#!/usr/bin/env bash
# Block until a pull request's CI reaches a terminal state, then emit one snapshot.
#
# Usage:
#   scripts/pr-await <PR> [options]
#
# Options:
#   --mode all-terminal | strict-deadline | first-failure
#                        all-terminal (default) waits until no check is pending,
#                        so every failure lands in a single report and a single
#                        fix pass. first-failure returns as soon as one check
#                        fails; it exists for the interactive case and costs an
#                        extra full CI cycle per failure. strict-deadline keeps
#                        collecting state until the deadline, even after checks
#                        become terminal; it is for an explicit fixed-duration
#                        wait.
#   --deadline-min N     Hard stop, in minutes (default 45).
#   --interval-sec N     Seconds between polls (default 60).
#   --format text | json Report format (default text).
#   --quiet              Suppress per-poll progress on stderr.
#
# Why this exists: `pr-state --summary` is a one-shot snapshot, so waiting for CI
# costs one model round-trip per snapshot. Measured over 6,658 such round-trips,
# each re-read ~177k of cached context at ~$0.26. This script moves the waiting
# below the model: one invocation, one report, regardless of how long CI runs.
#
# stdout carries ONLY the final report, so a caller can capture it verbatim.
# Progress goes to stderr.
#
# Exit codes:
#   0  terminal and clean      - no failed checks, no unresolved threads
#   1  terminal with findings  - failures and/or review threads to act on
#   2  deadline reached        - checks still pending or terminal rollup not confirmed
#   3  blocked                 - merged/closed, approval required, or access lost
#
# Requires: gh (authenticated), jq, and scripts/pr-state alongside this script.

set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
PR_STATE="${PR_AWAIT_PR_STATE:-$SCRIPT_DIR/pr-state}"
GH="${PR_AWAIT_GH:-gh}"
SLEEP="${PR_AWAIT_SLEEP:-sleep}"
ACTIVE_READ_PID=''
ACTIVE_READ_GROUP=''
ACTIVE_WATCHDOG_PID=''

# Blocked results leave the poll loop before a snapshot exists, but a caller
# that asked for JSON still needs to parse the answer.
emit_blocked() {
  local outcome=$1 message=$2
  if [[ "${FORMAT:-text}" == 'json' ]]; then
    local waited_sec="${WAITED:-0}"
    local polls="${POLLS:-0}"
    local head_changes="${HEAD_CHANGES:-0}"
    local deadline_sec="${DEADLINE_SEC:-0}"
    local interval_sec="${INTERVAL_SEC:-0}"
    local mergeable_value='UNKNOWN'
    local merge_status_value='UNKNOWN'
    if jq -e . >/dev/null 2>&1 <<<"${MERGE_JSON:-}"; then
      mergeable_value="$(jq -r '.mergeable // "UNKNOWN"' <<<"$MERGE_JSON")"
      merge_status_value="$(jq -r '.mergeStateStatus // "UNKNOWN"' <<<"$MERGE_JSON")"
    fi
    jq -n \
      --argjson pr "$PR" \
      --arg outcome "$outcome" \
      --arg mode "${MODE:-all-terminal}" \
      --argjson waited_sec "$waited_sec" \
      --argjson polls "$polls" \
      --argjson head_changes "$head_changes" \
      --argjson exit_code 3 \
      --arg message "$message" \
      --arg toolchain "${TOOLCHAIN:-unverified}" \
      --argjson deadline_sec "$deadline_sec" \
      --argjson interval_sec "$interval_sec" \
      --arg mergeable "$mergeable_value" \
      --arg merge_state_status "$merge_status_value" \
      '{pr: $pr, outcome: $outcome, mode: $mode, waited_sec: $waited_sec,
        polls: $polls, head_changes: $head_changes, exit_code: $exit_code,
        message: $message, toolchain: $toolchain, evidence_complete: false,
        deadline_sec: $deadline_sec, interval_sec: $interval_sec,
        mergeable: $mergeable, merge_state_status: $merge_state_status,
        summary: null}'
  else
    printf 'PR %s\n%s\n' "$PR" "$message"
    toolchain_hint
  fi
  exit 3
}

usage() {
  sed -n '2,/^# Requires:/p' "$0" | sed 's|^# \{0,1\}||'
  exit "${1:-1}"
}

PR=''
MODE='all-terminal'
DEADLINE_MIN=45
INTERVAL_SEC=60
FORMAT='text'
QUIET=0

while [[ $# -gt 0 ]]; do
  case "$1" in
    --mode)
      shift
      [[ $# -gt 0 ]] || usage 1
      case "$1" in
        all-terminal | strict-deadline | first-failure) MODE=$1 ;;
        *) usage 1 ;;
      esac
      shift
      ;;
    --deadline-min)
      shift
      [[ $# -gt 0 && "$1" =~ ^[0-9]+$ ]] || usage 1
      # 10# forces base 10. Without it "09" aborts arithmetic under set -e and
      # "010" silently means 8, because Bash reads a leading zero as octal.
      DEADLINE_MIN=$((10#$1))
      shift
      ;;
    --interval-sec)
      shift
      [[ $# -gt 0 && "$1" =~ ^[0-9]+$ && $((10#$1)) -gt 0 ]] || usage 1
      INTERVAL_SEC=$((10#$1))
      shift
      ;;
    --format)
      shift
      [[ $# -gt 0 ]] || usage 1
      case "$1" in
        text | json) FORMAT=$1 ;;
        *) usage 1 ;;
      esac
      shift
      ;;
    --quiet)
      QUIET=1
      shift
      ;;
    -h | --help) usage 0 ;;
    -*) usage 1 ;;
    *)
      [[ -z "$PR" ]] || usage 1
      PR=$1
      shift
      ;;
  esac
done

