helloagents-x 2.3.3b1

HelloAGENTS - 自主的高级智能伙伴

HelloAGENTS

Let AI go beyond analysis — keep pushing until implementation and verification are done.

---

Table of Contents

• Before and After
• Core Features
• Quick Start
• Configuration
• How It Works
• In-Chat Workflow Commands
• Usage Guide
• Repository Guide
• FAQ
• Troubleshooting
• Version History
• Contributing
• License

Before and After

Without HelloAGENTS

With HelloAGENTS

Challenge	Without HelloAGENTS	With HelloAGENTS
Stops at planning	Ends with suggestions	Pushes to implementation and validation
Output drift	Different structure every prompt	Unified routing and stage chain
Risky operations	Easier to make destructive mistakes	EHRB risk detection and escalation
Knowledge continuity	Context gets scattered	Built-in KB and session memory
Reusability	Prompt-by-prompt effort	Commandized reusable workflow

Core Features

RLM Sub-Agent Orchestration 5 specialized roles (reviewer / synthesizer / kb_keeper / pkg_keeper / writer) plus host CLI native sub-agents (explore / implement / test / design) are dispatched automatically based on task complexity. Tasks are scheduled via DAG dependency analysis with topological sort and layer-by-layer parallel dispatch. Supports cross-CLI parallel scheduling and Agent Teams collaboration. Your gain: complex tasks are broken down and handled by the right specialist, with parallel execution when possible.

Structured Workflow (Evaluate → Design → Develop) Every input is scored on five dimensions and routed to R0 direct response, R1 fast flow, R2 simplified flow, or R3 standard flow. R2/R3 enter the full stage chain with explicit entry conditions, deliverables, and verification gates. Supports interactive and fully delegated modes. Your gain: proportional effort — simple queries stay fast, complex tasks get full process with verification at every step.

Three-Layer Safety Detection (EHRB) Keyword scan, semantic analysis, and tool-output inspection catch destructive operations before execution. Interactive and delegated modes enforce user confirmation. Your gain: safer defaults with zero-config protection.

Three-Layer Memory Model L0 user memory (global preferences), L1 project knowledge base (structured docs synced from code), and L2 session summaries (auto-persisted at stage transitions). Your gain: context survives across sessions and projects.

Sub-Agent Native Mapping

CLI	Native Sub-Agent Mechanism	RLM Mapping
Claude Code	Task tool (explore / code / shell)	Direct mapping, supports Agent Teams
Codex CLI	spawn_agent / Collab (multi-thread)	spawn_agent parallel scheduling, CSV batch orchestration
OpenCode	Built-in agent mode	Fallback to sequential execution
Gemini CLI	Built-in tool calls	Fallback to sequential execution
Qwen CLI	Built-in tool calls	Fallback to sequential execution
Grok CLI	Built-in tool calls	Fallback to sequential execution

Additionally, HelloAGENTS provides: five-dimension routing scoring (action need, target clarity, decision scope, impact range, EHRB risk) to automatically determine processing depth for each input; 6 CLI targets (Claude Code / Codex CLI / OpenCode / Gemini CLI / Qwen CLI / Grok CLI) with one rule set across all; Hooks integration (Claude Code 9 lifecycle hooks + Codex CLI notify hook) with automatic graceful degradation when unavailable.

Quick Start

⚠️ Prerequisite: All AI CLIs (Codex CLI / Claude Code, etc.) should be upgraded to the latest version with relevant feature flags enabled (e.g., sub-agents, CSV orchestration) to access all HelloAGENTS capabilities. VSCode extensions for these CLIs update more slowly — some newer features may require waiting for the extension to catch up. See CLI-specific compatibility notes below.

⚠️ Windows PowerShell 5.1 does not support &&. Run commands on each side of && separately, or upgrade to PowerShell 7+.

Method A: One-line install script (recommended)

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/hellowind777/helloagents/main/install.sh | bash

Windows PowerShell:

irm https://raw.githubusercontent.com/hellowind777/helloagents/main/install.ps1 | iex

The script auto-detects uv or pip, installs the HelloAGENTS package, and launches an interactive menu for you to select target CLIs. Re-running performs an update.

