feat(agent): enforce durable long-horizon execution

.autopilot/PROGRESS.md

@@ -0,0 +1,32 @@
+# PROGRESS
+
+## 当前
+
+- 下一项：T11 full verify, merge-main, push, and CI
+- 最近绿 tag：autopilot/unify-long-horizon-green-010
+- 备注：Execution Graph 已接入 Plan、TaskRun、后台任务、CLI/TUI/Web/ACP；真实验收 long_task_loop 与 subagents_concurrency 均为 10/10。
+
+## 棒日志
+
+| 棒  | 时间       | 认领项                                     | 门禁  | 证据                                      | 备注              |
+| --- | ---------- | ------------------------------------------ | ----- | ----------------------------------------- | ----------------- |
+| 001 | 2026-08-18 | T1 Execution Graph contracts and stores    | GREEN | `test/execution-graph.spec.mjs`           | 4/4 focused cases |
+| 002 | 2026-08-18 | T2 DAG scheduler and recovery policy       | GREEN | `test/execution-graph-scheduler.spec.mjs` | 5/5 focused cases |
+| 003 | 2026-08-18 | T3 Workspace Lease and guarded patch merge | GREEN | `test/workspace-lease.spec.mjs`           | 5/5 focused cases |
+| 004 | 2026-08-18 | T4 Role routing, synthesis, completion     | GREEN | role/completion/plugin focused tests      | 11 focused cases  |
+| 005 | 2026-08-18 | T5 Instance-owned durable state            | GREEN | durable agent + Plan isolation tests      | 2 focused cases   |
+| 006 | 2026-08-18 | T6 CLI/TUI task control                    | GREEN | task controller/command tests             | 2 focused cases   |
+| 007 | 2026-08-18 | T7 Web graph control and view              | GREEN | control plane + Playwright                | browser passed    |
+| 008 | 2026-08-18 | T8 Plan/TaskRun graph projection           | GREEN | DAG + projection tests                    | 5 focused cases   |
+| 009 | 2026-08-18 | T9 Durable background task projection      | GREEN | async task lifecycle tests                | 2 focused cases   |
+| 010 | 2026-08-18 | T10 Legacy import + real acceptance        | GREEN | runtime evidence JSON                     | 10/10 + 10/10     |
+
+## 仲裁队列
+
+| 编号 | 类型 | 状态 | 摘要 |
+| ---- | ---- | ---- | ---- |
+
+## 打回记录
+
+| 时间 | 打回内容 | 蒸馏成的验收器 |
+| ---- | -------- | -------------- |

.autopilot/acceptance/accept-01-execution-graph.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+set -euo pipefail
+npm run build -w @rdk-moss/agent
+npm run test:filter -w @rdk-moss/agent -- --filter execution-graph
+

.autopilot/acceptance/accept-02-workspace-lease.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+set -euo pipefail
+npm run build -w @rdk-moss/agent
+npm run test:filter -w @rdk-moss/agent -- --filter workspace-lease
+

.autopilot/acceptance/accept-03-completion-arbiter.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+set -euo pipefail
+npm run build -w @rdk-moss/agent
+npm run test:filter -w @rdk-moss/agent -- --filter completion-arbiter
+

.autopilot/acceptance/accept-04-product-surface.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+set -euo pipefail
+npm run build -w @rdk-moss/agent
+(cd packages/moss-agent && node --input-type=module - <<'NODE'
+const orchestration = await import('@rdk-moss/agent/orchestration');
+for (const name of ['InMemoryExecutionStore', 'JsonlExecutionStore', 'ExecutionGraphScheduler', 'CompletionArbiter']) {
+  if (!(name in orchestration)) throw new Error(`missing orchestration export: ${name}`);
+}
+NODE
+)
+npm run test:filter -w @rdk-moss/agent -- --filter task-command --filter plugin-agent-role

