gabe-santana

gabe-santana/corollary

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

39 good first / help-wanted issues · Python · last activity Oct 4, 2026

16 stars 4 forks 16 watchers Python MIT License
llm python
View on GitHub Website
39 Open Issues Need Help Last updated: Oct 4, 2026

Open Issues Need Help

View All on GitHub
Harden agents against prompt injection from documents and tool results 17 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
kb.correct(): correct a fact in one call, with a reason 20 days ago
enhancement good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Accept Belief objects wherever a key or ref is accepted 20 days ago
bug good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Recognize Portuguese and Spanish scale words in citations (milhões, bilhões, millones) 20 days ago
enhancement good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Add an example that runs the agent on a local model with Ollama 20 days ago
documentation good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Snapshot and proof loaders accept malformed version fields 21 days ago
bug good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Explore competing hypotheses: an ATMS-style mode 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
A benchmark for how well agents recover from corrected inputs 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
A verifier check for natural-language contradictions 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Read numbers with decimal commas in citations and provenance checks 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Turn what the user says into premises: extract atomic assertions from conversation turns 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Make claims depend on the documents they were derived from 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Lay out model contexts for prompt caching 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Async agent runs: Agent.arun() and async models 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Record token usage and latency for each agent step 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Related work: add MemGPT/Letta and graph orchestration frameworks like LangGraph 21 days ago
documentation good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Proof.to_json() suggests default=str for values it can't fix 21 days ago
bug good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Add an example where two sources contradict each other 21 days ago
documentation good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Claim granularity: guidance and tooling for the right size of a belief 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Detect confabulated dependencies cheaply 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
An MCP server that exposes a belief base 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
LangGraph integration: a belief base as LangGraph state 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Test the claim contract against real models: a compatibility matrix 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Stream responses in AnthropicModel so large max_tokens works 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Use providers' structured outputs for the claim contract 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
OpenTelemetry tracing for runs, steps, tool calls and repairs 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Bound memory growth: compact old revisions and history 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Incremental persistence: a storage backend that doesn't rewrite the whole snapshot 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Concurrency: safe use of one belief base from several threads or workers 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Transactions: apply several changes atomically 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Tutorial: build a self-correcting agent end to end 21 days ago
documentation help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
An interactive HTML view of a proof or a whole belief base 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Typed keys: build belief keys without string formatting 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
A command-line tool to inspect snapshots 21 days ago
enhancement help wanted
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Say which belief can't be serialized when Proof.to_json() fails 21 days ago
enhancement good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Give Report a short repr instead of dumping every prompt 21 days ago
enhancement good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Export and document the Action type from corollary.contract 21 days ago
documentation good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Docs: propagate() re-derives in dependency order, not creation order 21 days ago
documentation good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python
Define __all__ in the corollary.verify, corollary.formula and corollary.contract modules 21 days ago
good first issue
gabe-santana/corollary
16

An agent runtime where state is beliefs, not messages. Every conclusion carries its justification; correct one fact and everything that depended on it updates automatically. Truth maintenance for LLM agents.

Python
#llm#python