agent-army

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Agent Army

Agent军团

Architecture

架构

Commander
 |
 |-- Layer 1: Team (3 to 50+, each = own 1M context)
 |    |
 |    |-- Agent A (1M) -- Sub-agent A1 (own context), A2 (own context), ...
 |    |-- Agent B (1M) -- Sub-agent B1 (own context), B2 (own context), ...
 |    |-- Agent C (1M) -- Sub-agent C1 (own context), C2 (own context), ...
 |    |-- ... (no cap)

Commander
 |
 |-- Layer 1: Team (3 to 50+, each = own 1M context)
 |    |
 |    |-- Agent A (1M) -- Sub-agent A1 (own context), A2 (own context), ...
 |    |-- Agent B (1M) -- Sub-agent B1 (own context), B2 (own context), ...
 |    |-- Agent C (1M) -- Sub-agent C1 (own context), C2 (own context), ...
 |    |-- ... (no cap)

Layer 1 Prompt Template

第一层提示词模板

This is what spawned agents actually see. It is the most important section in this skill.

You are [AGENT_NAME], specialist on [DOMAIN].

Objective: [One sentence]
Approved patterns: [Exact values -- hex codes, class names, etc.]
Forbidden patterns: [What to remove/avoid]
Your files: [Absolute paths with line counts]
Rules: [Constraints]. Skip files already using approved patterns. Flag issues outside your files in "Flags for Commander" -- do NOT fix them.

CRITICAL: You MUST use the Agent tool to spawn the sub-agents listed below. Do NOT do the work yourself. Do NOT skip spawning. Deploy ALL sub-agents in a single message with multiple Agent tool calls.

Sub-agents:
- "[NAME]": [files with line counts]
- "[NAME]": [files with line counts]

Pass each sub-agent: objective, their files, approved/forbidden patterns, rules, report format. After all complete, aggregate reports and verify no forbidden patterns remain.

这是被生成的Agent实际会看到的内容，是本技能中最重要的部分。

You are [AGENT_NAME], specialist on [DOMAIN].

Objective: [One sentence]
Approved patterns: [Exact values -- hex codes, class names, etc.]
Forbidden patterns: [What to remove/avoid]
Your files: [Absolute paths with line counts]
Rules: [Constraints]. Skip files already using approved patterns. Flag issues outside your files in "Flags for Commander" -- do NOT fix them.

CRITICAL: You MUST use the Agent tool to spawn the sub-agents listed below. Do NOT do the work yourself. Do NOT skip spawning. Deploy ALL sub-agents in a single message with multiple Agent tool calls.

Sub-agents:
- "[NAME]": [files with line counts]
- "[NAME]": [files with line counts]

Pass each sub-agent: objective, their files, approved/forbidden patterns, rules, report format. After all complete, aggregate reports and verify no forbidden patterns remain.

Layer 2 Prompt Template

第二层提示词模板

You are [SUB_AGENT_NAME], working under [TEAM_MEMBER_NAME].

Objective: [One sentence]
Files you own: [Absolute paths with line counts -- only touch these]
Approved/Forbidden patterns: [Exact values to use and remove]
Rules: [Constraints]. Skip files already correct. Flag issues outside your files -- do NOT fix them.

Report when done: Files Modified (file: N replacements), Files Skipped (already correct), Flags for Commander, Issues Encountered, Status (COMPLETE/PARTIAL/FAILED).

<mandatory-rules>

You are [SUB_AGENT_NAME], working under [TEAM_MEMBER_NAME].

Objective: [One sentence]
Files you own: [Absolute paths with line counts -- only touch these]
Approved/Forbidden patterns: [Exact values to use and remove]
Rules: [Constraints]. Skip files already correct. Flag issues outside your files -- do NOT fix them.

Report when done: Files Modified (file: N replacements), Files Skipped (already correct), Flags for Commander, Issues Encountered, Status (COMPLETE/PARTIAL/FAILED).

<mandatory-rules>

MANDATORY RULES

强制规则

EVERY Layer 1 agent MUST spawn 2+ sub-agents. No exceptions. If deploying without sub-agents, STOP and restructure.
NEVER silently reduce army size. Match the user's chosen tier.

Report as agents complete:

[Agent N/M complete] name: X files modified, Y flags

Show the army plan before deploying (Full Mode). Quick Mode: compose plan internally, skip confirmation.
Sub-agent deployment instructions go INSIDE the Layer 1 brief. If missing, sub-agents will not exist.
Build check after every wave. If it fails, trigger a fix wave.
"Keep going" / "don't stop" = continuous mode. Launch new agents as each completes. Do not wait. Do not ask.

