helloagents 2.3.5-beta.1 → 2.3.7-beta.1

package/README.md

@@ -8,7 +8,7 @@
 
 **Let AI go beyond analysis — keep pushing until implementation and verification are done.**
 
-[![Version](https://img.shields.io/badge/version-2.3.5-orange.svg)](./pyproject.toml)
+[![Version](https://img.shields.io/badge/version-2.3.7-orange.svg)](./pyproject.toml)
 [![npm](https://img.shields.io/npm/v/helloagents.svg)](https://www.npmjs.com/package/helloagents)
 [![Python](https://img.shields.io/badge/python-%3E%3D3.10-3776AB.svg)](./pyproject.toml)
 [![Commands](https://img.shields.io/badge/commands-15-6366f1.svg)](./helloagents/functions)
@@ -83,7 +83,7 @@
 
 **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.
+3 specialized roles (reviewer / writer / brainstormer) plus host CLI native sub-agents (explore / code / brainstorm) 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.
 </td>
@@ -110,9 +110,9 @@ Keyword scan, semantic analysis, and tool-output inspection catch destructive op
 <td width="50%" valign="top">
 <img src="./readme_images/05-feature-icon-compat.svg" width="48" align="left">
 
-**Three-Layer Memory Model**
+**Two-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).
+L0 user memory (global preferences) and L1 project knowledge base (structured docs synced from code).
 
 **Your gain:** context survives across sessions and projects.
 </td>
@@ -132,7 +132,7 @@ Voice notifications (5 event sounds), custom command extension (`.helloagents/co
 
 **Multi-CLI Support**
 
-One rule set works across Claude Code, Codex CLI, OpenCode, Gemini CLI, Qwen CLI, and Grok CLI. Automatic feature detection and graceful degradation ensure consistent experience regardless of CLI capabilities.
+One rule set works across Claude Code, Codex CLI, OpenCode, Gemini CLI, Qwen CLI, and Grok CLI (Experimental/Community). Automatic feature detection and graceful degradation ensure consistent experience regardless of CLI capabilities.
 
 **Your gain:** switch between CLIs without relearning workflows or reconfiguring rules.
 </td>
@@ -143,12 +143,12 @@ One rule set works across Claude Code, Codex CLI, OpenCode, Gemini CLI, Qwen CLI
 
 | CLI | Native Sub-Agent Mechanism | RLM Mapping |
 |-----|---------------------------|-------------|
-| Claude Code | Task tool (explore / code / shell) | Direct mapping, supports Agent Teams |
+| Claude Code | Agent 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 |
+| OpenCode | Task tool (build / plan / general / explore) | Direct sub-agent mapping |
 | 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 |
+| Grok CLI (Experimental) | 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 11 lifecycle hooks + Codex CLI notify hook + Gemini/Grok CLI hooks) with automatic graceful degradation when unavailable.
 
@@ -158,10 +158,10 @@ Additionally, HelloAGENTS provides: **five-dimension routing scoring** (action n
 |-----|-------------------|--------------|---------------------|
 | **Claude Code** | Latest | Agent Teams, 11 lifecycle hooks, auto-memory | Requires `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` for teams mode |
 | **Codex CLI** | 0.110+ | spawn_agent, CSV batch, collaboration_modes | Enable sub-agents, CSV orchestration, set `project_doc_max_bytes >= 131072` |
-| **OpenCode** | Latest | Built-in agent mode | Sequential execution fallback |
+| **OpenCode** | Latest | Task tool, custom agents, MCP | Supports primary agents (build/plan) + subagents (general/explore) |
 | **Gemini CLI** | Latest | Built-in tool calls | Sequential execution fallback |
 | **Qwen CLI** | Latest | Built-in tool calls | Sequential execution fallback |
-| **Grok CLI** | Latest | Built-in tool calls | Sequential execution fallback |
+| **Grok CLI** (Experimental) | Latest | Built-in tool calls | Community wrapper, hooks not fully verified |
 
 <details>
 <summary>📋 Detailed CLI-specific notes (click to expand)</summary>
@@ -180,7 +180,8 @@ Additionally, HelloAGENTS provides: **five-dimension routing scoring** (action n
 - Auto-memory feature enabled by default
 
 **Other CLIs:**
-- OpenCode/Gemini/Qwen/Grok use sequential execution fallback
+- OpenCode supports Task tool with primary agents (build/plan) and subagents (general/explore)
+- Gemini/Qwen/Grok use sequential execution fallback
 - All features work with graceful degradation
 - Hooks may not be available on all platforms
 
@@ -354,7 +355,7 @@ Additionally, HelloAGENTS provides: **five-dimension routing scoring** (action n
     helloagents update
 
 > 💡 **Claude Code sub-agent orchestration tips:**
-> - Sub-agents (Task tool) work out of the box, no extra configuration needed
+> - Sub-agents (Agent 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
 
@@ -579,9 +580,9 @@ 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
+    ~rlm spawn reviewer,writer "analyze and document the auth module"  # parallel
 
-Available roles: `reviewer` (code review), `synthesizer` (multi-source synthesis), `kb_keeper` (KB maintenance), `pkg_keeper` (plan package management), `writer` (documentation).
+Available roles: `reviewer` (code review), `writer` (documentation), `brainstormer` (multi-proposal comparison).
 
 ### Multi-Terminal Collaboration
 
@@ -607,9 +608,9 @@ Tasks are stored in `{KB_ROOT}/tasks/` with file locking to prevent concurrent c
 
 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 development stage, main 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
+- On session end (Claude Code Stop Hook), KB sync flag set 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.
 