.autopilot/config.sh

@@ -0,0 +1,8 @@
+BUILD_CMD="npm run build -w @rdk-moss/agent"
+TEST_CMD="npm run test:filter -w @rdk-moss/agent -- --filter execution-graph --filter workspace-lease"
+FULL_CMD="npm run verify"
+FULL_EVERY=2
+CONTRACT_PATHS="openspec/changes/unify-long-horizon-execution/"
+RUNNER_CMD="codex exec --full-auto"
+PROVIDER_PROBES=()
+REQUIRED_ENV=()

.autopilot/no-touch.txt

@@ -0,0 +1,5 @@
+packages/moss/src/
+packages/create-moss-app/
+:(glob)**/.env
+:(glob)**/*.lock
+

.autopilot/sprint-prompt.md

@@ -0,0 +1,3 @@
+你是 Moss 长程执行图实施任务的一棒。先读 qiaolong-autopilot Phase 3、`.autopilot/task.md`、
+`.autopilot/PROGRESS.md`、`.autopilot/state` 和最近五个提交；恰好认领一项，以测试先红后绿推进，
+保存真实命令证据，门禁绿后提交。禁止削弱 OpenSpec、工具安全、审批与插件原子生命周期。

.autopilot/state

@@ -0,0 +1,5 @@
+STATUS=RUNNING
+SPRINT=3
+CONSEC_RED=0
+SAME_ERR_COUNT=0
+LAST_ERROR_SIG=

.autopilot/task.md

@@ -0,0 +1,46 @@
+# Autopilot 合同：Moss 长程执行图与专家团
+
+## 1. 目标与 Done（机器可判定）
+
+- 问题本质：把分散、不可恢复、不可合并的 agent 工作转换为单一持久执行图与证据闭环。
+- 自治级别：L3（合并和推送是扳机，但用户已在任务中显式授权）。
+- Done = 下列验收器全绿 + `npm run verify` 全绿 + CI 全绿：
+  - [ ] `acceptance/accept-01-execution-graph.sh`
+  - [ ] `acceptance/accept-02-workspace-lease.sh`
+  - [ ] `acceptance/accept-03-completion-arbiter.sh`
+  - [ ] `acceptance/accept-04-product-surface.sh`
+
+## 2. Non-goals
+
+- 不实现分布式高可用、递归专家树、共享目录并行直写或外部工作流引擎。
+- 不削弱现有工具安全、审批、插件原子生命周期与错误边界。
+
+## 3. 改动范围与禁区
+
+- 允许改：`packages/moss-agent/`、相关 docs、OpenSpec、API reports、benchmarks、changelog。
+- 禁改清单：见 `no-touch.txt`。
+- 契约冻结：本 OpenSpec 在实现期间只做纠错，不为让测试变绿而弱化验收。
+
+## 4. Checklist
+
+- [x] T1 Execution Graph contracts and stores
+- [x] T2 DAG scheduler and recovery policy
+- [x] T3 Workspace Lease and guarded patch merge
+- [ ] T4 Expert routing, synthesis, and completion arbiter
+- [ ] T5 MossAgent, plugin, CLI/Web/TUI/ACP integration and migration
+- [ ] T6 public API, docs, benchmarks, full verification, merge, push, CI
+
+## 5. 风险与不确定
+
+- Windows 无法在本机真跑；以路径单测和 GitHub Windows CI 作为平台证据。
+- 真实模型场景依赖 provider 凭据；无凭据时必须明确 SKIP，不能用 fixture 冒充。
+
+## 6. 扳机操作
+
+- 合并到 `main` 与 push；用户已明确授权，仅在全量门全绿后执行。
+- 不执行 npm publish、删除用户数据或生产部署。
+
+## 7. 预算与阈值
+
+- 最大阶段：6；每个阶段 focused tests 后提交可恢复绿点。
+- 同一错误连续三次或连续两阶段全量门红即熔断并报告。

