Discovers .md files from skills_dir, bundled seed_skills/, and domain plugins.

Central shared artifact store for a single research session.

Manages pause flags and checkpoint serialisation for a single session. Directory layout:: ~/.eurekaclaw/sessions/<session_id>/

Retrieves top-k skills relevant to a task and formats for system prompt injection.

Unified read/write interface across all memory tiers.

Raised when a pause flag is detected mid-pipeline. The pipeline serialises its state before raising this exception so that a subsequent ``eur

User-provided input at the start of a research session. Three mutually exclusive modes control which pipeline stages are active: ``"detailed

A complete EurekaClaw research session. Usage: session = EurekaSession() result = asyncio.run(session.run_detailed("Prove that sa

Encapsulates the full Paper QA Gate flow for CLI (and later UI). Manages two nested loops: - Outer loop: review paper -> optionally rewrite -

Step 3 of the Theory Agent inner loop: LLM CoT + optional Lean4 dispatch.

Manages all available tools and provides dispatch by name.

Stage 7: verify theorem ↔ proof consistency.

Step 5 of the Theory Agent inner loop: adversarial falsification.

Step 1 of the Theory Agent inner loop: informal → formal notation. Uses the fast model (formalization is deterministic given a clear informal

Read a specific section from the paper LaTeX source by name or number.

Central brain. Drives the full pipeline from input spec to research output.

Drop-in replacement for anthropic.types.Message. Supports the access patterns used across all agents: response.content[0].text respon

Multi-turn QA agent with tool access for paper review.

One node in the proof plan produced by ProofArchitect.

Step 6 of the Theory Agent inner loop: update conjecture and prepare for retry.

Loads a proof pipeline spec from YAML and executes its stages. Usage: loop = TheoryInnerLoopYaml(bus) state = await loop.run(sess

Step 4 of the Theory Agent inner loop: formal or peer-review verification.

Manages the rolling conversation history for an agent's context window.

Human-on-the-loop or auto-approval at pipeline gate points. gate_mode: "none" — silent: no cards, no prompts (fully autonomous) "aut

A theorem, lemma, or technique extracted from an existing paper.

Step 2 of the Theory Agent inner loop: theorem → lemma DAG. Caching: if the formal statement has not changed since the last successful decomp

Stage 1: extract KnownResult objects from the session bibliography.

Tracks a running or completed session for UI polling.

Analyzes session failures and successes, generates new skill .md files.

Resolves a Task to the appropriate BaseAgent.

Stage 6: derive the formal theorem statement from the assembled proof.

Wraps anthropic.AsyncAnthropic and normalizes responses.

Stage 5: assemble proven lemmas into a coherent proof narrative.

Intercepts run outputs and improves skills/weights for future runs.

Cross-session persistent memory record.

Generates 5 research directions (diverge) then scores and selects 1 (converge).

Within-session episodic memory entry.

Bounded in-memory event buffer for a single research session.

Generates and executes Python experiments to validate theoretical bounds. Skips automatically when the proven theorem has no measurable numerical

Collects failures and proof trajectories during a run for post-run distillation.

Stage 2: identify the research gap from known results + direction.

Generates novel research hypotheses from survey findings. Uses a DeepInnovator-style approach: identifies gaps, makes cross-domain connection

Extract only the key technical lemmas implied by the proof skeleton.

Directed graph linking theorems and lemmas across research sessions.

Node in the cross-project theorem knowledge graph.

Iterative proof loop for 'adapted' and 'new' lemmas. Reuses Prover / Verifier / CounterexampleSearcher / Refiner from the existing stage file

Summarize how existing memory should shape the current proof attempt.

OpenAI-compatible Novita endpoint with provider defaults.

Backend for any OpenAI-spec endpoint (OpenRouter, vLLM, SGLang, LM Studio…).

Backend that calls the ChatGPT Codex endpoint (``/codex/responses``). Designed for Codex OAuth tokens obtained via ``codex auth login``. Thes

Persistent key-value store that survives across research sessions.

Builds a TaskPipeline from a declarative YAML spec. The spec lists stages with symbolic ``depends_on`` names; this class resolves those names

Scores proof trajectories for GRPO fine-tuning.

Build a proof skeleton before extracting technical lemmas.

Final output artifacts from a completed research session.

Runs in parallel with the Theory Agent inner loop to maintain math↔code mappings.