nsaspy/llm-agent-cl
1
0
Fork 0
Code Issues 1 Pull requests Projects Releases Packages Wiki Activity Actions
a llm agentic framework in lisp
5 commits 5 branches 0 tags 190 KiB
  • Python 100%
Find a file
Exact
nsaspy 5488f807df
 Complete auto-research and publish implementation designs (#3) 
Publish the completed research/design review set without authorizing runtime implementation.
2026-08-12 21:39:33 -04:00
.github/workflows Bootstrap auto-research workflow and LLM agent research 2026-08-12 14:05:32 -04:00
docs Complete research and publish implementation designs 2026-08-12 21:38:35 -04:00
roam Complete research and publish implementation designs 2026-08-12 21:38:35 -04:00
scripts Bootstrap auto-research workflow and LLM agent research 2026-08-12 14:05:32 -04:00
skills Complete research and publish implementation designs 2026-08-12 21:38:35 -04:00
tests Bootstrap auto-research workflow and LLM agent research 2026-08-12 14:05:32 -04:00
.gitignore Bootstrap auto-research workflow and LLM agent research 2026-08-12 14:05:32 -04:00
AGENTS.md Complete research and publish implementation designs 2026-08-12 21:38:35 -04:00
LICENSE Bootstrap auto-research workflow and LLM agent research 2026-08-12 14:05:32 -04:00
README.md Complete research and publish implementation designs 2026-08-12 21:38:35 -04:00

README.md

llm-agent-cl

A production-oriented Common Lisp library for multi-provider LLM, agent, graph and retrieval applications.

Current phase

Auto-research is complete enough for design review. Runtime implementation is not authorized yet.

The repository uses the document-driven workflow modeled on lost-rob0t/starintel-auto-research: canonical Org research/design nodes, approval tables, changelogs, stable IDs, one-active-design implementation handoff, skills and validation scripts.

The first provider is OpenRouter. The first major milestone requires behavioral capability parity with current LangChain + LangGraph, including provider-neutral models/messages, sync/async/batch invocation, streaming, tools, structured output, middleware, multi-turn state, graph reducers/routing/checkpoints/interrupts/time travel, MCP, document loaders, text splitters, embeddings, stores, vector stores, retrievers and RAG composition.

Dexador is the synchronous HTTP transport. lost-rob0t/carrier is the canonical maintained asynchronous HTTP fork. Carrier must be hardened according to the reviewed design before production provider use; current stock behavior is not treated as production-safe.

Review entrypoints

  • Research index: roam/indexes/llm-agent-cl/LLM-AGENT-INDEX-000-canonical-research-index.org
  • Design index: roam/design/llm-agent-cl/LLM-AGENT-DESIGN-INDEX-000-implementation-design-index.org
  • Carrier hardening design: roam/design/llm-agent-cl/LLM-AGENT-DESIGN-006-carrier-hardening.org

Merging research/design documents does not constitute implementation approval. Code begins only after the applicable canonical design is explicitly approved.

Configuration philosophy

Production use can be fully explicit, while quick REPL/tests may intentionally use scoped/process defaults or provider-standard environment variables. Explicit values always win. The library never ships hard-coded credentials and secrets are redacted from logs/conditions/traces.

Research workflow

python3 scripts/sync.py
python3 scripts/sync.py --check
python3 scripts/validate-docs.py
python3 scripts/validate_agent_pack.py
python3 scripts/list-unreviewed-research.py

Future approved designs use the one-active-design workflow:

python3 scripts/implement.py roam/design/<project>/<design>.org
python3 scripts/implement.py --status
python3 scripts/mark-design.py implemented --project <project> --summary <summary> --file <path> --test <observed-test>
python3 scripts/mark-design.py rejected --project <project> --reason <reason> --evidence <evidence>

License

MIT.