[[ -n "$PR" && "$PR" =~ ^[0-9]+$ ]] || usage 1
# Normalize so a value like 0012 is 12 everywhere, including `jq --argjson`.
PR=$((10#$PR))
# "0" and "00" both pass the pattern above and normalize to 0, which is not a
# PR number GitHub ever issues; reject it here instead of chasing it through
# three failed pr-state reads as a confusing blocked-access outcome.
[[ "$PR" -gt 0 ]] || usage 1

progress() {
  [[ "$QUIET" -eq 1 ]] || printf '%s\n' "$*" >&2
}

# A current-head check rollup can briefly contain only the first materialized
# workflows while reporting no failed or pending checks. Require two identical
# terminal snapshots so a later workflow registration is observed before the
# caller receives a clean verdict.
TERMINAL_ROLLUP_FINGERPRINT=''
TERMINAL_ROLLUP_STREAK=0
TERMINAL_ROLLUP_PROVISIONAL=0

terminal_rollup_fingerprint() {
  jq -c '
    def stable_checks:
      sort_by([
        (.workflow // ""),
        (.name // ""),
        (.status // ""),
        (.conclusion // ""),
        ((.run_id // "") | tostring),
        ((.job_id // "") | tostring)
      ]);
    {
      check_count: (.check_count // 0),
      passed_check_count: (.passed_check_count // 0),
      successful_checks: ((.successful_checks // []) | stable_checks),
      terminal_checks: ((.terminal_checks // []) | stable_checks),
      failed_checks: ((.failed_checks // []) | stable_checks),
      pending_checks: ((.pending_checks // []) | stable_checks)
    }' <<<"$1"
}

augment_required_status_checks() {
  # pr-state performs this augmentation in production. Keep this defensive
  # fallback for raw snapshots supplied by tests or another snapshot producer.
  jq -c '
    if (.required_status_checks_known // false) != true then
      .
    else
      ([.successful_checks[]?, .terminal_checks[]?, .failed_checks[]?, .pending_checks[]? | .name] | unique) as $reported
      | ([.required_status_checks[]?
          | . as $required
          | select(($reported | index($required)) == null)
          | {name: $required, workflow: null, status: "missing", run_id: null,
             job_id: null, details_url: null, target_url: null}
        ]) as $missing
      | .pending_checks = ((.pending_checks // []) + $missing)
      | .check_count = ((.check_count // 0) + ($missing | length))
    end
  '
}

# Mergeability is deliberately absent from `pr-state --summary`, so it is
# queried separately below rather than letting a caller assume a clean merge
# state. There is no `|| {}` fallback: a failed query is unknown state, and the
# caller must treat it that way instead of reading an empty object as
# "not conflicting".

TIMEOUT_BIN=''
TIMEOUT_RESOLVED=0
resolve_timeout_bin() {
  [[ "$TIMEOUT_RESOLVED" -eq 0 ]] || return 0
  TIMEOUT_RESOLVED=1
  # Set PR_AWAIT_NO_TIMEOUT=1 to exercise the shell fallback on a host that does
  # have timeout(1). The fallback is the path stock macOS takes, so it needs
  # coverage somewhere other than stock macOS.
  [[ -z "${PR_AWAIT_NO_TIMEOUT:-}" ]] || return 0
  local candidate
  for candidate in timeout gtimeout; do
    if command -v "$candidate" >/dev/null 2>&1; then
      TIMEOUT_BIN="$candidate"
      return 0
    fi
  done
}

stop_active_read() {
  local pid="${ACTIVE_READ_PID:-}"
  local group="${ACTIVE_READ_GROUP:-}"
  local watchdog="${ACTIVE_WATCHDOG_PID:-}"

  # The bounded command is a job-control process-group leader on the fallback
  # path. Stop the group first so a pr-state shell cannot leave its gh child
  # running. The KILL fallback makes cancellation deterministic if a child
  # ignores TERM. The direct kill also covers timeout(1), whose child group is
  # not portable across implementations.
  if [[ -n "$group" ]]; then
    kill -TERM -- "-$group" 2>/dev/null || true
    kill -KILL -- "-$group" 2>/dev/null || true
  fi
  [[ -z "$pid" ]] || kill -TERM "$pid" 2>/dev/null || true
  [[ -z "$pid" ]] || kill -KILL "$pid" 2>/dev/null || true
  [[ -z "$watchdog" ]] || kill -KILL "$watchdog" 2>/dev/null || true
}

# Run a command with a wall-clock bound, keeping its stdout. Used for the GitHub
# reads so a stalled `pr-state` or `gh pr view` cannot outlive the caller's
# deadline: the poll loop can only check the clock between reads, never during.
run_bounded_capture() {
  local secs=$1 outfile=$2
  shift 2
  [[ -z "$outfile" ]] || : >"$outfile"

  resolve_timeout_bin
  if [[ -n "$TIMEOUT_BIN" ]]; then
    local had_monitor=0
    case "$-" in
      *m*) had_monitor=1 ;;
    esac
    set -m
    if [[ -n "$outfile" ]]; then
      "$TIMEOUT_BIN" "$secs" "$@" >"$outfile" 2>/dev/null &
    else
      "$TIMEOUT_BIN" "$secs" "$@" >/dev/null 2>&1 &
    fi
    local pid=$!
    [[ "$had_monitor" -eq 1 ]] || set +m
    ACTIVE_READ_PID="$pid"
    ACTIVE_READ_GROUP="$pid"
    local rc=0
    wait "$pid" 2>/dev/null || rc=$?
    ACTIVE_READ_PID=''
    ACTIVE_READ_GROUP=''
    return "$rc"
  fi

