Swih/mistral-mcp

What's new

6 new Claude Code skills (claude-plugin/skills/)

Stateless (Mistral-native):

• contract-analyzer — mistral_ocr → mistral_chat (mistral-large-latest + json_schema) → risk-rated clause extraction
• pdf-invoice-extractor — mistral_ocr → mistral_chat (mistral-small-latest + json_schema) → structured invoice fields
• audio-dispatch — voxtral_transcribe (diarize:true) → per-speaker action plan (core path via json_schema, full path via mistral_classify)

Workflow-oriented (human-in-the-loop via Mistral Workflows):

• contract-review-workflow — durable contract review with workflow_interact(signal: human_approval) gates
• compliance-audit-workflow — multi-step audit with mid-run workflow_interact(query: get_findings) + decision signals
• research-pipeline-workflow — hypothesis-driven research with workflow_interact(query: get_hypotheses) + amendment injection

Prompt engineering improvements (src/prompts.ts)

All 6 prompt templates updated to follow Anthropic best practices:

• Data before instructions — long inputs wrapped in XML tags (<transcript>, <document>, diff blocks) placed before instructions
• XML tags for content separation
• Positive framing throughout (e.g. "Appuie-toi uniquement sur le contenu" vs "N'invente pas")

mistral-router skill updates

• Added core / full profile column to tools table
• Fixed reasoning_effort enum: only "none" | "high" (aligned with actual API)
• Added workflow tools + 6 new skills to decision rules and examples

Stats

• 174 tests passing (unit + contract + stdio e2e + live API)
• 22 MCP tools (8 core, full set on MISTRAL_MCP_PROFILE=full)
• 11 Claude Code skills total (5 legacy + 6 new)