# Third-party design & content attribution

`@sema-ai/core` is licensed under the Business Source License 1.1 (BUSL-1.1; Change License: Apache-2.0, Change Date: 2030-07-13 — see LICENSE). Versions ≤1.277.1 were published under MIT. It incorporates design and (where noted) prompt content
adapted from the following MIT-licensed projects. All are reused under the MIT License.

## openclaw — https://github.com/openclaw/openclaw (MIT)
Vendored source: the reusable agent core (`@openclaw/agent-core`) and `@openclaw/llm-core` live under
`src/vendor/` with minimal modifications; that tree has since been fully rewritten into first-party
code (design/118 de-vendoring, 2026-07-02). See `docs/vendor-history.md` for the baseline commit
and the complete modification history.

## CodeWhale (formerly DeepSeek CLI) — https://github.com/Hmbown/CodeWhale (MIT)
Copyright (c) 2024-2025 DeepSeek CLI Contributors.
Design adapted (re-implemented in TypeScript, not copied verbatim):
- `src/core/memory.ts` — the user-memory pattern (persistent notes, a `remember` tool, the
  `<user_memory>` system-prompt block, size truncation, timestamped bullets) is adapted from
  CodeWhale's `crates/tui/src/memory.rs`.
- `src/prompts/default.ts` — the default system prompt distills the *universal* agent disciplines
  from CodeWhale's `prompts/base.md` "Constitution" (truth, duty of action, verification, tool-use
  enforcement, hierarchy of authority) and `prompts/memory_guidance.md` (memory = declarative facts).
  CodeWhale's project branding and coding-/DeepSeek-specific content are not used; the text here is a
  domain-neutral rewrite.
- `test/prompt-cache.test.ts` — the offline byte-prefix-cache simulation (multi-turn hit-rate guard)
  is ported from CodeWhale's `crates/tui/tests/cache_guard.rs`.
- `src/brain/openai.ts` — parsing prefix-cache-hit tokens from the response usage
  (`prompt_tokens_details.cached_tokens` / `prompt_cache_hit_tokens`) follows CodeWhale's `client.rs`.

## Anthropic engineering guidance (referenced, not code)
Context-engineering and memory-tool design follow Anthropic's public engineering writing
(context engineering, memory tool + context editing, managed agents, multi-agent systems).