  # Fallback for hosts with no timeout(1), stock macOS among them.
  #
  # Job control makes the background job a process-group leader so the watchdog
  # can kill the whole tree; killing the pr-state shell alone leaves the `gh` it
  # is blocked on running. The watchdog polls in one-second steps and exits by
  # itself once the command finishes, because a shell blocked in `sleep N` does
  # not act on a signal until that sleep returns. Its fds are closed because
  # callers capture this with $(...), and any background process holding that
  # pipe keeps the substitution blocked afterwards.
  local had_monitor=0
  case "$-" in
    *m*) had_monitor=1 ;;
  esac
  set -m
  if [[ -n "$outfile" ]]; then
    "$@" >"$outfile" 2>/dev/null &
  else
    "$@" >/dev/null 2>&1 &
  fi
  local pid=$!
  [[ "$had_monitor" -eq 1 ]] || set +m
  ACTIVE_READ_PID="$pid"
  ACTIVE_READ_GROUP="$pid"

  (
    waited=0
    while [[ "$waited" -lt "$secs" ]]; do
      sleep 1
      kill -0 "$pid" 2>/dev/null || exit 0
      waited=$((waited + 1))
    done
    kill -9 -"$pid" 2>/dev/null || kill -9 "$pid" 2>/dev/null
  ) >/dev/null 2>&1 &
  ACTIVE_WATCHDOG_PID=$!

  local rc=0
  wait "$pid" 2>/dev/null || rc=$?
  kill "$ACTIVE_WATCHDOG_PID" 2>/dev/null || true
  wait "$ACTIVE_WATCHDOG_PID" 2>/dev/null || true
  ACTIVE_WATCHDOG_PID=''
  ACTIVE_READ_PID=''
  ACTIVE_READ_GROUP=''
  # 137 is SIGKILL, which here can only have come from the watchdog.
  [[ "$rc" -ne 137 ]] || return 124
  return "$rc"
}

# Seconds left before the caller's deadline, floored at zero.
remaining_budget() {
  local left=$((DEADLINE_SEC - (SECONDS - START_SEC)))
  [[ "$left" -gt 0 ]] || left=0
  printf '%s' "$left"
}

# Cap on a single GitHub read. A healthy `pr-state --summary` returns in about
# ten seconds; three minutes means it is wedged, and waiting out a 45 minute
# default deadline on one wedged read helps nobody.
READ_CAP=180

# A zero or already-spent deadline still gets exactly one snapshot, so the
# report can name the pending checks rather than saying nothing at all. That
# read is bounded tightly: a healthy read returns in about ten seconds, and a
# caller who asked for no waiting should not get a long one.
READ_FLOOR="${PR_AWAIT_READ_FLOOR:-15}"

# Recorded in every report as provenance, not as a gate. `pr-state` used to fail
# silently on jq 1.6 and on the bash 3.2 that macOS ships as /bin/bash, and this
# script refused to run there. Both are fixed at the source now, and blocking on
# a version would only turn away working installations.
#
# What replaced the gate is stronger than it was: the poll loop rejects a
# snapshot that reports `errors`, one whose unresolved-thread count is null, and
# one that does not parse. Those test the data rather than a proxy for it, so
# they catch a degraded pr-state whatever the cause, including a checkout that
# predates the fix.
record_toolchain() {
  local jq_bin="${PR_AWAIT_JQ:-jq}"
  local bash_bin="${PR_AWAIT_PROBE_BASH:-bash}"
  JQ_VERSION="$("$jq_bin" --version 2>/dev/null || printf 'unknown')"
  BASH_VERSION_SEEN="$("$bash_bin" -c 'printf "%s" "${BASH_VERSION%%(*}"' 2>/dev/null || printf 'unknown')"
  TOOLCHAIN="$JQ_VERSION, bash $BASH_VERSION_SEEN"
}

# Appended to a blocked report only. A reader whose pr-state produced nothing
# usable is the one person who needs to know their toolchain is a known cause.
toolchain_hint() {
  local hint=''
  case "$JQ_VERSION" in
    jq-1.[0-6] | jq-1.[0-6].* | jq-1.[0-6]-*)
      hint="$JQ_VERSION loops on an empty-matching gsub; a pr-state predating that fix hangs on it." ;;
  esac
  case "$BASH_VERSION_SEEN" in
    [123].*)
      hint="${hint:+$hint }bash $BASH_VERSION_SEEN errors on an empty array under set -u; a pr-state predating that fix loses review threads there." ;;
  esac
  [[ -z "$hint" ]] || printf '\nToolchain note: %s\n' "$hint"
}

TOOLCHAIN='unverified'
JQ_VERSION='unknown'
BASH_VERSION_SEEN='unknown'
record_toolchain

READ_TMP="$(mktemp "${TMPDIR:-/tmp}/pr-await-read-XXXXXX")"
cleanup_read_tmp() {
  [[ -z "${READ_TMP:-}" ]] || rm -f "$READ_TMP"
}

handle_signal() {
  local code=$1
  stop_active_read
  exit "$code"
}

trap cleanup_read_tmp EXIT
trap 'handle_signal 130' INT
trap 'handle_signal 143' TERM

