@kylecheng3146/agent-ops 0.6.0 → 0.7.0

package/dist/runtime/src/schema/validate.js

@@ -3,7 +3,7 @@ import { acceptanceError, runnerError, commitIdentity } from "./acceptance.js";
 const ID_PATTERN = /^[a-z][a-z0-9-]{0,127}$/;
 const HASH_PATTERN = /^[a-f0-9]{64}$/;
 const WINDOWS_RESERVED_SEGMENT = /^(?:aux|com[1-9]|con|lpt[1-9]|nul|prn)(?:\..*)?$/i;
-const PROFILE_VALUES = new Set(["advisory", "core", "guardrails", "loop"]);
+const PROFILE_VALUES = new Set(["advisory", "core", "guardrails", "loop", "run"]);
 const EVIDENCE_KINDS = new Set(["exit-code", "file", "test-count"]);
 const SCOPE_VALUES = new Set(["project", "user"]);
 const HARNESS_VALUES = new Set(["agy", "claude", "codex", "opencode"]);
@@ -330,10 +330,21 @@ export function validateConfig(value) {
     if (!isRecord(root.features)) {
         return failure("INVALID_TYPE", "$.features", "features must be an object.");
     }
-    const featuresUnknown = unknownFieldFailure(root.features, ["completionGate", "stopVerification"], "$.features");
+    const featuresUnknown = unknownFieldFailure(root.features, ["completionGate", "stopVerification", "office"], "$.features");
     if (featuresUnknown !== undefined) {
         return featuresUnknown;
     }
+    if (root.features.office !== undefined) {
+        if (!isRecord(root.features.office)) {
+            return failure("INVALID_TYPE", "$.features.office", "office must be an object.");
+        }
+        const unknown = unknownFieldFailure(root.features.office, ["enabled"], "$.features.office");
+        if (unknown !== undefined)
+            return unknown;
+        if (typeof root.features.office.enabled !== "boolean") {
+            return failure("INVALID_FEATURE", "$.features.office.enabled", "office.enabled must be a boolean.");
+        }
+    }
     if (!isRecord(root.features.stopVerification)) {
         return failure("INVALID_TYPE", "$.features.stopVerification", "stopVerification must be an object.");
     }

package/docs/en/spec/harness-adapters.md

@@ -41,7 +41,7 @@ capabilities and MUST track generated source as one whole-file artifact.
 - Action: For opencode, manage `.opencode/plugins/agent-ops.js` in a project or `.config/opencode/plugins/agent-ops.js` at user scope (`$XDG_CONFIG_HOME/opencode/plugins/agent-ops.js` when configured inside the managed user root, or `$OPENCODE_CONFIG_DIR/plugins/agent-ops.js` when that native config directory is configured), leave `opencode.json` untouched, and deduplicate a project `AGENTS.md` contribution by path.
 - Evidence: The manifest contains the plugin hash, the generated source contains only the selected hook registrations, and shared project markers occur once.
 - Positive: `codex,opencode` produces one project AGENTS route and one hashed opencode plugin.
-- Negative: `Add an opencode.json instructions entry or register a plugin for a core-only profile.`
+- Negative: `Add an opencode.json instructions entry or register a plugin for a core-only profile with Office disabled.`
 
 ## HARNESS-ADAPTER-004
 
@@ -106,6 +106,7 @@ The current registration matrix is intentionally asymmetric:
 
