@rune-kit/rune 2.4.0 → 2.7.0

package/skills/mcp-builder/SKILL.md

@@ -3,7 +3,7 @@ name: mcp-builder
 description: Build Model Context Protocol servers from specifications. Generates tool definitions, resource handlers, and test suites for MCP servers in TypeScript or Python (FastMCP).
 metadata:
   author: runedev
-  version: "0.3.0"
+  version: "0.4.0"
   layer: L2
   model: sonnet
   group: creation
@@ -321,6 +321,58 @@ mcp-server-<name>/
 - Installation: Claude Code, Cursor, Windsurf config snippets
 - Configuration reference (env vars with descriptions)
 
+## Reference Pattern: Multi-Provider Adapter
+
+When the MCP server needs to call multiple AI providers (e.g., both Anthropic and OpenAI), use the **Provider Adapter** pattern to normalize different APIs behind a unified interface.
+
+### Interface
+
+```typescript
+interface ProviderAdapter {
+  formatRequest(params: RequestParams): { url: string; init: RequestInit };
+  parseResponse(data: unknown): { content: string; usage: TokenUsage | null };
+  formatStreamRequest(params: RequestParams): { url: string; init: RequestInit };
+  parseSSEEvent(eventType: string, data: string): StreamChunk | null;
+}
+```
+
+### Discriminated Union for Stream Chunks
+
+```typescript
+type StreamChunk =
+  | { type: "thinking"; content: string }
+  | { type: "text"; content: string }
+  | { type: "done" }
+  | { type: "done_with_usage"; usage: TokenUsage }
+  | { type: "usage_delta"; inputTokens?: number; outputTokens?: number }
+  | { type: "error"; message: string };
+```
+
+### When to Apply
+
+- MCP server wraps multiple AI providers (e.g., a router server that dispatches to Claude, GPT, or local models)
+- MCP server aggregates responses from multiple APIs with different response formats
+- MCP server needs to support streaming from providers with different SSE event schemas
+
+### Key Implementation Notes
+
+- Each provider adapter handles its own SSE event types (Anthropic: `content_block_delta`, `message_start`; OpenAI: `response.output_text.delta`, `[DONE]`)
+- Buffer management for SSE: handle incomplete lines, track event types, manage abort signals
+- Provider-specific prompt tuning: some models benefit from additional constraints (e.g., "Maximum 2-3 paragraphs" for verbose models)
+- Per-provider token tracking: normalize different usage reporting formats into a single `TokenUsage` type
+
+### Cost-Aware Model Selection
+
+When building MCP servers that call AI providers, support **dual-model configuration** — allow users to specify a primary model for critical operations and a cheaper model for background tasks (summarization, classification, metadata extraction). This avoids burning expensive API credits on tasks that don't need maximum quality.
+
+```typescript
+// config.ts
+const config = {
+  primaryModel: process.env.PRIMARY_MODEL || 'claude-sonnet-4-20250514',
+  backgroundModel: process.env.BACKGROUND_MODEL || 'claude-haiku-4-5-20251001',
+};
+```
+
 ## Constraints
 
 1. MUST validate all tool inputs with Zod (TS) or Pydantic (Python) — never trust AI-provided inputs

package/skills/onboard/SKILL.md

@@ -3,11 +3,12 @@ name: onboard
 description: Auto-generate project context for AI sessions. Scans codebase, creates CLAUDE.md and .rune/ setup so every future session starts with full context.
 metadata:
   author: runedev
-  version: "0.3.0"
+  version: "0.4.0"
   layer: L2
   model: sonnet
   group: quality
   tools: "Read, Write, Edit, Glob, Grep"
+  emit: project.onboarded
 ---
 
