本文是面向用户的版本能力概览:每个版本「是什么、能做什么、如何升级」。 逐版本完整明细(Added / Changed / Fixed、技术细节、阶段 ID)见 CHANGELOG.md。 两者分工:VERSIONS = 用户视角摘要;CHANGELOG = Keep a Changelog 明细。本文不记录测试计数与内部审查编号。 覆盖范围:README 曾展示的版本(v1.1.2 起);更早版本(v1.0.0 / v1.1.0 / v1.1.1)见 CHANGELOG.md。
patch;含用户可见新能力/契约变更时 minor;发布前必须全门禁通过(测试 / dsh / tsc / i18n / kb / build ×2)。OpenFeel is a CLI for AI-agent development process governance. Recent releases:
openfeel-dsh npm page now leads with OpenFeel itself (what it is, the two problems, stages/requirements, the two modes, where records live) and then states that this package is OpenFeel’s adapter for DeepSeek Harness; adapter internals moved under an “Adapter details” section. Behavior is unchanged.openfeel-dsh on npm) are rewritten in plain language for non-programmers — pain points first (losing context, jumping steps), then stages/requirements, the two modes, and where records live. Behavior is unchanged.dsh/) is now also published as its own npm package openfeel-dsh, released in lockstep with the CLI and declaring its CLI floor via requiresCli, so community plugin markets (which require an npm identity) can list and install it via dsh plugin add openfeel-dsh. The CLI remains required (npm i -g openfeel); market installs should not also run openfeel setup --tools dsh (it rewrites the dependency to a local link:). Publishing uses npm publish ./dsh from the monorepo — no mirror repository, single source of truth.feel_mode states are now called Feel Mode (enabled, full process enforcement) and Free Mode (disabled, relaxed); the README gains a two-mode comparison (and clarifies that feel_mode is orthogonal to workflow_mode), the DSH prompt uses the same names, and the repo now ships an official format-1 DSH market index (docs/market/index.json, git source pinned to a commit with path:/dsh) that users can add as an index source in DSH’s Discover page; community-market paths and their npm-identity requirement are documented.feel_mode now defaults to disabled, so a fresh install / new user starts relaxed (no forced requirement/stage, preflight warns instead of blocking, the DSH prompt renders “free mode”), while an explicit enabled in the project config or the global profile still pins strict mode (project-over-global). Resolution, immediate effect, and every hard gate (state machine, done gate, stage verification, archive criteria, blocking-REV gate, write guard) are unchanged — only the unconfigured default flips; existing projects that already set feel_mode are unaffected.feel_mode: enabled|disabled (default enabled = unchanged). disabled softens process enforcement only — no forced requirement/stage, preflight downgrades from strict to warn, and the DSH prompt switches to “free mode” — while the state machine, done gate, stage verification, archive criteria, blocking-REV gate and the write guard stay in force. Resolution is project-over-global (.openfeel/config.yaml defaults.feel_mode > global ~/.config/openfeel/profile.yaml preferences.feel_mode > builtin enabled) and takes effect immediately (the adapter re-reads it on every turn). Also fixes requirement-ID reuse: flow archive now raises meta.last_req_seq, and req add / req split fold the archived maximum in before allocating.attempt now keeps checkpoints.exec.attempts in sync with the top-level counter, flow ops list stops warning about empty templates on done/failed ops (the fill value stays truthful; flow health still reports them as data hygiene), and docs now state that req claim only records the owner fields (use req advance to change status), that the fourth shell-gate class is a non-blocking warning with