 | Capability | agy | Codex | Claude Code | OpenCode |
 | --- | --- | --- | --- | --- |
+| office-presence (Preview) | degraded | degraded | supported | degraded |
 | lifecycle-summary | degraded | supported | supported | degraded |
 | command-policy | supported | unknown | supported | supported |
 | completion-gate | supported | unsupported | unsupported | unsupported |
@@ -137,3 +138,114 @@ stores only bounded local event metadata, returns bounded redacted session
 context, and preserves local goal, state, telemetry, and Codex TOML files on
 update or uninstall. A clearly parsed `[features]` / `hooks = false` in an
 existing Codex configuration MUST reject loop planning before any write.
+
+## HARNESS-ADAPTER-007
+
+The `run` profile MUST be opt-in, select `core` and `loop` with it, and add only
+the `auto-run` capability. With it, the Claude Code and Codex managed rules hand
+a change that needs more than five acceptance criteria to `agent-ops run`; five
+or fewer stay in the session. agy rules, and every rule file without `auto-run`,
+MUST stay byte-identical to the rules without the profile.
+
+- Trigger: A project selects `run`, and a Claude Code or Codex session meets a
+  change that needs more than five acceptance criteria.
+- Action: From the main checkout and before any `task create`, the agent writes
+  the user's prompt verbatim and its proposed acceptance criteria, marked as
+  proposals, to the gitignored `.agent-ops/state/run-goal.md`, then starts
+  `agent-ops run --goal-file .agent-ops/state/run-goal.md --host <claude|codex> --wait`
+  as a background shell command. Codex starts it like review: an outer request
+  with `sandbox_permissions: "require_escalated"` and
+  `env -u CODEX_SANDBOX_NETWORK_DISABLED`. An awaiting-input result is relayed
+  to the user and answered with `agent-ops run respond`. A start refused with
+  `RUN_TARGET_REQUIRED`, `RUN_BACKGROUND_UNSUPPORTED`, `RUN_TARGET_DIRTY` or
+  `WORKTREE_NESTED` is reported in one line naming the code and falls back to
+  the subtask flow; `RUN_REPO_UNTRUSTED` stops and asks the user. A run that
+  ends without completing (blocked, budget exhausted, stopped, failing review)
+  is reported with its `run status` state and code, its last `run logs` events,
+  and the exact `agent-ops run resume <id>` and `agent-ops run stop <id>`
+  commands, and is never resumed automatically or taken over by the session.
+  With worktree auto mode, the Claude `PreToolUse` worktree guard allows a
+  write to exactly `.agent-ops/state/run-goal.md` in the main checkout without
+  creating a worktree. Confirmed init/update pre-authorizes `agent-ops run`:
+  `Bash(agent-ops run *)` for Claude Code, and an escalated prefix rule for
+  `env -u CODEX_SANDBOX_NETWORK_DISABLED agent-ops run` for Codex.
+- Evidence: Profile, managed-rule digest, pre-authorization, and worktree-guard
+  tests cover the resolved profiles, the unchanged rules, the added entries, and
+  the single exempt path.
+- Positive: `A Claude Code session with the run profile writes .agent-ops/state/run-goal.md in the main checkout and starts agent-ops run --host claude --wait in the background.`
+- Negative: `Split an eight-criteria change into subtasks while the run profile is active, add auto-run text to GEMINI.md, or resume a blocked run without the user.`
+
+## HARNESS-ADAPTER-008
+
+`agent-ops office` (Preview) MUST show progress read-only from agent-ops state alone. The
+supervisor records an optional phase per run and worker (`planning`,
+`implementing`, `verifying`, `reviewing`, `integrating`) at deterministic
+transitions, and each task's latest verify and review result with passed/total
+criteria. Run state written before phases existed MUST still validate and show
+as `unknown`. Changed files come from each worktree's own `git diff` against its
+base; narration uses paths only (`docs/**` is "writing docs", `tests/**` is
+"writing tests", otherwise "editing <file>"). The office MUST NOT read native
+Claude or Codex transcripts or file contents.
+
+- Office Preview is opt-in via `features.office.enabled`. Init MUST offer an
+  off-by-default choice; interactive update MUST default to the existing choice.
+  Non-interactive update MUST preserve it unless `--office on|off` is supplied.
+  Older configs without the optional field remain valid and mean disabled.
+  Enabling Office MUST install its presence hooks even with only `core`; this
+  MUST NOT opt into command policy, Stop verification or lifecycle summaries.
