cowork-flow 0.0.23 → 0.0.24

package/README.md

@@ -55,7 +55,9 @@ cowork-flow 是一个用于新项目初始化协作流程的模板仓库。它
 项目级协作入口，包含编码前思考、简单优先、外科手术式改动、验证优先等基础原则。接入项目后，应把项目名称、技术栈、运行命令、测试命令和提交策略补齐。
 
 `template/.agents/skills/`
-本地技能入口，覆盖 start、before-dev、brainstorming、writing-plans、check、finish-work、continue、meta、python-design、update-spec 和 break-loop 等协作动作。这里的 skill 应保持通用，不承载某个业务项目的一次性细节。
+本地技能入口，覆盖 start、before-dev、brainstorming、party-mode、writing-plans、check、finish-work、continue、meta、python-design、update-spec 和 break-loop 等协作动作。这里的 skill 应保持通用，不承载某个业务项目的一次性细节。
+
+`party-mode` 是用户手动调用的 advisory roundtable。它通过当前 Host Adapter 协调真实子代理讨论，而不是模拟人设；默认 `max_agents=3`、`max_rounds=3`，配置优先级为本次调用、task/change、本地配置、skill 默认值。Party Mode 只产出建议、证据、分歧和验收信号，不能推进任务状态，也不能替代 `cowork-implement` 或 `cowork-check`。
 
 `template/.codex/agents/`
 固定角色 agent 定义，包含 `cowork-research`、`cowork-implement` 和 `cowork-check`，以及对 Codex 默认 `worker`、`default`、`explorer` 的项目级防漂移约束。主会话负责计划与收口，并通过 host adapter 派发固定 agent；正式派发必须先创建 runtime context，再把 `cowork_runtime_context_id` 和 `cowork_host_context_key` 传给子代理。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cowork-flow",
-  "version": "0.0.23",
+  "version": "0.0.24",
   "description": "CLI for installing and maintaining cowork-flow collaboration templates.",
   "type": "module",
   "bin": {

package/template/.agents/skills/party-mode/SKILL.md

@@ -0,0 +1,120 @@
+---
+name: party-mode
+description: Use when the user manually requests Party Mode or a real multi-agent roundtable for advisory discussion, option review, risk review, or decision convergence.
+---
+
+# Party Mode
+
+Use this skill as a manual advisory roundtable. Coordinate true child agents, not simulated personas, through the current Host Adapter.
+
+## Boundaries
+
+- Party Mode is advisory only.
+- It cannot mutate task status, start/review/complete/archive tasks, add sessions, commit, push, or satisfy formal workflow gates.
+- It cannot satisfy formal Implement or Check completion. Use `cowork-implement` and `cowork-check` for formal phases.
+- Discussion children are leaf executors. They do not dispatch, wait for, list, or close other agents.
+- Generic `worker`, `default`, or `explorer` agents may provide advisory views, but their output never proves implementation or checking done.
+
+## Defaults And Config
+
+Built-in defaults:
+
+- `max_agents=3`
+- `max_rounds=3`
+
+These defaults are configurable values, not hard-coded constants. Effective config order:
+
+```text
+call arguments > task/change config > `.cowork-flow/config.yaml` > skill defaults
+```
+
+Configurable fields:
+
+- `max_agents`
+- `max_rounds`
+- agent roster or review lenses
+- report enabled
+- report path
+- whether round 2 or round 3 is allowed
+
+Safety gates:
+
+- continue conditions can be tightened but not removed
+- stop conditions can be tightened but not removed
+- schema core fields can be extended but not removed
+- exceeding effective limits requires explicit user approval
+
+## Round Model
+
+1. Frame the question, decision needed, scope, and evidence packet.
+2. Round 1 uses fresh child contexts. Send each child the same packet and one lens. Child agents cannot see each other.
+3. Synthesize only evidence-backed positions, disagreements, rejected options, and acceptance signals.
+4. Continue only when a continue condition is met. Send narrow follow-up prompts to the smallest useful set of children.
+5. Stop when any stop condition is met. Close all live children through the Host Adapter close primitive.
+
+Round intent:
+
+- Opening round: independent first judgments.
+- Challenge rounds: rebuttal, risk drilldown, or evidence repair on specific disagreements.
+- Convergence rounds: decision check only. Verify, narrow, or choose. Do not open new directions.
+
+Default mapping:
+
+- Round 1 = Opening.
+- Round 2 = Challenge.
+- Round 3+ = Convergence.
+
+Only the user can restart exploration after convergence begins.
+
+## Continue Conditions
+
+Run another round only if at least one condition holds:
+
+- A disagreement could change the recommended decision.
+- A high risk lacks enough evidence.
+- Acceptance criteria are still not testable.
+- A child found new file, command, rule, or user-scenario evidence.
+- The coordinator cannot write one recommended direction.
+
+## Stop Conditions
+
+Stop when any condition holds:
+
+- Recommendation, rejected options, and measurable acceptance criteria are clear.
+- Remaining issue is user value preference, not missing evidence.
+- A full round adds no evidence and does not narrow scope.
+- Effective `max_rounds` is reached.
+- Output fails schema, and one repair prompt does not fix it.
+
+## Child Output Schema
+
+Each child must return these core fields. Extra fields are allowed only after them.
+
+```text
+position:
+evidence:
+risk:
+tradeoff:
+rejected_option:
+acceptance_signal:
+what_would_change_my_mind:
+```
+
+Reject unsupported opinion. Evidence should name files, commands, rules, observed behavior, user scenarios, or concrete assumptions.
+
+## Coordinator Output Schema
+
+Final synthesis must return these core fields. Extra fields are allowed only after them.
+
+```text
+consensus:
+disagreements:
+evidence:
+decision:
+rejected_options:
+acceptance_criteria:
+open_questions:
+stop_reason:
+```
+
+Keep the final decision traceable to child evidence. Do not count votes as validation.

