path: root/.ai/scripts/tests/task-review-staleness.bats

#!/usr/bin/env bats
#
# Tests for claude-templates/.ai/scripts/task-review-staleness.sh —
# counts top-level todo.org tasks whose review has gone stale.
#
# Strategy: write a synthetic todo.org into a temp dir per test, with
# LAST_REVIEWED dates generated relative to the real `date` (never
# hardcoded). Run the real script against it and assert the count it
# prints on stdout.
#
# Staleness rule under test:
#   - A qualifying task is a depth-2 (**) heading with a TODO/DOING/VERIFY
#     keyword and an [#A]/[#B]/[#C] priority cookie.
#   - It is stale when LAST_REVIEWED is missing/malformed (NIL → oldest),
#     or when its age strictly exceeds the threshold (age > N days).
#   - age == N exactly is fresh (the spec's wording is ">N days").

# The script under test is always the sibling-of-parent of this test file
# (scripts/task-review-staleness.sh next to scripts/tests/). This holds in
# both the canonical claude-templates/ tree and the rsync'd project mirror,
# so the suite runs identically from either location.
SCRIPT="$(cd "$(dirname "$BATS_TEST_FILENAME")/.." && pwd)/task-review-staleness.sh"

setup() {
  TEST_DIR="$(mktemp -d -t task-review-bats.XXXXXX)"
  TODO="$TEST_DIR/todo.org"

  TODAY="$(date +%F)"
  D5="$(date -d '5 days ago' +%F)"
  D30="$(date -d '30 days ago' +%F)"
  D31="$(date -d '31 days ago' +%F)"
  D40="$(date -d '40 days ago' +%F)"
}

teardown() {
  rm -rf "$TEST_DIR"
}

# Emit a qualifying task with an explicit LAST_REVIEWED date.
task_reviewed() {
  local keyword="$1" prio="$2" title="$3" date="$4"
  printf '** %s [#%s] %s\n:PROPERTIES:\n:LAST_REVIEWED: %s\n:END:\nBody.\n\n' \
    "$keyword" "$prio" "$title" "$date" >> "$TODO"
}

# Emit a qualifying task with no PROPERTIES drawer at all.
task_unreviewed() {
  local keyword="$1" prio="$2" title="$3"
  printf '** %s [#%s] %s\nBody.\n\n' "$keyword" "$prio" "$title" >> "$TODO"
}

# Emit a qualifying task whose LAST_REVIEWED is an org-native inactive
# timestamp — [YYYY-MM-DD Day] — matching the CREATED:/CLOSED: cookies that
# sit in the same drawer. The date is derived from an ISO date via `date`.
task_reviewed_org() {
  local keyword="$1" prio="$2" title="$3" isodate="$4"
  local org="[$(date -d "$isodate" '+%F %a')]"
  printf '** %s [#%s] %s\n:PROPERTIES:\n:LAST_REVIEWED: %s\n:END:\nBody.\n\n' \
    "$keyword" "$prio" "$title" "$org" >> "$TODO"
}

# ---- Normal cases ----------------------------------------------------

@test "staleness: empty file reports zero" {
  : > "$TODO"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "0" ]
}

@test "staleness: all tasks fresh reports zero" {
  task_reviewed TODO A "Fresh one" "$D5"
  task_reviewed TODO B "Fresh two" "$D5"
  task_reviewed DOING A "Fresh three" "$TODAY"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "0" ]
}

@test "staleness: all tasks stale reports full count" {
  task_reviewed TODO A "Stale one" "$D40"
  task_reviewed TODO B "Stale two" "$D40"
  task_reviewed VERIFY C "Stale three" "$D40"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "3" ]
}

@test "staleness: mixed fresh, stale, and unreviewed counts only the latter two" {
  task_reviewed TODO A "Fresh" "$D5"
  task_reviewed TODO B "Stale" "$D40"
  task_unreviewed DOING A "Never reviewed"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "2" ]
}

@test "staleness: org-native bracketed LAST_REVIEWED parses — recent is fresh" {
  task_reviewed_org TODO A "Reviewed five days ago, org stamp" "$D5"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "0" ]
}

@test "staleness: org-native bracketed LAST_REVIEWED parses — old is stale" {
  task_reviewed_org TODO A "Reviewed forty days ago, org stamp" "$D40"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "1" ]
}

# ---- Boundary cases --------------------------------------------------

@test "staleness: age exactly equal to threshold is fresh" {
  task_reviewed TODO A "Exactly at cutoff" "$D30"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "0" ]
}

@test "staleness: age one day past threshold is stale" {
  task_reviewed TODO A "One day over" "$D31"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "1" ]
}

@test "staleness: unreviewed task (no drawer) counts as stale" {
  task_unreviewed TODO A "Never reviewed"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "1" ]
}

@test "staleness: threshold of 7 is softer than 30 on the same list" {
  task_reviewed TODO A "Reviewed five days ago" "$D5"
  task_reviewed TODO B "Reviewed thirty-one days ago" "$D31"
  run bash "$SCRIPT" "$TODO" 7
  [ "$status" -eq 0 ]
  [ "$output" = "1" ]
}

# ---- Error / exclusion cases -----------------------------------------

@test "staleness: DONE and CANCELLED tasks are excluded even when old" {
  task_reviewed DONE A "Shipped long ago" "$D40"
  task_reviewed CANCELLED B "Abandoned long ago" "$D40"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "0" ]
}

