SAP Router Skill — Multi-IDE Agent Instructions

89 skills mirrored across 4 IDEs — same SKILL.md content in each.

All skills auto-trigger by file context and keyword. See SKILL.md for master dispatch.

Implemented v6 operating model

Canonical source is .agents/skills plus .agents/registries/mcp-capabilities.json. Claude and Gemini skill trees are generated mirrors; Codex and Cursor reuse .agents/skills through AGENTS/rules/agent assets instead of maintaining duplicate skill trees.

Generate/check IDE assets:

python scripts/generate_ide_assets.py generate --targets all
python scripts/generate_ide_assets.py check

Capability routing is fail-closed:

python scripts/validate_catalog.py --strict
python scripts/mcp_launcher.py list --capability sap.cpi.message.read
python scripts/mcp_launcher.py list --capability sap.apim.proxy.read

Dynamic local asset discovery

Before selecting an SAP skill or MCP for a task that is not an exact static route, query the repository-owned catalog. The search ranks canonical skills, reviewed MCPs, and bundled snapshots together:

python scripts/source_catalog.py search "task description"
python scripts/mcp_launcher.py search --query "task description"

Use the highest relevant enabled/reviewed result. A bundled MCP with status disabled_candidate is evidence for review, never authorization to execute it. All bundled code lives under bundled/; routing does not fetch GitHub at runtime.

CPI/APIM/Fiori/UI5/CAP now route through API/CLI/plugin MCPs first and use Web UI MCP fallbacks (integration-suite-ui-mcp, apim-ui-mcp) only when background access is blocked. The Web UI bridge uses the logged-in browser session and does not accept credential values.

Mutating capabilities use plan/approval/commit semantics. Local approval state is handled by:

python scripts/approval_broker.py plan --capability sap.apim.proxy.deploy --target DEV

What this project is

40 Python operational scripts plus Node.js review gates — no live SAP system required credentials required. Everything runs offline.

Routing rules (v4.2.0 — ADT-first, GUI-fallback, caveman-optimized)

Decision tree:

1. CAVEMAN DELEGATION? (fix/find/review, 1-2 files) → cavecrew-investigator/builder/reviewer
2. ADT AVAILABLE? (read_source, search, activate...) → ARC-1 / aibap / mcp-abap-adt
3. GUI FALLBACK? (SPRO, SM30, SU01, MM01, VA01...) → mcp-sap-gui (25 transactions)
4. FUNCTIONAL WRITE? (BAPI batch) → needs explicit --functional context (else needs-functional-context, NO BAPI fired); then BAPI-first / SAP GUI dispatch. ZROUTER RFC only if opted in (zrouter accept).
5. SPEC→PIPELINE? → sap_router.py pipeline (8 stages)
6. LLM / RAG? (prompt optimization, doc retrieval) → sap-llm-engineering skill + RAG routing (Pinecone, Supabase pgvector, Azure AI Search)

Memory session commands

python scripts/memory_manager.py init   --input MEMORY.md --sys S4H --client 100 --env DEV --usr DEVELOPER
python scripts/memory_manager.py add    --input MEMORY.md --module MM --action-name CreateMaterial --status OK
python scripts/memory_manager.py verify --input MEMORY.md
python scripts/memory_manager.py compact --input MEMORY.md

Caps: 20 active blocks max, 100 lines total. Older blocks auto-archive.

XLS converter commands

python scripts/xls_to_bapi.py template --output tmpl.csv --module MM --action CREATE_MATERIAL
python scripts/xls_to_bapi.py convert  --input data.csv  --module MM --action CREATE_MATERIAL

Action-to-BAPI mapping (29 actions / 9 modules)

Functional WRITE actions need explicit --functional context (else needs-functional-context, no BAPI fired); they then dispatch BAPI-first / SAP GUI. ZROUTER RFC is opt-in only (zrouter accept) — never the default. ADT ops and code_search → ARC-1 ADT. sf_* → sf-mcp.

RTK and Context Mode Optimization Rules (MANDATORY)

To keep token consumption low (saving 60-90% on terminal commands and 98% on tool outputs):

RTK (Rust Token Killer) Rules

1. Always prefix CLI/shell commands with rtk in powershell/bash sessions (e.g., rtk git status, rtk cargo test, rtk npm install).
2. Do NOT prefix commands that are interactive, or those specified in exclude_commands (e.g., curl, playwright).
3. Use rtk gain to check token savings and rtk discover to find optimization opportunities.

Context Mode Rules

1. Think in Code: Always write short scripts to analyze large outputs instead of reading raw files into context. Use ctx_execute (or ctx_execute_file) to execute scripts and print only the final answer.
2. Sandbox Execution: Use ctx_execute or ctx_batch_execute to execute any tool calls that generate large outputs (like python scripts/healthcheck.py or ADT/RAG searches) to reduce raw output size by 98%.
3. Fetch Hardening: Use ctx_fetch_and_index instead of curl/wget to index URLs into SQLite and search them via ctx_search.
4. Session Continuity: Use ctx_checkpoint to save session state, and ctx_restore to restore the state when the conversation is compacted.