@seanyao/roll 2026.522.2 → 2026.523.2

package/CHANGELOG.md

@@ -1,5 +1,51 @@
 # Changelog
 
+## v2026.523.2
+
+### Added
+
+- **每轮 cycle 实时打出 7 阶段进入退出与耗时** — `startup` / `preflight` / `worktree_setup` / `claude_invoke` / `publish_push` / `publish_wait_merge` / `cleanup`，长静默阶段 30–60s 一次心跳，再也不用盯着空屏猜 loop 卡在哪 `[loop]`
+- **cycle 结束打一份按耗时降序的阶段面板** — 同时把每阶段秒数固化到 `runs.jsonl` 的 `phases` 字段，跨轮对比"今天哪一步最慢"有完整数据 `[loop]`
+- **`roll loop runs` 每行末尾追加 `slowest=<阶段> <占比>%`，新子命令 `roll loop runs --detail <cycle_id>` 打开完整阶段面板** — 不必翻 ndjson 就能定位单轮瓶颈 `[dashboard]`
+- **`tests/run.sh --affected [base-ref]`** — TCR micro-step 只跑被本次改动覆盖的测试文件，反馈秒级；改动 `tests/helpers/*` 等保守降级到全套 `[testing]`
+- **`tests/run.sh --tier=fast|slow|all`** — 本地 TCR 默认 `fast`，CI 跑 `--tier=all`；`ROLL_TEST_TIME_CAP=1` 时 `fast` 全跑 60s 硬上限，性能漂移立刻打红 `[testing]`
+- **测试质量评分卷（6 类反模式 rubric）** — `roll-.dream` Scan 7 按 rubric 输出 `REFACTOR-XXX [test-quality:❶|❷|...]`，单轮上限 5 条，存量 ❶ 类反模式可结构化治理 `[dream]`
+- **`roll lang [zh|en|--list]`** — i18n 地基命令：查看 / 切换 / 列出可用语言，按本机 locale 自动选默认；`ROLL_LANG` 环境变量临时覆盖 `[i18n]`
+- **`roll prices show / refresh`** — 价格表挪进带版本号的 snapshot，能一键拉官方文档对比；调价不再要改源码、提 PR，每个 version 的价格都查得到 `[pricing]`
+- **cycle 结束时按当时价格固化成本** — 以后调价或升级都不会回头改写历史轮次的数字；老事件继续可读，渲染时打 `[legacy]` 标记 `[loop]`
+
+### Fixed
+
+- **dashboard 把已合并 PR 错显示成"开放"** — `roll loop status` 兜底走 git log 找 squash merge subject，原正则只认 `loop/cycle-LABEL` 旧分支名，对新的 `loop cycle LABEL (#N)` 默认 subject 匹配不上；现在两种都认 `[dashboard]`
+- **`model_prices.bats` 改用 fixture 价格表注入** — 之前对生产 PRICES 算术断言把调价误伤成测试红，重写后只断结构不变量（cache_read < input、out ≥ in 等）`[testing]`
+- **已合入临时分支堆在 origin 不被清理** — `_loop_cleanup_stale_cycle_branches` 漏过 agent / abort 两条路径，现在 cycle 入口先扫一遍再走主流程 `[loop]`
+- **临时目录复现 bug 时留下幽灵 launchd 服务** — 自动检测临时路径并把所有 `launchctl` 调用走 sandbox 注册表，主用户 launchd domain 不再被污染 `[launchd]`
+- **dashboard 把 story 列错显示成上一轮的 ID** — `_STORY_ID_PAT` 老正则每段只认纯字母，`US-I18N-001` 这种字母+数字混合 segment 匹配不上 → 新 `pick_todo` 被静默丢掉，仍展示前一轮的旧 story 名；正则更新为 `[A-Z][A-Z0-9]*(?:-[A-Z][A-Z0-9]*)*-\d+` 后 I18N / K8S / D2 / S3 等命名都能识别 `[dashboard]`
+
+### Docs
+
+- **新增 `guide/{en,zh}/loop.md` "Cycle phases" 小节** — 七阶段表 + 触发时机 + 典型耗时 + 收尾面板样例；FAQ 增 C6 "某阶段慢怎么定位"
+- **新增 `guide/{en,zh}/testing.md` "TCR Test Strategy" 小节** — `--affected` 与 `--tier` 两个杠杆的语义、匹配规则、CI/本地默认组合；新增 "测试质量评分卷" 小节列六类反模式与修复方向
+
+## v2026.523.1
+
+### Added
+
+- **`roll loop branches`** — 一眼看见本机残留的 loop 分支；每轮入口先 GC 一次，半途中止的 cycle 也会被收掉 `[loop]`
+
+### Changed
+
+- **dashboard token 列拆成 input / output / cache 写 / cache 读** — cache 是真花钱的，账单终于解释得清 `[loop]`
+
+### Fixed
+
+- **每日 dream / brief 在 macOS 26.4 上从来没真跑过** — 换成 interval 触发，从今天起稳定每天产出 `[loop]`
+- **dashboard 上 tcr 次数、built 列表、ALERT 文案不再显示假零或别故事的旧标签** `[loop]`
+- **选一个故事不再把别的依赖它的故事也标成"在做"** — dashboard 不再骗你说有人在干活 `[loop]`
+- **`roll setup` / `roll update` 不再在隐藏的覆盖提示上无声卡死**
+- **`$roll-notes` 现在写到 `.roll/notes/`** — 和 dream / brief 一致，不再扔到项目根目录 `[loop]`
+- **loop CI 网关不再把"排队中 / 进行中"误判成失败** `[loop]`
+
 ## v2026.522.2
 
