[" ]]; then
    echo "$text"
  else
    local json
    json=$(echo "$text" | grep -oP '(\[\s*\].*|\[\s*[^\]]+\])' | tail -n1)
    if [[ -z "$json" || "$json" == "[]" ]]; then
      echo "[]"
    else
      echo "$json"
    fi
  fi
}

# ── Question handling ─────────────────────────────────────────────────────
get_pending_questions() {
  local result
  run_timed result 15 curl -sf "${BASE}/question"
  echo "$result"
}

answer_question() {
  local qid="$1"
  local text="$2"
  curl -sf -X POST "${BASE}/question/${qid}/answer" \
    -H "Content-Type: application/json" \
    -d "{\"body\":\"${text}\"}" >/dev/null 2>&1
}

reject_question() {
  local qid="$1"
  curl -sf -X POST "${BASE}/question/${qid}/reject" >/dev/null 2>&1
}

# ── Forgejo configuration from environment + git remote fallback ─────────────
resolve_forgejo_config() {
  local url="${FORGEJO_URL:-}"
  local owner="${FORGEJO_OWNER:-}"
  local repo="${FORGEJO_REPO:-}"
  local pat="${FORGEJO_PAT:-}"
  local git_name="${GIT_USER_NAME:-}"
  local git_email="${GIT_USER_EMAIL:-}"

  if [[ -z "$owner" || -z "$repo" ]]; then
    local remote
    remote=$(git remote get-url origin 2>/dev/null)
    if [[ -n "$remote" ]]; then
      local proto_host path_seg remote_owner remote_repo
      proto_host=$(echo "$remote" | sed -n 's|^\(https\?://[^/]\+\).*|\1|p')
      path_seg=$(echo "$remote" | sed -n 's|.*://[^/]\+/\([^/]\+/[^/]\+\).*|\1|p' | sed 's/\.git$//')
      if [[ -z "$path_seg" ]]; then
        log_warn "resolve_forgejo_config: could not parse owner/repo from git remote '${remote}'"
      else
        remote_owner="${path_seg%%/*}"
        remote_repo="${path_seg#*/}"
        url="${url:-${proto_host}}"
        owner="${owner:-${remote_owner}}"
        repo="${repo:-${remote_repo}}"
      fi
    else
      log_warn "resolve_forgejo_config: FORGEJO_OWNER/REPO not set and 'git remote get-url origin' failed — forgejo config unavailable"
    fi
  fi

  FORGEJO_URL="$url"
  FORGEJO_OWNER="$owner"
  FORGEJO_REPO="$repo"
  FORGEJO_PAT="$pat"
  GIT_USER_NAME="$git_name"
  GIT_USER_EMAIL="$git_email"
}

# ── Work-group fetchers ────────────────────────────────────────────────────
fetch_list_script() {
  local script_path="$1"
  local forgejo_url="$2"
  local forgejo_owner="$3"
  local forgejo_repo="$4"
  local forgejo_pat="$5"

  if [[ -z "$forgejo_url" || -z "$forgejo_pat" ]]; then
    echo "[]"
    return
  fi

  local result
  result=$(timeout "$SCRIPT_TIMEOUT_SEC" npx --yes tsx "$script_path" \
    --url "$forgejo_url" \
    --pat "$forgejo_pat" \
    --owner "$forgejo_owner" \
    --repo "$forgejo_repo" 2>/dev/null)
  local status=$?
  if [[ $status -eq 139 || $status -eq 137 || $status -eq 143 || $status -eq 124 ]]; then
    log "[fetch_list_script] WARNING: $(basename "$script_path") timed out after ${SCRIPT_TIMEOUT_SEC}s (exit ${status})"
    echo "[]"
    return
  fi
  if [[ $status -ne 0 ]]; then
    log "[fetch_list_script] WARNING: $(basename "$script_path") failed with exit ${status}"
    echo "[]"
    return
  fi
  local first_char="${result%"${result#?}"}"
  if [[ "$first_char" == "[" ]]; then
    echo "$result"
  else
    log "[fetch_list_script] WARNING: $(basename "$script_path") produced non-JSON output (starts with '${first_char}'); treating as empty"
    echo "[]"
  fi
}

# ── Worker prompt builders ─────────────────────────────────────────────────
# All fields (title, body) have newlines collapsed so the prompt file stays
# valid as a single-line-ish text file. The full item JSON is embedded compact
# (jq -c . → one line, no literal newlines) at the end of the prompt.
# session_start.ts reads this file as plain text and sends it to the agent.
build_impl_prompt() {
  local item_json="$1"
  local kind="$2"
  local work_num="$3"

  local work_type; work_type=$([ "$kind" == "ISSUE" ] && echo "issue_impl" || echo "pr_fix")
  local compact_item
  compact_item=$(echo "$item_json" | jq -c .)

  jq -n \
    --arg wn "$work_num" \
    --arg wt "$work_type" \
    --argjson ij "$compact_item" \
    '{
  issue_number: $wn,
  pr_number: $wn,
  work_type: $wt,
  item_json: $ij
}'
}

build_groom_prompt() {
  local item_json="$1"
  local kind="$2"
  local work_num="$3"

  local work_type; work_type=$([ "$kind" == "ISSUE" ] && echo "issue_groom" || echo "pr_groom")
  local compact_item
  compact_item=$(echo "$item_json" | jq -c .)

