Revised Sprint 11e from training data export to workflow annotation foundation (routing types, skeleton AST, inference-to-routing bridge). Detailed plans for Sprints 12-19 covering workflow model, GUI overhaul, 6 new languages, orchestration engine, C++ self-hosting, Claude Code plugin, and workflow visualization. Created ARCHITECT.md with core thesis and design principles. Training data deferred to post-25 where real workflow decisions provide superior signal. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
6.3 KiB
Sprint 18 Plan: Claude Code Plugin + MCP Workflow Tools
Context
This is the payoff sprint. Sprints 12 and 15 built the workflow model and orchestration engine. Sprint 18 builds the client side — the thing an engineer actually installs and uses.
The deliverable: a Claude Code MCP tool configuration that an engineer adds to their project, and suddenly Claude Code can orchestrate annotation-driven workflows. The engineer says "modernize this module" and the system:
- Creates a skeleton with inferred annotations
- Routes tasks (deterministic for simple stuff, Claude for complex stuff)
- Executes what it can automatically
- Surfaces the hard decisions to the human with full context
- Tracks progress and handles review
This sprint doesn't build a custom plugin binary. It builds the MCP server
tools and the prompt templates that make the existing whetstone_mcp binary
work as an effective Claude Code tool server. The "plugin" is a configuration
file pointing Claude Code at the whetstone_mcp server.
Phase 18a: Claude Code Integration (Steps 417-422)
Step 417: MCP Server Configuration + Discovery (12 tests)
- Generate
.mcp.jsonconfiguration file for Claude Code - Server manifest: all 68+ tools organized by category with descriptions
- Tool categories: AST, Diagnostics, Workflow, Routing, Project, File ops
- Auto-detection:
whetstone_mcpbinary path resolution - Workspace detection: auto-find project root and language
Step 418: Workflow Prompt Templates (12 tests)
- MCP prompt templates that guide Claude through the workflow protocol:
architect_project— "Analyze this codebase and create an annotated skeleton"modernize_module— "Modernize this legacy code with safety annotations"port_to_language— "Port this module from {source} to {target}"add_feature— "Add {feature} to this project using the workflow system"review_pending— "Review and approve/reject pending work items"
- Each prompt includes the workflow protocol steps the agent should follow
- Prompts reference specific MCP tools by name
Step 419: Agent Workflow Loop (12 tests)
- The core interaction loop for an agent using the MCP tools:
whetstone_infer_annotationson existing codewhetstone_create_skeletonfor new functionalitywhetstone_create_workflowto plan executionwhetstone_orchestrate_run_deterministicfor auto taskswhetstone_get_blockersto find what needs the agent- For each agent task: read context bundle, generate code,
whetstone_submit_result whetstone_get_progressto track completion
- Test that this loop works end-to-end through MCP
Step 420: Human Review Interface via MCP (12 tests)
- Tools for presenting review decisions to the engineer:
whetstone_get_review_queue— items awaiting human reviewwhetstone_get_review_context— full context for a review item (original skeleton, generated code, annotations, confidence, reasoning)whetstone_approve_item— approve with optional feedbackwhetstone_reject_item— reject with required feedback
- Review context formatted for human readability (not just raw JSON)
Step 421: Workspace Onboarding Flow (12 tests)
- First-time setup experience for a new workspace:
- Detect project languages from file extensions
- Index workspace files
- Run annotation inference on key files
- Suggest initial annotations and workflow
- Generate
.whetstone/directory with sidecar files
whetstone_onboard_workspacetool that does this in one call
Step 422: Phase 18a Integration (8 tests)
- Full end-to-end: configure MCP → onboard workspace → create workflow → orchestrate → agent completes tasks → human reviews → complete
- All MCP tools documented with clear descriptions
- 75+ MCP tools total
Phase 18b: Multi-Model Dispatch (Steps 423-427)
The orchestration engine routes tasks to "slm", "llm", or "human" — but which specific model? Phase 18b adds model selection awareness.
Step 423: Model Profile Registry (12 tests)
- ModelProfile struct: name, contextWindow, costPerToken, capabilities, speed
- Pre-configured profiles: claude-opus, claude-sonnet, claude-haiku, local-slm
- Routing engine maps: "llm" → claude-sonnet (default), "slm" → claude-haiku
- Configurable overrides in
.whetstone/config.json
Step 424: Context Window Optimization (12 tests)
- Given a model's context window, optimize the context bundle:
- Small model (8k context) → aggressive truncation, local scope only
- Medium model (32k) → file scope with budget
- Large model (200k) → project scope, include related files
- Context assembly adapts automatically based on model profile
Step 425: Batch Submission for Agent Tasks (12 tests)
- When multiple agent tasks are ready, batch them for efficient submission
- Group by context overlap (tasks in same file share context)
- Batch format: shared context once + per-task instructions
- Reduces total token cost by ~30-50% for related tasks
Step 426: Cost Tracking + Reporting (12 tests)
- Track actual tokens used per task (from submitExternalResult metadata)
- Compare estimated vs actual costs
- Running cost total for the workflow
- Cost report: breakdown by worker type, language, task complexity
Step 427: Phase 18b Integration + Sprint Summary (8 tests)
- Model profiles select appropriate context window
- Batch submission reduces total tokens vs individual
- Cost tracking matches expected for a 10-task workflow
- Sprint 18 totals: 75+ tools, workflow operational end-to-end
Step & Test Summary
| Phase | Steps | Tests | Theme |
|---|---|---|---|
| 18a | 417-422 | 68 | Claude Code config, prompts, agent loop, review, onboarding |
| 18b | 423-427 | 56 | Model profiles, context optimization, batching, cost tracking |
| Total | 417-427 | ~124 | 11 steps |
Key Architectural Decision
The "plugin" is not a custom binary or VS Code extension. It is:
- The existing
whetstone_mcpbinary (headless MCP server) - A
.mcp.jsonconfiguration file that tells Claude Code where to find it - MCP prompt templates that guide the agent through the workflow protocol
- MCP tools organized into a coherent workflow
This means any MCP-compatible client (not just Claude Code) can drive Whetstone workflows. The protocol is the product, not a proprietary integration.