+  Office-only OpenCode runtime failure MUST remain fail-open; adding Office to
+  command policy MUST preserve policy denial and unavailable-runtime behavior.
+  Agy startup is per invocation, OpenCode startup is app-scoped, and Codex
+  Stop delivery is unavailable in some execution modes; stale presence expires
+  using the bounded hook-activity timeout rather than assuming native Stop.
+  Features retain their existing whole-object layer precedence. Disabled hooks
+  MUST NOT record Office presence or start/open Office; the explicit command
+  MUST explain how to enable it. A running server MUST close after observing a
+  disabled choice on its next 15-second lifecycle check.
+- Trigger: When enabled, a managed ordinary session starts or reports activity, or a user runs
+  `agent-ops office`, `agent-ops run start` or `agent-ops run status`.
+- Action: The command reuses the live server recorded in
+  `<git common dir>/agent-ops/office.json`, or starts one, and prints its URL.
+  Ordinary session hooks also start or reuse it, including before a task or
+  worktree exists. Only the successful new-server owner opens the browser;
+  other sessions share the same Office. Session metadata contains bounded
+  identifiers, known work state and activity times, never native transcripts,
+  prompts or tool/file contents. Startup and observation failures are advisory.
+  The server binds
+  127.0.0.1 on a random port, requires the URL's unguessable token, refuses any
+  Host other than that address and port, answers every non-GET with 405, and
+  exits 10 minutes after neither a run nor an ordinary session is active.
+  Ordinary-session liveness follows hooks, with stale presence expiring after
+  30 minutes without activity; it does not inspect native host processes.
+  Its single inline page uses cozy cream, oak and sage in-code 16-color scene
+  sprites with no image assets. Each 32×48 character uses its own fixed skin,
+  hair and clothing palette, four-direction walking, and localized role labels.
+  The viewer's badge displays `You` in English and `你` in Chinese; its overview size matches the largest room's
+  characters. The fixed-viewport overview
+  fills the available page below the header and shows all furnished work-unit
+  rooms, shrinking rooms as necessary without document scrolling. A shared
+  hallway connects the rooms. The user controls a local Supervisor with arrow
+  keys or WASD, can walk into every room and return to the overview, without
+  model requests or network mutations. Each run team shares a room; each ordinary session has its own,
+  and running reviews remain visible. Clicking a room enters a larger detail
+  view with a return control. Every room contains planning, development,
+  verification, review and integration areas. Characters walk to their known
+  phase's zone, use light phase-driven animation, and have short name/status
+  labels; unknown state is not inferred from dialogue. Whiteboards show task
+  criteria progress, verify/review outcomes and pending questions, with `!` for
+  unanswered questions. Clicking a whiteboard or progress board opens the room
+  work list; workstations and phase areas show the known work in that phase.
+  Pending markers open complete questions with the existing copyable commands.
+  The same paged views are available with `L` (work list), `1–5` (phases) and
+  `Q` (questions). Clicking an actor opens bounded, paged status details
+  with copyable commands. Completed rooms are available through Recently
+  completed for at most two hours. Chinese/English switching remembers the
+  choice, and keyboard navigation and reduced motion are supported. A failure
+  to start the office never fails a hook, `run start` or `run status`.
+- Evidence: Phase, snapshot, server, scene and CLI tests cover legacy state,
+  aggregation and narration, ordinary-session registration and startup-once,
+  the two-hour completion boundary, token/Host/method guards, reuse and the
+  injected-clock idle exit, viewport/room navigation, localization, and the
+  printed URL. A browser read-back checks the rendered single-page layout.
+- Positive: `An ordinary session opens the shared Office, where its room is visible before it creates a worktree; entering a team room shows the coordinator at the verification bench and a whiteboard with 2 of 6 criteria passed.`
+- Negative: `Serve the office on 0.0.0.0, accept a POST, read a worker's transcript to narrate it, or bundle a PNG sprite sheet.`