  jq -n \
    --arg wn "$work_num" \
    --arg wt "$work_type" \
    --argjson ij "$compact_item" \
    '{
  issue_number: $wn,
  pr_number: $wn,
  work_type: $wt,
  item_json: $ij
}'
}

build_merge_prompt() {
  local pr_json="$1"
  local pr_num="$2"

  local compact_item
  compact_item=$(echo "$pr_json" | jq -c .)

  jq -n \
    --arg pn "$pr_num" \
    --argjson ij "$compact_item" \
    '{
  pr_number: $pn,
  item_json: $ij
}'
}

build_review_prompt() {
  local pr_json="$1"
  local pr_num="$2"

  local title head_sha base_sha merge_base is_stale has_conflicts
  title=$(echo "$pr_json" | jq -r '.title' | tr '\n' ' ' | sed 's/[[:space:]]\+/ /g')
  head_sha=$(echo "$pr_json" | jq -r '.head.sha')
  base_sha=$(echo "$pr_json" | jq -r '.base.sha')
  merge_base=$(echo "$pr_json" | jq -r '.merge_base // empty')
  [[ "$merge_base" == "empty" || -z "$merge_base" || "$merge_base" == "null" ]] && merge_base="$base_sha"
  is_stale=$(echo "$pr_json" | jq -r 'if (.stale_state // "not_stale") == "not_stale" then "false" else "true" end')
  has_conflicts=$(echo "$pr_json" | jq -r 'if .mergeable == false then "true" else "false" end')

  local approvals_count ci_status compact_item
  approvals_count=$(echo "$pr_json" | jq -r '.approvals_count // 0')
  ci_status=$(echo "$pr_json" | jq -r '.ci_status // "unknown"')
  compact_item=$(echo "$pr_json" | jq -c .)

  cat <<PROMPT
pr_number: ${pr_num}
pr_title: ${title}
head_sha: ${head_sha}
base_sha: ${base_sha}
merge_base_sha: ${merge_base}
is_stale: ${is_stale}
has_conflicts: ${has_conflicts}
approvals_count: ${approvals_count}
ci_status: ${ci_status}
item_json: ${compact_item}

Process the indicated Pull Request or Issue.
PROMPT
}

# ── Fleet analysis ────────────────────────────────────────────────────────
get_busy_worker_tags() {
  local prefix="$1"
  find_sessions_by_prefix "$prefix" "busy" | jq -r '.[].tag // empty'
}

extract_work_number() {
  local tag="$1"
  echo "$tag" | sed -n 's/.*-\(PR\|ISSUE\)-\([0-9]\+\)$/\2/p'
}

# ── Duplicate worker guard ────────────────────────────────────────────────
deduplicate_workers() {
  local prefix="$1"
  local workers_json
  workers_json=$(find_sessions_by_prefix "$prefix" "busy")
  if [[ -z "$workers_json" ]]; then
    log_warn "[deduplicate_workers:${prefix}] find_sessions_by_prefix returned empty (timeout or error) — skipping deduplication"
    return
  fi
  local duplicates
  duplicates=$(echo "$workers_json" | jq -r '
    group_by(.tag)
    | map(select(length > 1))
    | map(sort_by(.last_active) | reverse | .[1:])
    | flatten
    | .[].id
  ' 2>/dev/null)
  if [[ -n "$duplicates" ]]; then
    local count
    count=$(echo "$duplicates" | grep -c . 2>/dev/null || echo 0)
    log "[deduplicate_workers:${prefix}] found ${count} duplicate session(s) — deleting older copies"
    while IFS= read -r dup_id; do
      [[ -n "$dup_id" ]] || continue
      log "  deleting duplicate session ${dup_id}"
      delete_session "$dup_id"
    done <<< "$duplicates"
  fi
}

# ── Disk cleanup ──────────────────────────────────────────────────────────
cleanup_orphan_tmp_dirs() {
  local prefix="$1"
  local sessions_json
  sessions_json=$(find_sessions_by_prefix "$prefix" "all")
  if [[ -z "$sessions_json" ]]; then
    log_warn "[cleanup_orphan_tmp_dirs:${prefix}] find_sessions_by_prefix failed (timeout/error) — skipping disk cleanup"
    return
  fi
  local active_tags
  active_tags=$(echo "$sessions_json" | jq -r '.[].tag // empty' 2>/dev/null | sort -u)
  # Note: active_tags can legitimately be empty when there are zero sessions
  # for this prefix — in that case every matching tmp dir is an orphan.