 ### Changed

package/bin/dream-test-quality-scan

@@ -0,0 +1,110 @@
+#!/usr/bin/env bash
+# dream-test-quality-scan — ad-hoc helper for roll-.dream Scan 7.
+#
+# Walks bats files and flags ❶-class anti-patterns (hardcoded business data
+# in assertion bodies). Emits structured REFACTOR-shaped lines so the
+# maintainer can sanity-check the rubric against the current suite without
+# waiting for the nightly dream cycle.
+#
+# Usage:
+#   dream-test-quality-scan [--category N] [--path PATH] [--max N]
+#   dream-test-quality-scan --help
+#
+# Only category 1 (❶ hardcoded business data) is implemented as a deterministic
+# heuristic; categories ❷..❻ stay with the dream skill (AI agent applies the
+# rubric). The helper exists so a smoke test and a maintainer dry-run can
+# confirm ❶ detection keeps working as the suite evolves.
+
+set -euo pipefail
+
+CATEGORY=1
+TARGET=""
+MAX=5
+
+usage() {
+  cat <<'EOF'
+dream-test-quality-scan — Scan 7 ❶ dry-run helper
+
+Usage:
+  dream-test-quality-scan [--category N] [--path PATH] [--max N]
+  dream-test-quality-scan --help
+
+Options:
+  --category N   Anti-pattern category (only 1 is implemented; default 1)
+  --path PATH    File or directory to scan (default: tests/)
+  --max N        Maximum entries to emit (default: 5; matches dream skill cap)
+  --help         Show this message
+
+Output:
+  One line per finding:
+    [test-quality:❶] <file>:<line> — <one-line description>
+  Exit code is 0 even when nothing is found (dry-run is informational).
+EOF
+}
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --category) CATEGORY="${2:-1}"; shift 2 ;;
+    --path)     TARGET="${2:-}"; shift 2 ;;
+    --max)      MAX="${2:-5}"; shift 2 ;;
+    --help|-h)  usage; exit 0 ;;
+    *)          echo "unknown flag: $1" >&2; usage >&2; exit 2 ;;
+  esac
+done
+
+if [[ "$CATEGORY" -ne 1 ]]; then
+  echo "category $CATEGORY not yet implemented — only ❶ is mechanical" >&2
+  exit 0
+fi
+
+# Default scan root.
+if [[ -z "$TARGET" ]]; then
+  repo_root=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
+  TARGET="${repo_root}/tests"
+fi
+
+if [[ ! -e "$TARGET" ]]; then
+  echo "path not found: $TARGET" >&2
+  exit 2
+fi
+
+scan_file() {
+  local file="$1"
+  # ❶ heuristic — assertion lines whose RHS contains a numeric literal:
+  #   - lines containing `[[` or `[ ` (bats assertion syntax)
+  #   - AND containing `==` or `=` (equality)
+  #   - AND containing a decimal/integer literal of length ≥ 1 inside quotes
+  # Emits one entry per file (not per line) to stay under the rate cap.
+  local first_hit
+  first_hit=$(grep -nE '\[\[.*"[^"]*[0-9]+(\.[0-9]+)?[^"]*"' "$file" 2>/dev/null \
+              | head -1 || true)
+  [[ -z "$first_hit" ]] && return 1
+
+  local lineno
+  lineno=$(echo "$first_hit" | cut -d: -f1)
+  local rel
+  rel=$(python3 -c "import os,sys; print(os.path.relpath(sys.argv[1]))" "$file" 2>/dev/null || echo "$file")
+  printf '[test-quality:❶] %s:%s — assertion body hardcodes a numeric literal that likely owns its value elsewhere\n' \
+    "$rel" "$lineno"
+  return 0
+}
+
+emitted=0
+if [[ -d "$TARGET" ]]; then
+  # Iterate over .bats files under TARGET; stop after MAX hits.
+  # Exclude vendored bats-core helpers — those are framework tests, not ours.
+  while IFS= read -r f; do
+    case "$f" in
+      */tests/helpers/bats-*/*) continue ;;
+    esac
+    if scan_file "$f"; then
+      emitted=$((emitted + 1))
+      [[ "$emitted" -ge "$MAX" ]] && break
+    fi
+  done < <(find "$TARGET" -type f -name '*.bats' | sort)
+else
+  scan_file "$TARGET" && emitted=1 || true
+fi
+
+# Always succeed — dry-run is informational, the dream cycle decides what to do.
+exit 0