@seanyao/roll 4.629.2 → 4.630.2

package/CHANGELOG.md

@@ -2,6 +2,36 @@
 
 ## Unreleased
 
+## v4.630.2 — 2026-06-30
+
+### 自动化流水线
+
+- **评分能读懂真实 agent 的输出了**：评审给分时夹带的终端控制字符、启动横幅、JSONL 包装或项目符号前缀都会被自动清理，只要回复里有一段完整的分数/结论/理由就能采信；同时，只要装了别的 agent，干活的 agent 就绝不会给自己的交付打分——要么由独立评审打分，要么诚实地不打分并保留卡片不上线。(FIX-1044) `[loop-engine]`
+
+### 可见性
+
+- **v4 角色日志入口补齐**：文档说明 Builder、Peer Reviewer、Evaluator、采信分数和解析失败原始产物分别该去哪里看，方便 supervisor 诊断多 agent 协作链路。(US-OBS-036) `[loop-observability]`
+
+## v4.630.1 — 2026-06-30
+
+### 自动化流水线
+
+- **Reasonix builder 可以在 linked worktree 里正常提交了**：loop 现在会给每个 Reasonix cycle 生成临时本地 sandbox 配置，把已计算出的 git common dir 写入允许写目录；不需要永久放宽 `~/.reasonix/config.toml`，也不会把 API key 写进 argv 或配置文件。(FIX-1036) `[loop-engine]`
+- **Loop 交付不会再误报完成**：PR、CI 或验收链路缺证时，Roll 会停在可恢复状态，不再把没真正上线的卡记成已交付。(FIX-1032a, FIX-1032c, FIX-1039) `[loop-engine]`
+- **Scoped `roll loop go` 更守边界**：单卡运行不会被旧计数或定时 tick 带到下一张卡，适合人工盯一张关键修复。(FIX-1034, FIX-1040) `[goal-mode]`
+- **Agent auth 诊断更准**：日志里的 `403` 卡号不再被误判成登录失效，真正的鉴权失败仍会明确阻塞。(FIX-1033) `[loop-engine]`
+- **发版前检查更可信**：`roll release --json` 和一致性闸现在会真实读取待发布内容，旧交付缓存不会再误拦或误放。(FIX-1030, FIX-1041) `[release-management]`
+
+### 新功能
+
+- `roll design` 不再在已有 backlog 的项目里意外启动完整设计 session，而是先给出可选的几种目标方式（`--from-file`、直接需求文本等）(FIX-1031) `[loop]`
+
+### 可见性
+
+- 验收报告和故事档案开始显示执行阵容（Execution Cast），每位 Builder、Peer Reviewer、Evaluator 和验收门禁在交付页面上一眼可见（US-OBS-034）`[loop]`
+- **Cycle 角色链可直接查看**：`roll cycle --roles` 会展示 Builder、Peer Reviewer、Evaluator 和采信结果，不用再翻事件流。(US-OBS-033) `[loop]`
+- **评审解析失败不再消失**：有返回但格式不对的 reviewer/scorer 会显示原因和原始产物位置。(US-OBS-035) `[loop]`
+
 ## v4.629.2 — 2026-06-29
 
 ### Agent 管理

package/README.md

@@ -65,6 +65,18 @@ Roll V4 separates project coordination from story delivery:
 - **Skills remain** the capability layer. Roles invoke `$roll-design`, `$roll-build`, `$roll-fix`, `$roll-peer`, `$roll-.qa`, and related skills instead of rewriting those contracts into TypeScript.
 - **Fallback is fail-loud**. If a requested agent or rig is unavailable, Roll records that unavailability and pauses or asks for owner action; it does not silently pretend another agent was used.
 
+### Supervisor backlog-clearing standard
+
+When the owner asks Roll to clear a backlog, Supervisor treats the scope as every
+live non-Hold `FIX-*`, `US-*`, and `REFACTOR-*` row unless the owner narrows it.
+Before scheduling another card it reconciles backlog status, open PRs, recent
+cycle endings, CI/evaluator gates, manual-merge PRs, and `.roll` meta state.
+Each card gets its own Builder and, when required by the execution profile, an
+independent Evaluator/Scorer selected from the current Agent roster. Repeated
+failure, zero TCR, missing PR/CI/evaluator evidence, parser failures, auth
+blocks, permission blocks, and `[roll:manual-merge]` PRs stop new scheduling and
+surface an owner action through `roll supervisor status/next/why`.
+
 ### Operating modes
 
 Roll has two product modes over the same backlog, truth, route profile, execution
@@ -186,6 +198,7 @@ with `roll loop resume` when ready.
 | `roll pair [init\|status\|score]` | Cross-agent pairing: heterogeneous peer re-checks and scores deliveries |
 | `roll cycles [--since 1d\|3d\|7d\|all]` | The cycle ledger — one line per cycle, failures never swallowed |
 | `roll cycle <id>` | One cycle's full trace tape (cycle→story→build→peer→ci→pr→end); the same trace is persisted as `.roll/loop/cycle-<id>.signals.jsonl` for replay |
+| `roll cycle <id> --roles [--json]` | The execution cast — who was Builder and who was Evaluator, which peers were consulted, and the single accepted score; also written to `summary.md`/`summary.json` and the Execution Cast report block ([guide](guide/en/loop.md#cycle-role-visibility)) |
 | `roll peer [--reviewer <agent>] (--prompt <text>\|--file <path>)` | One-shot structured external-provider review; records `.roll/peer/runs.jsonl` |
 | **Config & machine** | |
 | `roll ls [--json] [--stale-days <n>]` | List the cross-project registry (`~/.roll/projects.json`): name · tag · verdict · path; missing/stale rows flagged, never dropped |