  local tag_pattern="${prefix}"
  local find_result
  run_timed find_result 10 find /tmp -maxdepth 1 -type d -name "*${tag_pattern}*" -print0 2>/dev/null
  if [[ -z "$find_result" ]]; then
    log "[cleanup_orphan_tmp_dirs:${prefix}] no tmp dirs found matching ${tag_pattern}"
    return
  fi

  local removed=0
  while IFS= read -r -d '' dir; do
    [[ -d "$dir" ]] || continue
    local dir_id
    dir_id=$(basename "$dir" | grep -oP '\d+' | tail -n1)
    if [[ -z "$dir_id" ]]; then
      continue
    fi
    local is_active=false
    while IFS= read -r atag; do
      [[ -n "$atag" ]] || continue
      if [[ "$atag" == *"${dir_id}"* ]]; then
        is_active=true
        break
      fi
    done <<< "$active_tags"
    if ! $is_active; then
      log "[cleanup_orphan_tmp_dirs:${prefix}] removing orphan ${dir}"
      rm -rf "$dir"
      ((removed++))
    fi
  done < <(printf '%s\0' "$find_result")

  if [[ $removed -eq 0 ]]; then
    log "[cleanup_orphan_tmp_dirs:${prefix}] no orphan tmp dirs found"
  else
    log "[cleanup_orphan_tmp_dirs:${prefix}] removed ${removed} orphan tmp dir(s)"
  fi
}

# run_single_doom_check — Evaluate the health of a single busy session by
# spinning up a worker-health-evaluator eval session, waiting for it to
# finish (up to 120s), and printing one of:
#   "unrecoverable"  — eval reports unrecoverable=true; caller should delete
#   "ok"             — anything else; session is fine
#   "create_failed"  — couldn't create the eval session
# Designed to be run in the background so many checks can proceed in parallel.
run_single_doom_check() {
  local sid="$1"
  local eval_tag="BUILDER-HEALTH-${sid: -6}"
  local eval_title="[${eval_tag}] worker-health-evaluator"
  local eval_prompt="session_id: ${sid}

Evaluate the health of this session."
  local eval_session_json
  eval_session_json=$(create_session "$eval_title")
  local eval_sid
  eval_sid=$(echo "$eval_session_json" | jq -r '.id')
  if [[ -z "$eval_sid" || "$eval_sid" == "null" ]]; then
    echo "create_failed"
    return
  fi
  local eval_tmp="/tmp/builder-prompt-${eval_tag}-$$.txt"
  printf '%s' "$eval_prompt" > "$eval_tmp"
  local eval_body
  eval_body=$(jq -nc --arg a "worker-health-evaluator" --rawfile t "$eval_tmp" '{agent: $a, parts: [{type: "text", text: $t}]}')
  if [[ -z "$eval_body" ]]; then
    # jq failed (likely couldn't read eval_tmp) — bail rather than send empty
    rm -f "$eval_tmp"
    delete_session "$eval_sid"
    echo "ok"
    return
  fi
  curl -sf -X POST "${BASE}/session/${eval_sid}/prompt_async" -H "Content-Type: application/json" -d "$eval_body" >/dev/null 2>&1
  local eval_deadline=$(( $(date +%s) + 120 ))
  while [[ $(date +%s) -lt $eval_deadline ]]; do
    local eval_status
    eval_status=$(curl -sf "${BASE}/session/status" 2>/dev/null | jq -r --arg sid "$eval_sid" '.[$sid].type // "idle"')
    [[ "$eval_status" == "idle" ]] && break
    sleep 2
  done
  local eval_msgs
  eval_msgs=$($SESSION_MESSAGES --server "${BASE}" --session-id "$eval_sid" --limit 10 2>/dev/null)
  local eval_text
  eval_text=$(extract_last_assistant_text "$eval_msgs")
  delete_session "$eval_sid"
  rm -f "$eval_tmp"
  local is_unrecoverable
  is_unrecoverable=$(echo "$eval_text" | jq -r '.unrecoverable // false' 2>/dev/null || echo "false")
  if [[ "$is_unrecoverable" == "true" ]]; then
    echo "unrecoverable"
  else
    echo "ok"
  fi
}

# ── Health checks ─────────────────────────────────────────────────────────
run_health_cycle() {
  local now
  now=$(date +%s)
  local did_something=false

  # 1) Question handling
  local questions_json
  questions_json=$(get_pending_questions)
  if [[ -n "$questions_json" && "$questions_json" != "[]" && "$questions_json" != "null" ]]; then
    local qids sids
    qids=$(echo "$questions_json" | jq -r '.[].id')
    sids=$(echo "$questions_json" | jq -r '.[].sessionID')
    while IFS= read -r qid && IFS= read -r sid <&3; do
      [[ -n "$qid" ]] || continue
      [[ -n "$sid" ]] || continue
      local cnt="${QUESTION_COUNT[$sid]:-0}"
      if [[ "$cnt" -eq 0 ]]; then
        log "Question from ${sid} (first time) — answering with anti-question text"
        answer_question "$qid" "you are never under any circumstances to ask questions or pass along questions from subagents, use your best judgement and do not ask questions again"
        QUESTION_COUNT[$sid]=1
      else
        log "Question from ${sid} (repeat) — deleting session"
        delete_session "$sid"
        unset QUESTION_COUNT[$sid]
      fi
      did_something=true
    done <<< "$qids" 3<<< "$sids"
  fi