@@ -637,7 +638,7 @@ On the first response of each session, the system silently checks for new versio
 
 - AGENTS.md: router and workflow protocol
 - SKILL.md: skill discovery metadata for CLI targets
-- pyproject.toml: package metadata (v2.3.5)
+- pyproject.toml: package metadata (v2.3.7)
 - helloagents/cli.py: CLI entry point
 - helloagents/_common.py: shared constants and utilities
 - helloagents/core/: CLI management modules (install, uninstall, update, status, dispatcher, hooks settings)
@@ -648,8 +649,8 @@ On the first response of each session, the system silently checks for new versio
 - helloagents/rlm: role library and orchestration helpers
 - helloagents/hooks: Claude Code, Codex CLI, Gemini CLI, and Grok CLI hooks configs
 - helloagents/scripts: automation scripts (sound notify, progress snapshot, safety guard, etc.)
-- helloagents/agents: sub-agent definitions (5 RLM roles)
-- helloagents/commands: custom command templates
+- helloagents/agents: sub-agent definitions (3 RLM roles)
+- helloagents/user/commands: custom command templates
 - helloagents/assets: sound resources (5 event sounds)
 - helloagents/templates: KB and plan templates
 
@@ -669,7 +670,7 @@ A: HelloAGENTS automatically backs up non-HelloAGENTS files before replacement.
 
 **Q: What is RLM?**
 
-A: Role Language Model — HelloAGENTS's sub-agent orchestration system. It includes 5 specialized roles (reviewer, synthesizer, kb_keeper, pkg_keeper, writer) plus native CLI sub-agents. Tasks are scheduled via DAG dependency analysis with parallel execution when possible. Learn more in [Usage Guide](#usage-guide).
+A: Role Language Model — HelloAGENTS's sub-agent orchestration system. It includes 3 specialized roles (reviewer, writer, brainstormer) plus native CLI sub-agents. Tasks are scheduled via DAG dependency analysis with parallel execution when possible. Learn more in [Usage Guide](#usage-guide).
 
 **Q: Where does project knowledge go?**
 
@@ -677,7 +678,7 @@ A: In the project-local `.helloagents/` directory. The knowledge base auto-syncs
 
 **Q: Does memory persist across sessions?**
 
-A: Yes, through three layers: L0 user memory (global preferences in `~/.helloagents/`), L1 project KB (per-project in `.helloagents/`), and L2 session summaries (auto-saved at stage transitions). Context survives even if you close and reopen your CLI.
+A: Yes, through two layers: L0 user memory (global preferences in `~/.helloagents/`) and L1 project KB (per-project in `.helloagents/`). Context survives even if you close and reopen your CLI.
 
 **Q: What are Hooks?**
 
@@ -769,7 +770,7 @@ A: An experimental Claude Code feature where multiple Claude Code instances coll
     helloagents install claude    # reinstall to specific CLI target
     helloagents update            # update + auto-sync all installed targets
 
-**Prevention:** v2.3.5 adds automatic configuration integrity check on session start — if HelloAGENTS config is missing or corrupted, a warning is displayed with recovery instructions
+**Prevention:** Since v2.3.5, automatic configuration integrity check on session start — if HelloAGENTS config is missing or corrupted, a warning is displayed with recovery instructions
 
 **Verification:** Run `helloagents status` to confirm all targets show as installed
 
@@ -787,7 +788,35 @@ A: An experimental Claude Code feature where multiple Claude Code instances coll
 
 ## Version History
 
-### v2.3.5 (current)
+### v2.3.7 (current)
+
+**Bug Fixes:**
+- Fixed non-coding tasks incorrectly creating knowledge base when KB_CREATE_MODE=2 (added programming task check in design.md Phase1 step 1)
+- Fixed R2 standard flow redirecting to archive instead of DEVELOP after proposal selection (constrained overview type to ~exec entry only)
+- Fixed non-coding tasks incorrectly creating plan packages (added programming task precondition to package.md create() interface)
+
+**Improvements:**
+- Optimized implementation plan state recovery after context compression
+- Optimized overall design flow
+
+### v2.3.6
+
+**New Features:**
+- Sub-agent orchestration overhaul: added brainstormer sub-agent for independent parallel proposal ideation during DESIGN multi-proposal comparison
+- Sub-agent blocking mechanism: auto-block and fallback to main agent on sub-agent failure or timeout
+
+**Improvements:**
+- Tool/Shell constraint optimization: allow fallback to Shell when built-in tools fail (fixes Codex CLI Windows apply_patch repeated failures)
+- Shell encoding constraint refinement: explicit UTF-8 no-BOM requirement, separate read/write encoding handling for PowerShell
+- Added batch file write rule (merge ≥3 files into a single temp script to avoid sandbox per-command blocking)
+- Removed session memory features that couldn't be fully implemented in CLI (session_summary template, SessionEnd memory sync), streamlined service layer
+- Sub-agent consolidation: removed 3 redundant sub-agents (kb-keeper, pkg-keeper, synthesizer), functionality returned to main agent and RLM roles
+- Sub-agent voice notification skip, task stability fixes
+- Unified user config directory structure (user/memory/, user/commands/, user/sounds/)
+- Uninstall script enhancements
+- Visual verification gap and UI quality fixes
+
+### v2.3.5
 
 **New Features:**
 - Voice notification system with 5 event sounds (complete, idle, confirm, error, warning) across Windows/macOS/Linux, with smart two-layer routing (stop_reason + G3 format icon detection)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "helloagents",
-  "version": "2.3.5-beta.1",
+  "version": "2.3.7-beta.1",
   "type": "module",
   "description": "HelloAGENTS - AI-native sub-agent orchestration framework for multi-CLI environments",
   "author": "HelloWind",