DandreYang · Dandre126 · Aug 15, 2026 · Aug 4, 2026 · Aug 4, 2026 · Aug 4, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -2,6 +2,18 @@
 
 ## Unreleased
 
+## 0.6.9 - 2026-08-15
+
+- Make multi-harness writes a Core peer wave: each task worktree has one
+  executor, overlapping slices share a `conflict_group`, and wave members are
+  not live supervisors. Objective budgets now default to `max_parallel = 3`;
+  when write-capable Providers are ready the effective wave size is
+  `min(requested, ready_count)`. Empty `conflict_group` values warn in a
+  parallel wave. `task.executor` can run a ready dispatch adapter inside the
+  existing task worktree (Cursor edit stays fail-closed); scratch
+  `dispatch run --mode edit` remains detached and patch-only. `objective tick`
+  and `task daemon` preview idle-harness bindings and honor per-backend caps.
+
 ## 0.6.8 - 2026-08-15
 
 - Fix Console inspection falsely timing out healthy workspaces by budgeting for

diff --git a/README.md b/README.md
@@ -298,8 +298,10 @@ preview with `dyro integration install dispatch --dry-run`, then install with
 For two to four different roles, `dyro dispatch batch-plan` produces a
 side-effect-free, context-bound plan. Review its digest before
 `batch-start --expect-plan-sha256 …`, then recover through `batch-status`,
-`batch-result`, or `batch-cancel`. Batch V1 is independent fan-out with at most
-one edit writer; it is not a dependency DAG, retry queue, or automatic judge.
+`batch-result`, or `batch-cancel`. Batch V1 is advisory independent fan-out with
+at most one scratch edit writer; it is not the delivery write plane, a
+dependency DAG, retry queue, or automatic judge. Simultaneous writes use Core
+peer wave: one executor per task worktree, serialized only by `conflict_group`.
 Use explicit synchronous `panel --members all` only for a full-ready-Provider
 same-task comparison.
 

diff --git a/README.zh-CN.md b/README.zh-CN.md
@@ -293,8 +293,10 @@ Dyro Skill 的机器不会被后台静默写入。
 需要 2–4 个不同角色时，先用 `dyro dispatch batch-plan` 生成不创建状态、绑定
 上下文的计划；审阅摘要后再执行 `batch-start --expect-plan-sha256 …`，随后可用
 `batch-status`、`batch-result` 或 `batch-cancel` 恢复生命周期。Batch V1 是最多
-一个编辑者的独立 fan-out，并非依赖 DAG、重试队列或自动终裁器。只有用户明确要求
-全 Harness 同题比较时，才使用同步的 `panel --members all`。
+一个 scratch 编辑者的建议面 fan-out，不是交付写路径，也不是依赖 DAG、重试队列
+或自动终裁器。要同时改多块，走 Core Peer Wave：每条 task worktree 一个执行位，
+只按 `conflict_group` 串行。只有用户明确要求全 Harness 同题比较时，才使用同步
+的 `panel --members all`。
 
 交互运行 `dyro`、`dyro home` 或 `dyro start` 时，Dyro 每个本地自然日最多访问一次官方 PyPI；断网、超时或状态目录不可写都不会阻塞进入工作区。默认仍由用户确认更新：
 

diff --git a/docs/agent-orchestration-discipline.md b/docs/agent-orchestration-discipline.md
@@ -78,7 +78,7 @@ docs/reviews/YYYY-MM-DD-<topic>-adversarial-board.md
 