  # 2) Idle worker check — collect all idle SIDs across prefixes, then fetch
  # their messages in parallel before processing each one.
  local all_prefixes=("AUTO-IMP" "AUTO-MRG" "AUTO-REV" "AUTO-GROOM")
  local -a idle_sids=()
  for prefix in "${all_prefixes[@]}"; do
    local sessions_json
    sessions_json=$(find_sessions_by_prefix "$prefix" "all")
    if [[ -z "$sessions_json" ]]; then
      log_warn "[health:idle-check:${prefix}] find_sessions_by_prefix returned empty (timeout or error) — skipping idle check for this pool"
      continue
    fi
    local idle_workers
    idle_workers=$(echo "$sessions_json" | jq -r \
      --arg cutoff "$(( (now * 1000) - IDLE_THRESHOLD_MS ))" \
      '.[] | select(.last_active < ($cutoff | tonumber)) | .id // empty')
    while IFS= read -r sid; do
      [[ -n "$sid" ]] || continue
      idle_sids+=("$sid")
    done <<< "$idle_workers"
  done

  if [[ ${#idle_sids[@]} -gt 0 ]]; then
    log "Idle check: fetching messages for ${#idle_sids[@]} idle worker(s) in parallel"
    local -a msg_pids=()
    local -a msg_files=()
    for sid in "${idle_sids[@]}"; do
      local msg_file="/tmp/idle-msgs-${sid}-$$.json"
      ( $SESSION_MESSAGES --server "${BASE}" --session-id "$sid" --limit 10 > "$msg_file" 2>/dev/null ) &
      msg_pids+=($!)
      msg_files+=("$msg_file")
    done
    for pid in "${msg_pids[@]}"; do
      wait "$pid" 2>/dev/null
    done

    # Sequential processing (CONTINUE_TIMES is shared state)
    for ((i = 0; i < ${#idle_sids[@]}; i++)); do
      local sid="${idle_sids[i]}"
      local msgs
      msgs=$(cat "${msg_files[i]}" 2>/dev/null)
      rm -f "${msg_files[i]}"
      if [[ -z "$msgs" || "$msgs" == "[]" ]]; then
        log "Idle worker ${sid} has no messages — deleting"
        delete_session "$sid"
        did_something=true
        continue
      fi
      local last_text
      last_text=$(extract_last_assistant_text "$msgs")
      if echo "$last_text" | grep -qiE '(done|finished|complete|exiting|all tests pass|committed|merged|review submitted)'; then
        log "Idle worker ${sid} appears finished — deleting"
        delete_session "$sid"
        did_something=true
        continue
      fi
      local now_s
      now_s=$(date +%s)
      local times_str="${CONTINUE_TIMES[$sid]:-}"
      local cutoff=$((now_s - CONTINUE_BURST_WINDOW))
      local new_times_arr=()
      local count=0
      if [[ -n "$times_str" ]]; then
        for ts in $times_str; do
          [[ -n "$ts" ]] || continue
          if (( ts > cutoff )); then
            new_times_arr+=("$ts")
            ((count++))
          fi
        done
      fi
      new_times_arr+=("$now_s")
      ((count++))
      CONTINUE_TIMES[$sid]=$(printf '%s ' "${new_times_arr[@]}")
      if [[ "$count" -gt $CONTINUE_BURST_LIMIT ]]; then
        log "Idle worker ${sid} received ${count} continues in ${CONTINUE_BURST_WINDOW}s — deleting"
        delete_session "$sid"
        unset CONTINUE_TIMES[$sid]
        did_something=true
      else
        log "Idle worker ${sid} — sending continue"
        local cont_body
        cont_body=$(jq -nc '{agent: "", parts: [{type: "text", text: "continue"}]}')
        send_message_capture "$sid" "$cont_body" "/tmp/continue-${sid}.txt"
        did_something=true
      fi
    done
  fi

  # 3) Doom-loop check (every 15 minutes) — runs ALL evaluations in parallel
  if [[ $((now - LAST_DOOM_CHECK)) -ge $DOOM_CHECK_INTERVAL ]]; then
    LAST_DOOM_CHECK=$now

    # Collect all busy sessions across all prefixes.
    local -a doom_sids=()
    for prefix in "${all_prefixes[@]}"; do
      local sessions_json busy_workers
      sessions_json=$(find_sessions_by_prefix "$prefix" "all")
      if [[ -z "$sessions_json" ]]; then
        log_warn "[health:doom-check:${prefix}] find_sessions_by_prefix returned empty (timeout or error) — skipping doom check for this pool"
        continue
      fi
      busy_workers=$(echo "$sessions_json" | jq -r \
        --arg cutoff "$(( (now - DOOM_AGE_THRESHOLD) * 1000 ))" \
        '.[] | select(.last_active >= ($cutoff | tonumber)) | .id // empty')
      while IFS= read -r sid; do
        [[ -n "$sid" ]] || continue
        doom_sids+=("$sid")
      done <<< "$busy_workers"
    done

    if [[ ${#doom_sids[@]} -gt 0 ]]; then
      log "Doom-loop check: evaluating ${#doom_sids[@]} session(s) in parallel"
      local -a doom_pids=()
      local -a doom_result_files=()
      for sid in "${doom_sids[@]}"; do
        local result_file="/tmp/doom-result-${sid}-$$.txt"
        ( run_single_doom_check "$sid" > "$result_file" 2>/dev/null ) &
        doom_pids+=($!)
        doom_result_files+=("$result_file")
      done