OPENFEEL_SHELL_ALLOW as the exact-command allowlist, and that removing a done op needs --force."user": "unknown" sentinel in .openfeel/.info.json no longer blocks the fallback to git config user.name, so feel_req claim works without an explicit --owner again; ensureInfoJson also heals such existing files when a git identity is available (never writes unknown as a result).flow attempt / flow retry accept short op IDs, and mismatch/ambiguity fails loudly with candidates instead of a silent no-op), the preflight window is configurable (OPENFEEL_PREFLIGHT_MAX_AGE_MS) and the DSH adapter self-heals a stale ledger once before a write, an archive-idle git commit channel (git-commit-idle) covers the post-archive cleanup commit, flow next prefers the pass path (review_passed / test_passed) and warns before archiving → done (run stage verify) and when all linked requirements are terminal, OPENFEEL_SHELL_ALLOW whitelists exact undecidable shell commands, and guard audit/lastDecision commands keep 500 chars instead of 200.git commit is now allowed in exec_running|review_passed|test_passed|archiving, so you can commit right after review or test pass instead of first advancing to archiving (the write side is unchanged — source writes still require exec_running, and review_pending / test_pending / *_failed / plan_* / scheme_* / done stay rejected, with a commit-specific denial message that lists the four allowed phases). Also fixes the openfeel-cli-usage skill frontmatter being truncated at a bare # (“exit #43” → “exit No.43”): the parsed description grows from 2455 to 2766 chars, so the v1.5.10 / v1.5.11 / v1.5.12 capability chains finally appear in the skill’s frontmatter, with a new S6 regression test using the production splitFrontmatter and a frontmatter-only /\s#/ check.guard.stale-never-recovers bug reported by an external project): the guard marker’s identity keys (pid / hostStartedAt / loaded_hash) now come only from the current runtime and stale is decided by pid + loaded_hash alone (the host-started-at false positive is gone; an HMAC mismatch downgrades to hmacWarn instead of stale) — deleting guard-state.json is no longer needed. Supported self-heal repairs only dead-pid states (live foreign pids and loaded_hash mismatches stay untouched) and is also available explicitly on the DSH side via feel_verify action=health fix=true (missing/incomplete marker → marker-incomplete, zero writes). Also: dsh test isolation via OPENFEEL_GUARD_DIR, a seven-condition allowance for pure bookkeeping git commit (all pending paths under .openfeel/**), a whitelist for the git external-execution surface (-c / --config-env and GIT_*), a multi-profile notice for setup --tools dsh, and feel_status / feel_next headlines that no longer say “no active stage” for guard reasons.1.5.10 was not published on its own): shell tools (bash / pwsh) are now gated with four outcomes — hard-deny protected paths and openfeel write commands, judge source / git-workspace writes by target, exempt build artifacts (npx vitest / npm run build / git status keep working), and warn when undecidable; the protected-path list grows (chore.json, dev/.preflight.json, dev/sessions.json, dev/guard-state.json, pipeline.yaml, config.yaml, metrics.json, flow-archive/**, checkpoints/**, requirements/*.md, log/guard-audit.jsonl, dsh/cordis.patch.yml); doc/asset exemptions no longer apply inside code directories; Win32 path normalization (trailing dots / ADS / 8.3 short names); six-link cwd resolution with fail-closed and a guard-state.json self-check (pid / loaded_hash / HMAC) surfaced as feel_status.guard; chore shrinks