 | 规则 | 说明 |
 | --- | --- |
-| 一仓一 branch 同时仅一个写 agent | 并行写同一 checkout 会导致已验证修复被静默覆盖 |
+| 一棵 task worktree / 同一 `conflict_group` 同时仅一个写 agent | 并行写同一 checkout 会覆盖已验证修复；不同 task 树或不同冲突组可以同时写 |
 | 改前 `git log -5 -- <path>` | 避免回退到已废弃的 workaround |
 | 精确 stage | 禁止盲目 `git add .`；同文件无关 hunk 必须拆分 |
 | 「验证后又坏了」 | 先查是否有后续 commit 覆盖，再查运行时 |
@@ -112,5 +112,5 @@ docs/reviews/YYYY-MM-DD-<topic>-adversarial-board.md
 5. 验证者是否被要求「打开真产物并尝试证伪」？  
 6. 若 resume：substrate 是否变化？  
 7. 是否声明缩放妥协？  
-8. 写路径是否保证单写 agent？  
-9. 结果是否只回收摘要/契约字段，而非完整事件流？  
+8. 写路径是否保证每棵 task worktree / 每个 conflict_group 只有一个写 agent，且波次成员都是执行位？
+9. 结果是否只回收摘要/契约字段，而非完整事件流？
diff --git a/docs/architecture.md b/docs/architecture.md
@@ -234,4 +234,4 @@ Dyro 的交付拓扑与之**实质相近**：TaskGraph（`depends_on` / conflict
 
 未来的 adapter、通知、签名规则、发布平台与审批系统应使用 Python entry point 或独立 Profile 扩展包接入；不要把某个组织的策略加入 core 默认行为。
 