</mandatory-rules>

每个第一层Agent必须生成2个以上的Sub-agent，无例外。如果部署时没有Sub-agent，请立即停止并重新调整结构。
绝对不能悄悄缩减军团规模，必须匹配用户选择的档位。

当Agent完成任务时按以下格式上报：

[Agent N/M complete] name: X files modified, Y flags

部署前展示军团计划（完整模式）；快速模式：内部生成计划，无需确认。
Sub-agent部署指令必须写在第一层Agent的任务说明内部，如果缺失，Sub-agent将无法生成。
每轮执行后都要进行构建检查，如果失败，触发修复轮次。
"Keep going" / "don't stop" = 持续模式，每个Agent完成后立即启动新的Agent，无需等待，无需询问。

</mandatory-rules>

Army Size

军团规模

Confirm tier before starting. Present this table:

| Tier | Agents | Total with Sub-agents | Est. Tokens |
|------|--------|----------------------|-------------|
| Conservative | 3 | ~9 | ~200-500K |
| Standard | 5-10 | ~15-30 | ~500K-1.5M |
| Aggressive | 10-20 | ~30-60 | ~1.5-4M |
| Maximum | 20-50+ | ~60-100+ | ~4M+ |
| Custom | You pick | You pick | Varies |

Token estimates vary by task complexity. Pick a tier or enter a custom number:

Default to Standard if user says "just do it." Tier question applies to BOTH Quick and Full modes.

After recon (Step 3), recommend a specific number based on what you found. Example: "Found 35 files across 6 domains. I recommend Aggressive tier: 8 Layer 1 agents with 2-3 sub-agents each (~22 total, ~2M tokens). Want to adjust?"

开始前确认档位，向用户展示下表：

| Tier | Agents | Total with Sub-agents | Est. Tokens |
|------|--------|----------------------|-------------|
| Conservative | 3 | ~9 | ~200-500K |
| Standard | 5-10 | ~15-30 | ~500K-1.5M |
| Aggressive | 10-20 | ~30-60 | ~1.5-4M |
| Maximum | 20-50+ | ~60-100+ | ~4M+ |
| Custom | You pick | You pick | Varies |

Token estimates vary by task complexity. Pick a tier or enter a custom number:

如果用户说"just do it"则默认使用标准档。档位确认规则同时适用于快速模式和完整模式。

摸排（第3步）完成后，根据摸排结果推荐具体的数量。示例："Found 35 files across 6 domains. I recommend Aggressive tier: 8 Layer 1 agents with 2-3 sub-agents each (~22 total, ~2M tokens). Want to adjust?"

Deployment Gate

部署准入检查

Output this checklist in your response before deploying. Do NOT skip it. Do NOT just check it mentally.

DEPLOYMENT GATE:
[ ] Every L1 brief contains "You MUST spawn N sub-agents"
[ ] Every sub-agent named with specific files assigned
[ ] Every file owned by exactly one sub-agent
[ ] L1 briefs include full sub-agent deployment instructions
[ ] Tier matches user's selection

All must PASS. If any FAIL: fix the plan first.

部署前在响应中输出以下检查清单，不得跳过，不得仅在内部核对。

DEPLOYMENT GATE:
[ ] Every L1 brief contains "You MUST spawn N sub-agents"
[ ] Every sub-agent named with specific files assigned
[ ] Every file owned by exactly one sub-agent
[ ] L1 briefs include full sub-agent deployment instructions
[ ] Tier matches user's selection

所有项必须通过，如有任何一项不通过，请先调整计划。

Protocol

执行流程

Mode: If scope is clear and specific (file paths, exact changes), skip to Step 3. Otherwise start at Step 1.

模式： 如果范围清晰明确（文件路径、具体修改内容），直接跳到第3步，否则从第1步开始。

Step 1: Intake (Full Mode)

第1步：信息收集（完整模式）

Confirm with user:

Goal (one sentence)
Scope (files, directories, or "everything")
Constraints (don't touch X, match Y style)
Tier (show table above)

If user already provided context, confirm in one line: "Goal: [X]. Scope: [Y]. Tier: [Z]. Starting."

向用户确认：

目标（一句话描述）
范围（文件、目录或"全部内容"）
约束（不能修改X、匹配Y风格等）
档位（展示上述表格）

如果用户已经提供了上下文，用一句话确认："Goal: [X]. Scope: [Y]. Tier: [Z]. Starting."

Step 2: Git Checkpoint

第2步：Git检查点

Run
```
git status
```
. Warn if uncommitted changes.

Create checkpoint:

git checkout -b agent-army/checkpoint-{timestamp}

, switch back.

No git repo? Warn user and get confirmation.

运行
```
git status
```
，如果有未提交的变更则发出警告。

创建检查点：

git checkout -b agent-army/checkpoint-{timestamp}

，然后切回原分支。

没有Git仓库？向用户发出警告并获取确认。

Step 3: Recon

第3步：摸排

Scan -- Grep, Glob, Read to find all affected files
Count units of work
```
wc -l
```
each file. Flag 500+ line files as heavy (assign solo)
Classify into domains (by directory, imports, file type, naming pattern)
Identify shared dependencies (foundation files -- handle first)
Identify build command (check package.json, Makefile, etc.)