to a human-only channel (30-minute cap, --source required for source writes, --ttl-min 1..30).openfeel dashboard [--json] [--out] [--open] aggregates requirements / stages / release gates / archives / blocking REVs / health / metrics into a single self-contained HTML file (default .openfeel/tmp/dashboard.html; zero external links, zero network, zero executable scripts); dashboard --json is JSON exit #43 (contract at 43 success exits / 44 registered --json options).sessions.stage clearing semantics, chore.json ignored).docs/releases/<version>.md and harness-aware skills; stages hold multiple requirements (requirements[]); multi-session P0 (sessions registry, auto-commit pollution guard, current-stage rule, feel_status.sessions); completion requires review + verification evidence at both requirement and stage level; completed requirements travel into version archives with a monotonic last_req_seq counter, plus req release-check / req rollover (carried_from) and an --allow-unfinished archive gate.exec_running, holds many requirements); requirement completion is decoupled from stage closure (req advance --to done works with zero linked stages, or while linked stages are at/after exec_running); stages are archived/released only after the user confirms. Adapter prompt and gates now say so, and feel_status defaults to a compact snapshot (detail: "full" for the full one). Also: plan stage add / flow stage add refuse to silently reuse an archived stage’s plan directory.--dry-run is never gated (10 commands), flow repair / flow checkpoint restore stay usable when flow.json is unreadable, and strict rejections are zero-write (including flow.json.bak and config.yaml). Also new: openfeel lint req (JSON exit #41) validates command names in requirement bodies against the real command tree (A-class blocking / B-class advisory), with in-line and region suppression markers. Process-health proxies (flow metrics): resumed ratio / plan_pending → plan_review dwell / last-preflight age in the human block only (--json unchanged), plus adapter observability (feel_status/feel_next), with zero runtime overhead and no self-measurement.flow next / flow status record .preflight.json; write commands check a 10-minute window), adapter defaults to strict preflight (feel_status/feel_next expose preflight), prompt bans shell state-write commands and requires independent review (self-review must be marked).flow log: last transition / last error / resume count), adapter snapshot revision / stale, observability metrics (flow metrics: log entries/bytes), and a gate-abuse hint when a stage sits in plan_pending for more than 10 minutes. No JSON contract changes.feel_next.dispatch: goal/context/boundaries/returnFormat; subagents never read state files, return ≤10 lines), recovery bundle (feel_status.recovery: lastTransition / handoff / nextAction), structured handoff (done / next / block) and no-op sync when state is unchanged. Release cadence is now batched (see “发布节奏”).feel_status/feel_next expose workflowGate when no stage is active; docs/research/** writes are rejected without an active stage (chore cannot bypass), with a denial audit event. Includes the dispatch/persistence research doc and the 1.5.3/1.5.4 iteration plan.natural-compare in package-lock.json to a non-existent 1.5.0, breaking CI npm ci; the lock is repaired (functionality identical to v1.5.0).dev_last* / dev/current.md / status.md / flow.json); subagents never read or write them; logs are no longer mandatory (brief records only, flow log capped at 200); status records update at start / turning point / end or every 30 minutes via