package/template/.claude/skills/party-mode/SKILL.md

@@ -0,0 +1,120 @@
+---
+name: party-mode
+description: Use when the user manually requests Party Mode or a real multi-agent roundtable for advisory discussion, option review, risk review, or decision convergence.
+---
+
+# Party Mode
+
+Use this skill as a manual advisory roundtable. Coordinate true child agents, not simulated personas, through the current Host Adapter.
+
+## Boundaries
+
+- Party Mode is advisory only.
+- It cannot mutate task status, start/review/complete/archive tasks, add sessions, commit, push, or satisfy formal workflow gates.
+- It cannot satisfy formal Implement or Check completion. Use `cowork-implement` and `cowork-check` for formal phases.
+- Discussion children are leaf executors. They do not dispatch, wait for, list, or close other agents.
+- Generic `worker`, `default`, or `explorer` agents may provide advisory views, but their output never proves implementation or checking done.
+
+## Defaults And Config
+
+Built-in defaults:
+
+- `max_agents=3`
+- `max_rounds=3`
+
+These defaults are configurable values, not hard-coded constants. Effective config order:
+
+```text
+call arguments > task/change config > `.cowork-flow/config.yaml` > skill defaults
+```
+
+Configurable fields:
+
+- `max_agents`
+- `max_rounds`
+- agent roster or review lenses
+- report enabled
+- report path
+- whether round 2 or round 3 is allowed
+
+Safety gates:
+
+- continue conditions can be tightened but not removed
+- stop conditions can be tightened but not removed
+- schema core fields can be extended but not removed
+- exceeding effective limits requires explicit user approval
+
+## Round Model
+
+1. Frame the question, decision needed, scope, and evidence packet.
+2. Round 1 uses fresh child contexts. Send each child the same packet and one lens. Child agents cannot see each other.
+3. Synthesize only evidence-backed positions, disagreements, rejected options, and acceptance signals.
+4. Continue only when a continue condition is met. Send narrow follow-up prompts to the smallest useful set of children.
+5. Stop when any stop condition is met. Close all live children through the Host Adapter close primitive.
+
+Round intent:
+
+- Opening round: independent first judgments.
+- Challenge rounds: rebuttal, risk drilldown, or evidence repair on specific disagreements.
+- Convergence rounds: decision check only. Verify, narrow, or choose. Do not open new directions.
+
+Default mapping:
+
+- Round 1 = Opening.
+- Round 2 = Challenge.
+- Round 3+ = Convergence.
+
+Only the user can restart exploration after convergence begins.
+
+## Continue Conditions
+
+Run another round only if at least one condition holds:
+
+- A disagreement could change the recommended decision.
+- A high risk lacks enough evidence.
+- Acceptance criteria are still not testable.
+- A child found new file, command, rule, or user-scenario evidence.
+- The coordinator cannot write one recommended direction.
+
+## Stop Conditions
+
+Stop when any condition holds:
+
+- Recommendation, rejected options, and measurable acceptance criteria are clear.
+- Remaining issue is user value preference, not missing evidence.
+- A full round adds no evidence and does not narrow scope.
+- Effective `max_rounds` is reached.
+- Output fails schema, and one repair prompt does not fix it.
+
+## Child Output Schema
+
+Each child must return these core fields. Extra fields are allowed only after them.
+
+```text
+position:
+evidence:
+risk:
+tradeoff:
+rejected_option:
+acceptance_signal:
+what_would_change_my_mind:
+```
+
+Reject unsupported opinion. Evidence should name files, commands, rules, observed behavior, user scenarios, or concrete assumptions.
+
+## Coordinator Output Schema
+
+Final synthesis must return these core fields. Extra fields are allowed only after them.
+
+```text
+consensus:
+disagreements:
+evidence:
+decision:
+rejected_options:
+acceptance_criteria:
+open_questions:
+stop_reason:
+```
+
+Keep the final decision traceable to child evidence. Do not count votes as validation.