      # Wait for all evaluations and process results
      for ((i = 0; i < ${#doom_pids[@]}; i++)); do
        wait "${doom_pids[i]}" 2>/dev/null
        local result
        result=$(cat "${doom_result_files[i]}" 2>/dev/null)
        if [[ "$result" == "unrecoverable" ]]; then
          log "Session ${doom_sids[i]} is unrecoverable — deleting"
          delete_session "${doom_sids[i]}"
          did_something=true
        fi
        rm -f "${doom_result_files[i]}"
      done
    fi
  fi

  if $did_something; then
    return 0
  else
    return 1
  fi
}

# ── Preflight checks ─────────────────────────────────────────────────────
for cmd in curl jq opencode npx; do
  command -v "$cmd" &>/dev/null || die "'$cmd' is required but not found in PATH"
done

# ── Start the opencode server ─────────────────────────────────────────────
if check_opencode_running; then
  echo "┌─────────────────────────────────────────────────────────────────┐"
  echo "│ ⚠  WARNING: OpenCode server already running on ${HOST}:${PORT}"
  echo "│ 🔗 Connecting to existing server instead of starting new one"
  echo "│ 💡 The script will NOT manage this server's lifecycle"
  echo "└─────────────────────────────────────────────────────────────────┘"
  echo ""
  log "Using existing OpenCode server at ${HOST}:${PORT}"
  OWN_SERVER=false
else
  log "Starting opencode server on ${HOST}:${PORT}..."
  opencode serve --port "$PORT" --hostname "$HOST" &
  SERVER_PID=$!
  OWN_SERVER=true

  log "Waiting for server to become healthy (up to ${HEALTH_TIMEOUT}s)..."
  healthy=false
  for (( i = 1; i <= HEALTH_TIMEOUT; i++ )); do
    if curl -sf "${BASE}/global/health" >/dev/null 2>&1; then
      healthy=true
      break
    fi
    kill -0 "$SERVER_PID" 2>/dev/null || die "Server process exited unexpectedly"
    sleep 1
  done
  $healthy || die "Server did not become healthy within ${HEALTH_TIMEOUT}s"
  log "Server is ready."
fi

# ── Main orchestration loop ───────────────────────────────────────────────
main_loop() {
  log "Starting orchestration loop."
  log "Worker pools: merge=${merge_max_workers} imp=${imp_max_workers} rev=${rev_max_workers} groom=${groom_max_workers}"

  while ! $STOP_LOOP; do
    N=$((N + 1))

    idle_count=$(count_idle_events_last_minute)
    if (( idle_count > MAX_IDLE_PER_MINUTE )); then
      log "── iteration #${N} ── TOO MANY IDLE EVENTS (${idle_count}) — skipping cycle"
      sleep 5
      continue
    fi

    log "── iteration #${N} ──"
    local did_anything=false

    local all_sessions_json
    all_sessions_json=$(query_all_sessions)
    if [[ -z "$all_sessions_json" ]]; then
      if is_timeout "${RUN_TIMED_EXITCODE:-0}"; then
        log_warn "query_all_sessions timed out — OpenCode server may be unresponsive; skipping this cycle"
        sleep "$SLEEP_WHEN_IDLE"
        continue
      fi
      log "No sessions found."
    fi

    # ── Step 2: Per-type analysis & direct worker launch ──────────────────
    log "[step2] resolving Forgejo config..."
    resolve_forgejo_config
    log "[step2] Forgejo config: url=${FORGEJO_URL} owner=${FORGEJO_OWNER} repo=${FORGEJO_REPO} pat_set=$([[ -n "$FORGEJO_PAT" ]] && echo yes || echo no)"