openfeel dev sync (CLI auto-syncs on flow advance); feel_next/feel_status expose statusSync.due.chore micro-change channel: static presentational assets (.css/.scss/.html/.svg/.png/.jpg/.jpeg/.webp, assets/**, public/**) are writable in any phase like docs (mixed targets stay strict); openfeel chore "<note>" opens a 1-step, 60-minute write window without creating a stage (no flow.json.stages entry; --end/--status), and guard writes an audit line for every exemption.bash (Git Bash), then pwsh (PowerShell 7); macOS/Linux use bash; avoid defaulting to cmd.exe. Added to the DSH Feel system prompt and the OpenFeel global constraint templates (agents-md / claude constraint / openfeel-tool-usage skill).feel_flow review_add now forwards --stage (stage-level REVs work without ops); openfeel setup --check gains a restart-required state (adapter fingerprint deployed vs loaded) so upgrades no longer loop; setup --tools dsh now refreshes update_state.json even when profiles are already up to date.flow review add --stage) so reviews work without ops; rework edge review_passed → exec_running (flow advance --reason); risk-tiered write guard (new files in an uninitialized workspace, docs/** and doc/note paths allowed; source still requires exec_running); plan stage add focuses atomically; status.md becomes a derived view; stage-ID ↔ plan-directory mapping documented.ctx.tools.guard() denies edit / write / apply_patch / str_replace_editor; non-OpenFeel workspaces get a bootstrap prompt, and enforce=always blocks writes until openfeel init (project / off opt-out).setup --tools dsh now runs pnpm install with --no-frozen-lockfile, so stale profile lockfiles no longer block deployment.openfeel package; openfeel setup --tools dsh installs it into DSH profiles. README was slimmed: version notes moved here.archive <stage> overwrite protection (ARCHIVE_EXISTS / --force), JSON output for the remaining read commands, and config get --all.flow next, error envelopes, exit codes 0/1/2/3) and its concurrency/test-seam fixes.req), three workflow modes (full / fast / general), review tiers.flow archive) and requirement flow (req).npm i -g openfeel@latest → openfeel setup → restart the harness.openfeel-dsh 页面重排:先介绍 OpenFeel 本身(是什么、解决两个问题、版本与阶段、待办清单、两种模式、记录位置),然后说明「这是 OpenFeel 在 DeepSeek Harness 上的适配版本」;安装章节后移。openfeel-dsh 的 npm 页面(dsh/README.md)改成同样的通俗说明,讲清它做什么、不做什么,以及两种安装方式。openfeel-dsh 独立包:适配器 dsh/ 随 CLI 同步发布为独立 npm 包(版本 lockstep、requiresCli 声明所需 CLI 下限),社区插件市场可凭 npm 身份收录并一键安装;单仓打包发布(npm publish ./dsh),不需要镜像子仓库。dsh plugin add openfeel-dsh(市场/npm)或 openfeel setup --tools dsh(本地 link);两者都会写 profile 依赖,不要叠加(setup 会覆盖为 link)。CLI 始终需要全局安装(门禁在 CLI 内)。docs/market/index.json(git 源,索引固定 commit);社区市场(awesome-dsh-plugin / dshmarket / 1024Store / dshbase / dsh.market)用 npm 包 openfeel-dsh。npm i -g openfeel@latest → openfeel setup → 重启宿主。feel_mode: enabled = Feel Mode(流程模式),disabled = Free Mode(自由模式)。配置值、解析层级与门禁行为不变;DSH 提示词状态行改用这两个称呼。feel_mode 与 workflow_mode 是两个正交开关——前者决定是否强制流程,后者决定走流程时的关卡密度。docs/market/index.json(git 源 + 固定 commit + path:/dsh,版本随每次发布生成),配合 GitHub Pages 即可在 DSH「发现 → 添加索引源」中添加;dsh.market 等社区市场的收录方式与限制见 docs/research/dsh-plugin-market-landscape.md。npm i -g openfeel@latest → openfeel setup → 重启宿主。feel_mode 有效值为 disabled(宽松):不强制建需求/阶段、preflight 只提示、DSH 提示词渲染「自由模式」;是否建需求/工作流由 AI 自行判断。.openfeel/config.yaml 的 defaults.feel_mode: enabled 或全局 preferences.feel_mode: enabled 仍优先(项目 > 全局),严格门禁与提示词行为不变;init 模板默认仍不写键,注释说明如何钉死严格。done 门 / 阶段验证门 / 归档判据 / blocking REV / 写入门禁全部保持;已显式配置 feel_mode 的项目行为完全不变。npm i -g openfeel@latest → openfeel setup → 重启宿主。feel_mode: enabled|disabled:enabled(默认)= 现状(要求/催促创建需求与阶段、preflight 先读后写严格生效)。