-开发者侧的可选本地多 Agent 派发（五段式任务契约、注入前机密守卫、locator 核验、隔离 patch）与上述控制面分层并列，随 `dyro` 安装包分发（`dyro dispatch` / `import experiments.local_agent_dispatch`），但**不**替代 gates/合并；见 [`ADR-0002`](adr/0002-optional-local-agent-dispatch.md)、[`多智能体编排纪律`](agent-orchestration-discipline.md) 与 [`可选本地 Agent 派发设计`](designs/optional-local-agent-dispatch.md)。
+开发者侧的可选本地多 Agent 派发（五段式任务契约、注入前机密守卫、locator 核验、隔离 patch）与上述控制面分层并列，随 `dyro` 安装包分发（`dyro dispatch` / `import experiments.local_agent_dispatch`），但**不**替代 gates/合并。同时写多块走 Core Peer Wave（task worktree + `conflict_group`），见 [`peer-wave-execution.md`](designs/peer-wave-execution.md)、[`ADR-0002`](adr/0002-optional-local-agent-dispatch.md)、[`多智能体编排纪律`](agent-orchestration-discipline.md) 与 [`可选本地 Agent 派发设计`](designs/optional-local-agent-dispatch.md)。
diff --git a/docs/designs/native-continuation-engine.md b/docs/designs/native-continuation-engine.md
@@ -61,7 +61,7 @@ max_actions = 20
 max_attempts_per_task = 2
 max_failures = 3
 max_no_progress_cycles = 2
-max_parallel = 1
+max_parallel = 3
 deadline = "2026-10-02T12:00:00Z"
 ```
 

diff --git a/docs/designs/optional-local-agent-dispatch.md b/docs/designs/optional-local-agent-dispatch.md
@@ -288,7 +288,8 @@ judge 或多 edit writer。显式 `panel --members all` 仍是同步的全 ready
 
 | 场景 | 用哪条路径 |
 | --- | --- |
-| 开发者要第二意见 / 大调研 / patch 竞赛 | ADR-0002 本设计 |
+| 同时改多块交付代码 | Core Peer Wave：多 task + worktree + `conflict_group`；见 [peer-wave-execution.md](peer-wave-execution.md) |
+| 开发者要第二意见 / 大调研 / scratch patch | ADR-0002 本设计 |
 | 高风险设计评审 | 本设计 panel + 对抗评审板 |
 | 生产 evidence / merge | 仅 Dyro 控制面；派发 harness 不可越权 |
 

diff --git a/docs/designs/peer-wave-execution.md b/docs/designs/peer-wave-execution.md
@@ -0,0 +1,59 @@
+# Peer Wave：多 Harness 并行执行面
+
+状态：已接受（随 0.6.x 落地）  
+范围：Dyro Core 写路径；不替代 gates / review / merge  
+关联：[多智能体编排纪律](../agent-orchestration-discipline.md) · [ADR-0002](../adr/0002-optional-local-agent-dispatch.md) · [可选本地 Agent 派发](optional-local-agent-dispatch.md)
+
+## 1. 产品判断
+
+多 Harness 的默认画面是 **一波 peer 任务同时写**，不是一路改、其余盯梢。
+
+写隔离已经在 Core：
+
+- 一条 Task 同一时刻只有一个 `executor`
+- 每个 Task 有自己的 `task/<id>` worktree
+- 重叠切片用 `conflict_group` 串行；不重叠的进入同一波
+
+0.6.8 的九路 headless 适配器接到这条写路径上：在 **既有 task worktree** 里跑，不再为交付执行另开 detached patch 树。
+
+## 2. 硬规则
+
+1. 一棵 task worktree 同一时刻只有一个写 harness。不是全工作区只有一个写。
+2. 同一 `conflict_group` 同一波只进一个 Task。空 `conflict_group` 在 `max_parallel > 1` 时警告，但不互斥。
+3. 波次里每个成员都是执行位。只读意见走 `dyro dispatch panel`；复核是该 Task 的下一阶段，审冻结 HEAD，不看 live 树。
+4. `cursor-agent` 不能进入写波次；其 edit 仍 fail-closed。
+5. Dispatch 仍然不能 merge / push / signoff。交付证据只走 Core gates。
+
+## 3. 产品路由
+
+| 意图 | 路径 |
+| --- | --- |
+| 同时改多块 | 拆成 N 条 Task，填 `conflict_group` 与 `executor`，`task daemon --parallel` 或 Objective `max_parallel` |
+| 只要第二意见 | `dyro dispatch panel` |
+| 还不成 Task 的试改 | `dyro dispatch run`（只读或 detached patch） |
+| 禁止默认 | 一个 edit writer + 其余角色陪跑 |
+
+Batch V1 保持建议面：2–4 个独立角色、最多一个 edit。不要在 Batch 里重做 TaskGraph。
+
+## 4. 执行桥
+
+`task.executor` 若是已就绪且支持 edit 的 dispatch Provider，`run_task` 调用该适配器：
+
+- cwd = `_ensure_task_worktrees()` 的根
+- 分支 = `task/<id>`
+- 使用适配器自己的隔离 Home 与进程监督
+- 不创建 `EditWorkspace`，不回写源工作区之外的 detached tree
+- 终态仍走 receipt → gates → review → merge
+
+未就绪或非 dispatch 的 executor 回退到 Profile `write` argv，避免没有本机登录时阻断既有工作区。`auto` 只从就绪的可写 Provider 里确定性分配。
+
+## 5. 异构波次
+
+`objective tick` 与 `task daemon` 预览为每个 ready Task 绑定空闲就绪 harness：
+
+- 钉死的 `executor` 优先；达到每后端上限则推迟，不静默换人
+- `auto` 按 Provider id 排序领取空闲可写 harness
+- 遵守 Objective / daemon 并行容量、`conflict_group`、每后端上限
+- Cursor 作为写 executor 被拒绝
+
+Objective 默认 `max_parallel = 3`。本机有就绪可写 Provider 时，有效容量为 `min(requested, ready_write_count)`；没有就绪 Provider 时保持 requested，以便 Profile-only 工作区继续跑。
diff --git a/docs/diagrams.en.md b/docs/diagrams.en.md
@@ -269,7 +269,7 @@ flowchart TB
   Host -->|"explicit dyro commands"| Dyro
 ```
 
-See `docs/agent-orchestration-discipline.md`. The optional harness under `experiments/local_agent_dispatch/` ships with the installed package but never replaces Core gates or merge.
+See `docs/agent-orchestration-discipline.md` and `docs/designs/peer-wave-execution.md`. `dyro dispatch` is advisory; simultaneous writes use Core task worktrees, not Batch supervisors. The optional harness under `experiments/local_agent_dispatch/` ships with the installed package but never replaces Core gates or merge.
 