DEADLINE_SEC=$((DEADLINE_MIN * 60))
# Tests can use a short deterministic deadline without making the public
# minute-based interface less clear. This is intentionally not documented as a
# user-facing option.
if [[ -n "${PR_AWAIT_DEADLINE_SEC:-}" ]]; then
  [[ "${PR_AWAIT_DEADLINE_SEC}" =~ ^[0-9]+$ ]] || usage 1
  DEADLINE_SEC=$((10#${PR_AWAIT_DEADLINE_SEC}))
fi

strict_deadline_reached() {
  if [[ "$MODE" == 'strict-deadline' && "$TERMINAL_ROLLUP_STREAK" -ge 2 ]]; then
    OUTCOME='strict-terminal'
  else
    OUTCOME='deadline'
  fi
}

# Sleep until the next poll unless the deadline arrives first, capping the last
# sleep to whatever time is left. Returns 1 when the caller must stop, so a
# retry path cannot outlive a limit the caller asked for.
sleep_or_stop() {
  local remaining=$((DEADLINE_SEC - (SECONDS - START_SEC)))
  [[ "$remaining" -gt 0 ]] || return 1
  if [[ "$remaining" -lt "$INTERVAL_SEC" ]]; then
    "$SLEEP" "$remaining"
  else
    "$SLEEP" "$INTERVAL_SEC"
  fi
}
START_SEC=$SECONDS
POLLS=0
CONSECUTIVE_ERRORS=0
MERGE_ERRORS=0
MERGE_TIMED_OUT=0
EVIDENCE_ERRORS=0
MERGE_UNKNOWN=0
REQUIRED_STATUS_UNKNOWN=0
HEAD_CHANGES=0
FIRST_HEAD=''
SUMMARY_JSON=''
MERGE_JSON=''
OUTCOME=''

while :; do
  POLLS=$((POLLS + 1))
  ELAPSED=$((SECONDS - START_SEC))

  # A spent deadline still buys one snapshot, but one for the whole iteration:
  # granting the floor again to the second query would let --deadline-min 0 wait
  # out two floor-length reads.
  floor_start=''
  budget="$(remaining_budget)"
  if [[ "$budget" -le 0 ]]; then
    if [[ -n "$SUMMARY_JSON" ]]; then
      strict_deadline_reached
      break
    fi
    floor_start=$SECONDS
    budget="$READ_FLOOR"
  fi
  [[ "$budget" -le "$READ_CAP" ]] || budget="$READ_CAP"

  read_rc=0
  run_bounded_capture "$budget" "$READ_TMP" "$PR_STATE" --summary "$PR" || read_rc=$?
  if [[ "$read_rc" -eq 124 ]]; then
    progress "poll ${POLLS}: pr-state exceeded its ${budget}s bound"
    OUTCOME='deadline'
    [[ -n "$SUMMARY_JSON" ]] || OUTCOME='blocked-access'
    break
  fi
  summary_body="$(cat "$READ_TMP")"
  # A non-zero read is a failed read, whatever it printed. Checking only the
  # parse would let a pr-state that exited 1 through on an empty snapshot.
  # Staged in a local first: assigning straight into SUMMARY_JSON would
  # clobber the last good snapshot with this failed read's garbage, so a
  # deadline reached right after would lose evidence a working poll had
  # already produced.
  if [[ "$read_rc" -ne 0 ]] || ! jq -e . >/dev/null 2>&1 <<<"$summary_body"; then
    CONSECUTIVE_ERRORS=$((CONSECUTIVE_ERRORS + 1))
    progress "poll ${POLLS}: pr-state failed (${CONSECUTIVE_ERRORS} in a row)"
    # Three consecutive failures is a lost connection, not a transient blip.
    # Reporting "blocked" beats looping silently against revoked access.
    if [[ "$CONSECUTIVE_ERRORS" -ge 3 ]]; then
      OUTCOME='blocked-access'
      break
    fi
    if ! sleep_or_stop; then
      OUTCOME='blocked-access'
      break
    fi
    continue
  fi
  summary_body="$(augment_required_status_checks <<<"$summary_body")"
  SUMMARY_JSON="$summary_body"
  CONSECUTIVE_ERRORS=0

  budget="$(remaining_budget)"
  if [[ "$budget" -le 0 ]]; then
    if [[ -n "$MERGE_JSON" ]]; then
      OUTCOME='deadline'
      break
    fi
    if [[ -n "$floor_start" ]]; then
      # Whatever the first read left of this iteration's single floor.
      budget=$((READ_FLOOR - (SECONDS - floor_start)))
      [[ "$budget" -gt 0 ]] || budget=1
    else
      floor_start=$SECONDS
      budget="$READ_FLOOR"
    fi
  fi
  [[ "$budget" -le "$READ_CAP" ]] || budget="$READ_CAP"

  read_rc=0
  run_bounded_capture "$budget" "$READ_TMP" \
    "$GH" pr view "$PR" --json state,mergeable,mergeStateStatus,reviewDecision,headRefOid || read_rc=$?
  if [[ "$read_rc" -eq 124 ]]; then
    progress "poll ${POLLS}: gh pr view exceeded its ${budget}s bound"
    OUTCOME='blocked-merge-state'
    MERGE_TIMED_OUT=1
    break
  fi
  merge_body="$(cat "$READ_TMP")"
  # Staged in a local first, same reason as the pr-state read above: a failed
  # read must not overwrite the last good MERGE_JSON with unusable text.
  if [[ "$read_rc" -ne 0 ]] || ! jq -e . >/dev/null 2>&1 <<<"$merge_body"; then
    MERGE_ERRORS=$((MERGE_ERRORS + 1))
    progress "poll ${POLLS}: gh pr view failed (${MERGE_ERRORS} in a row)"
    if [[ "$MERGE_ERRORS" -ge 3 ]]; then
      OUTCOME='blocked-merge-state'
      break
    fi
    if ! sleep_or_stop; then
      OUTCOME='blocked-merge-state'
      break
    fi
    continue
  fi
  MERGE_JSON="$merge_body"
  MERGE_ERRORS=0