Update: re-run the install command above.

Uninstall: uv tool uninstall helloagents or pip uninstall helloagents (depends on what the script detected)

Switch branch:

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/hellowind777/helloagents/beta/install.sh | HELLOAGENTS_BRANCH=beta bash

# Windows PowerShell
$env:HELLOAGENTS_BRANCH="beta"; irm https://raw.githubusercontent.com/hellowind777/helloagents/beta/install.ps1 | iex

Method B: npx (Node.js >= 16)

npx helloagents

Installs the Python package via pip and launches an interactive menu. You can also specify directly: npx helloagents install codex (or use npx -y to auto-download without prompting)

Requires Python >= 3.10. After first install, use the native helloagents command directly.

Acknowledgment: Thanks to @setsuna1106 for generously transferring the npm helloagents package ownership.

Update: npx helloagents@latest

Uninstall: pip uninstall helloagents

Switch branch: npx helloagents@beta

Method C: UV (isolated environment)

Step 0 — Install UV first (skip if already installed):

# Windows PowerShell
irm https://astral.sh/uv/install.ps1 | iex

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

After installing UV, restart your terminal to make the uv command available.

Install and select targets (one command):

uv tool install --from git+https://github.com/hellowind777/helloagents helloagents && helloagents

Installs the package and launches an interactive menu for you to select target CLIs. You can also specify directly: helloagents install codex

Update: uv tool install --from git+https://github.com/hellowind777/helloagents helloagents --force

Uninstall: uv tool uninstall helloagents

Switch branch: uv tool install --from git+https://github.com/hellowind777/helloagents@beta helloagents --force

Method D: pip (Python >= 3.10)

Install and select targets (one command):

pip install git+https://github.com/hellowind777/helloagents.git && helloagents

Installs the package and launches an interactive menu for you to select target CLIs. You can also specify directly: helloagents install codex

Update: pip install --upgrade git+https://github.com/hellowind777/helloagents.git

Uninstall: pip uninstall helloagents

Switch branch: pip install --upgrade git+https://github.com/hellowind777/helloagents.git@beta

HelloAGENTS commands (after installation)

⚠️ These commands depend on the installed package. If a remote update causes issues, use the native install/update/uninstall commands from your installation method above.

helloagents                  # interactive menu
helloagents install codex    # specify target directly
helloagents install --all    # install to all detected CLIs
helloagents status           # check installation status
helloagents version          # check version
helloagents update           # update + auto-sync all targets
helloagents update beta      # switch branch + auto-sync
helloagents uninstall codex  # uninstall from a CLI target
helloagents uninstall --all  # uninstall from all targets
helloagents clean            # clean caches

Codex CLI example

First install:

# One-line script (recommended, auto-launches interactive menu after install)
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/hellowind777/helloagents/main/install.sh | bash

# Windows PowerShell
irm https://raw.githubusercontent.com/hellowind777/helloagents/main/install.ps1 | iex

# npx (or use npx -y to auto-download without prompting)
npx helloagents install codex

# UV
uv tool install --from git+https://github.com/hellowind777/helloagents helloagents && helloagents install codex

# pip
pip install git+https://github.com/hellowind777/helloagents.git && helloagents install codex

Update later (auto-syncs installed targets):

helloagents update

⚠️ Codex CLI config.toml compatibility notes: The following settings may affect HelloAGENTS:

• [features] child_agents_md = true — experimental, injects extra instructions that may conflict with HelloAGENTS
• project_doc_max_bytes too low — default 32KB, AGENTS.md will be truncated (auto-set to 131072 during install)
• agent_max_depth = 1 — limits sub-agent nesting depth, recommend keeping default or ≥2
• agent_max_threads too low — default 6, lower values limit parallel sub-agent scheduling (CSV batch mode recommends ≥16)
• [features] multi_agent = true — must be enabled for sub-agent orchestration to work
• [features] sqlite = true — must be enabled for CSV batch orchestration (spawn_agents_on_csv)
• Collab sub-agent scheduling requires Codex CLI feature gate to be enabled

