KDP Code
An automotive software AI agent inside KDP Hub that turns engineering intent into reviewable, signable, and replayable cross-tool DSL Blueprints.
Why KDP Code?
KDP Code is not a generic chatbot. It is an agent platform for automotive software engineering workflows. It shares project context, Git repositories, tool data, and engineering knowledge with KDP Hub, so specifications, A2L, ARXML, DBC, source code, and test artifacts can be updated in one traceable workflow. The LLM never executes actions directly. It produces typed DSL Blueprints that are reviewed and signed before the Executor runs them.
-
Plan, Execute, and Reflect agent loop with ContextManager-controlled token and context budgets.
-
DSL Blueprint plus Ed25519 approval flow, so AI actions are compiled, reviewed, signed, and then executed.
-
MCP, Skill, and Playbook orchestration for both single-tool tasks and cross-tool engineering workflows.
Key Features
ContextManager and Token Budget
Rebuilds each LLM call from system prompts, project tree, DSL docs, playbooks, memory, RAG, and messages, reducing stale context and oversized prompt risk.
Deterministic DSL Pipeline
The LLM writes DSL, and the system compiles it through tokenizer, parser, AST, compiler, and Blueprint stages. Execution remains controlled by deterministic software rather than direct AI edits.
Project-level Prompt System
Hub provides base prompts and tool rules. Projects can add engineering conventions, naming rules, and review requirements through .kdp/prompts and .kdp/AGENT.md.
Skill and Playbook Orchestration
Skills automate repeatable tasks for one tool. Playbooks orchestrate two or more tools, such as keeping specifications, SWCs, network databases, and validation artifacts aligned.
Memory and Codebase RAG
Supports short-term conversation memory, session summaries, and cross-session memory. Codebase RAG can retrieve complete functions or classes using AST-aware chunking.
Transcript and Replayability
Context snapshots, LLM calls, and tool executions can be recorded as JSONL transcripts for audit, review, and replay of AI-assisted engineering operations.
From One Requirement to Cross-tool Engineering Output
An engineer describes the intent. KDP Code creates DSL, compiles it into a Blueprint, routes it through human review and approval, then synchronizes outputs across KDP Library, Calibrator, SWC Designer, and network database tools.
Intent
Describe the requirement, constraint, ASIL level, signal range, or deliverable.
ContextManager
Assemble project, tool, RAG, memory, and conversation context.
LLM / Skill
Emit a DSL block or expand a deterministic Skill template.
Compiler
Tokenizer -> Parser -> AST -> Blueprint.
Approval
Human review and signed executable Blueprint.
Executor
Execute through Builtin, MCP, CLI, or API adapters.
Blueprint outputs can feed
KDP Library: SYS / TECH / DD specs and trace links
Calibrator: A2L characteristics and DAQ maps
SWC Designer: AUTOSAR ports, interfaces, and ARXML
Network Designer: CAN DBC, signal matrix, and routing rules
FAQ
How is KDP Code different from GitHub Copilot or Cursor?
KDP Code is not focused on generic code completion. It is built for cross-tool orchestration in automotive software workflows, converting intent into typed DSL Blueprints that KDP Hub tools and MCP servers can execute.
Can an LLM mistake directly damage my project?
No. The LLM only emits DSL. DSL is compiled into a Blueprint, reviewed by a human, signed, and then executed by the Executor. Each step can be captured in a transcript for review and rollback.
Can KDP Code work offline?
The DSL compiler, Blueprint model, Skill templates, and Playbooks can run locally. Steps that require LLM inference are provided through the KDP platform service.