Output scope report:

Files: N | Heavy: [list] | Domains: [list] | Shared deps: [list] | Build cmd: [cmd]

扫描——通过Grep、Glob、读取操作找出所有受影响的文件
统计工作单元数量
对每个文件执行
```
wc -l
```
统计行数，标记500行以上的文件为大文件（单独分配）
按目录、导入、文件类型、命名模式分类到不同领域
识别共享依赖（基础文件——优先处理）
识别构建命令（检查package.json、Makefile等）

输出范围报告：

Files: N | Heavy: [list] | Domains: [list] | Shared deps: [list] | Build cmd: [cmd]

Step 4: Compose

第4步：编排组建

Foundation Agent for shared deps (runs first, before parallel wave)
Layer 1: 1 agent per domain. Split large domains across multiple agents.
Layer 2: 2+ sub-agents per L1 agent based on file weight
Name every agent. Assign every file to exactly one sub-agent.

Output army plan. Full Mode: pause for "Proceed?" Quick Mode: show one-line summary, deploy.

优先为共享依赖配置基础Agent（先执行，早于并行批次）
第一层：每个领域配置1个Agent，大型领域拆分给多个Agent处理
第二层：根据文件重量为每个第一层Agent配置2个以上的Sub-agent
为每个Agent命名，每个文件恰好分配给一个Sub-agent

输出军团计划。完整模式：暂停等待用户回复"Proceed?"；快速模式：展示一行总结后直接部署。

Step 5: Deploy

第5步：部署

Foundation Agent first (if needed). Wait for completion.
Launch ALL Layer 1 agents in parallel (
```
run_in_background: true
```
, single message).
L1 agents spawn L2 sub-agents in parallel (per their brief).
Report each completion:
```
[Agent N/M complete] name: results
```

优先启动基础Agent（如有需要），等待其完成。
并行启动所有第一层Agent（
```
run_in_background: true
```
，单条消息触发）。
第一层Agent根据其任务说明并行生成第二层Sub-agent。
每个Agent完成后上报结果：
```
[Agent N/M complete] name: results
```

Step 6: Verify

第6步：校验

Re-scan for remaining violations (same searches as Step 3)
Run build command
Resolve cross-team flags
Output army report:

Agents: N total | Files modified: N | Skipped: N | Build: PASS/FAIL | Flags: [list] | Rollback: git checkout agent-army/checkpoint-{timestamp}

重新扫描是否有剩余违规项（和第3步使用相同的搜索规则）
执行构建命令
解决跨团队标记的问题
输出军团报告：

Agents: N total | Files modified: N | Skipped: N | Build: PASS/FAIL | Flags: [list] | Rollback: git checkout agent-army/checkpoint-{timestamp}

Waves

执行波次

After Wave 1: build + re-scan. If clean, done. If issues remain, deploy Wave 2 (fix wave). If still issues, Wave 3 (propagate to tests/docs). Max 3 waves. Pause for user approval before each.

第1波完成后：执行构建+重新扫描，如果没有问题则结束；如果仍有问题，部署第2波（修复波）；如果还有问题，部署第3波（同步到测试/文档），最多3波。每波部署前都需要暂停获取用户批准。

Shared Scratchpad (optional)

共享暂存区（可选）

For complex multi-wave tasks, write

.army-state.md

after Wave 1: files modified, flags, issues, decisions. Wave 2 agents read it directly for full context. Skip for single-wave tasks.

针对复杂的多波次任务，第1波完成后写入

.army-state.md

：修改的文件、标记的问题、遇到的问题、决策内容。第2波的Agent直接读取该文件获取完整上下文，单波次任务可跳过。

Report Format

报告格式

Every agent returns:

undefined

每个Agent返回内容如下：

undefined

Report: [Name]

Files Modified: [file: N replacements] Files Skipped: [file: reason] Flags for Commander: [issue or "None"] Issues: [issue or "None"] Status: COMPLETE / PARTIAL / FAILED

undefined

Files Modified: [file: N replacements] Files Skipped: [file: reason] Flags for Commander: [issue or "None"] Issues: [issue or "None"] Status: COMPLETE / PARTIAL / FAILED

undefined

Error Handling

错误处理

Sub-agent fails: L1 agent retries or spawns replacement
L1 agent fails: Commander spawns replacement with same brief
Build failure: compare against checkpoint, diagnose, fix wave

Sub-agent执行失败：第一层Agent重试或生成新的替代Sub-agent
第一层Agent执行失败：指挥官生成相同任务说明的替代Agent
构建失败：和检查点版本对比，诊断问题，触发修复波