disabled = 软关闭流程强制:不强制建需求/阶段(workflowGate 的 no-active-stage 换为 feel-mode-off,不催办、无标题告警)、preflight 从 strict 降为 warn、DSH <feel_mode> 提示切「自由模式」——是否建需求/工作流由 AI 自行判断。disabled 的边界(刻意保留):状态机合法性、done 门、阶段验证门、版本归档判据、blocking REV 门、写入门禁(禁止手改 flow.json / status.md / pipeline.yaml / config.yaml)、guard-* 门禁失活告警。归档 / 发布仍需用户确认。.openfeel/config.yaml 的 defaults.feel_mode > 全局 ~/.config/openfeel/profile.yaml 的 preferences.feel_mode > builtin=enabled;config get --all 变五键、config get --global 变七键、config effective feel_mode 带来源;适配器每轮现读,改完当轮生效(feel_status / feel_next 输出 feelMode)。该键无阶段级字段(--sync-stages 会报 skipped-no-field)。meta.last_req_seq 原不感知归档,归档过的需求 ID 会被复用(本仓 v1.5.8 归档 REQ-025~030 后计数器仍为 24)。现在 flow archive 在移除需求前抬升计数器,req add / req split 分配前并入 .openfeel/flow-archive/*.json 的最大号。checkpoints.exec.attempts 同步:attempt 后不再出现「顶层 attempts=1、checkpoints.exec.attempts=0」的漂移;pass / fail / 重试耗尽三条路径一致。flow ops list 只对可推进状态提示「模板未填充」;fill 仍如实反映模板填充度(不伪装已填充);flow health 的空模板项保留数据卫生口径(含 done)。plan scheme remove 保护收紧:尝试过一次(exec.attempts>0)的 op 默认拒绝注销(--force 可越过)——此前该子句因计数恒 0 是死代码。req claim 只写 owner/claimed_at,不改状态(推进用 req advance);第 4 类 shell 判定是非阻断告警(审计 + 一行 stderr),已知安全命令用 OPENFEEL_SHELL_ALLOW(|| 分隔、整条全等)白名单;plan scheme remove 删除 done 或存在 checkpoint 进展的 op 需 --force。.openfeel/.info.json 一旦持久化成 {"user":"unknown"}(init 时 git 未配置),getUserName 会把它当真实用户名直接返回,不再回落 git config user.name,于是 feel_req claim 永远要求显式 --owner。unknown(trim + 大小写不敏感)与空串都视为「未解析」,继续读 git config user.name;两源皆无时才返回哨兵 unknown,req claim 仍给出「请显式传 –owner」的可操作提示。ensureInfoJson 遇到已存在的 user 为空/unknown 且 git 身份可用时写回真实用户名(lang 与正常用户值不覆盖;git 不可用时不写 unknown)。flow attempt --op op-001 / flow retry --op op-001 支持当前阶段短 ID(当前阶段优先 → 全 stages 唯一命中);未找到或歧义会列出候选并非 0 退出,不再「打印成功但什么都没写」。flow advance --op 回显解析后的 stage.op。flow archive 人类模式会在写盘前提示候选阶段内未完成的 op 与工作区未提交改动;--json 载荷不变。OPENFEEL_PREFLIGHT_MAX_AGE_MS 放宽(缺省仍 10 分钟),CLI / flow health / 适配器同源;适配器写命令前若账本过期,会先真实读一次状态再写。git-commit-idle——无活跃阶段(stages 空或 pipeline.phase=done)时,单条直连且无仓库改写的 git commit 可提交 docs/源码收尾(纯 .openfeel/** 仍走书账窄口)。flow next 建议改走通过路径(review_passed / test_passed),archiving 前提示先 stage verify,关联需求全部终态时提示阶段可收口;OPENFEEL_SHELL_ALLOW 可白名单整条不可判定命令;门禁审计命令截断 200 → 500 字符。git commit 允许相位扩展:exec_running|archiving → exec_running|review_passed|test_passed|archiving。审查通过或测试通过后即可提交,不必先推进到 archiving;写侧未放宽(源码写与整树 git 写仍只认 exec_running),未通过相位(review_pending / test_pending / *_failed / plan_* / scheme_* / done)仍拒,拒绝时会按 commit 分叉并列出四个允许相位。openfeel-cli-usage skill frontmatter 修复:SKILL.md 的 description 是未加引号 YAML 标量,出口 #43 里的裸 # 被当成注释起点,解析值停在 2455 字符——v1.5.10 / v1.5.11 / v1.5.12 三段能力串与触发末句全部丢失。改为 出口 No.43 后解析长度回到 2766,能力串全部可见;并新增 S6 回归测试(生产 splitFrontmatter + 2766 长度锚 + frontmatter 段专属 /\s#/ 阴性断言 + 注入副本同步)。guard-stale 修复)guard.stale 重启后永久为 true(外部项目报告的缺陷):guard-state.json 的身份三键(pid / hostStartedAt / loaded_hash)只来自当前运行时,不再继承旧文件;stale 判定只认 pid + loaded_hash,删除 host-started-at 误报(多 profile / 多宿主下时间比较必然误报),HMAC 失配只降级为 hmacWarn 可见提示、不再判 stale。此刻起不再需要删掉 .openfeel/dev/guard-state.json 来恢复门禁。feel_* 自动修复(只针对 pid 缺失 / pid 已死的陈旧状态;活异 pid 与 loaded_hash 失配保持不动——前者防多宿主互写,后者是「适配器已更新但宿主未重启」的信号);DSH 侧显式入口 feel_verify action=health fix=true(dry_run 零写盘,结果见 data.health.guardRepair);marker 缺失 / 不完整 → 明确报 marker-missing / marker-incomplete 且零写盘;OPENFEEL_GUARD_SELF_HEAL=off 可关闭自动自愈。OPENFEEL_GUARD_DIR(沿用 OPENFEEL_ADAPTER_HOST_PATH),四套 dsh 测试不再覆写真实宿主的 marker / guard-state。git commit 纯书账放行:单语句、直连 git、无重定向与进程替换、无 -c / -C / --git-dir / --work-tree、无改写仓库的 env、无不可判定形态、且待提交路径全在 .openfeel/** —— 七条件全满足才放行。-c / --config-env 与 GIT_* 环境前缀改为白名单(只放 user.name / user.email / color.* / core.quotepath / core.autocrlf / core.safecrlf / i18n.