@test "staleness: deeper headings and cookie-less headings are excluded" {
  # Depth-3 child with an old review date — not a review unit.
  printf '*** TODO [#A] Child task\n:PROPERTIES:\n:LAST_REVIEWED: %s\n:END:\n\n' "$D40" >> "$TODO"
  # Depth-2 but no priority cookie — not a review unit.
  printf '** TODO Cookie-less task\nBody.\n\n' >> "$TODO"
  run bash "$SCRIPT" "$TODO" 30
  [ "$status" -eq 0 ]
  [ "$output" = "0" ]
}

@test "staleness: malformed LAST_REVIEWED warns to stderr and is not counted" {
  task_reviewed TODO A "Bad date" "not-a-date"
  # stdout carries only the count — the malformed stamp is not folded in.
  run bash -c "bash '$SCRIPT' '$TODO' 30 2>/dev/null"
  [ "$status" -eq 0 ]
  [ "$output" = "0" ]
  # stderr carries the loud warning naming the offending value.
  run bash -c "bash '$SCRIPT' '$TODO' 30 2>&1 1>/dev/null"
  [ "$status" -eq 0 ]
  [[ "$output" == *"not-a-date"* ]]
  [[ "$output" == *"LAST_REVIEWED"* ]]
}

@test "staleness: malformed stamp is excluded while real stale tasks still count" {
  task_reviewed TODO A "Real stale" "$D40"
  task_reviewed TODO B "Broken stamp" "garbage"
  run bash -c "bash '$SCRIPT' '$TODO' 30 2>/dev/null"
  [ "$status" -eq 0 ]
  [ "$output" = "1" ]
}

@test "staleness: missing todo file exits non-zero" {
  run bash "$SCRIPT" "$TEST_DIR/does-not-exist.org" 30
  [ "$status" -ne 0 ]
}

# ---- --list mode -----------------------------------------------------

@test "staleness --list: orders oldest-reviewed first, unreviewed before dated" {
  task_reviewed TODO A "Reviewed recently" "$D5"
  task_reviewed TODO B "Reviewed long ago" "$D40"
  task_unreviewed DOING A "Never reviewed"
  run bash "$SCRIPT" --list "$TODO" 10
  [ "$status" -eq 0 ]
  [[ "${lines[0]}" == *"Never reviewed"* ]]
  [[ "${lines[1]}" == *"Reviewed long ago"* ]]
  [[ "${lines[2]}" == *"Reviewed recently"* ]]
}

@test "staleness --list: org-native bracketed stamp sorts by its real date" {
  task_reviewed TODO A "Bare recent" "$D5"
  task_reviewed_org TODO B "Org-stamped old" "$D40"
  run bash "$SCRIPT" --list "$TODO" 10
  [ "$status" -eq 0 ]
  # The org-bracketed old stamp must sort ahead of the bare recent one —
  # proof it parsed to a real date rather than falling to 0000-00-00.
  [[ "${lines[0]}" == *"Org-stamped old"* ]]
  [[ "${lines[1]}" == *"Bare recent"* ]]
}

@test "staleness --list: takes only the requested count" {
  task_unreviewed TODO A "First"
  task_reviewed TODO B "Second" "$D40"
  task_reviewed TODO C "Third" "$D30"
  run bash "$SCRIPT" --list "$TODO" 2
  [ "$status" -eq 0 ]
  [ "${#lines[@]}" -eq 2 ]
}

@test "staleness --list: count larger than available returns all candidates" {
  task_unreviewed TODO A "Only one"
  run bash "$SCRIPT" --list "$TODO" 10
  [ "$status" -eq 0 ]
  [ "${#lines[@]}" -eq 1 ]
}

@test "staleness --list: excludes DONE, deeper headings, and cookie-less headings" {
  task_reviewed DONE A "Shipped" "$D40"
  printf '*** TODO [#A] Child task\n\n' >> "$TODO"
  printf '** TODO Cookie-less\n\n' >> "$TODO"
  task_unreviewed TODO B "Real one"
  run bash "$SCRIPT" --list "$TODO" 10
  [ "$status" -eq 0 ]
  [ "${#lines[@]}" -eq 1 ]
  [[ "${lines[0]}" == *"Real one"* ]]
}

@test "staleness --list: emits line, review value, and heading tab-separated" {
  task_unreviewed TODO A "Never reviewed"
  run bash "$SCRIPT" --list "$TODO" 10
  [ "$status" -eq 0 ]
  line_no=$(printf '%s' "${lines[0]}" | cut -f1)
  value=$(printf '%s' "${lines[0]}" | cut -f2)
  heading=$(printf '%s' "${lines[0]}" | cut -f3)
  [ "$line_no" = "1" ]
  [ "$value" = "NONE" ]
  [ "$heading" = "** TODO [#A] Never reviewed" ]
}

@test "staleness --list: empty file produces no output" {
  : > "$TODO"
  run bash "$SCRIPT" --list "$TODO" 10
  [ "$status" -eq 0 ]
  [ -z "$output" ]
}

@test "staleness --list: missing file exits non-zero" {
  run bash "$SCRIPT" --list "$TEST_DIR/nope.org" 10
  [ "$status" -ne 0 ]
}