  pr_state="$(jq -r '.state // ""' <<<"$MERGE_JSON")"
  head_oid="$(jq -r '.headRefOid // ""' <<<"$MERGE_JSON")"
  [[ -n "$head_oid" ]] || head_oid="$(jq -r '.pr.head_ref_oid // ""' <<<"$SUMMARY_JSON")"
  [[ -n "$FIRST_HEAD" ]] || FIRST_HEAD="$head_oid"

  # A push during the wait restarts CI, so every check counted so far belongs to
  # a head that no longer exists. Discard that evidence, but not the caller's
  # clock: --deadline-min is a hard stop, and a PR receiving periodic commits
  # would otherwise defer it forever.
  if [[ -n "$head_oid" && "$head_oid" != "$FIRST_HEAD" ]]; then
    HEAD_CHANGES=$((HEAD_CHANGES + 1))
    progress "poll ${POLLS}: head moved ${FIRST_HEAD:0:9} -> ${head_oid:0:9}; discarding stale evidence"
    FIRST_HEAD="$head_oid"
    EVIDENCE_ERRORS=0
    MERGE_UNKNOWN=0
    # The snapshot in hand describes the previous head. Keeping it would let a
    # deadline reached before the next read print those checks under the new
    # SHA; with no snapshot the run reports blocked, which is the truth.
    SUMMARY_JSON=''
    TERMINAL_ROLLUP_FINGERPRINT=''
    TERMINAL_ROLLUP_STREAK=0
    TERMINAL_ROLLUP_PROVISIONAL=0
    if ! sleep_or_stop; then
      OUTCOME='deadline'
      break
    fi
    continue
  fi

  if [[ "$pr_state" == 'MERGED' || "$pr_state" == 'CLOSED' ]]; then
    OUTCOME='blocked-pr-state'
    break
  fi

  failed="$(jq -r '.failed_checks | length' <<<"$SUMMARY_JSON")"
  pending="$(jq -r '.pending_checks | length' <<<"$SUMMARY_JSON")"
  check_count="$(jq -r '.check_count // 0' <<<"$SUMMARY_JSON")"
  passed="$(jq -r '.passed_check_count // 0' <<<"$SUMMARY_JSON")"
  complete="$(jq -r '.checks_snapshot_complete' <<<"$SUMMARY_JSON")"
  checks_head="$(jq -r '.checks_head_sha // ""' <<<"$SUMMARY_JSON")"

  progress "poll ${POLLS} (+$((SECONDS - START_SEC))s): ${passed} passed, ${failed} failed, ${pending} pending"

  # Everything below decides on check evidence, so a snapshot describing a head
  # that has already been replaced is discarded outright rather than guarded
  # against case by case. Every snapshot retained past this point is therefore
  # current-head by construction, and a deadline reached with none in hand
  # reports blocked instead of printing the previous head's checks under the
  # new SHA.
  if [[ -z "$checks_head" || "$checks_head" != "$head_oid" ]]; then
    progress "poll ${POLLS}: snapshot is for ${checks_head:0:9}, head is ${head_oid:0:9}; discarding"
    SUMMARY_JSON=''
    if ! sleep_or_stop; then
      OUTCOME='deadline'
      break
    fi
    continue
  fi

  required_status_known="$(jq -r 'if .required_status_checks_known == true then "yes" else "no" end' <<<"$SUMMARY_JSON")"
  if [[ "$required_status_known" != 'yes' ]]; then
    REQUIRED_STATUS_UNKNOWN=$((REQUIRED_STATUS_UNKNOWN + 1))
    progress "poll ${POLLS}: required status policy is unavailable (${REQUIRED_STATUS_UNKNOWN} in a row)"
    if [[ "$REQUIRED_STATUS_UNKNOWN" -ge 3 ]]; then
      OUTCOME='blocked-required-statuses'
      break
    fi
    if ! sleep_or_stop; then
      OUTCOME='blocked-required-statuses'
      break
    fi
    continue
  fi
  REQUIRED_STATUS_UNKNOWN=0

  if [[ "$(jq -r '.approval_required_runs | length' <<<"$SUMMARY_JSON")" -gt 0 ]]; then
    # action_required reads as green-or-skipped in a naive check count. It is
    # blocked verification: waiting for it to finish would never terminate.
    OUTCOME='blocked-approval'
    break
  fi

  # GitHub can expose a new head before Actions has registered any check. An
  # empty rollup satisfies "nothing pending" without any evidence behind it;
  # in this repository the required jobs always run, so zero means not yet.
  if [[ "$check_count" -eq 0 ]]; then
    TERMINAL_ROLLUP_FINGERPRINT=''
    TERMINAL_ROLLUP_STREAK=0
    TERMINAL_ROLLUP_PROVISIONAL=0
    progress "poll ${POLLS}: no checks registered at this head yet; waiting"
    if ! sleep_or_stop; then
      OUTCOME='deadline'
      break
    fi
    continue
  fi

  if [[ "$pending" -gt 0 || "$complete" != 'true' ]]; then
    TERMINAL_ROLLUP_FINGERPRINT=''
    TERMINAL_ROLLUP_STREAK=0
    TERMINAL_ROLLUP_PROVISIONAL=0
  fi