    if [[ -z "$FORGEJO_PAT" ]]; then
      log "FORGEJO_PAT not set — skipping worker launch"
    else
      declare -a IMP_WORK_GROUPS=(
        "CI Failing PRs|${SCRIPT_DIR}/list_prs_ci_failing.ts|PR"
        "Addressed Changes CI Passing|${SCRIPT_DIR}/list_prs_addressed_changes_ci_passing.ts|PR"
        "Addressed Changes CI Failing|${SCRIPT_DIR}/list_prs_addressed_changes_ci_failing.ts|PR"
        "Changes Requested|${SCRIPT_DIR}/list_prs_changes_requested.ts|PR"
        "CI Missing|${SCRIPT_DIR}/list_prs_missing_ci_checks.ts|PR"
        "No Active Review CI Passing|${SCRIPT_DIR}/list_prs_no_active_review_ci_passing.ts|PR"
        "No Active Review CI Failing|${SCRIPT_DIR}/list_prs_no_active_review_ci_failing.ts|PR"
        "Needs Review Not Stale|${SCRIPT_DIR}/list_prs_needs_review_not_stale.ts|PR"
        "Needs Review Stale Clean|${SCRIPT_DIR}/list_prs_needs_review_stale_clean.ts|PR"
        "Needs Review Stale Conflicts|${SCRIPT_DIR}/list_prs_needs_review_stale_conflicts.ts|PR"
        "Open Issues|${SCRIPT_DIR}/list_issues.ts|ISSUE"
      )
      declare -a MRG_WORK_GROUPS=(
        "Ready to Merge|${SCRIPT_DIR}/list_prs_ready_to_merge.ts|PR"
        "Stale No Conflicts|${SCRIPT_DIR}/list_prs_stale_clean.ts|PR"
        "Stale Has Conflicts|${SCRIPT_DIR}/list_prs_stale_conflicts.ts|PR"
      )
      declare -a REV_WORK_GROUPS=(
        "No Active Review CI Passing|${SCRIPT_DIR}/list_prs_no_active_review_ci_passing.ts|PR"
        "No Active Review CI Failing|${SCRIPT_DIR}/list_prs_no_active_review_ci_failing.ts|PR"
        "Needs Review Not Stale|${SCRIPT_DIR}/list_prs_needs_review_not_stale.ts|PR"
        "Needs Review Stale Clean|${SCRIPT_DIR}/list_prs_needs_review_stale_clean.ts|PR"
        "Needs Review Stale Conflicts|${SCRIPT_DIR}/list_prs_needs_review_stale_conflicts.ts|PR"
        "Addressed Changes CI Passing|${SCRIPT_DIR}/list_prs_addressed_changes_ci_passing.ts|PR"
        "Addressed Changes CI Failing|${SCRIPT_DIR}/list_prs_addressed_changes_ci_failing.ts|PR"
        "Changes Requested|${SCRIPT_DIR}/list_prs_changes_requested.ts|PR"
        "CI Failing|${SCRIPT_DIR}/list_prs_ci_failing.ts|PR"
      )
      # GROOM mirrors IMP: same priority order (PRs first, issues second), same
      # work groups. Re-uses IMP's cached results within the same iteration
      # via the shared cache_dir, so grooming pre-fetch is effectively free.
      declare -a GROOM_WORK_GROUPS=(
        "CI Failing PRs|${SCRIPT_DIR}/list_prs_ci_failing.ts|PR"
        "Addressed Changes CI Passing|${SCRIPT_DIR}/list_prs_addressed_changes_ci_passing.ts|PR"
        "Addressed Changes CI Failing|${SCRIPT_DIR}/list_prs_addressed_changes_ci_failing.ts|PR"
        "Changes Requested|${SCRIPT_DIR}/list_prs_changes_requested.ts|PR"
        "CI Missing|${SCRIPT_DIR}/list_prs_missing_ci_checks.ts|PR"
        "No Active Review CI Passing|${SCRIPT_DIR}/list_prs_no_active_review_ci_passing.ts|PR"
        "No Active Review CI Failing|${SCRIPT_DIR}/list_prs_no_active_review_ci_failing.ts|PR"
        "Needs Review Not Stale|${SCRIPT_DIR}/list_prs_needs_review_not_stale.ts|PR"
        "Needs Review Stale Clean|${SCRIPT_DIR}/list_prs_needs_review_stale_clean.ts|PR"
        "Needs Review Stale Conflicts|${SCRIPT_DIR}/list_prs_needs_review_stale_conflicts.ts|PR"
        "Open Issues|${SCRIPT_DIR}/list_issues.ts|ISSUE"
      )

      declare -a pools=(
        "AUTO-IMP|implementation-worker|${imp_max_workers}|IMP_WORK_GROUPS"
        "AUTO-MRG|pr-merge-worker|${merge_max_workers}|MRG_WORK_GROUPS"
        "AUTO-REV|pr-review-worker|${rev_max_workers}|REV_WORK_GROUPS"
        "AUTO-GROOM|grooming-worker|${groom_max_workers}|GROOM_WORK_GROUPS"
      )

      # Shared cache dir for this iteration — pools reuse each other's results
      # (e.g. REV's scripts are all a subset of IMP's, so REV gets them for free)
      local cache_dir="/tmp/opencode-builder-cache-$$/iter-${N}"
      mkdir -p "$cache_dir"

      for pool_spec in "${pools[@]}"; do
        IFS='|' read -r prefix agent max wg_array_name <<< "$pool_spec"

        log "[pool:${prefix}] starting pool processing (agent=${agent}, max=${max})"
        log "[pool:${prefix}] deduplicate_workers..."
        deduplicate_workers "$prefix"
        log "[pool:${prefix}] cleanup_orphan_tmp_dirs..."
        cleanup_orphan_tmp_dirs "$prefix"

        local exclude_numbers=","
        log "[pool:${prefix}] find_sessions_by_prefix '${prefix}' 'all'..."
        local all_sessions_json
        all_sessions_json=$(find_sessions_by_prefix "$prefix" "all")
        if [[ -z "$all_sessions_json" ]]; then
          if is_timeout "${RUN_TIMED_EXITCODE:-0}"; then
            log_warn "[pool:${prefix}] find_sessions_by_prefix timed out after ${SESSION_SCRIPT_TIMEOUT}s — skipping pool"
            continue
          fi
          log_warn "[pool:${prefix}] find_sessions_by_prefix returned empty — proceeding with 0 known sessions"
          all_sessions_json="[]"
        fi
        local total_count
        total_count=$(echo "$all_sessions_json" | jq 'length' 2>/dev/null || echo 0)
        log "[pool:${prefix}] total sessions found: ${total_count}"

        log "[pool:${prefix}] get_busy_worker_tags..."
        local busy_tags
        busy_tags=$(get_busy_worker_tags "$prefix")
        local busy_count=0
        while IFS= read -r tag; do
          [[ -n "$tag" ]] || continue
          local num
          num=$(echo "$tag" | sed -n 's/.*-\(PR\|ISSUE\)-\([0-9]\+\)$/\2/p')
          if [[ -n "$num" ]]; then
            exclude_numbers="${exclude_numbers}${num},"
            ((busy_count++))
          fi
        done <<< "$busy_tags"
        log "[pool:${prefix}] busy_count=${busy_count} (from ${total_count} sessions)"

        local available_slots=$((max - busy_count))
        if [[ "$available_slots" -le 0 ]]; then
          log "Pool ${prefix}: ${busy_count}/${max} active — full, skipping"
          continue
        fi

        log "Pool ${prefix}: ${busy_count}/${max} active (${total_count} total sessions, ${available_slots} slots free)"
        did_anything=true

        local wg_array_ref="${wg_array_name}[@]"
        local work_groups=("${!wg_array_ref}")

