feat: metacognitive-guard v0.5.0 + design-studio plugin (#154)

* feat: add design-studio plugin (merges frontend-design + ui-ux-pro-max)

Combines Anthropic's frontend-design creative philosophy with ui-ux-pro-max's
data-driven search engine into a single unified plugin. Single workflow:
creative direction -> design system generation -> domain/stack searches ->
implementation with aesthetic precision.

- 50 styles, 97 palettes, 57 font pairings, 99 UX guidelines, 25 chart types
- 13 framework stacks (React, Next.js, Vue, Svelte, Flutter, etc.)
- BM25 search engine with design system generator (Python stdlib-only)
- Pre-delivery checklist with creative distinctiveness check
- Master + Overrides persistence pattern for multi-session projects

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat(metacognitive-guard): agent_type filtering, InstructionsLoaded hook, git instructions opt-out

- Add agent_type=subagent early exit to struggle-detector.sh and ralph-loop.sh
- Add InstructionsLoaded hook firing truth-beacon after CLAUDE.md/rules load
- Add includeGitInstructions: false in .claude/settings.json (saves ~2K tokens/turn)
- Update truth-beacon with new Claude Code conventions (CLAUDE_SKILL_DIR, /reload-plugins, agent_type)
- Soften CLAUDE.md version bump docs to mention /reload-plugins for dev workflows
- Bump metacognitive-guard 0.4.5 -> 0.5.0

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Author: ANcpLua

.claude/settings.json

@@ -0,0 +1,3 @@
+{
+  "includeGitInstructions": false
+}

CHANGELOG.md

@@ -8,6 +8,8 @@ and the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 ### Added
 
+- **metacognitive-guard `InstructionsLoaded` hook**: Truth beacon now fires on both SessionStart AND InstructionsLoaded — ground truth re-injected when CLAUDE.md/rules are loaded, ensuring authoritative facts arrive after instructions context
+- **metacognitive-guard `agent_type` filtering**: Struggle detector and Ralph Loop now skip subagents via `agent_type` field in hook events — prevents wasted haiku calls and false positives from subagent responses
 - **`design-studio` plugin (1.0.0)**: Design intelligence studio merging creative direction with data-driven recommendations. Combines `frontend-design` (Anthropic plugin — bold aesthetic philosophy, anti-generic guidelines) with `ui-ux-pro-max` (local skill — BM25 search engine, 750+ CSV rows, design system generator). Single unified workflow: creative direction → `--design-system` CLI → domain/stack searches → implementation with aesthetic precision. 50 styles, 97 palettes, 57 font pairings, 99 UX guidelines, 25 chart types, 13 stacks. Pre-delivery checklist includes creative distinctiveness check. Python stdlib-only, no pip dependencies
 
 ### Added
@@ -17,6 +19,9 @@ and the project follows [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 ### Changed
 
+- **`includeGitInstructions: false`**: Disabled built-in git instructions in `.claude/settings.json` — we have our own commit/PR workflows in CLAUDE.md, saves ~2K tokens/turn
+- **CLAUDE.md version bump docs**: Softened "DO NOT SKIP" to acknowledge `/reload-plugins` for dev workflows while keeping version bumps mandatory for releases
+- **metacognitive-guard (0.4.5 → 0.5.0)**: InstructionsLoaded hook + agent_type filtering
 - **`claude-self-obs` (2.0.0 → 3.0.0)**: Deleted standalone TypeScript MCP server — qyl.mcp already provides `qyl.claude_code_sessions`, `qyl.claude_code_timeline`, `qyl.claude_code_tools`. Plugin is now 38 lines of declarative JSON: HTTP hooks POST raw event JSON to `qyl.collector:5100/api/v1/claude-code/hooks`. Zero TypeScript, zero npm, zero processes to manage
 - **`claude-self-obs` (1.0.0 → 2.0.0)**: Complete rewrite from bash/jq/curl hooks to HTTP hooks + MCP server. Deleted 210 lines of shell scripts (emit-span.sh, emit-agent-start.sh, emit-agent-stop.sh). Replaced with `type: "http"` hooks that POST directly to a dual-mode MCP server (stdio for Claude tools + HTTP for hook events). Claude can now query its own telemetry via 4 MCP tools: `get_status`, `get_session_timeline`, `get_tool_stats`, `search_events`. In-memory ring buffer stores last 10k events per session
 - **`marketplace.json`**: Added claude-self-obs to registry, updated plugin count (9 → 10) and command count (24 → 25)

CLAUDE.md

@@ -463,15 +463,16 @@ For EVERY non-trivial change:
 
 **NO EXCEPTIONS.** Forgetting to update CHANGELOG = incomplete task.
 
-### ⚠️ PLUGIN VERSION BUMP - DO NOT SKIP
+### ⚠️ PLUGIN VERSION BUMP
 
 **When you modify a plugin's code (hooks, commands, skills, agents), you MUST bump its version:**
 
 1. Open `plugins/<name>/.claude-plugin/plugin.json`
 2. Increment the `version` field (patch for fixes, minor for features)
 3. The plugin cache (`~/.claude/plugins/cache/`) is version-keyed — without a bump, new sessions won't pick up changes
 
-**NO EXCEPTIONS.** Changed plugin code without version bump = users stuck on stale cache.
+**During development:** Use `/reload-plugins` to activate changes in the current session without bumping.
+**For releases:** Version bump is mandatory — other users' sessions won't pick up changes without it.
 
