cowork-flow 0.0.11 → 0.0.13

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cowork-flow",
-  "version": "0.0.11",
+  "version": "0.0.13",
   "description": "CLI for installing and maintaining cowork-flow collaboration templates.",
   "type": "module",
   "bin": {

package/template/.agent/skills/agent-team-execution/SKILL.md

@@ -19,17 +19,30 @@ Use this skill after a task is active and an approved `.cowork-flow/plans/*.md`
 8. Repeat `next -> dispatch -> record` until no assignments remain.
 9. Run `./.cowork-flow/run agent-team complete <task-dir>` before claiming the agent team work is done.
 
+## Dispatch Protocol
+
+Do not invoke `superpowers:subagent-driven-development` from this skill. Agent-team has its own persisted state machine; use Codex subagent orchestration directly and keep `agent-team` as the source of truth for ready assignments, results, reviews, retries, and completion.
+
 ### Codex Orchestration Prompt
 
 - Ask Codex explicitly to spawn subagents. Use this wording for each ready batch: `Spawn one <agent_type> agent per ready assignment, wait for all requested results, and summarize each result by assignment id.`
 - Codex handles spawning, routing follow-up instructions, waiting for results, and closing completed agent threads. Do not look for an `agent-team` spawn subcommand; the Python script only prepares state and prompts.
 - Use `/agent` in interactive CLI sessions when you need to inspect, steer, stop, or close active agent threads.
 
+### Subagent Evidence Gate
+
+- Do not treat wording in the final answer as evidence that a subagent actually ran. Phrases like "worker reported" or "explorer agent result" can be produced by the parent agent without a real child thread.
+- Before recording any assignment result, confirm real subagent evidence from the Codex host: visible subagent activity in the app or CLI, `/agent` showing the child thread, JSON event output whose item type shows an agent thread or agent job, or a successful `spawn_agents_on_csv` output written by child workers.
+- If no subagent evidence appears, stop before `record-result` or `record-review`. Report that the Codex runtime did not expose or start subagents for this batch, include the exact dispatch prompt or command used, and keep the assignment ready for retry in a working runtime.
+- In `codex exec --json` experiments, inspect the event stream. A run that only shows the parent thread plus ordinary `command_execution` items is not a successful subagent dispatch, even if the final answer claims an agent result.
+
 ### Fresh Worker Per Assignment
 
 - Spawn one agent per ready assignment. Use the assignment's `agent_type` as the Codex spawn target; built-in Codex agent types include `default`, `worker`, and `explorer`, and project or personal custom agents may add more.
+- Do not request a full-history fork when you also specify `agent_type`. In Codex hosts that expose real subagent dispatch tools, full-history fork cannot be combined with `agent_type` and the spawn call will be rejected.
 - Treat `recommended_agent` as the agent-team registry match and prompt source, not as the Codex spawn target unless it also names a real Codex custom agent.
 - Spawn a fresh worker for each ready assignment. Do not reuse the main agent context as the worker context.
+- When the host offers fork controls, prefer a fresh child thread with explicit prompt context over inherited full history. Put the assignment Markdown, write boundary, and any needed task facts directly in the worker prompt.
 - Use `agent-team/assignments/<assignment-id>.md` as the worker prompt body.
 - Add this scene-setting to every worker prompt: it is not alone in the codebase, it must respect the listed write boundary, it must not revert other agents' edits, and it must report changed files plus exact verification commands.
 - Dispatch multiple ready assignments in parallel only after confirming their write files do not overlap. If they overlap, dispatch them sequentially or repair the dependency graph.

package/template/.cowork-flow/.version

@@ -1 +1 @@
-0.0.11
+0.0.13

package/template/.cowork-flow/workflow.md

@@ -224,13 +224,13 @@ done
 - 子任务边界不清，拆分后的主要成本会落在集成。
 - 任务规模小于 subagent 调度成本。
 
-若不适合使用 subagent，必须简要记录理由，并在当前会话中按 plan 顺序执行。
+若不适合使用 subagent，必须简要记录理由，并按 `superpowers`方法论执行。
 
 若适合使用 subagent：
 
 - 当 `.cowork-flow/config.yaml` 中设置 `agent_team.enabled: true` 时，优先使用 `./.cowork-flow/run agent-team prepare <task-dir> --plan <plan-file>` 生成调度图，再用 `./.cowork-flow/run agent-team next <task-dir>` 获取 assignments，由主 agent 审核并调度 subAgent。
-- 当 `agent_team.enabled: false` 但宿主支持 subAgent 时，使用 `superpowers:subagent-driven-development` 或宿主等价机制开启多 agent 执行。
-- 当宿主不支持 subAgent 或任务不适合拆分时，使用 `superpowers:executing-plans` 或当前会话顺序执行。
+- 当 `agent_team.enabled: false` 或者`agent_team`执行失败，使用 `superpowers`并行执行 或宿主等价机制开启多 agent 执行。
+- 当宿主不支持 subAgent 或任务不适合拆分时，按 `superpowers`方法论执行。
 
 无论是否使用 subagent，都必须严格遵循 `superpowers:test-driven-development`，不得先写生产代码再补测试。