        # ── PRE-FETCH: launch ALL list scripts in parallel, cache results ────
        # Each list_prs_*.ts script fetches ALL open PRs from Forgejo (~90s) and
        # applies a filter. Running them sequentially = 11×90s = 990s.
        # Running them in parallel  = ~90s total (one round of Forgejo API calls).
        log "[pool:${prefix}] pre-fetching ${#work_groups[@]} work-group scripts in parallel..."
        local -A script_cache
        cache_pids=()
        cache_keys=()

        for type_spec in "${work_groups[@]}"; do
          IFS='|' read -r wg_label wg_scripts wg_kind <<< "$type_spec"
          local scripts_arr
          IFS=',' read -ra scripts_arr <<< "$wg_scripts"
          for script_path in "${scripts_arr[@]}"; do
            [[ -n "$script_path" ]] || continue
            local cache_key="${script_path##*/}"  # just filename
            cache_key="${cache_key%.ts}"
            local out_file="${cache_dir}/${cache_key}.json"
            # Skip if a prior pool in this iteration already fetched this script
            # and produced a valid JSON array.
            if [[ -s "$out_file" ]] && [[ "$(head -c1 "$out_file" 2>/dev/null)" == "[" ]]; then
              log "[pool:${prefix}]   $(basename "$script_path"): reusing cached result from this iteration"
              continue
            fi
            cache_keys+=("$cache_key")
            log "[pool:${prefix}]   launching $(basename "$script_path") (cache key: ${cache_key})..."
            timeout "$SCRIPT_TIMEOUT_SEC" npx --yes tsx "$script_path" \
              --url "$FORGEJO_URL" \
              --pat "$FORGEJO_PAT" \
              --owner "$FORGEJO_OWNER" \
              --repo "$FORGEJO_REPO" \
              > "$out_file" 2> "${out_file}.err" &
            cache_pids+=($!)
          done
        done

        log "[pool:${prefix}] waiting for ${#cache_pids[@]} pre-fetch jobs to complete..."
        local cache_idx=0
        local timed_out=0
        local failed=0
        for pid in "${cache_pids[@]}"; do
          wait "$pid" 2>/dev/null
          local ec=$?
          local ck="${cache_keys[$cache_idx]}"
          local out_file="${cache_dir}/${ck}.json"
          if is_timeout $ec; then
            log_warn "[pool:${prefix}]   ${ck}: TIMEOUT after ${SCRIPT_TIMEOUT_SEC}s — treating as empty"
            echo "[]" > "$out_file"
            ((timed_out++))
          elif [[ $ec -ne 0 ]]; then
            log_warn "[pool:${prefix}]   ${ck}: FAILED (exit ${ec}) — treating as empty"
            local snippet
            snippet=$(head -n 20 "${out_file}.err" 2>/dev/null | tr '\n' '|' | sed 's/|$//')
            if [[ -n "$snippet" ]]; then
              log_warn "  stderr: ${snippet}"
            fi
            echo "[]" > "$out_file"
            ((failed++))
          elif [[ -s "$out_file" ]]; then
            local first_char
            first_char=$(head -c1 "$out_file" 2>/dev/null)
            if [[ "$first_char" == "[" ]]; then
              log "[pool:${prefix}]   ${ck}: ready ($(wc -l < "$out_file" | tr -d ' ') lines)"
            else
              log_warn "[pool:${prefix}]   ${ck}: produced non-JSON output (starts with '${first_char}') — treating as empty"
              echo "[]" > "$out_file"
              ((failed++))
            fi
          else
            log_warn "[pool:${prefix}]   ${ck}: empty output with no error — treating as empty"
            echo "[]" > "$out_file"
          fi
          ((cache_idx++))
        done
        if (( timed_out > 0 || failed > 0 )); then
          log_warn "[pool:${prefix}] pre-fetch summary: ${timed_out} timed out, ${failed} failed (of ${#cache_pids[@]} total)"
        else
          log "[pool:${prefix}] pre-fetch complete: all ${#cache_pids[@]} scripts succeeded"
        fi

        # ── Worker selection: iterate cached work-groups ─────────────────────
        local launched=0
        for type_spec in "${work_groups[@]}"; do
          if (( launched >= available_slots )); then
            break
          fi

          IFS='|' read -r wg_label wg_scripts wg_kind <<< "$type_spec"
          log "[pool:${prefix}] work-group '${wg_label}': selecting workers..."