  # A failed job can be visible while its parent workflow is still running, so
  # in all-terminal mode a failure is never on its own a reason to stop.
  if [[ "$MODE" == 'first-failure' && "$failed" -gt 0 ]]; then
    snapshot_errors="$(jq -r '.errors // [] | length' <<<"$SUMMARY_JSON")"
    threads_known="$(jq -r 'if .unresolved_review_thread_count == null then "no" else "yes" end' <<<"$SUMMARY_JSON")"
    if [[ "$complete" != 'true' || "$snapshot_errors" -gt 0 || "$threads_known" == 'no' ]]; then
      EVIDENCE_ERRORS=$((EVIDENCE_ERRORS + 1))
      progress "poll ${POLLS}: first failure visible but evidence incomplete (${EVIDENCE_ERRORS} in a row)"
      if [[ "$EVIDENCE_ERRORS" -ge 3 ]]; then
        OUTCOME='blocked-evidence'
        break
      fi
      if ! sleep_or_stop; then
        OUTCOME='blocked-evidence'
        break
      fi
      continue
    fi
    EVIDENCE_ERRORS=0
    OUTCOME='terminal'
    break
  fi

  if [[ "$pending" -eq 0 && "$complete" == 'true' ]]; then
    # Terminal by check state, but only trustworthy if the snapshot is whole.
    # A partial fetch reports `errors`, and a failed thread query leaves the
    # unresolved count null - which `// 0` would otherwise launder into "clean".
    snapshot_errors="$(jq -r '.errors // [] | length' <<<"$SUMMARY_JSON")"
    threads_known="$(jq -r 'if .unresolved_review_thread_count == null then "no" else "yes" end' <<<"$SUMMARY_JSON")"
    if [[ "$snapshot_errors" -gt 0 || "$threads_known" == 'no' ]]; then
      TERMINAL_ROLLUP_FINGERPRINT=''
      TERMINAL_ROLLUP_STREAK=0
      TERMINAL_ROLLUP_PROVISIONAL=0
      EVIDENCE_ERRORS=$((EVIDENCE_ERRORS + 1))
      progress "poll ${POLLS}: checks terminal but evidence incomplete (${EVIDENCE_ERRORS} in a row)"
      if [[ "$EVIDENCE_ERRORS" -ge 3 ]]; then
        OUTCOME='blocked-evidence'
        break
      fi
      if ! sleep_or_stop; then
        OUTCOME='blocked-evidence'
        break
      fi
      continue
    fi
    EVIDENCE_ERRORS=0

    # GitHub reports UNKNOWN while it is still computing mergeability. Deciding
    # on it would call a PR clean before the answer exists.
    if [[ "$(jq -r '.mergeable // "UNKNOWN"' <<<"$MERGE_JSON")" == 'UNKNOWN' ]]; then
      TERMINAL_ROLLUP_FINGERPRINT=''
      TERMINAL_ROLLUP_STREAK=0
      TERMINAL_ROLLUP_PROVISIONAL=0
      MERGE_UNKNOWN=$((MERGE_UNKNOWN + 1))
      progress "poll ${POLLS}: mergeability still UNKNOWN (${MERGE_UNKNOWN} in a row)"
      if [[ "$MERGE_UNKNOWN" -ge 3 ]]; then
        OUTCOME='blocked-merge-unknown'
        break
      fi
      if ! sleep_or_stop; then
        OUTCOME='blocked-merge-unknown'
        break
      fi
      continue
    fi
    MERGE_UNKNOWN=0

    terminal_fingerprint="$(terminal_rollup_fingerprint "$SUMMARY_JSON")"
    if [[ "$terminal_fingerprint" == "$TERMINAL_ROLLUP_FINGERPRINT" ]]; then
      TERMINAL_ROLLUP_STREAK=$((TERMINAL_ROLLUP_STREAK + 1))
    else
      TERMINAL_ROLLUP_FINGERPRINT="$terminal_fingerprint"
      TERMINAL_ROLLUP_STREAK=1
    fi
    if [[ "$TERMINAL_ROLLUP_STREAK" -lt 2 ]]; then
      TERMINAL_ROLLUP_PROVISIONAL=1
      progress "poll ${POLLS}: terminal check rollup is provisional; waiting for a matching snapshot"
      if ! sleep_or_stop; then
        OUTCOME='deadline'
        break
      fi
      continue
    fi
    TERMINAL_ROLLUP_PROVISIONAL=0

    if [[ "$MODE" != 'strict-deadline' ]]; then
      OUTCOME='terminal'
      break
    fi
    progress "poll ${POLLS}: terminal evidence collected; continuing through the strict deadline"
    if ! sleep_or_stop; then
      strict_deadline_reached
      break
    fi
    continue
  fi

  if ! sleep_or_stop; then
    OUTCOME='deadline'
    break
  fi
done

WAITED=$((SECONDS - START_SEC))