💡 Best practices:

• HelloAGENTS is optimized for Codex CLI — supports high and below reasoning effort levels. xhigh reasoning is not supported and may cause instruction-following issues
• Use the terminal/CLI version of Codex for the best experience. The VSCode extension updates lag behind the CLI — newer features (e.g., CSV batch orchestration, Collab multi-agent) may require waiting for the extension to catch up

Claude Code example

First install:

# One-line script (recommended, auto-launches interactive menu after install)
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/hellowind777/helloagents/main/install.sh | bash

# Windows PowerShell
irm https://raw.githubusercontent.com/hellowind777/helloagents/main/install.ps1 | iex

# npx (or use npx -y to auto-download without prompting)
npx helloagents install claude

# UV
uv tool install --from git+https://github.com/hellowind777/helloagents helloagents && helloagents install claude

# pip
pip install git+https://github.com/hellowind777/helloagents.git && helloagents install claude

Update later (auto-syncs installed targets):

helloagents update

💡 Claude Code sub-agent orchestration tips:

• Sub-agents (Task tool) work out of the box, no extra configuration needed
• Agent Teams collaboration mode requires environment variable: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
• Parallel sub-agent count is managed automatically by the model, no user-side limit config needed

Configuration

Customize workflow behavior via config.json after installation. Only include keys you want to override — missing keys use defaults.

Storage locations (highest priority first):

1. Project-level: {project_root}/.helloagents/config.json — current project only
2. Global: ~/.helloagents/config.json — all projects
3. Built-in defaults

Available keys:

Key	Type	Default	Description
OUTPUT_LANGUAGE	string	zh-CN	Language for AI output and KB files
KB_CREATE_MODE	int	2	KB creation: 0=OFF, 1=on-demand (prompt ~init), 2=auto on code changes, 3=always auto
BILINGUAL_COMMIT	int	1	Commit language: 0=OUTPUT_LANGUAGE only, 1=OUTPUT_LANGUAGE + English
EVAL_MODE	int	1	Clarification mode: 1=progressive (1 question/round, max 5), 2=one-shot (all at once, max 3)
UPDATE_CHECK	int	72	Update check cache TTL in hours: 0=OFF
CSV_BATCH_MAX	int	16	CSV batch max concurrency: 0=OFF, cap 64 (Codex CLI only)

Example:

{
  "KB_CREATE_MODE": 0,
  "EVAL_MODE": 2
}

File missing or unparseable is silently skipped with defaults applied. Unknown keys produce a warning and are ignored.

How It Works

1. Install the package (script/pip/uv) and run helloagents to launch an interactive menu for selecting target CLIs (or specify directly with helloagents install <target>). Hooks and SKILL.md are auto-deployed during installation.
2. In AI chat, every input is scored on five dimensions and routed to R0–R3.
3. R2/R3 tasks enter the stage chain: EVALUATE → DESIGN → DEVELOP. R1 fast flow handles single-point operations directly.
4. RLM dispatches native sub-agents and specialized roles based on task complexity. Tasks with dependencies are scheduled via DAG topological sort with layer-by-layer parallel dispatch.
5. EHRB scans each step for destructive operations; risky actions require explicit user confirmation. Hooks provide additional pre-tool safety checks when available.
6. Three-layer memory (user / project KB / session) preserves context across sessions.
7. Stage chain completes with verified output and optional knowledge base sync.

In-Chat Workflow Commands

These commands run inside AI chat, not your system shell.