          local scripts_arr
          IFS=',' read -ra scripts_arr <<< "$wg_scripts"
          for script_path in "${scripts_arr[@]}"; do
            if (( launched >= available_slots )); then
              break 2
            fi
            [[ -n "$script_path" ]] || continue

            local cache_key="${script_path##*/}"
            cache_key="${cache_key%.ts}"
            local out_file="${cache_dir}/${cache_key}.json"
            local json
            json=$(cat "$out_file" 2>/dev/null || echo "[]")

            local raw_count
            raw_count=$(echo "$json" | jq 'length' 2>/dev/null || echo 0)
            log "[pool:${prefix}]   script $(basename "$script_path"): ${raw_count} items in cache"

            if [[ "$raw_count" -eq 0 || "$raw_count" == "null" || "$raw_count" == "" ]]; then
              continue
            fi

            local -a items_arr=()
            local -a item_work_nums=()
            local -a item_kinds=()
            while IFS= read -r item; do
              [[ -n "$item" ]] || continue

              local work_num
              work_num=$(echo "$item" | jq -r 'if type == "object" then .number // empty else empty end')
              if [[ -z "$work_num" || ! "$work_num" =~ ^[0-9]+$ ]]; then
                continue
              fi
              if echo ",${exclude_numbers}," | grep -q ",${work_num},"; then
                continue
              fi

              local kind="$wg_kind"
              if [[ "$kind" == "AUTO" ]]; then
                if echo "$item" | jq -e '.head // empty' >/dev/null 2>&1; then
                  kind="PR"
                else
                  kind="ISSUE"
                fi
              fi

              items_arr+=("$item")
              item_work_nums+=("$work_num")
              item_kinds+=("$kind")
            done < <(jq -c '.[]' "$out_file" 2>/dev/null)

            if [[ ${#items_arr[@]} -eq 0 ]]; then
              continue
            fi

            local batch_count
            if [[ ${#items_arr[@]} -lt $((available_slots - launched)) ]]; then
              batch_count=${#items_arr[@]}
            else
              batch_count=$((available_slots - launched))
            fi

            log "[pool:${prefix}]   '${wg_label}': launching ${batch_count} workers in parallel (${#items_arr[@]} candidates)"

            local -a pids=()
            local -a tmp_files=()
            local -a work_nums_arr=()

            for ((i = 0; i < batch_count; i++)); do
              local item="${items_arr[i]}"
              local work_num="${item_work_nums[i]}"
              local kind="${item_kinds[i]}"

              local worker_tag="${prefix}-${kind}-${work_num}"
              local tmp_prompt
              tmp_prompt=$(mktemp "/tmp/ocb-${worker_tag}-${$}.XXXXXX.txt") || continue

              local prompt_text=""
              case "$agent" in
                *implementation*) prompt_text=$(build_impl_prompt "$item" "$kind" "$work_num") ;;
                *merge*)          prompt_text=$(build_merge_prompt "$item" "$work_num") ;;
                *review*)         prompt_text=$(build_review_prompt "$item" "$work_num") ;;
                *grooming*)       prompt_text=$(build_groom_prompt "$item" "$kind" "$work_num") ;;
              esac

              # Skip if the build function returned an empty prompt — sending
              # an empty body to prompt_async results in "no user message
              # found in stream" errors on the server.
              if [[ -z "$prompt_text" ]]; then
                log_warn "[pool:${prefix}]   built empty prompt for ${kind} ${work_num} (${agent}) — skipping launch"
                rm -f "$tmp_prompt"
                continue
              fi

              printf '%s' "$prompt_text" > "$tmp_prompt"

              # Belt-and-braces: confirm the file actually has content on disk
              # before launching the session.
              if [[ ! -s "$tmp_prompt" ]]; then
                log_warn "[pool:${prefix}]   prompt file empty after write for ${kind} ${work_num} (${agent}) — skipping launch"
                rm -f "$tmp_prompt"
                continue
              fi

              $SESSION_START --server "${BASE}" --tag "$worker_tag" --agent "$agent" --prompt-file "$tmp_prompt" --restart \
                >/dev/null 2>&1 &
              pids+=($!)
              tmp_files+=("$tmp_prompt")
              work_nums_arr+=("$work_num")
            done

            for ((i = 0; i < ${#pids[@]}; i++)); do
              if wait "${pids[i]}" 2>/dev/null; then
                exclude_numbers="${exclude_numbers}${work_nums_arr[i]},"
                ((launched++))
                did_anything=true
                log "[${prefix}]   launched ${work_nums_arr[i]} (${agent})"
              else
                log "[${prefix}]   FAILED ${work_nums_arr[i]}"
              fi
              rm -f "${tmp_files[i]}"
            done
          done
        done

        log "Pool ${prefix}: done — ${launched} workers launched this iteration"
      done

      # Clean up the iteration's shared cache dir
      rm -rf "$cache_dir"
    fi

    # ── Step 3: Health & maintenance cycle ────────────────────────────────
    if run_health_cycle; then
      did_anything=true
    fi

    # ── Step 4: Idle sleep if nothing happened ────────────────────────────
    if ! $did_anything; then
      log "Nothing to do — sleeping ${SLEEP_WHEN_IDLE}s"
      record_idle_event
      sleep "$SLEEP_WHEN_IDLE"
    fi
  done

  log "Loop stopped. Shutting down."
  exit 0
}

main_loop