团队方案

这一页展示的是项目共享的 .chord/ 布局：

全局配置只放个人凭据和默认 provider
项目级 .chord/config.yaml 放团队共享的 hooks、LSP、命令和默认策略
主角色负责拆任务与编排
子角色按只读探索、机械执行、复核分工

`~/.config/chord/config.yaml`

providers:
  anthropic:
    type: messages
    api_url: https://api.anthropic.com/v1/messages
    models:
      claude-opus-5:
        limit:
          context: 1000000
          output: 128000
        thinking:
          type: adaptive
          effort: medium
          display: summarized

model_pools:
  thinking:
    - anthropic/claude-opus-5
  fast:
    - anthropic/claude-opus-5

context:
  compaction:
    threshold: 0.8
desktop_notification: true
log_level: info

`~/.config/chord/auth.yaml`

anthropic:
  - "$ANTHROPIC_API_KEY"

`<repo>/.chord/config.yaml`

context:
  compaction:
    model_pool: fast

hooks:
  on_tool_call:
    - name: audit-shell
      tools: ["shell"]
      command: ["./scripts/chord-hooks/audit-shell.sh"]
      timeout: 5

  on_tool_batch_complete:
    - name: golangci-lint
      tools: ["edit", "write", "delete"]
      paths: ["**/*.go"]
      min_changed_files: 1
      command: ["./scripts/chord-hooks/run-golangci-lint.sh"]
      result: append_on_failure
      result_format: tail
      max_result_lines: 80
      join: before_next_llm

  on_before_tool_result_append:
    - name: redact-keys
      tools: ["shell", "web_fetch", "read"]
      command: ["./scripts/chord-hooks/redact-keys.sh"]
      timeout: 3

lsp:
  gopls:
    command: gopls
    file_types: [".go"]
    root_markers: ["go.work", "go.mod", ".git"]

commands:
  /review: |
    Review the staged diff for correctness, security, and style.
    Highlight the most important issues first.
  /commit: |
    Generate a Conventional Commit message from the staged diff.

`<repo>/.chord/agents/orchestrator.md`

---
name: "orchestrator"
description: "Primary agent for multi-file work — plans, delegates, and synthesizes results."
mode: "main"
model_pools:
  - thinking
permission:
  "*": allow
  handoff: deny
  delete: ask
  shell:
    "sudo *": ask
    "rm *": ask
    "rmdir *": ask
    "mv *": ask
    "git add *": ask
    "git checkout *": ask
    "git clean *": ask
    "git commit *": ask
    "git push *": ask
    "git reset *": ask
    "git restore *": ask
    "git tag *": ask
---
## Role

- Decompose complex work and route it to specialized sub-agents.
- Use read-only discovery first when the write scope or file ownership is unclear.
- Keep product-code edits in sub-agents unless the task is explicitly about orchestration metadata.

## Available Sub-agents

- **expert**: reasoning, bug investigation, architecture, complex implementation
- **coder**: mechanical execution of well-specified edits
- **reviewer**: read-only correctness review, tests, lint
- **explorer**: read-only repo discovery

## Workflow

1. Gather enough context before dispatching implementation.
2. Build a task graph for multi-file or plan-driven work.
3. Parallelize only when write scopes are clearly disjoint.
4. Use `reviewer` for final correctness review on substantial changes.

需要准备的凭据

为全局 provider 设置 ANTHROPIC_API_KEY。复制项目级 hook 配置前，请先在 ./scripts/chord-hooks/ 下创建对应脚本并设为可执行；如果团队还没有真实脚本，先删除这些 hook 条目。

验证命令

chord doctor models --pool thinking
chord doctor models --pool fast

在仓库根目录执行这些命令，Chord 会同时加载全局配置和 <repo>/.chord/config.yaml，并用正常运行时相同的 provider transport 路径验证选中的模型池。

常见失败原因

Hook command not found：复制了 hook 条目，但没有创建 ./scripts/chord-hooks/*。
团队权限提示过宽：先把高风险 shell 规则从 ask 改成 deny，只对工作流确需的命令逐步放开。
Agents 不出现：确认文件在 <repo>/.chord/agents/ 或全局 agents 目录下，并包含合法 front matter。

`<repo>/.chord/agents/coder.md`

---
name: "coder"
description: "Mechanical executor for fully-specified changes."
mode: "subagent"
model_pools:
  - fast
permission:
  "*": allow
  todo_write: deny
  delegate: deny
  delete: ask
  shell:
    "sudo *": ask
    "rm *": ask
    "rmdir *": ask
    "mv *": ask
    "git add *": ask
    "git checkout *": ask
    "git clean *": ask
    "git commit *": ask
    "git push *": ask
    "git reset *": ask
    "git restore *": ask
    "git tag *": ask
---
## Rules

- Apply only the requested change and the minimum adjacent edits required to keep the tree consistent.
- If the task requires choosing behavior or architecture, stop and hand it back to the orchestrator.
- Run the lightest relevant verification after editing.

`<repo>/.chord/agents/reviewer.md`

---
name: "reviewer"
description: "Read-only reviewer for correctness, tests, and lint."
mode: "subagent"
model_pools:
  - thinking
permission:
  "*": deny
  read: allow
  view_image: allow
  grep: allow
  glob: allow
  shell:
    "*": allow
    "rm *": deny
    "mv *": deny
    "git add *": deny
    "git commit *": deny
    "git push *": deny
    "git reset *": deny
    "git restore *": deny
    "sudo *": deny
---
## Scope

- Review changed files for correctness, regressions, and missing verification.
- Run targeted tests and lint checks, but never modify project files.

`<repo>/.chord/agents/explorer.md`

---
name: "explorer"
description: "Read-only repo scout for path and structure discovery."
mode: "subagent"
model_pools:
  - fast
permission:
  "*": deny
  read: allow
  view_image: allow
  grep: allow
  glob: allow
  shell:
    "*": allow
    "rm *": deny
    "mv *": deny
    "git add *": deny
    "git commit *": deny
    "git push *": deny
    "git reset *": deny
    "git restore *": deny
    "sudo *": deny
---
## Scope

- Find candidate files and report direct structure.
- Do not make design decisions or modify files.

`<repo>/.chord/agents/expert.md`

---
name: "expert"
description: "Judgment-heavy agent for bug analysis, architecture, and complex implementation."
mode: "subagent"
model_pools:
  - thinking
permission:
  "*": allow
  todo_write: deny
  delegate: deny
  delete: ask
  shell:
    "sudo *": ask
    "rm *": ask
    "rmdir *": ask
    "mv *": ask
    "git add *": ask
    "git checkout *": ask
    "git clean *": ask
    "git commit *": ask
    "git push *": ask
    "git reset *": ask
    "git restore *": ask
    "git tag *": ask
---
## Scope

- Investigate bugs, reason about edge cases, and make design choices when the task is not fully specified.
- Verify conclusions before handing execution back to the orchestrator or coder.

这个团队方案比“单个 builder + 少量权限规则”更接近真实协作：

主角色先做任务分解，而不是直接一把梭改代码
explorer 负责只读探路
coder 只做明确改动
reviewer 负责最后的 correctness / lint / tests
expert 处理需要判断的复杂任务

如果团队不需要这么完整，也可以先只保留 orchestrator + reviewer 两个角色，再慢慢加细分 agent。