 ---
 

diff --git a/docs/diagrams.md b/docs/diagrams.md
@@ -269,7 +269,7 @@ flowchart TB
   Host -->|"显式 dyro 命令"| Dyro
 ```
 
-见 [`agent-orchestration-discipline.md`](agent-orchestration-discipline.md)。能力：`dyro dispatch` / `experiments.local_agent_dispatch`（随 `dyro` 安装；**不**替代 gates/merge）。
+见 [`agent-orchestration-discipline.md`](agent-orchestration-discipline.md) 与 [`peer-wave-execution.md`](designs/peer-wave-execution.md)。`dyro dispatch` 是建议面；同时写多块走 Core task worktree 波次，不在 Batch 里陪跑监工。能力：`dyro dispatch` / `experiments.local_agent_dispatch`（随 `dyro` 安装；**不**替代 gates/merge）。
 
 ---
 

diff --git a/experiments/local_agent_dispatch/README.md b/experiments/local_agent_dispatch/README.md
@@ -53,9 +53,10 @@ python3 -m experiments.local_agent_dispatch doctor
 
 State home: `~/.dyro/local-agent-dispatch/` (override with `--home` or `DYRO_LOCAL_AGENT_DISPATCH_HOME`).
 
-Batch V1 accepts two to four independent member contracts, resolves installed
-Provider candidates without starting authentication CLIs, and allows at most one
-edit writer. `batch-start` actively authenticates every selected Provider before
+Batch V1 is the advisory plane: two to four independent member contracts,
+installed Provider candidates without starting authentication CLIs, and at most
+one scratch edit writer. Simultaneous delivery writes use Core peer wave
+(`docs/designs/peer-wave-execution.md`), not a second Batch writer. `batch-start` actively authenticates every selected Provider before
 creating state. Its reviewed
 plan digest binds the canonical project root, normalized contracts, Provider
 choices and non-secret execution profiles, guarded context digests, timeouts, and

diff --git a/pyproject.toml b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "dyro"
-version = "0.6.8"
+version = "0.6.9"
 description = "DyroEngineeringFlow: local-first automation and delivery control for multi-repository teams"
 readme = "README.md"
 requires-python = ">=3.11"

diff --git a/src/dyro/cli.py b/src/dyro/cli.py
@@ -36,8 +36,8 @@
 )
 from .continuation.engine import (
     build_scheduler_tick,
-    render_scheduler_tick_json,
     render_scheduler_tick_text,
+    scheduler_tick_payload,
 )
 from .continuation.models import Operation, RequestedMode
 from .continuation.planner import (
@@ -3192,13 +3192,34 @@ def cmd_objective_tick(args: argparse.Namespace) -> None:
         )
     )
     plan = build_continuation_plan(snapshot)
+    from .peer_wave import (
+        annotate_objective_tick,
+        discover_available_write_providers,
+        recommended_max_parallel,
+    )
+
+    available_write = discover_available_write_providers()
     tick = build_scheduler_tick(
-        snapshot, plan, max_parallel=record.objective.budget.max_parallel
+        snapshot,
+        plan,
+        max_parallel=recommended_max_parallel(
+            record.objective.budget.max_parallel, len(available_write)
+        ),
     )
+    overlay = annotate_objective_tick(snapshot, plan, tick, available_write)
     if args.format == "json":
-        print(render_scheduler_tick_json(tick))
+        payload = scheduler_tick_payload(tick)
+        payload.update(overlay)
+        print(json.dumps(payload, ensure_ascii=False, sort_keys=True, indent=2))
         return
     print(render_scheduler_tick_text(tick))
+    for note in overlay.get("peer_wave", {}).get("warnings", []):
+        print(f"Warning: {note}")
+    for binding in overlay.get("peer_wave", {}).get("executor_bindings", []):
+        print(
+            f"Harness: {binding['task_id']} -> {binding['executor']} "
+            f"({binding['source']})"
+        )
 
 
 def cmd_objective_attention(args: argparse.Namespace) -> None:
@@ -3458,7 +3479,20 @@ def cmd_task_daemon(args: argparse.Namespace) -> None:
