v6.0.0

Scope reduction (breaking). Driven by a cross-project usage analysis of 8
projects with .rulebook/: features with near-zero real adoption are removed so
the toolkit ships only what people actually use. Delivered in phases; this
section is appended as each phase lands.

Removed — over-provisioned agents & niche MCP modules

• Pruned the agent template library from 21 flat agents + 5 project-type subdirs
  to the 11 actually used (delegation core + performance-engineer,
  quality-gatekeeper, team-lead). The 10 never-dispatched role agents and
  their delegation-table entries were removed, along with the orphaned
  compiler/game-engine/generic/mobile/web-app agent subdirs (the
  installer only ever copied the flat *.md).
• Removed the niche, never-detected MCP module docs/skills figma, grafana,
  notion, atlassian (and their ModuleId/init-prompt entries). The
  remaining modules (vectorizer, synap, context7, github, playwright, supabase,
  serena, rulebook-mcp, sequential-thinking) and the language/dev/core skill
  library are retained — they are detected and heavily used.

Removed — persistent-memory & indexer subsystem

• Removed the persistent-memory and background-indexer subsystem. Cross-project
  usage showed it was effectively dev-repo-only, and it carried an always-on
  indexing process plus 9 of the 35 MCP tools. Deleted src/memory/*
  (MemoryManager, file store/search, codegraph, legacy migrator),
  src/core/indexer/* (BackgroundIndexer, file parser), the rulebook memory
  CLI commands, and the memory/indexer config blocks. Removed the MCP tools
  memory_save/search/get/timeline/stats/cleanup, codebase_search,
  codebase_graph, indexer_status, and the memory-backed
  rulebook_workspace_search — the MCP surface drops 35 → 25.
  session_start/session_end keep working on PLANS.md (their optional
  memory branches were removed). The lightweight, file-based knowledge /
  learnings / decisions features are retained. Existing .rulebook/memory/
  data on disk is left untouched.

Removed — all IDE/CLI adapters (tool-agnostic via AGENTS.md)

• Rulebook no longer ships per-tool adapters. It generates only the
  tool-agnostic AGENTS.md (+ AGENTS.override.md, CLAUDE.md, and .claude/
  for Claude Code) — every other agent reads AGENTS.md natively, so coverage
  is unchanged with far less surface. Removed:
  • OpenCode integration (generator + templates).
  • multi-tool generator (Gemini/Windsurf/Copilot/Continue config files) +
    templates/ides/.
  • IDE/CLI rule generators in workflow-generator.ts (cursor/windsurf/
    vscode/copilot rules, cursor commands, CODEX/GEMINI CLI files) and the
    non-Claude templates/cli/*.
  • IDE detection (detectCursor/GeminiCli/ContinueDev/Windsurf/ GithubCopilot/Opencode) and the detection types.
  • ProjectConfig.ides and the "Select IDEs/tools" init prompt.
  • rule-engine IDE projection (projectToCursor/Gemini/Copilot/Windsurf/ ContinueDev); projectRules now targets Claude Code only.
  • The IDE/CLI skill matrix: templates/skills/ides/* and the non-Claude
    templates/skills/cli/*.
• rulebook agent removed: the external-CLI runner (cli-bridge.ts + the
  cursor/gemini/claude stream parsers in src/agents/) was orphaned dead code
  with no CLI command wiring it up.

Removed — Claude Code plugin distribution

• Dropped the Claude Code plugin/marketplace distribution. Rulebook is used via
  its MCP server, not as a Claude plugin, so the plugin payload and installer
  were dead weight. Removed the root commands/ and skills/ plugin payload,
  .claude-plugin/ (plugin.json + marketplace.json), the root marketplace.json
  duplicate, the setup:plugin CLI command and its setupClaudeCodePlugin
  installer (which wrote into ~/.claude/plugins), and the plugin test suite.
  package.json files now ships only dist + templates.

Build

• Pinned @modelcontextprotocol/sdk to 1.22.0. SDK ≥1.26 (the first
  security-patched line) regressed registerTool's generic types into a
  TS2589 "type instantiation excessively deep" error in the per-domain tool
  modules. The advisory the newer line fixes (GHSA-345p-7cg4-v4c7 — cross-client
  data leak via a shared server/transport instance) does not apply to rulebook's
  per-process stdio MCP server (one client per process). Tracked for an upgrade
  once the SDK type regression is resolved.

Internal — readability

• Rewrote the README for the 6.0.0 direction: removed the multi-IDE matrix, the
  "23 AI tools" list, persistent-memory, terse-mode and VSCode-extension
  sections, and trimmed the hook/MCP tables. The framing is now tool-agnostic
  via AGENTS.md + Claude Code.
• Reformatted the codebase to 4-space indentation (prettier tabWidth: 4).
• Split the 2400-line src/mcp/rulebook-server.ts into per-domain tool modules
  under src/mcp/tools/* (task, skill, workspace, decision, knowledge, learn,
  rules), each exporting registerXxxTools(server, ctx). No file exceeds 1500
  lines. Pure structural refactor — no behavior change.

Removed — dead feature flags (phase 1)

• Removed six features flags that were OFF in every sampled project and gated
  no behavior: watcher, agent, notifications, dryRun, repl, plugins.
  The RulebookFeatures type now exposes only logging, gitHooks,
  templates, context, health, parallel, smartContinue. Legacy
  rulebook.json files carrying the removed keys are normalized (the keys are
  dropped) on load/migrate — no error.

Changed — Ralph cleanup (phase 1)

• The legacy-Ralph purge that rulebook update runs is now an exported,
  unit-tested function (purgeLegacyRalphArtifacts). The Ralph subsystem was
  already removed; this finishes the cleanup and guards it with a regression
  test. Archived task records under .rulebook/archive/ are retained as project
  history.