* 与 GIT_AUTHOR_* / GIT_COMMITTER_* 六项),其余硬拒。setup --tools dsh 在显式 --dsh-profile 且还有其他 profile 时于 stderr 提示「其余保持原样」,SetupResult.dsh.profileScope 记录 deployed / skipped;feel_status / feel_next 不再把 guard-stale 等门禁原因误报成「无活跃阶段」。1.5.10 未单独发布。bash / pwsh 四类判定——受保护路径与 openfeel 写命令硬拒;源码与 git 工作区写按目标判定;构建产物(dist / build / out / target / .next 等)豁免;不可判定只告警。npx vitest / npm run build / git status 不误杀。chore.json、dev/.preflight.json、dev/sessions.json、dev/guard-state.json、pipeline.yaml、config.yaml、metrics.json、flow-archive/**、checkpoints/**、requirements/*.md、log/guard-audit.jsonl、dsh/cordis.patch.yml 只能经 CLI 写入,堵住 agent 手写状态文件自提权。src / test / scripts / dsh / bin)内 .md / .css / assets/* 不再按文档放行——先判目录、再判扩展名。guard-state.json 落 os.tmpdir() 全局路径 + stderr 告警),不再静默放行;.openfeel/dev/guard-state.json 记录 pid / loaded_hash / HMAC,feel_status / feel_next 的 guard 段对不上即 stale。chore 收窄为纯人类通道:窗口上限 30 分钟(--ttl-min 1..30),源码写入须显式 --source。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主(guard 只在宿主重启时重新注册)。详见 CHANGELOG。openfeel dashboard [--json] [--out <file>] [--open]:聚合需求看板(7 状态)/ 阶段进度 / 版本发布门与归档历史 / 阻塞 REV 与流程健康,默认生成单文件自包含 HTML(.openfeel/tmp/dashboard.html),零外链、零网络请求、零可执行脚本,双击 file:// 即可打开。dashboard --json(成功出口 #43):10 键固定序;与 --out / --open 互斥;快照命令(health.ok=false 仍 exit 0)。--json 选项;既有 42 个出口零改动。1.5.10 未单独发布。npm i -g openfeel@latest → openfeel setup。详见 CHANGELOG。exec_running,一个阶段容纳多个需求;不再「每做一个需求就建/推进/关闭阶段」。无开启阶段时才开新阶段。req advance --to done 在零关联阶段时可直接完成,或关联阶段已进入 exec_running 及之后相位(plan_* / scheme_* 仍拒绝);聚合自动收口语义不变(零关联仍不误判)。feel_archive create;AI 可在需求规模足够时建议归档,但不得自行归档或发布。workflowGate 与写入门禁文案改为需求导向;feel_status 默认 compact(detail:"full" 取全量,转移表用 feel_flow action=phases|mode 现取)。plan stage add / flow stage add 目标目录已被其它阶段(完整 stageId)占用或该阶段已归档时拒绝,并给出建议新编号。openfeel lint req [--json] 扫 .openfeel/requirements/*.md 里的行内反引号命令名,剥 openfeel 前缀后沿真实命令树解析。A 类(命令路径无法解析,如 stage deps / flow review_add)→ 报错并 exit 1;B 类(疑似命令组 / 参数写成命令名,如 op create / archive create)→ 只提示、exit 0(advisory 永不阻塞)。带位置参数的正常写法不会误报(openfeel stage set stage-01 / config get workflow_mode / model get openfeel-executor)。--json 恒七键(比其它 lint 出口多 suspects / unknownCount / suspectCount)。<!-- lint-req:ignore -->、区间 <!-- lint-req:ignore-start --> … <!-- lint-req:ignore-end -->;只用于正文按定义要引用错误命令名的场合(bug 例证 / 历史差异记录),不用于掩盖真实错误名。漏写 ignore-end 会在 stderr 告警(不静默、不改退出码),人类输出首行显示「抑制 N 行」。src/commands/shared/preflight-gate.ts);src/core/preflight.ts 判定源零改动。REV-999 / op-999 / 阶段 ID / --to 先报参数错误);stage task --add 与 flow review update 等此前不可达 / 被遮蔽的门位补齐。--dry-run 永不被门拒绝,保住「试探 → flow next → 真执行」路径;其中 flow repair --dry-run --backup 会跳过备份(预览模式不写任何文件,也避免用损坏的 flow.json 覆盖原本可用的 .bak)。flow repair / flow checkpoint restore 在 flow.json 不可读时放行(否则 strict 永久死锁);状态可读时仍受门。flow.json / status.md / config.yaml / 归档与 .bak 全部不落盘(含 repair --backup、config set --sync-stages)。flow metrics 人类输出追加「流程健康」三行——resumed 比例(分母只含带布尔 resumed 的 advance_stage_phase,reset_stage_phase 不计入)、plan_pending → plan_review 平均停留(窗口内逐事件配对,锚点复用唯一实现)、最近 preflight 年龄(只读账本,不刷新账本)。窗口 = 最近 200 条日志;0 样本输出「无数据 / 无样本」而非 0.0%(「没测到」≠「测到为零」),resumed 行恒外显「另有 N 条不可判定」。observability(只增不改):feel_status / feel_next 追加 observability(stale + preflightAgeMs / preflightOk 复制副本 + statusSync 三键 + 一句摘要)。preflight 账本在 snapshot 之前读取,年龄反映「上一次读状态 → 现在」,不会被本次调用刷成 ~0。lint req(41 成功出口 / 42 注册选项),既有 40 个出口键集/键序/类型零改动、错误码表恒 13 行;flow metrics --json 键集恒 schemaVersion,agents(健康块仅人类输出);适配器既有字段全部保留、新字段尾部追加。36 个门位中 4 个所属命令注册了 --json(flow advance / flow archive / req split / flow health --fix;后两者为 legacy 三键,拒绝值 preflight-required);chore / dev sync / config set-lang / 顶层 archive <stage> / 顶层 migrate 不在接入面(排除清单见 commands.md)。