         tasks = list_tasks(config)
         assert_legacy_scheduler_allowed(config, (task.id for task in tasks))
         queued = _daemon_select_runnable(config, tasks, limit=max(1, args.parallel))
-        if queued:
+        from .peer_wave import apply_harness_bindings, discover_available_write_providers
+        from .tasks import ScheduleWave
+
+        available_write = discover_available_write_providers()
+        bound, decision = apply_harness_bindings(
+            ScheduleWave(tasks=tuple(queued), deferred=()),
+            available_write,
+        )
+        for note in decision.warnings:
+            print(f"warning: {note}")
+        for item in decision.deferred:
+            print(f"defer {item.task.id}: {item.reason}")
+        if bound:
+            overrides = {item.task_id: item.executor for item in decision.bindings}
             with ThreadPoolExecutor(
                 max_workers=max(1, args.parallel), thread_name_prefix="dyro-dispatch"
             ) as pool:
@@ -3469,8 +3503,9 @@ def cmd_task_daemon(args: argparse.Namespace) -> None:
                         task,
                         dry_run=args.dry_run,
                         legacy_scheduler=True,
+                        executor_override=overrides.get(task.id),
                     ): task
-                    for task in queued
+                    for task in bound
                 }
                 for future in as_completed(futures):
                     task = futures[future]

diff --git a/src/dyro/continuation/contracts.py b/src/dyro/continuation/contracts.py
@@ -19,7 +19,7 @@
     max_attempts_per_task=2,
     max_failures=3,
     max_no_progress_cycles=2,
-    max_parallel=1,
+    max_parallel=3,
     deadline=None,
 )
 MAX_TITLE_LENGTH = 240

diff --git a/src/dyro/continuation/supervision.py b/src/dyro/continuation/supervision.py
@@ -178,7 +178,19 @@ def build_supervised_wave(
         raise DyroError("当前 Objective 不是 supervised 模式；拒绝使用受监督 apply")
     snapshot = build_scheduler_snapshot(config, objective=record, clock=clock)
     plan = build_continuation_plan(snapshot)
-    tick = build_scheduler_tick(snapshot, plan, max_parallel=record.objective.budget.max_parallel)
+    from ..peer_wave import (
+        discover_available_write_providers,
+        recommended_max_parallel,
+    )
+
+    tick = build_scheduler_tick(
+        snapshot,
+        plan,
+        max_parallel=recommended_max_parallel(
+            record.objective.budget.max_parallel,
+            len(discover_available_write_providers()),
+        ),
+    )
     actions = tuple(action for action in tick.wave if action.kind in _SUPPORTED_ACTIONS)
     if len(actions) != len(tick.wave):
         raise DyroError("当前 wave 含尚未获受监督执行支持的 Action；拒绝部分执行")
@@ -246,7 +258,19 @@ def _current_action(
         raise DyroError("Objective 模式已变化；拒绝执行已确认 Action")
     snapshot = build_scheduler_snapshot(config, objective=record, clock=clock)
     plan = build_continuation_plan(snapshot)
-    tick = build_scheduler_tick(snapshot, plan, max_parallel=record.objective.budget.max_parallel)
+    from ..peer_wave import (
+        discover_available_write_providers,
+        recommended_max_parallel,
+    )
+
+    tick = build_scheduler_tick(
+        snapshot,
+        plan,
+        max_parallel=recommended_max_parallel(
+            record.objective.budget.max_parallel,
+            len(discover_available_write_providers()),
+        ),
+    )
     actual = next((action for action in tick.wave if action == expected), None)
     if actual is None:
         raise DyroError("确认后的 Action 已不再位于当前安全 wave；请重新运行 objective apply")
@@ -356,7 +380,15 @@ def _task_result_status(action: PlannedAction, result: object) -> tuple[ActionSt
 