 # onboard
@@ -44,6 +45,7 @@ project/
     ├── progress.md        # Empty, ready for session-bridge
     ├── session-log.md     # Empty, ready for session-bridge
     ├── instincts.md       # Empty, ready for session-bridge instinct learning
+    ├── contract.md        # Project invariants enforced by cook/sentinel
     └── DEVELOPER-GUIDE.md # Human-readable onboarding for new developers
 ```
 
@@ -105,6 +107,11 @@ Use `Write` to create each file:
 - `.rune/progress.md` — create with header `# Progress Log` and one placeholder entry
 - `.rune/session-log.md` — create with header `# Session Log` and current date as first entry
 - `.rune/instincts.md` — create with header `# Project Instincts` and a description: "Learned trigger→action patterns. Managed by session-bridge. See session-bridge SKILL.md Step 5.7 for format."
+- `.rune/contract.md` — generate a starter contract based on the detected tech stack:
+  - Copy structure from `docs/CONTRACT-TEMPLATE.md`
+  - Customize rules based on Step 2 findings (e.g., Python → add `no bare except`, Node.js → add `no console.log`, SQL database → add parameterized queries rule)
+  - Remove sections that don't apply (e.g., `contract.operations` for a library with no deployed service)
+  - The contract is a starting point — tell the user to review and customize it
 
 ### Step 5.5 — Load Existing Instincts
 
@@ -220,6 +227,56 @@ Audit the project's baseline context cost from MCP servers and agent configurati
 
 **Skip if**: Total MCP tools ≤80 AND CLAUDE.md ≤150 lines (healthy baseline).
 
+### Step 6e — AI-Driven Interview (Optional, User-Initiated)
+
+When invoked as `/rune onboard --interview` or when the project is too ambiguous for automated detection (e.g., no package.json, no clear entry point, mixed languages), switch to **conversational onboarding** — the AI asks targeted questions instead of relying solely on file scanning.
+
+#### Interview Flow
+
+Ask 5-8 questions in sequence, adapting based on answers. Start broad, narrow based on responses:
+
+```
+Q1: "What does this project do in one sentence?"
+    → Captures purpose (README may be missing or outdated)
+
+Q2: "Who uses this — internal team, external users, or both?"
+    → Determines audience, affects DEVELOPER-GUIDE.md tone
+
+Q3: "What's the main entry point — where does execution start?"
+    → Bypasses file scanning for complex monorepos
+
+Q4: "What commands do you use daily? (dev server, tests, build)"
+    → Gets verified commands instead of guessing from config files
+
+Q5: "Any areas of the codebase you'd warn a new developer about?"
+    → Captures tribal knowledge that no scan can detect
+
+Q6: "Are there external services this depends on? (databases, APIs, queues)"
+    → Maps integration points for Architecture Map
+
+Q7: "What's the deployment story — how does code get to production?"
+    → Captures CI/CD context
+
+Q8 (conditional): "Anything else a new session should know that's not in the code?"
+    → Catches edge cases, workarounds, known issues
+```
+
+#### Interview Rules
+
+- **Adapt**: Skip questions that were already answered by earlier responses. If Q1 reveals "it's a Next.js app", don't ask about the framework.
+- **Validate**: Cross-reference answers with actual file scan results. If user says "we use Jest" but `vitest.config.ts` exists, ask to clarify.
+- **Merge**: Interview answers supplement (not replace) automated scan. Scan provides facts, interview provides context and intent.
+- **Store**: Save interview responses as high-confidence entries in `.rune/conventions.md` and `.rune/cumulative-notes.md` (tagged `[from-interview]`)
+
+#### When to Auto-Suggest Interview
+
+Suggest switching to interview mode (but don't force it) when:
+- Step 2 produces 3+ "unknown" fields in tech stack detection
+- Project has no README.md and no package.json/pyproject.toml/Cargo.toml
+- Project appears to be a monorepo with 3+ distinct sub-projects
+
+Output: `"ℹ️ This project is hard to auto-detect. Run /rune onboard --interview for guided setup."`
+
 ### Step 7 — Commit
 Use `Bash` to stage and commit the generated files:
 ```bash

package/skills/perf/SKILL.md

@@ -3,7 +3,7 @@ name: perf
 description: Performance regression gate. Detects N+1 queries, sync-in-async, missing indexes, memory leaks, and bundle bloat before they reach production.
 metadata:
   author: runedev
-  version: "0.2.0"
+  version: "0.3.0"
   layer: L2
   model: sonnet
   group: quality
@@ -242,6 +242,39 @@ Emit structured report:
 ### Verdict: PASS | WARN | BLOCK
 ```
 
+### Step 8.5 — Token Budget Tracking (AI-Powered Apps)
+
+For projects that call AI APIs (detected via imports of `anthropic`, `openai`, `@anthropic-ai/sdk`, `@ai-sdk/core`, `langchain`, `llamaindex`, or `fastmcp`), audit token usage patterns per operation type.
+
+**Scan for:**
+
+| Pattern | Finding | Impact |
+|---------|---------|--------|
+| AI call inside a loop without batching | `TOKEN_LOOP — [file:line] — AI call in loop over [collection] — batch or parallelize` | Cost scales linearly with collection size |
+| No token usage tracking | `NO_TOKEN_TRACKING — [file:line] — AI response usage not captured — add cost logging` | Invisible spend, no budget control |
+| Expensive model for simple tasks | `MODEL_MISMATCH — [file:line] — using [opus/gpt-4] for [classification/extraction] — use [haiku/gpt-4.1-mini]` | 10-30x cost difference for same result |
+| Missing max_tokens on open-ended prompts | `UNBOUNDED_TOKENS — [file:line] — no max_tokens on [call] — add limit to prevent runaway cost` | Single call can consume entire budget |
+| Duplicate AI calls for same input | `DUPLICATE_AI_CALL — [file:line] — same prompt sent to [provider] without caching — add response cache` | Wasted tokens on redundant calls |
+
+**Per-Operation Cost Awareness:**
+
+When token tracking IS present, analyze the operation type breakdown:
+
+```
+Operation Type          Avg Tokens    Frequency    Monthly Est.
+─────────────────────────────────────────────────────────────
+Chat (primary)          2,500 in/800 out    high         $X.XX
+Background notes        500 in/200 out      per-chat     $X.XX
+Summarization           1,500 in/300 out    periodic     $X.XX
+Classification          200 in/50 out       high         $X.XX
+─────────────────────────────────────────────────────────────
+Total estimated monthly                                  $X.XX
+```
+
+**Report this under a `### AI Token Budget` subsection** in the Perf Report. Only include when AI API usage detected — skip entirely for non-AI projects.
+
+**Key insight**: The most impactful optimization is often **model selection per operation** — using a cheaper model for background tasks (summarization, classification, metadata extraction) while reserving expensive models for primary user-facing interactions. A 10x cost reduction on 60% of calls = 6x overall savings.
+
 ## Output Format
 
 ```