# A failed read can leave a partial, unparseable body behind, and that value
# survives the `continue` that follows it. Normalize before the report rather
# than letting `jq --argjson` abort under `set -e` with no output at all.
jq -e . >/dev/null 2>&1 <<<"$MERGE_JSON" || MERGE_JSON='{}'

if [[ -z "$SUMMARY_JSON" ]] || ! jq -e . >/dev/null 2>&1 <<<"$SUMMARY_JSON"; then
  emit_blocked 'blocked-access' \
"BLOCKED: no usable pr-state snapshot was obtained; treat CI state as unknown."
fi

if [[ "$OUTCOME" == 'blocked-access' ]]; then
  emit_blocked 'blocked-access' \
"BLOCKED: pr-state failed $CONSECUTIVE_ERRORS times in a row; treat CI state as unknown."
fi

if [[ "$OUTCOME" == 'blocked-merge-state' ]]; then
  if [[ "$MERGE_TIMED_OUT" -eq 1 ]]; then
    emit_blocked 'blocked-merge-state' \
"BLOCKED: gh pr view exceeded its read bound, so merge state is unknown.
A conflicted or closed PR is indistinguishable from a clean one here."
  fi
  emit_blocked 'blocked-merge-state' \
"BLOCKED: gh pr view failed $MERGE_ERRORS times in a row, so merge state is unknown.
A conflicted or closed PR is indistinguishable from a clean one here."
fi

failed_n="$(jq -r '.failed_checks | length' <<<"$SUMMARY_JSON")"
pending_n="$(jq -r '.pending_checks | length' <<<"$SUMMARY_JSON")"
passed_n="$(jq -r '.passed_check_count // 0' <<<"$SUMMARY_JSON")"
# `unresolved_review_thread_count` is every unresolved thread (pr-state:922);
# `hidden_unresolved_threads` is the subset filtered out of `unresolved_threads`,
# not an addition to it. Summing the two counts the hidden ones twice.
threads_total_n="$(jq -r '.unresolved_review_thread_count // 0' <<<"$SUMMARY_JSON")"
visible_n="$(jq -r '.unresolved_threads | length' <<<"$SUMMARY_JSON")"
hidden_n="$(jq -r '.hidden_unresolved_threads | length' <<<"$SUMMARY_JSON")"
comments_n="$(jq -r '.actionable_issue_comment_count // 0' <<<"$SUMMARY_JSON")"
changes_req_n="$(jq -r '.review_evidence.active_changes_requested_count // 0' <<<"$SUMMARY_JSON")"
blocked_rev_n="$(jq -r '.review_evidence.blocked_exact_current_head_review_count // 0' <<<"$SUMMARY_JSON")"
evidence_errors_n="$(jq -r '.errors // [] | length' <<<"$SUMMARY_JSON")"
error_sources="$(jq -r '[.errors // [] | .[].source] | join(", ")' <<<"$SUMMARY_JSON")"
threads_known_final="$(jq -r 'if .unresolved_review_thread_count == null then "no" else "yes" end' <<<"$SUMMARY_JSON")"
base_advanced="$(jq -r 'if .pr.base_advanced_since_head == true then 1 else 0 end' <<<"$SUMMARY_JSON")"
mergeable="$(jq -r '.mergeable // "UNKNOWN"' <<<"$MERGE_JSON")"
merge_status="$(jq -r '.mergeStateStatus // "UNKNOWN"' <<<"$MERGE_JSON")"
review_decision="$(jq -r '.reviewDecision // "UNKNOWN"' <<<"$MERGE_JSON")"

conflicted=0
[[ "$mergeable" == 'CONFLICTING' || "$merge_status" == 'DIRTY' ]] && conflicted=1

# A CHANGES_REQUESTED review, or one whose body carries an explicit blocker,
# can exist with no unresolved thread and no issue comment attached. Counting
# only threads would report such a PR clean.
findings=$((failed_n + threads_total_n + comments_n + conflicted + changes_req_n + blocked_rev_n + base_advanced))

case "$OUTCOME" in
  blocked-pr-state | blocked-approval | blocked-evidence | blocked-merge-unknown | blocked-required-statuses) code=3 ;;
  deadline) code=2 ;;
  strict-terminal) [[ "$findings" -gt 0 ]] && code=1 || code=0 ;;
  *) [[ "$findings" -gt 0 ]] && code=1 || code=0 ;;
esac

if [[ "$FORMAT" == 'json' ]]; then
  jq -n \
    --argjson pr "$PR" \
    --argjson summary "$SUMMARY_JSON" \
    --argjson merge "$MERGE_JSON" \
    --arg outcome "$OUTCOME" \
    --arg mode "$MODE" \
    --argjson waited_sec "$WAITED" \
    --argjson polls "$POLLS" \
    --argjson head_changes "$HEAD_CHANGES" \
    --argjson exit_code "$code" \
    --arg toolchain "$TOOLCHAIN" \
    --argjson deadline_sec "$DEADLINE_SEC" \
    --argjson interval_sec "$INTERVAL_SEC" \
    --argjson evidence_complete "$([[ "$evidence_errors_n" -eq 0 && "$threads_known_final" == 'yes' ]] && echo true || echo false)" \
    '{pr: $pr, outcome: $outcome, mode: $mode, waited_sec: $waited_sec, polls: $polls,
      head_changes: $head_changes, exit_code: $exit_code,
      message: null, toolchain: $toolchain, evidence_complete: $evidence_complete,
      deadline_sec: $deadline_sec, interval_sec: $interval_sec,
      mergeable: ($merge.mergeable // "UNKNOWN"),
      merge_state_status: ($merge.mergeStateStatus // "UNKNOWN"),
      summary: $summary}'
  exit "$code"
fi

