@hivehub/rulebook 5.5.0 → 5.5.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hivehub/rulebook",
-  "version": "5.5.0",
+  "version": "5.5.1",
   "description": "Tool-agnostic AI development framework. Standardize projects across Claude Code, Cursor, Gemini, Codex, Windsurf, Copilot with automated templates, quality gates, persistent memory, and framework detection for 28 languages, 17 frameworks, 13 MCP modules, and 20 services",
   "type": "module",
   "main": "dist/index.js",

package/templates/core/AGENTS_LEAN.md

@@ -194,7 +194,23 @@ If this is part of a multi-project workspace, every Rulebook MCP call MUST pass
 | Build/CI | build-engineer | sonnet |
 | Security | security-reviewer | haiku |
 
-Rules: never write code directly in main conversation — delegate. Use haiku for read-only tasks. Launch independent agents in parallel.
+### Mandatory delegation rules
+
+- **Never implement directly in the main conversation when an agent fits.** The main thread orchestrates; agents do the work.
+- **Read-only work → haiku researcher.** Cheap, isolates context.
+- **Independent work runs in parallel.** Multiple `Agent` tool-use blocks in a single message. Sequential dispatching of independent units is wrong by default.
+- **Multi-specialist work uses Teams.** Background `Agent` calls without `team_name` are blocked by the enforcement hook (see Multi-Agent Teams above).
+- **Foreground agents** when the result blocks your next step. **Background agents** only inside Teams so `SendMessage` works.
+
+### When to create a new skill or agent
+
+Lift to a skill / agent instead of repeating instructions:
+
+- **Same multi-step prompt twice in one session** → make a skill (`templates/skills/<category>/<name>/SKILL.md`).
+- **Recurring class of work across projects** → make an agent definition (`.claude/agents/<role>.md`) and add a row to the delegation table above.
+- **A workflow needs a specific persona / tool set** → agent. **A behavior modifier the user invokes on demand** → skill.
+
+Default to creating, not improvising. The template scaffolding is cheap; ad-hoc context-window churn is expensive.
 
 ## Plans & Session Continuity
 

package/templates/core/CLAUDE_MD_v2.md

@@ -31,6 +31,16 @@ automatically at session start (see [Anthropic memory docs](https://code.claude.
 7. **Capture learnings**: at the end of significant work, save patterns and anti-patterns to `.rulebook/knowledge/` and insights to `.rulebook/learnings/`.
 8. **Never archive a task** without docs updated, tests written, and tests passing — the task tail enforces this structurally.
 
+## Delegation & parallelism (highest precedence — apply on every turn)
+
+**Default behavior: delegate, don't do it yourself. Parallelize, don't serialize. Create new agents/skills when the gap is real.**
+
+1. **Delegate by default.** If a step matches an agent in the delegation table, dispatch it via `Agent` instead of doing it inline. Implementation → `implementer` (sonnet). Research / read-only exploration → `researcher` (haiku). Tests → `tester`. Docs → `docs-writer` (haiku). Architecture / cross-cutting → `architect` (opus). Reserve the main conversation for orchestration + decisions.
+2. **Parallelize independent work.** When a turn requires multiple independent investigations or edits, dispatch every independent piece in **a single message with multiple `Agent` tool-use blocks**. Sequential `Agent` calls are a smell — every time you catch yourself writing "first X, then Y", check whether the two halves are independent.
+3. **Use Teams for multi-specialist work.** Anything that needs ≥2 background agents to coordinate MUST go through a Team (`TeamCreate` + `team_name` on dispatch). Standalone background `Agent` calls without `team_name` are blocked by the enforcement hook.
+4. **Create skills + agents when the gap is real.** If you write the same multi-step instructions twice in one session, lift it into a skill (`templates/skills/<category>/<name>/SKILL.md`). If a class of work repeats across projects, create an agent definition under `.claude/agents/`. Default to creating, not improvising.
+5. **Foreground vs background.** Use foreground `Agent` when you need the result to inform your next step. Use background only with `team_name` so messages can flow.
+
 ## Editing discipline (Karpathy-inspired)
 
 Behavioral guidelines that reduce common LLM coding mistakes. Adapted from [forrestchang/andrej-karpathy-skills](https://github.com/forrestchang/andrej-karpathy-skills), grounded in [Andrej Karpathy's observations](https://x.com/karpathy/status/2015883857489522876).