Skip to content

LoopX v0.2.13

Choose a tag to compare

View all tags
@huangruiteng huangruiteng released this 24 Jul 04:13
v0.2.13
7232dca
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

LoopX v0.2.13

Monitor transitions now follow through

Material monitor writeback now resolves the exact todo before target-key fallback, reloads status after mutation, and immediately exposes a newly runnable successor. If the persisted write succeeds but projection reload fails, LoopX returns a structured stale-projection warning through one narrow normalization helper instead of reporting that the write itself failed (#2455).

Continuous-monitor todos may no longer carry resume_when. Invalid persisted rows must be migrated directly; there is no runtime legacy bypass:

loopx todo update \
  --goal-id <goal-id> \
  --todo-id <todo-id> \
  --clear-resume-when \
  --execute

Agent lanes keep their own recommendations

Status and quota now select agent-lane run recommendations by exact agent id, preventing one peer's monitor or replan action from leaking into another peer's lane. Projection warnings are removed when their corresponding scoped recommendation is absent, while goal-level status remains separate (#2491, #2494).

Agent-bound bootstrap guidance now emits explicit agent identity and agent_lane scope for refresh-state, using the shared command renderer rather than a parallel command shape (#2486).

Durable vision and replan contracts

Local waits preserve omitted durable vision fields, and autonomous durable-vision changes require an explicit goal-path delta. Repeated repeat_until_closed replans stalled on the same successor must now provide a runnable successor or an explicit supersede delta instead of acknowledging another watch-only loop (#2485, #2492).

Acknowledged PR-review reconciliation

Issue-fix workflows can persist an exact owner acknowledgement and reconcile it with an exact terminal GitHub PR observation. Reconciliation is host-neutral, idempotent per todo revision, and fail-closed for stale or mismatched bindings; terminal PR state alone never completes the reminder (#2479).

Opt-in SkillsBench split control

SkillsBench gains an optional reverse-channel Codex provider with structural runner binding, fixed pre-agent receipts, workspace-write enforcement, and redacted public output. Exact-host execution remains the default (#2489).

This release does not change benchmark scoring, task semantics, leaderboard, upload, submission, or permission boundaries, launches no benchmark jobs, and makes no benchmark-uplift claim.

Toolchain and release consistency

The test extra is capped below Ruff 0.16 so CI stays on the qualified lint behavior (#2490). The package version, generated manpage, and public tag are all 0.2.13 / v0.2.13.

The exact pre-release main commit passed GitHub Python Tests and all seven Full Public Smoke shards plus smoke-fleet-health. On the release branch, version, manpage, release-readiness docs, build artifacts, update/install paths, Python compile, diff hygiene, public/private boundaries, and the risk-based premerge canary passed. The low-frequency live-model gate and complete exact-release qualification reducer were not rerun for this version-only commit.

Update

loopx update --check
loopx update --execute
loopx doctor

Full changelog

中文摘要

Monitor 写回会继续推进后继任务

物化 monitor 变更会先按精确 todo id 定位，写回后重载 status，并让新就绪的 successor 立即进入下一次 quota 决策。若状态已经成功写入、但投影重载失败，LoopX 会返回结构化 stale-projection warning，而不会误报写入失败（#2455）。

continuous_monitor 不再允许 resume_when，旧的非法状态需要直接执行上面的 todo update --clear-resume-when --execute 迁移；运行时不保留 legacy bypass。

Agent lane 推荐严格隔离

Status/quota 会按精确 agent id 读取 lane-run 推荐，避免其他 peer 的 monitor/replan 动作串入当前 lane；当 scoped recommendation 不存在时，对应的陈旧 warning 也会一起移除（#2491、#2494）。Agent-bound 启动指引现在会显式携带 agent identity 与 agent_lane scope（#2486）。

Durable vision 与 PR review 闭环

本地等待不会丢失 durable vision 字段；重复 repeat_until_closed replan 必须给出可运行 successor 或显式 supersede delta，不能继续用 watch-only acknowledgement 原地循环（#2485、#2492）。

Issue-fix workflow 现在可将精确 owner acknowledgement 与精确 GitHub terminal PR 观察进行 host-neutral reconciliation；仅有 terminal PR 状态不会自动完成提醒（#2479）。

可选 SkillsBench split control

SkillsBench 新增默认不启用的 reverse-channel Codex provider；exact-host 仍是默认路径，且 public output 会隐藏本地连接细节（#2489）。本版本不改变 benchmark 评分、任务语义、排行榜、上传、提交或权限边界，不启动 benchmark job，也不宣称 benchmark 提升。

发布一致性

包版本、生成 manpage 与公开 tag 统一为 0.2.13 / v0.2.13。发布前 main 的 GitHub Python Tests、7 个 Full Public Smoke 分片及 smoke-fleet-health 全部通过；release 分支的版本、文档、构建、更新/安装、编译、diff、边界扫描与 risk-based premerge canary 也已通过。

Assets 2
Loading