preflight 拒绝次数 / feel_next 调用次数 / 主会话直写与自审占比——都需要新增埋点或写盘计数器,违反「零运行时开销」边界,本批次不做。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主。详见 CHANGELOG。flow next / flow status 记录 .openfeel/dev/.preflight.json;flow advance / plan stage add / req add / req claim 写前校验 10 分钟窗口。warn(仅 TTY 提示,不破坏脚本);DSH 适配器 strict(未读状态拒绝,detail.reason=preflight-required)。feel_*;禁止 shell 跑 openfeel flow/stage/req 写命令;fast 单轮审查也应由独立 reviewer,自审标 self-review。INVALID_ARGUMENT);feel_status/feel_next 新增 preflight 字段。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主。详见 CHANGELOG。flow log 人类输出追加当前阶段的最近转换 / 最近失败 / resume 次数(从最近 200 条日志派生,不新增存储)。feel_status / feel_next 返回 revision / snapshotAt / stateUpdatedAt / stale;flow next 缺失降级时明确提示。flow metrics 人类输出追加 flow.log 条目/字节与活跃阶段数。plan_pending 停留 >10 分钟时提示推进或合并/取消,避免为过闸建空阶段。flow next --json / flow metrics --json 键集不变。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主。详见 CHANGELOG。feel_next.dispatch 给出 goal/context/boundaries/returnFormat;子 agent 禁读 dev_last* / current.md / status.md / flow.json,返回 ≤10 行;主会话只保留摘要。feel_status.recovery 聚合 lastTransition、交接摘要(≤600 字)、nextAction;新会话一次调用即可恢复,无需先读全文。dev_last 交接摘要使用 done: / next: / block: 固定键;恢复包直接读取。dev_last,只刷新同步时间,减少无意义写。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主。详见 CHANGELOG。feel_status 与阶段创建;docs/** 原本任意相位可写,研究/方案产物没有阶段要求,写入门禁也无摩擦。feel_status / feel_next 在无活跃阶段时返回 workflowGate(required/reason/message/actions),summary 前置警告。docs/research/** 被拒;错误信息给出建阶段命令;chore 不能绕过;拒绝事件写 guard-audit.jsonl(event=guard_deny)。feel_status;不得以调研/只读为由跳过阶段;纯只读问答例外)。docs/research/dispatch-persistence-2026-10-31.md 给出 1.5.3(统一派发/返回契约 + 恢复包)与 1.5.4(结构化交接 + 日志索引)顺序。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主(适配器为 host module)。详见 CHANGELOG。package-lock.json 中 natural-compare 的 1.4.0 替换为不存在的 1.5.0,CI npm ci 失败。dev_last*、dev/current.md、status.md、flow.json 只由 Feel(主会话)更新;子 agent 一律不读不写,上下文由派发 prompt / 交接文档提供。flow.json.log 由 CLI 自动维护(最近 200 条)。跨会话信息写 ≤5 行交接摘要;openfeel-logging / openfeel-memory 不再要求逐操作 / 会话结束写档。openfeel dev sync 写 dev_last.md 的 CLI 受管简要状态块;flow advance 在首次 / 超过 30 分钟 / 节点相位自动同步;适配器 feel_next / feel_status 返回 statusSync.due,到期时只提示更新交接摘要。dev_last.md 手写内容保留。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主。详见 CHANGELOG。.css/.scss/.html/.svg/.png/.jpg/.jpeg/.webp 与 assets/**、public/** 在任意相位(含全部阶段 done)可写,不要求活跃阶段;混合目标从严(同时含源码与静态资源 → 按源码规则)。.rs/.ts/.js/.py 等仍严格要求 exec_running;.openfeel/flow.json 与 **/status.md 仍禁止手改。.openfeel/log/guard-audit.jsonl(谁、何时、写了哪些路径、相位/原因)。openfeel chore "<说明>":承认「单文件、无接口变化、可肉眼验证」的微变更——无活跃阶段时开启 60 分钟写窗口(源码也可写),--end 收口、--status 查看;不写入 flow.json 的阶段列表,只留日志与 ephemeral chore.json;存在活跃阶段时拒绝(不绕过相位门)。chore 无 --json);注册选项 41 → 43;相位机 / REV / 归档语义不变。npm i -g openfeel@latest → openfeel setup → 完全退出并重启宿主(DSH 的 host module 不热加载)。详见 CHANGELOG。bash(Git Bash),其次 pwsh(PowerShell 7);macOS / Linux 使用 bash;默认不用 cmd.exe;同一任务内保持同一 shell,路径与转义按该 shell 书写。<feel_mode> 中文/英文各新增「命令执行」节);② OpenFeel 全局约束模板(agents-md 与 claude/constraint 的跨 Agent 工具使用约束第 5 条);③ openfeel-tool-usage skill 的工具名映射表注记。npm i -g openfeel@latest → openfeel setup(刷新全局约束)→ 完全退出并重启宿主(DSH 的 host module 不热加载)。详见 CHANGELOG。feel_flow review_add 透传 --stage:DSH 适配器此前只透传 --op,导致 feel_flow action=review_add --stage <id> 始终报「请提供 –op 或 –stage」。v1.3.4 起 op / stage 二选一(皆缺 → FEEL_BAD_USAGE),fast 无 op 场景可直接在工具层挂阶段级 REV。setup --check 三态化:新增 restart-required——区分「适配器文件旧」(mismatch,重跑 setup)与「文件已新但宿主仍加载旧适配器」(restart-required,只需完全退出并重启宿主)。检测方式:setup 把适配器 lib/** 指纹写入 update_state.adapters.dsh,适配器在宿主启动时把实际加载指纹写入 ~/.openfeel/adapter-host-dsh.json,check 比对两者。setup --tools dsh(含「已是最新、跳过」分支)部署成功后刷新 openfeel_version / last_update 并写入 adapters.dsh,消除「setup 成功但 check 永远 mismatch」的死循环。