# Package Thoughtbox Workflow System as a Claude Code Plugin
## Problem
The entire workflow system (OODA loops, commands, skills, agents) lives in `.claude/` in the project directory. Every agent has write access. Agents routinely overwrite, truncate, or fail to preserve user-authored content in these files. This has been happening for months — confirmed by git history showing zero loop references in skills/commands/agents despite a supervised refactor that should have wired them in.
The compound-engineering plugin (Every) doesn't have this problem because it's installed as a plugin at `~/.claude/plugins/cache/`. Agents don't write there. The system is effectively immutable during normal operation.
## Solution
Package the Thoughtbox workflow system as a Claude Code plugin. Once installed, agents can read and invoke the commands/skills/agents but cannot modify them.
## What Goes Into the Plugin
### Commands (~50 files)
- `loops/` — 15 OODA loop building blocks (exploration, authoring, refinement, verification, orchestration) + meta docs
- `specifications/` — spec-designer, spec-orchestrator, spec-validator, specification-suite
- `hdd/` — Hypothesis-Driven Development workflow (init, research, stage-adr, validate, decide, state, overview, quick-reference)
- `development/` — implement-spec, hub-tdd, agentops-tdd, multi-agent-tdd, profiles-tdd, parallel-executor, spec-to-test, implementation-variants
- `meta/` — capture-learning, dgm-evolve, code-review-game, refactoring-game, ulysses-protocol, virgil-protocol, complexity-navigator, output-gap-analysis, learning-accelerator, feature-discovery
- `research/` — docs-researcher
- `startup/` — metaclaude, metacursor, prime
- `utils/` — post-edit-workflow
### Skills (~12 directories)
- `capture-learning/`
- `distill/`
- `escalate/`
- `implement/`
- `research-task/`
- `status/`
- `synthesize/`
- `taste/`
- `hub-collab/`
- `diagram/`
- `team/`
- `deploy-team-hub/`
### Agents (~11 files)
- `fact-checking-agent.md`
- `meta-agent.md`
- `hub-architect.md`, `hub-debugger.md`, `hub-manager.md`
- `architecture-diagrammer.md`
- `coordination-momentum.md`
- `research-taste.md`
- `triage-fix.md`
- `verification-judge.md`
- `dependency-verifier.md`
### Rules (move to plugin CLAUDE.md or keep as project rules — decide at build time)
- `ooda-foundation.md`
- `spiral-detection.md`
- `escalation-protocol.md`
- `continual-calibration.md`
- `git-workflow.md`
- `post-edit.md`
## What Stays in the Project `.claude/`
- Project-specific CLAUDE.md instructions
- Team-prompts (spawning context for specific workspaces)
- Any project-specific overrides or additions
## Critical Step: Wire the Loops FIRST
Before packaging, complete the refactor that was lost: every command, skill, and agent that should compose the OODA loops must actually reference them. The loops index already documents the intended compositions:
```
Spec Design: exploration/problem-space → authoring/spec-drafting → refinement/requirement-quality → verification/acceptance-gate
Implementation: orchestration/dependency-resolver → orchestration/queue-processor → authoring/code-generation → orchestration/spiral-detector → verification/integration-test
Code Review: exploration/codebase-discovery → refinement/code-quality → refinement/consistency-check → verification/acceptance-gate
Fact-Checking: orchestration/queue-processor → verification/fact-checking → refinement/consistency-check
```
Do this wiring manually or verify every change before committing. Do NOT trust an agent to do a bulk refactor of these files without verifying each one.
## Structure Ideas from Compound-Engineering
Their plugin structure that works well:
```
plugin-name/
.claude-plugin/
plugin.json # name, version, description, mcpServers
agents/
review/ # grouped by function
research/
design/
workflow/
docs/
commands/
workflows/ # core loop: plan, work, review, compound
(utility commands)
skills/
skill-name/
SKILL.md
references/ # deep-dive docs
templates/
assets/
CLAUDE.md # plugin development guidelines
README.md # component inventory
CHANGELOG.md # version history
```
Things to adopt:
1. **`plugin.json`** — version, description, MCP server registration (Thoughtbox server itself)
2. **Agent grouping by function** — not flat, organized into subdirectories
3. **Schema validation for compounding** — their `compound-docs` skill has a `schema.yaml` with enum-validated YAML frontmatter for searchable documentation. Adopt this pattern.
4. **The `/compound` step** — explicit "stop and document what you learned" workflow that produces a file at a known path. Currently missing from our system.
5. **Known-path file artifacts** — every workflow produces a file at a predictable location. Plans at `docs/plans/`, solutions at `docs/solutions/`, todos at `todos/`. Makes success ls-verifiable.
## Verification Plan
After building the plugin:
1. `grep -r '@loops' commands/ skills/ agents/` — every non-trivial command should reference loops
2. Install the plugin in a fresh project
3. Run each command and verify it works
4. Confirm agents CANNOT write to the plugin directory
5. Confirm agents CAN still create project-specific files (plans, solutions, todos) in the working directory
## Priority
High. This is a prerequisite for everything else. Until the system is write-protected, any agent session can silently destroy months of work.