 def _dispatch(config: Config, action: PlannedAction, task: Task, *, expected_contract_sha256: str) -> object:
     if action.kind is ActionKind.EXECUTE_TASK:
-        return run_task(config, task, expected_contract_sha256=expected_contract_sha256)
+        from ..peer_wave import bind_wave_executors, discover_available_write_providers
+
+        decision = bind_wave_executors((task,), discover_available_write_providers())
+        return run_task(
+            config,
+            task,
+            expected_contract_sha256=expected_contract_sha256,
+            executor_override=decision.executor_for(task.id),
+        )
     if action.kind is ActionKind.REVIEW_TASK:
         return review_task(config, task, expected_contract_sha256=expected_contract_sha256)
     raise DyroError("受监督执行只支持 execute_task 与 review_task")

diff --git a/src/dyro/integrations/assets/dyro-control-plane/SKILL.md b/src/dyro/integrations/assets/dyro-control-plane/SKILL.md
@@ -22,7 +22,7 @@ When the request already supplies a workspace alias, skip global discovery and u
 - Objective explanation: `dyro --workspace <alias> objective explain <id> --format json`
 - Objective blockers or human attention: `dyro --workspace <alias> objective attention <id> --format json`
 - Objective dependency graph: `dyro --workspace <alias> objective graph <id> --format json`
-- Objective next-wave preview: `dyro --workspace <alias> objective tick <id> --format json`
+- Objective next-wave preview: `dyro --workspace <alias> objective tick <id> --format json`. Treat `peer_wave.executor_bindings` as the intended peer executors for that wave, and `peer_wave.warnings` as missing `conflict_group` or harness-capacity notes. A wave member is an executor, not a live supervisor.
 - Objective plan: `dyro --workspace <alias> objective plan <id> --format json`
 
 Use only an existing Objective or Change Set ID returned by Dyro or supplied by the user. A non-zero exit, unavailable workspace, pending transaction, failed finding, missing field, or partial observation is unknown or blocked—not ready.
Original file line number	Diff line number	Diff line change
Expand Up		@@ -234,4 +234,4 @@ Dyro 的交付拓扑与之实质相近：TaskGraph（`depends_on` / conflict

		未来的 adapter、通知、签名规则、发布平台与审批系统应使用 Python entry point 或独立 Profile 扩展包接入；不要把某个组织的策略加入 core 默认行为。

		开发者侧的可选本地多 Agent 派发（五段式任务契约、注入前机密守卫、locator 核验、隔离 patch）与上述控制面分层并列，随 `dyro` 安装包分发（`dyro dispatch` / `import experiments.local_agent_dispatch`），但不替代 gates/合并；见 [`ADR-0002`](adr/0002-optional-local-agent-dispatch.md)、[`多智能体编排纪律`](agent-orchestration-discipline.md) 与 [`可选本地 Agent 派发设计`](designs/optional-local-agent-dispatch.md)。
		开发者侧的可选本地多 Agent 派发（五段式任务契约、注入前机密守卫、locator 核验、隔离 patch）与上述控制面分层并列，随 `dyro` 安装包分发（`dyro dispatch` / `import experiments.local_agent_dispatch`），但不替代 gates/合并。同时写多块走 Core Peer Wave（task worktree + `conflict_group`），见 [`peer-wave-execution.md`](designs/peer-wave-execution.md)、[`ADR-0002`](adr/0002-optional-local-agent-dispatch.md)、[`多智能体编排纪律`](agent-orchestration-discipline.md) 与 [`可选本地 Agent 派发设计`](designs/optional-local-agent-dispatch.md)。
-Original file line number
+Diff line change
@@ Expand Up / @@ -269,7 +269,7 @@ flowchart TB @@
       Host -->|"explicit dyro commands"| Dyro
     ```
-    See `docs/agent-orchestration-discipline.md`. The optional harness under `experiments/local_agent_dispatch/` ships with the installed package but never replaces Core gates or merge.
+    See `docs/agent-orchestration-discipline.md` and `docs/designs/peer-wave-execution.md`. `dyro dispatch` is advisory; simultaneous writes use Core task worktrees, not Batch supervisors. The optional harness under `experiments/local_agent_dispatch/` ships with the installed package but never replaces Core gates or merge.
     ---
@@ Expand Down @@