Command	Purpose
~auto	full autonomous workflow
~plan	planning and package generation
~exec	execute existing package
~init	initialize knowledge base
~upgradekb	upgrade knowledge structure
~clean / ~cleanplan	cleanup workflow artifacts
~test / ~review / ~validatekb	quality checks
~commit	generate commit message from context
~rollback	rollback workflow state
~rlm	role orchestration (spawn / agents / resume / team)
~status / ~help	status and help

Usage Guide

Three Workflow Modes

Mode	Description	When to use
~auto	Full autonomous flow from requirement to verified implementation (Evaluate → Design → Develop → Verify)	Clear requirement, want end-to-end delivery
~plan	Planning only, generates a proposal package then stops — no code written	Want to review the plan before committing
~exec	Skip evaluation and design, execute an existing plan package directly	After ~plan review, ready to implement

Typical pattern: ~plan first → review → ~exec to implement. Or just ~auto for one-shot delivery.

Interactive vs Delegated Mode

When ~auto or ~plan presents its confirmation, you choose:

• Interactive (default): pauses at key decision points (plan selection, failure handling)
• Delegated (fully automatic): auto-advances all stages, auto-selects recommended options, only pauses on EHRB risk
• Plan-only delegated: fully automatic but stops after design, never enters development

Without ~ commands, plain-text input is automatically routed to R0–R3 based on complexity.

Requirement Evaluation

Before R2/R3 tasks enter execution, the system scores requirements on four dimensions (scope 0–3, deliverable spec 0–3, implementation conditions 0–2, acceptance criteria 0–2, total 10). Score ≥ 8 proceeds to confirmation; < 8 triggers clarifying questions:

• EVAL_MODE=1 (default, progressive): asks 1 lowest-scoring dimension per round, up to 5 rounds
• EVAL_MODE=2 (one-shot): asks all low-scoring dimensions at once, up to 3 rounds

Context inferred from the existing codebase counts toward the score automatically. Say "skip evaluation / just do it" to bypass the questioning phase.

Parallel Design Proposals

In the R3 standard path, the design stage dispatches 3–6 sub-agents to independently generate competing implementation proposals. The main agent evaluates all proposals across four dimensions: user value, solution soundness, risk (including EHRB), and implementation cost. Weights are dynamically adjusted based on project characteristics (e.g., performance-critical systems weight soundness higher; MVPs weight cost higher).

• Interactive mode: user selects a proposal or requests re-generation (max 1 retry)
• Delegated mode: recommended proposal is auto-selected
• R2 simplified path skips multi-proposal comparison and goes directly to planning

Auto Dependency Management

During development, the system auto-detects the project's package manager via lockfiles (yarn.lock → yarn, uv.lock → uv, Gemfile.lock → bundler, etc.) and handles dependencies:

• Declared but missing dependencies: auto-installed
• New dependencies required by tasks: auto-added with declaration file updated
• Ambiguous dependencies: user is asked before installing

Quality Verification (Ralph Loop & Break-loop)

Ralph Loop (Claude Code, via SubagentStop Hook): after a sub-agent completes code changes, the project's verification command runs automatically. On failure, the sub-agent is blocked from exiting and must fix the issue (max 1 retry loop). Verification command priority: .helloagents/verify.yaml → package.json scripts → auto-detected.

Break-loop (deep root cause analysis): triggered when a task fails repeatedly (after Ralph Loop + at least 1 manual fix attempt), performing five-dimension root cause analysis:

1. Root cause classification (logic error / type mismatch / missing dependency / environment / design flaw)
2. Why previous fixes didn't work
3. Prevention mechanism suggestions
4. Systemic scan — same issue in other modules?
5. Lessons learned recorded in the acceptance report

Custom Command Extension

Create .helloagents/commands/ in your project and drop in Markdown files — the filename becomes the command name:

.helloagents/commands/deploy.md  →  ~deploy
.helloagents/commands/release.md →  ~release

File content defines the execution rules. The system applies a lightweight gate (requirement understanding + EHRB check).

Smart Commit (~commit)

~commit does more than generate a message:

• Analyzes git diff to auto-generate Conventional Commits formatted messages
• Pre-commit quality checks (code-doc consistency, test coverage, verification commands)
• Auto-excludes sensitive files (.env, *.pem, *.key, etc.) — never runs git add .
• Shows file list before staging, supports exclusion
• Options: local commit only / commit + push / commit + push + create PR
• Bilingual commit messages when BILINGUAL_COMMIT=1

Manual Sub-Agent Invocation

Beyond automatic dispatch, you can manually invoke specific roles:

~rlm spawn reviewer "review src/api/ for security issues"
~rlm spawn writer "generate API reference docs"
~rlm spawn reviewer,synthesizer "analyze and summarize the auth module"  # parallel

Available roles: reviewer (code review), synthesizer (multi-source synthesis), kb_keeper (KB maintenance), pkg_keeper (plan package management), writer (documentation).

