A single callable tool available to an agent.

Result of a single preflight check.

A cross-section consistency issue found in the paper.

Configuration for a single pipeline stage.

Review of a single paper section.

Simple token-bucket rate limiter.

Prepare Python runtimes for local experiment execution.

Concrete pip-install strategy for a generated experiment project.

A hypothesis for a single iteration round.

Complete output of the review stage.

Complete record for one iteration round.

Record of a single pipeline stage execution.

Registry of tools that can be passed to an LLM via function calling.

Persistent long-term memory store under ``~/.nanoresearch/memory``.

Dispatches LLM calls via the OpenAI-compatible API. All stages use the same base_url + api_key (your self-hosted endpoint). Each stage has it

Parsed output from a single RAM invocation.

Runs the standard 6-stage research pipeline.

Run static preflight checks on a generated code project.

Main research pipeline orchestrator for new runs.

Generates a full LaTeX research paper from experiment results.

Generates runnable training code + SLURM scripts based on cloned repos and experiment plan.

Estimated compute resources needed for the experiment.

Collection of quantitative evidence extracted from literature.

Deep pipeline with explicit skill, memory, and policy evolution enabled. The evo mode intentionally reuses the stable deep stage order and only

Submits SLURM training jobs, monitors them, debugs failures, and collects results.

Analyze experiment results and produce structured feedback.

Check cross-file import consistency using AST parsing. Scans all .py files in a directory and detects: 1. ``from X import Y`` where Y doesn't

Raised when an invalid stage transition is attempted.

Top-level state tracking across all iteration rounds.

Error from LLM interaction (JSON parse failure, empty response, etc.).

Result from a single LLM call, including usage metadata.

Contract thresholds used to determine whether a run is complete enough.

Aggregated preflight check report.

Extracted static skill context ready for prompt injection.

Accumulated cost for one pipeline stage.

Heuristic analysis of training behavior from training_log.

Compose K-Dense static skills with evolved NL/script skills.

Analyzes real experiment results and generates publication figures from actual data.

A registered output artifact.

Per-chat persistent memory: conversation history + condensed summary + facts.

Execute experiments on a SLURM cluster (local or remote).

Manages zip-based snapshots of a code directory.

A single contribution claim extracted from the Introduction.

Structured claims extracted from Introduction, injected into later sections. Ensures consistency: every claim in Introduction has matching conten

Event-sourced conversation memory with LLM-based condensation. Design principles (inspired by OpenHands context condenser): 1. Append-only

Accumulates LLM costs across all pipeline stages.

Reads error context, diagnoses failures, and patches code.

High-level execution behavior presets for the unified pipeline.

Central execution-time remediation policy for local experiment runs.

A single quantitative metric extracted from a published paper.

Structured feedback from one iteration round.

飞书 NanoResearch 机器人 -- AI 对话 + Pipeline 管理。

Structured summary of all evidence available for paper writing. Built once from execution/analysis outputs, consumed by every section generat

Complete output of the ideation stage.

LLM-powered interactive code editor with backup/rollback.

Formats log records as single-line JSON objects.

Summarize stage traces into research-facing direction/strategy memories.

ReAct agent that orchestrates paper editing via tool calls.

Paper type/mode for the research pipeline.

A single paper from literature search.

Complete paper structure for LaTeX generation.

Zip-based backup for the drafts/ + figures/ directories.

Stages of the research pipeline.

Manages state transitions for the research pipeline.

Emits pipeline progress events to a JSON file. The file is atomically overwritten on each event so that external consumers can poll it safely

Local manifest inspection snapshot shared across execution backends.

Collects (x, y, o) triples for SDPO training. Lifecycle per triple: 1. ``record_input()`` — called when RAM receives input, returns tripl

Local language model for reflection-augmentation. The model is lazily loaded on first ``generate()`` call. If ``enabled`` is ``False``, ``gen

One complete (x, y, o) interaction triple.

Top-level research configuration.

Run the SDPO router online, or produce a schema-compatible fallback.

Router-layer write-back for long-term memory and skill evolution.

A single section of the paper.

Searches for relevant code repos, clones them, and downloads required resources.

Index entry for a single static skill.

Persistent adaptive skill registry under ``~/.nanoresearch/skills``.

Index, match, and extract static skills from one or more directories.

Combined static and evolved skill context for prompt injection.

Master manifest for a research session workspace.