ARCHITECTURE.md

@@ -127,6 +127,7 @@ State has an explicit owner and scope:
 | -------------------- | ----------------------------------------------------- | -------------------------------------------------- | ----------------------------------------- |
 | Conversation         | Session store and event contracts                     | JSONL or in-memory stores, depending on entry path | Inject durable, tenant-scoped storage     |
 | Active run           | Run/session identity, abort and steering state        | Process-local orchestration                        | Project events into a host job system     |
+| Execution graph      | Goal, DAG nodes, budgets, evidence, leases, verdict   | `ExecutionStore` under `.moss/runtime/executions`  | Subscribe/project; never infer completion |
 | Context              | Messages, budgets, pruning and compaction checkpoints | Runtime-managed per run                            | Configure budgets and observe usage       |
 | Skills and knowledge | Registries, loaders, capability packs                 | Bundled/workspace discovery                        | Inject product or domain sources          |
 | Memory               | Scoped memory interfaces and stores                   | CLI/runtime-selected local stores                  | Provide tenant-aware persistence          |

CHANGELOG.md

@@ -10,12 +10,35 @@ Categories: **Added** · **Changed** · **Fixed** · **Removed** · **Internal**
 
 ### Fixed
 
+- **Windows execution-store durability**: atomic lease and snapshot publication now retains file
+  fsync while avoiding unsupported Windows directory-handle fsync, so durable long-horizon runs can
+  acquire and renew graph ownership on Windows.
+- **Clean-checkout long-horizon acceptance**: the acceptance command now builds the agent before
+  importing its distribution entrypoints, including on macOS, Ubuntu, and Windows CI.
+- **Repeated legacy TaskFrame recovery**: migration markers now resolve their persisted graph ID,
+  so a later turn with a new run ID reuses the original execution graph instead of failing the Web
+  conversation before approval or tool execution.
 - **Cross-platform plugin process lifecycle**: active Worker RPC calls now keep their worker alive
   until every pending promise settles, and exact-version plugin installation runs npm's JavaScript
   CLI directly on Windows instead of failing through the `npm.cmd` shim.
 
 ### Added
 
+- **Durable long-horizon execution and expert teams**: the beta
+  `@rdk-moss/agent/orchestration` surface adds a CAS event-sourced execution graph, JSONL snapshots
+  and owner leases, dependency-aware parallel scheduling, explicit budgets/retries, Git/copy
+  workspace leases with guarded patch merge, capability-routed advisor/implementer/verifier roles,
+  deterministic evidence-bound completion arbitration, and safe legacy checkpoint import. CLI/TUI
+  `/tasks` and `/task inspect|resume|retry|stop` plus the Web details panel expose the same graph
+  identity, revision, nodes, evidence, recovery state, and verdict. Recorded runtime acceptance now
+  requires `long_task_loop` 10/10 and `subagents_concurrency` 10/10. Scheduler owner leases now fence
+  every event for the complete execution wave and abort workers on stop or lease loss; capability
+  routing and structured synthesis run through `MossAgent`; routed coding work provisions an
+  isolated lease, requires explicit host merge approval and a distinct zero-exit verifier receipt;
+  direct store callers cannot forge arbiter terminal events. Isolated patches merge only through
+  adapter-verified approval paths. The acceptance gate now exercises the packaged CLI and real
+  loopback Web task API while emitting reproducible tracked evidence.
+
 - **Complete local Web workbench**: `moss web` now provides the responsive three-column Moss design
   system, durable searchable sessions, resume/rename/export/delete/fork/rewind, cursor-based SSE,
   rich conversation and tool evidence, bounded text/image attachments, generated-file downloads,
@@ -60,6 +83,12 @@ Categories: **Added** · **Changed** · **Fixed** · **Removed** · **Internal**
 
 ### Changed
 
+- **Plan, background-task, and TaskRun state converge on Execution Graph**: Plan dependencies now
+  admit independent siblings concurrently while omitted dependencies retain ordered compatibility;
+  Plan tools, background tasks, and TaskRun v1 shadow-write graph facts. A TaskRun `run.completed`
+  claim cannot complete a graph until a verifier binds evidence. Full sub-agents require declared
+  write paths and return retained lease/patch references instead of working in deleted empty folders.
+
 - **Audited Cordis vendoring direction**: the agent package now carries a licensed, revision-pinned
   effect-ownership kernel and an OpenSpec migration plan for service seams, shared skill catalogs,
   full-core evaluation, controlled loading, and eventual quiescent HMR.