Multi-Terminal Collaboration

Multiple terminals (across different CLIs) can share a task list:

# Terminal A
hellotasks=my-project codex

# Terminal B
hellotasks=my-project claude

Commands once enabled:

~rlm tasks                  # view shared task list
~rlm tasks available        # see unclaimed tasks
~rlm tasks claim <id>       # claim a task
~rlm tasks complete <id>    # mark done
~rlm tasks add "task title" # add a new task

Tasks are stored in {KB_ROOT}/tasks/ with file locking to prevent concurrent conflicts.

KB Auto-Sync & CHANGELOG

The knowledge base syncs automatically at these points:

• After every development stage, kb_keeper sub-agent syncs module docs to reflect actual code
• After every R1/R2/R3 task completion, CHANGELOG is auto-appended
• On session end (Claude Code Stop Hook), KB sync + L2 session summary write triggered asynchronously

CHANGELOG uses semantic versioning (X.Y.Z). Version source priority: user-specified → project file (package.json, pyproject.toml, etc., supporting 15+ languages/frameworks) → git tag → last CHANGELOG entry → 0.1.0. R1 fast-path changes are recorded under a "Quick Modifications" category with file:line range.

KB_CREATE_MODE controls automatic behavior: 0=off, 1=prompt on demand, 2=auto on code changes (default), 3=always auto.

Worktree Isolation

When multiple sub-agents need to modify different regions of the same file simultaneously (Claude Code only), the system automatically uses Task(isolation="worktree") to create an independent git worktree for each sub-agent, preventing Edit tool conflicts. The main agent merges all worktree changes in the consolidation phase. Only activated when sub-agents have overlapping file writes; read-only tasks don't use it.

CSV Batch Orchestration (Codex CLI)

When ≥6 structurally identical tasks exist in the same execution layer, the system auto-converts tasks.md into a task CSV and dispatches via spawn_agents_on_csv. Each worker receives its row data + instruction template, executes independently, and reports results.

• Progress tracked in real-time via agent_job_progress events (pending/running/completed/failed/ETA)
• State persisted in SQLite for crash recovery
• Partial failures still export results with failure summary
• Heterogeneous tasks automatically fall back to spawn_agent sequential dispatch
• Configure concurrency via CSV_BATCH_MAX (default 16, max 64, set to 0 to disable)

Update Check

On the first response of each session, the system silently checks for new versions. Results are cached at ~/.helloagents/.update_cache, valid for the duration set by UPDATE_CHECK (default 72 hours, set to 0 to disable). When a new version is available, ⬆️ New version {version} available appears in the response footer. Any errors during the check are silently skipped and never interrupt normal usage.

Repository Guide

• AGENTS.md: router and workflow protocol
• SKILL.md: skill discovery metadata for CLI targets
• pyproject.toml: package metadata (v2.3.3)
• helloagents/cli.py: installer entry
• helloagents/functions: workflow commands (15)
• helloagents/stages: design, develop
• helloagents/services: knowledge, package, memory and support services
• helloagents/rules: state, cache, tools, scaling, evaluation, sub-agent protocols
• helloagents/rlm: role library and orchestration helpers
• helloagents/hooks: Claude Code and Codex CLI hooks configs
• helloagents/scripts: automation scripts
• helloagents/templates: KB and plan templates

FAQ

• Q: Is this a Python CLI tool or prompt package? A: Both. CLI manages installation; workflow behavior comes from AGENTS.md and helloagents docs.

• Q: Which target should I install? A: Use the CLI you run: codex, claude, gemini, qwen, grok, or opencode.