package/template/.cowork-flow/.version

@@ -1 +1 @@
-0.0.23
+0.0.24

package/template/.cowork-flow/spec/subagent-dispatch.md

@@ -12,6 +12,12 @@ classification or a prompt handshake.
 - Generic `worker`, `default`, or `explorer` dispatch is advisory only and
   cannot satisfy formal Implement or Check completion.
 
+## Advisory Party Mode
+
+Party Mode discussion children are advisory leaf executors. They use fresh child contexts for evidence gathering, disagreement surfacing, risk review, and acceptance-signal review. They cannot mutate task status, write code, archive, commit, or coordinate other agents. Their output cannot satisfy formal Implement or Check completion.
+
+The `party-mode` skill owns round limits, continuation gates, stop gates, and output schemas. This document remains the formal `cowork-*` dispatch protocol source.
+
 ## Runtime Context
 
 Before spawning a formal child, the main session creates:

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

@@ -92,27 +92,33 @@ changes -> brainstorming -> read spec -> plan -> tasks -> implement -> check ->
 - 固定 `cowork-*` 代理是叶子执行者；不得再派发、等待、列出或取消其他代理。
 - 通用 `worker`、`default` 或 `explorer` 只能作为 advisory work，不能满足正式实现或检查完成条件。
 
-## 3.2 并行会话
+## 3.2 手动 Party Mode
+
+Party Mode 是用户手动触发的 advisory roundtable。主会话可通过当前宿主适配器创建 fresh child contexts，收集真实讨论子代理的证据、分歧、风险和可测验收信号，再由主会话综合结论。
+
+Party Mode 不能推进任务状态，不能满足正式实现或检查完成条件，也不能替代 `cowork-implement` 或 `cowork-check`。轮次上限、继续/停止条件、输出 schema 和可配置默认值由 party-mode skill 定义；正式子代理协议仍以 `.cowork-flow/spec/subagent-dispatch.md` 为准。
+
+## 3.3 并行会话
 
 并行执行采用干净隔离的并行会话模型。并行只是执行策略，不改变固定代理叶子边界。
 
-### 3.2.1 并行决策
+### 3.3.1 并行决策
 
 - 用户无需在需求输入时声明是否并行；计划阶段由主会话评估并行可行性。
 - 开发计划必须明确执行策略：串行执行，或列出可并行的低冲突切片。
 - 同一文件、同一行为链、依赖未合并或验收标准不清的工作不得并行，改为串行。
 
-### 3.2.2 隔离方式
+### 3.3.2 隔离方式
 
 - 多个独立任务优先拆成多个 cowork-flow 会话。
 - 只要存在写入冲突风险，就用独立 `git worktree` 隔离。
 
-### 3.2.3 单任务内切片
+### 3.3.3 单任务内切片
 
 - 单个任务内只允许低冲突切片并行。
 - 每个切片必须写清文件归属、依赖关系、预期产物和验证命令。
 
-### 3.2.4 协调与验收
+### 3.3.4 协调与验收
 
 - 主会话是唯一协调者：派发所有子代理后逐个等待，核对子代理汇报的文件、命令和产物，再用适配器列表/取消原语收口。
 - 多个实现切片合并后必须再执行最终集成验证；不能把各子代理的局部通过当成整体通过。