branch="$(jq -r '.pr.head_ref_name // "?"' <<<"$SUMMARY_JSON")"
base="$(jq -r '.pr.base_ref_name // "?"' <<<"$SUMMARY_JSON")"
printf 'PR %s @ %s (%s -> %s)\n' "$PR" "${FIRST_HEAD:0:9}" "$branch" "$base"
printf 'gate: %s | outcome: %s | waited %dm%02ds over %d poll(s)\n' \
  "$MODE" "$OUTCOME" "$((WAITED / 60))" "$((WAITED % 60))" "$POLLS"
[[ "$HEAD_CHANGES" -gt 0 ]] && printf 'note: head moved %d time(s) during the wait\n' "$HEAD_CHANGES"
printf 'checks: %s passed, %s failed, %s pending\n' "$passed_n" "$failed_n" "$pending_n"
printf 'merge: %s / %s | review: %s\n' "$mergeable" "$merge_status" "$review_decision"
printf 'toolchain: %s\n' "$TOOLCHAIN"

if [[ "$failed_n" -gt 0 ]]; then
  printf '\nFAILED CHECKS:\n'
  jq -r '.failed_checks[] | "  \(.name)  run=\(.run_id // "?") job=\(.job_id // "?")  \(.conclusion // "?")"' <<<"$SUMMARY_JSON"
fi

if [[ "$pending_n" -gt 0 ]]; then
  printf '\nSTILL PENDING:\n'
  jq -r '.pending_checks[] | "  \(.name)  run=\(.run_id // "?")"' <<<"$SUMMARY_JSON"
fi

if [[ "$threads_total_n" -gt 0 ]]; then
  printf '\nUNRESOLVED THREADS (%s total: %s visible, %s hidden):\n' \
    "$threads_total_n" "$visible_n" "$hidden_n"
  jq -r '(.unresolved_threads + .hidden_unresolved_threads)[]
         | "  \(.thread_id)  c=\(.comment_id)  \(.author)  \(.path // "-")"' <<<"$SUMMARY_JSON"
fi

[[ "$comments_n" -gt 0 ]] && printf '\nACTIONABLE ISSUE COMMENTS: %s\n' "$comments_n"

if [[ "$base_advanced" -eq 1 ]]; then
  # code-review/SKILL.md: mergeable=MERGEABLE proves conflict compatibility, not
  # that the merged result was tested. Green checks here ran against an older base.
  printf '\nBASE ADVANCED since this head: checks did not run against the current base.\n'
  printf '  Validate the merge result before treating this PR as ready.\n'
fi

if [[ "$changes_req_n" -gt 0 || "$blocked_rev_n" -gt 0 ]]; then
  printf '\nBLOCKING REVIEWS: %s active changes requested, %s blocked at this head\n' \
    "$changes_req_n" "$blocked_rev_n"
  jq -r '.review_evidence.blocked_exact_current_head_reviews // [] | .[]
         | "  \(.author)  \(.verdict // .state // "?")"' <<<"$SUMMARY_JSON" 2>/dev/null || true
fi

if [[ "$evidence_errors_n" -gt 0 || "$threads_known_final" == 'no' ]]; then
  printf '\nINCOMPLETE EVIDENCE: %s\n' "${error_sources:-unresolved review thread count is null}"
fi

printf '\nNEXT: '
case "$OUTCOME" in
  blocked-pr-state) printf 'PR is %s. Do not push a stale head.\n' "$pr_state" ;;
  blocked-approval) printf 'workflow approval required; CI cannot complete unattended.\n' ;;
  blocked-evidence)
    printf 'checks finished but review evidence is incomplete (%s). Treat as unknown, not clean.\n' \
      "${error_sources:-null unresolved thread count}" ;;
  blocked-merge-unknown)
    printf 'GitHub still reports mergeability as UNKNOWN. Re-run before treating this head as clean.\n' ;;
  blocked-required-statuses)
    printf 'required status check policy is unavailable. Treat the check rollup as unknown, not clean.\n' ;;
  deadline)
    if [[ "$TERMINAL_ROLLUP_PROVISIONAL" -eq 1 ]]; then
      printf 'deadline hit while confirming the terminal check rollup. Re-run or raise --deadline-min.\n'
    elif [[ "$(jq -r '.check_count // 0' <<<"$SUMMARY_JSON")" -eq 0 ]]; then
      printf 'deadline hit with no checks registered yet at this head. Re-run or raise --deadline-min.\n'
    else
      printf 'deadline hit with %s check(s) pending. Re-run or raise --deadline-min.\n' "$pending_n"
    fi
    ;;
  strict-terminal)
    printf 'fixed-duration wait ended with terminal evidence; review the findings above.\n'
    ;;
  *)
    if [[ "$conflicted" -eq 1 ]]; then
      printf 'resolve the merge conflict before triaging checks or threads.\n'
    elif [[ "$findings" -eq 0 && "$review_decision" == 'REVIEW_REQUIRED' && "$merge_status" == 'BLOCKED' ]]; then
      printf 'checks and reviews are clean, but human approval is required before merge.\n'
    elif [[ "$findings" -eq 0 ]]; then
      printf 'clean at this head.\n'
    elif [[ "$failed_n" -eq 0 && "$threads_total_n" -eq 0 ]]; then
      # The finding is entirely from a category the "N failed / M thread"
      # phrasing below cannot name: an issue comment, a blocking review, or an
      # advanced base. Saying "fix 0 failed check(s) and 0 thread(s)" here
      # would tell the caller there is nothing to do while findings > 0.
      printf 'review the findings reported above (comments, blocking reviews, or an advanced base), then push once.\n'
    else
      printf 'fix %s failed check(s) and %s thread(s) in one pass, then push once.\n' \
        "$failed_n" "$threads_total_n"
    fi
    ;;
esac

exit "$code"