package/docs/zh-TW/spec/harness-adapters.md

@@ -1,6 +1,6 @@
 # Harness Adapter
 
-English source version: 2026-08-03. Revalidate: when the English specification or any vendor reference changes.
+English source version: 2026-10-07. Revalidate: when the English specification or any vendor reference changes.
 
 本文件所述 OpenCode plugin 行為已於 2026-07-31 依據[官方 plugin 文件](https://opencode.ai/docs/plugins/)與[Bun shell 文件](https://bun.sh/docs/runtime/shell)檢查；Codex 與 Claude Code loop-hook 行為已於 2026-08-03 依據 [Codex hook 文件](https://developers.openai.com/codex/config-advanced#hooks) 與 [Claude Code hook 文件](https://code.claude.com/docs/en/hooks) 檢查；agy hook 行為已於 2026-08-28 依據[官方 Antigravity hook 文件](https://antigravity.google/docs/hooks) 檢查。任何 vendor 參考變更時都必須重新驗證。
 
@@ -90,6 +90,7 @@ script，也不得改變一般 permission request。
 
 | Capability | agy | Codex | Claude Code | OpenCode |
 | --- | --- | --- | --- | --- |
+| office-presence (Preview) | degraded | degraded | supported | degraded |
 | lifecycle-summary | degraded | supported | supported | degraded |
 | command-policy | supported | unknown | supported | supported |
 | completion-gate | supported | unsupported | unsupported | unsupported |
@@ -118,3 +119,94 @@ Stop verification 必須明確啟用、具備 trust、為 report-only 且預設
 event metadata、回傳有界且 redacted 的 session context，並在 update 或 uninstall
 時保留 local goal、state、telemetry 與 Codex TOML file。既有 Codex configuration
 中清楚解析出的 `[features]` / `hooks = false` MUST 在任何 write 前拒絕 loop planning。
+
+## HARNESS-ADAPTER-007
+
+`run` profile MUST 是 opt-in，會一併選取 `core` 與 `loop`，且只新增 `auto-run`
+capability。啟用後，Claude Code 與 Codex 的 managed rules 會將需要超過五項
+acceptance criteria 的變更交給 `agent-ops run`；五項以內仍留在 session 中處理。
+agy rules，以及所有未含 `auto-run` 的 rule file，MUST 與未啟用此 profile 時
+byte-identical。
+
+- Trigger: Project 選擇 `run`，且 Claude Code 或 Codex session 遇到需要超過五項
+  acceptance criteria 的變更。
+- Action: Agent 在 main checkout、且在任何 `task create` 之前，將使用者 prompt
+  原文與其提議的 acceptance criteria（標示為提議）寫入已 gitignore 的
+  `.agent-ops/state/run-goal.md`，再以 background shell command 啟動
+  `agent-ops run --goal-file .agent-ops/state/run-goal.md --host <claude|codex> --wait`。
+  Codex 的啟動方式與 review 相同：outer request 使用
+  `sandbox_permissions: "require_escalated"` 與
+  `env -u CODEX_SANDBOX_NETWORK_DISABLED`。Awaiting-input 結果會轉達給使用者，並以
+  `agent-ops run respond` 回覆。啟動遭 `RUN_TARGET_REQUIRED`、
+  `RUN_BACKGROUND_UNSUPPORTED`、`RUN_TARGET_DIRTY` 或 `WORKTREE_NESTED` 拒絕時，以
+  一行說明該 code，並退回 subtask flow；`RUN_REPO_UNTRUSTED` 則停止並詢問使用者。
+  未完成即結束的 run（blocked、budget exhausted、stopped、review 失敗）須回報其
+  `run status` state 與 code、最後的 `run logs` event，以及確切的
+  `agent-ops run resume <id>` 與 `agent-ops run stop <id>` command，且永不自動
+  resume 或由 session 接手。在 worktree auto mode 下，Claude `PreToolUse`
+  worktree guard 允許在 main checkout 寫入恰為 `.agent-ops/state/run-goal.md`
+  的路徑，且不建立 worktree。確認後的 init/update 會預先授權 `agent-ops run`：
+  Claude Code 為 `Bash(agent-ops run *)`，Codex 為
+  `env -u CODEX_SANDBOX_NETWORK_DISABLED agent-ops run` 的 escalated prefix rule。
+- Evidence: Profile、managed-rule digest、pre-authorization 與 worktree-guard
+  test 涵蓋解析後的 profile、未變更的 rules、新增的 entry，以及唯一豁免的路徑。
+- Positive: `啟用 run profile 的 Claude Code session 在 main checkout 寫入 .agent-ops/state/run-goal.md，並在 background 啟動 agent-ops run --host claude --wait。`
+- Negative: `run profile 啟用時仍將八項 criteria 的變更拆成 subtask、在 GEMINI.md 加入 auto-run 文字，或未經使用者即 resume blocked run。`
+
+## HARNESS-ADAPTER-008
+
+`agent-ops office`（Preview）MUST 只依據 agent-ops state 以 read-only 方式呈現進度。
+Supervisor 在確定性的轉換點記錄每個 run 與 worker 的 optional phase
+（`planning`、`implementing`、`verifying`、`reviewing`、`integrating`），以及每個