• Q: What if a rules file already exists? A: Non-HelloAGENTS files are backed up before replacement.

• Q: What is RLM? A: Role Language Model — a sub-agent orchestration system with 5 specialized roles + native CLI sub-agents, DAG-based parallel scheduling, and standardized prompt/return format.

• Q: Where does project knowledge go? A: In the project-local .helloagents/ directory, auto-synced when code changes.

• Q: Does memory persist across sessions? A: Yes. L0 user memory is global, L1 project KB is per-project, L2 session summaries are auto-saved at stage transitions.

• Q: What are Hooks? A: Lifecycle hooks auto-deployed during installation. Claude Code gets 9 event hooks (safety checks, progress snapshots, KB sync, etc.); Codex CLI gets a notify hook for update checks. All optional — features degrade gracefully without hooks.

• Q: What is Agent Teams? A: An experimental Claude Code multi-agent collaboration mode. Multiple Claude Code instances work as teammates with shared task lists and mailbox communication, mapped to RLM roles. Falls back to standard Task sub-agents when unavailable.

Troubleshooting

• command not found: ensure install path is in PATH
• package version unknown: install package first for metadata
• target not detected: launch target CLI once to create config directory
• custom rules overwritten: restore from timestamped backup in CLI config dir
• images not rendering: keep relative paths and commit readme_images files

Version History

v2.3.3 (current)

• AGENTS.md slimmed from 1186 to ~700 lines (-41%), reducing risk of LLMs selectively ignoring rules
• G4 evaluation details, G9/G10 sub-agent protocols extracted to on-demand module files via G7
• Sub-agent rule file (subagent.md) reduced from 26KB to 1.4KB (-94%)
• Eliminated cross-module reference risks by inlining critical rules at usage sites
• inject_context.py gains module-load signals to ensure rule recovery after context compaction

v2.3.0

• Comprehensive cross-audit fix: unified role output format, normalized path references, code-doc consistency alignment
• Quality verification loop (Ralph Loop): auto-verify after sub-agent completion, block and feedback on failure
• Auto context injection for sub-agents and rule reinforcement for main agent
• Deep 5-dimension root cause analysis on repeated failures (break-loop)
• Auto-inject project technical guidelines before sub-agent development
• Pre-commit quality checks (code-doc consistency, test coverage, verification commands)
• Worktree isolation for parallel editing
• Custom command extension (.helloagents/commands/)
• Auto-append Git author info to CHANGELOG entries

v2.2.16

• Refactored evaluation dimension system with dimension isolation rule, pass threshold tuned to 8/10. Options are user-need-driven, organized by style direction rather than complexity tiers (e.g. UI design offers different styles instead of simple/medium/complex); recommended option points to the most complete deliverable, derived from recommendation principles and scoring criteria rather than hardcoded
• Proposal design requires both implementation path and deliverable design direction to differ across alternatives, each sub-agent independently outputs a complete proposal including presentation direction, style, and experience
• Proposal evaluation criteria optimized: user value weight is always no less than any other single dimension, evaluation dimensions dynamically adjusted by project context
• Universal task type support: generalized evaluation, follow-up, and proposal design terminology from programming-specific to documents, design, general tasks and more
• Added sub-agent DAG dependency scheduling with topological sort, layer-by-layer parallel dispatch, and failure propagation
• Dynamic sub-agent parallel count based on independent work units, eliminated hardcoded limits
• Unified output format: structured display for score breakdowns, follow-up options, and confirmation messages
• Streamlined execution paths with shorter stage chains and step-level on-demand module loading
• Adjusted Codex CLI memory limit to 128 KiB to prevent rules file truncation
• Improved recommendation option generation rules, proposal differentiation requirements, and evaluation scoring criteria

Contributing

See CONTRIBUTING.md for contribution rules and PR checklist.

License

This project is dual-licensed: Code under Apache-2.0, Documentation under CC BY 4.0. See LICENSE.md.

---

If this project helps your workflow, a star is always appreciated.

Thanks to codexzh.com / ccodezh.com for supporting this project