 </EXTREMELY_IMPORTANT>
 

plugins/metacognitive-guard/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "metacognitive-guard",
   "description": "Cognitive amplification stack: epistemic hooks, competitive review, commit integrity checks, CI verification scripts, and deep-thinking agents.",
-  "version": "0.4.5",
+  "version": "0.5.0",
   "author": {
     "name": "AncpLua",
     "url": "https://github.com/ANcpLua"

plugins/metacognitive-guard/CLAUDE.md

@@ -7,7 +7,7 @@ agents amplify thinking. Absorbs completion-integrity and autonomous-ci.
 
 | Hook | Event | Script | Purpose |
 |------|-------|--------|---------|
-| Truth Beacon | SessionStart | `truth-beacon.sh` | Injects `blackboard/assertions.yaml` as authoritative facts |
+| Truth Beacon | SessionStart + InstructionsLoaded | `truth-beacon.sh` | Injects `blackboard/assertions.yaml` as authoritative facts |
 | Epistemic Guard | PreToolUse (Write/Edit) | `epistemic-guard.sh` | Blocks writes with wrong versions, banned APIs, AGENTS.md in plugins |
 | Commit Integrity | PreToolUse (Bash) | `commit-integrity-hook.sh` | Blocks `git commit` with suppressions, commented tests, deleted assertions |
 | Struggle Detector | Stop (async) | `struggle-detector.sh` | Scores response for uncertainty, writes to blackboard |
@@ -56,6 +56,7 @@ agents amplify thinking. Absorbs completion-integrity and autonomous-ci.
 - Struggle detector is a two-part system: Stop hook (async) does analysis + blackboard writes,
   UserPromptSubmit hook reads blackboard and injects `additionalContext` so Claude actually sees the
   suggestion. No latency on responses.
+- Struggle detector and Ralph Loop skip subagents via `agent_type` filtering (only lead agent matters).
 - TaskCompleted prompt hook fires on every task completion in team contexts (haiku, 15s timeout).
 - Ralph Loop fires PostToolUse on Write/Edit — two layers run in parallel:
   (1) Haiku prompt analyzes context for deep drift (over-engineering, complexity creep, premature

plugins/metacognitive-guard/hooks/hooks.json

@@ -12,6 +12,18 @@
         ]
       }
     ],
+    "InstructionsLoaded": [
+      {
+        "hooks": [
+          {
+            "type": "command",
+            "command": "bash ${CLAUDE_PLUGIN_ROOT}/hooks/scripts/truth-beacon.sh",
+            "timeout": 5,
+            "statusMessage": "Re-injecting ground truth after instructions loaded..."
+          }
+        ]
+      }
+    ],
     "PreToolUse": [
       {
         "matcher": "Write|Edit",

plugins/metacognitive-guard/hooks/scripts/ralph-loop.sh

@@ -20,6 +20,10 @@ INPUT=$(cat 2>/dev/null || true)
 # Requires jq for JSON parsing
 command -v jq &>/dev/null || exit 0
 
+# Skip subagents — drift detection only matters for the lead agent
+AGENT_TYPE=$(echo "$INPUT" | jq -r '.agent_type // empty' 2>/dev/null)
+[[ "$AGENT_TYPE" == "subagent" ]] && exit 0
+
 # Extract both fields in a single jq call
 eval "$(echo "$INPUT" | jq -r '
   "FILE_PATH=" + (.tool_input.file_path // .tool_input.file // "" | @sh) +

plugins/metacognitive-guard/hooks/scripts/struggle-detector.sh

@@ -36,6 +36,10 @@ mkdir -p "$BLACKBOARD" 2>/dev/null || true
 INPUT=$(cat 2>/dev/null || true)
 [[ -z "$INPUT" ]] && exit 0
 
+# Skip subagents — only analyze the lead agent's responses
+AGENT_TYPE=$(echo "$INPUT" | jq -r '.agent_type // empty' 2>/dev/null)
+[[ "$AGENT_TYPE" == "subagent" ]] && exit 0
+
 # Extract transcript path from hook input
 TRANSCRIPT=$(echo "$INPUT" | jq -r '.transcript_path // empty' 2>/dev/null)
 [[ -z "$TRANSCRIPT" || ! -f "$TRANSCRIPT" ]] && exit 0

plugins/metacognitive-guard/hooks/scripts/truth-beacon.sh

@@ -60,6 +60,9 @@ CLAUDE CODE CONVENTIONS:
 - AGENTS.md is NOT auto-loaded by plugins (dead weight - use skill descriptions)
 - hooks/hooks.json auto-loads by convention (do NOT declare in plugin.json)
 - Passive context = CLAUDE.md + skill description frontmatter + SessionStart hooks
+- Use \${CLAUDE_SKILL_DIR} for skill-local resources, \${CLAUDE_PLUGIN_ROOT} for plugin-wide
+- /reload-plugins activates plugin changes without restart (no version bump needed during dev)
+- Hook events include agent_id and agent_type — filter subagents to avoid wasted work
 
 If uncertain about versions/APIs:
 - WebSearch to verify current state