+task 最近一次 verify 與 review 結果與通過/總數 criteria。Phase 出現前寫入的 run
+state MUST 仍能通過驗證，並顯示為 `unknown`。變更檔案來自各 worktree 自身相對
+base 的 `git diff`；旁白只使用路徑（`docs/**` 為「writing docs」、`tests/**` 為
+「writing tests」，其他為「editing <file>」）。Office MUST NOT 讀取 Claude 或 Codex
+的 native transcript 或檔案內容。
+
+- Office Preview 透過 `features.office.enabled` 明確啟用。Init MUST 提供預設關閉
+  的選項；互動式 update MUST 以既有選擇為預設。非互動式 update MUST 保留選擇，
+  除非傳入 `--office on|off`。即使只選 core，啟用 Office 也 MUST 安裝 presence
+  hook，但 MUST NOT 同時啟用 command policy、Stop verification 或 lifecycle
+  summary。Office-only OpenCode 的 runtime 失敗 MUST fail-open；與 command
+  policy 同時啟用時 MUST 保留原有拒絕與 runtime unavailable 行為。Agy 啟動事件
+  以 invocation 為單位，OpenCode 以 app 為單位，部分 Codex 模式沒有 Stop；過期
+  presence 依有界限的 hook 活動逾時處理。舊設定可省略此欄位，省略代表關閉；features 維持
+  原有的整個物件 layer 優先順序。關閉時 hook MUST NOT 記錄 Office presence、
+  啟動 server 或開啟瀏覽器，明確的 office 指令 MUST 說明啟用方法。運行中的
+  server MUST 在下一次 15 秒生命週期檢查觀察到關閉選擇後結束。
+- Trigger: 啟用後，Managed 一般 session 啟動或回報活動，或使用者執行
+  `agent-ops office`、`agent-ops run start` 或 `agent-ops run status`。
+- Action: Command 重用記錄於 `<git common dir>/agent-ops/office.json` 的 live
+  server，否則啟動一個，並印出 URL。一般 session hook 也會自動啟動或重用
+  Office，包含尚未建立 task 或 worktree 的 session。只有成功取得新 server
+  所有權的 process 開啟瀏覽器，後續 session 共用同一個 Office。Session metadata
+  只記錄有界限的識別資訊、已知工作狀態與活動時間，不包含 native transcript、
+  prompt、tool 或檔案內容。啟動與活動觀察失敗均為 advisory。
+  Server 綁定 127.0.0.1 的隨機 port，要求 URL 中不可猜測的 token，拒絕該位址與
+  port 以外的 Host，所有非 GET 回應 405，並在 run 與一般 session 都沒有活動
+  10 分鐘後結束。唯一的 inline page 使用奶油白、淺木與鼠尾草綠的程式內 16 色
+  場景 sprite，沒有圖片 asset。32×48 人物各自使用固定的膚色、髮色與服裝色盤，
+  支援四方向走路，並顯示本地化的職責文字。使用者標籤在中文顯示「你」、英文顯示「You」，總覽大小與
+  最大房間的成員一致。總覽填滿頂部工具列下方的單一
+  viewport，同時呈現全部有家具的工作房間，必要時自動縮小且不捲動頁面。
+  房間以共用走廊連接；使用者可以用方向鍵或 WASD 操控本機 Supervisor 角色，
+  走進每間房並返回總覽，不會呼叫模型或透過網路修改狀態。每個 run 團隊共用一間房，
+  每個一般 session 各有一間，執行中的 review 也保持可見。點房間進入放大的
+  詳細畫面，並可返回總覽。每間房包含規劃、開發、驗證、審查與整合區。角色依
+  已知 phase 移動與播放輕量動作，並有簡短名牌與狀態；不從對話推測未知狀態。
+  白板顯示 criteria 進度、verify/review 結果與待回覆問題，`!` 標示未回答的問題。
+  點白板或進度板開啟房間工作清單，點工位或階段區查看該階段的已知工作。
+  點待回覆標記查看完整問題與現有可複製指令；同一套分頁視窗也可用
+  `L`（工作清單）、`1–5`（階段）與 `Q`（問題）開啟。
+  點角色開啟有界限且分頁的 status 詳情，包含可複製的 command。完成房間移到
+  最近完成，最多保留兩小時。支援中英文切換並記住選擇、鍵盤操作與 reduced
+  motion。一般 session 的存活狀態依 hook 活動判定，30 分鐘沒有事件就會移除
+  過期的 presence，不會讀取 native host process。Office 啟動失敗絕不使
+  hook、`run start` 或 `run status` 失敗。
+- Evidence: Phase、snapshot、server、scene 與 CLI test 涵蓋 legacy state、彙整與
+  旁白、一般 session 登記與只開啟一次、完成後兩小時的邊界、token/Host/method
+  防護、重用與 injected clock 的 idle 結束、viewport 與房間切換、中英文，以及
+  印出的 URL。Browser read-back 檢查實際渲染的單頁布局。
+- Positive: `一般 session 自動開啟共用 Office，尚未建立 worktree 就能看到自己的房間；進入團隊房時看到 coordinator 在驗證桌旁，白板顯示 6 項 criteria 已通過 2 項。`
+- Negative: `在 0.0.0.0 提供 office、接受 POST、讀取 worker transcript 來產生旁白，或打包 PNG sprite sheet。`

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kylecheng3146/agent-ops",
-  "version": "0.6.0",
+  "version": "0.7.0",
   "description": "Evidence-driven development loops for agy, Codex, Claude Code, and opencode",
   "type": "module",
   "license": "SEE LICENSE IN LICENSE",

package/schemas/config.schema.json

@@ -24,7 +24,8 @@
           "core",
           "advisory",
           "guardrails",
-          "loop"
+          "loop",
+          "run"
         ]
       }
     },
@@ -243,6 +244,14 @@
         "stopVerification"
       ],
       "properties": {
+        "office": {
+          "type": "object",
+          "additionalProperties": false,
+          "required": ["enabled"],
+          "properties": {
+            "enabled": { "type": "boolean" }
+          }
+        },
         "completionGate": {
           "type": "object",
           "additionalProperties": false,