update_state.json 仅追加可选 adapters 字段(旧文件零迁移);setup --check --json 键集/键序不变(status 取值扩展);CLI 相位机 / REV 语义 / 写入门禁不变。npm i -g openfeel@latest → openfeel setup(会刷新部署标记)→ 完全退出并重启宿主;若 setup --check 报 restart-required,重启即可,无需重跑 setup。详见 CHANGELOG。flow review add --op 改可选,新增 --stage <id>;fast 阶段没有 op 时挂阶段级 REV(op = stageId,新增 ReviewItem.stage),blocking 门与 archive <stage> 摘要都能正确归属。review_passed → exec_running 成为合法转移(flow advance --reason "<整改原因>" 落审计);review_pending → exec_running 的同深度多义有显式 tie-break(仍优先经 review_failed,行为不回归)。full 表不变。apply_patch 须全部 *** Add File:)→ 放行并提示补账本;docs/** 与文档类扩展(.md/.mdx/.txt/.rst/.adoc)在无活跃阶段 / 非 exec_running 也可写(混合目标从严);源码仍要求 exec_running。plan stage add 注册后原子设置 pipeline.current.stage;flow next 在 current 为空时从 stages 只读派生;advance 将受管字段(状态 / 责任 Agent / 更新时间)派生写入 status.md;DSH 序列化器只对真环标记 [循环引用](重复引用正常输出)。instructions 增加包内置 schema 回退(空白项目可用)且 --change 可选;execution_mode=auto + auto_advance=disabled 的静默无效组合在 flow advance / flow next 告警;flow advance 成功输出回显生效审查档位。init 目录惰性化(核心 4 + dev 2,其余按需创建);阶段 ID 归一化与 plan/ 目录映射文档化,plan stage add 回显实际目录;归档时非 git 仓库显式告警(不再静默「跳过 commit」)。npm i -g openfeel@latest → openfeel setup;已有项目无需迁移(ReviewItem.stage 为可选字段,旧 REV 照常)。详见 CHANGELOG。ctx.tools.guard 写入门禁:同步拒绝 edit / write / apply_patch / str_replace_editor。.openfeel/** 流程产物任意阶段可写(flow.json / status.md 除外);外部源码要求 exec_running;无活跃阶段拒绝(general 需求例外)。activation=auto 时未初始化工作区注入短提示(只读可进行;写文件按 enforce 说明),不再静默。enforce=always:无 .openfeel 时也拒绝文件写入,防止绕过 openfeel init;project / off opt-out。pwsh / bash 默认不守,仍是显式逃生通道;apply_patch 混合/无头保守拒绝。npm i -g openfeel@latest → openfeel setup --tools dsh → 重启 DSH。详见 CHANGELOG。setup --tools dsh 修复:profile 的 openfeel-dsh link: 目标变化会使 pnpm-lock.yaml 过期;内部 pnpm install 增加 --no-frozen-lockfile,不再被 ERR_PNPM_OUTDATED_LOCKFILE 阻断。package.json、幂等跳过与多 profile fail-fast;失败时 lockfile / node_modules 可能保留部分更新,下一次 setup 收敛。npm i -g openfeel@latest → openfeel setup --tools dsh → 重启 DSH。详见 CHANGELOG。openfeel npm 包现已包含 dsh/ 运行文件;不再需要单独的 openfeel-dsh 包。openfeel setup --tools dsh:把适配器以 link: 依赖写入 DSH profile、追加 dsh.profile.bundles、按需执行 pnpm install,并提示重启 DSH 宿主。支持 --dsh-profile <name> 窄化;默认部署全部已检测 profile。openfeel update --tools dsh:明确拒绝并指引使用 setup(DSH 是全局部署,不属于项目 update)。feel_op create / register、checkpoint list、config effective、scheme list、roadmap show 带位置参数时不再错误)。npm i -g openfeel@latest → openfeel setup --tools dsh(可选 --tools opencode,dsh)→ 重启 DSH。详见 CHANGELOG。openfeel archive <stage> 对已存在的归档文件默认拒绝覆盖(ARCHIVE_EXISTS、零写盘),显式 --force 才覆盖。plan stage list --json:阶段调度视图(七键)。model list|get --json:模型配置的 provider / model / 生效来源。flow retry / lint i18n|kb / roadmap show --json:低频只读面结构化。config get --all --json:一次取全量项目配置,不再要求 key。npm i -g openfeel@latest;JSON 契约与退出码语义保持兼容。详见 CHANGELOG。--review-tier help 文案订正:缺省 = full(零裁剪)。req 过滤单点化:行为不变,减少双实现。flow next --json:负责 Agent、合法下一相位、建议命令、守卫与阻塞审查。--project <path>:任意命令锚定项目。--json:结构化输出全面补全。{schemaVersion,error:{code,kind,message,detail}};退出码 0/1/2/3。--version --json、flow archive --dry-run --json、reviewTier 消歧。full / fast / general:general 仅需求级;模式可逐阶段覆盖与切回。req split 拆分即发布;并行认领;需求与版本解绑。--review-tier:与模式正交,裁剪审查关但恒保留代码审查 / 测试 / 归档关。openfeel-explorer:只读调研路由。model 键不被覆盖。general 仅需求级合法。详见 CHANGELOG。flow archive:按版本把 flow.json 切分到 .openfeel/flow-archive/{version}.json;判据 = 全部阶段 done 且无 blocking REV。openfeel req:需求 7 状态与 10 子命令;全部关联阶段 done 时自动收口;认领无 TTL。flow.json 无需迁移。flow phases [--json] 输出合法相位与转移表。flow stage remove、plan stage add --deps、stageId 校验与冲突检测。config effective 输出来源链。flow status/current/health/metrics/overview --json。更早版本(v1.0.x / v1.1.0 / v1.1.1 等)见 CHANGELOG.md 与 docs/index.md。