One or more tool calls supplied through a single OpenHands confirmation callback and resolved together by Heartwood.
Air-gapped environment
A compute environment whose deployment boundary prevents outbound and external network access; moving artifacts into it requires a separately approved process.
Agent
Software that uses a model and tools iteratively to pursue a task rather than returning only one text response.
Agent Skills
An open directory format for agent instructions and optional scripts, references, assets, and tool declarations; Heartwood loads it through OpenHands.
Application programming interface (API)
A structured way for software components to communicate.
Authorization Domain
A Terra access restriction that limits a workspace and its data to members of named groups.
Audit checkpoint
A canonical audit export whose metadata and signer identity are authenticated by a deployment-approved public-key signature and bound to its deployment identifier, session, creation time, and retention declaration.
Audit export
A scrubbed JSON Lines representation of the tamper-evident session audit chain.
Carina
Stanford's research-computing platform using project storage and Slurm-managed compute.
ChatGPT subscription connection
Account-based access to supported Codex models through OpenHands, separate from OpenAI API billing and API keys.
Codex Responses API
The Responses API transport used by OpenHands for supported models reached through a ChatGPT subscription connection.
Container
A packaged Linux filesystem and process environment run by software such as Docker.
Context window
The token budget available to a model for instructions, conversation history, project content, tool results, and output.
Credential binding
A non-secret identifier that tells Heartwood where an authorized process or platform supplies a provider credential.
CUDA
NVIDIA's software platform and application binary interface for GPU computing.
Dependency lock
A file recording exact dependency versions and artifact hashes; Heartwood uses pylock.toml to reconstruct Python analysis environments.
Device-code sign-in
An OAuth flow in which one interface displays a short-lived code that the user enters on a separate provider sign-in page.
Ed25519
A public-key signature algorithm supported by Heartwood checkpoint signers and used by the explicit local fallback.
Global Alliance for Genomics and Health (GA4GH)
A standards organization whose interfaces include the Task Execution Service, Workflow Execution Service, and Data Repository Service used by research platforms.
GGUF
A single-file model format commonly used with llama.cpp and quantized CPU inference.
Graphics processing unit (GPU)
Specialized compute hardware used to accelerate model inference.
Hardware security module (HSM)
A protected hardware or managed service boundary that performs cryptographic operations without exporting private keys.
Heartwood-managed model
A model whose files and inference runtime Heartwood manages on the compute environment where Heartwood is running, such as a workstation container, Terra runtime, or Carina allocation.
Hugging Face
A model repository service used by Heartwood to inspect public metadata and download supported model artifacts.
Ingress
The validated network route through which a browser or API client reaches the Heartwood gateway, such as direct loopback, a Jupyter proxy, or an explicitly trusted platform proxy.
JSON Lines
A text format that stores one JSON value on each line so records can be appended and processed incrementally.
Key management service (KMS)
A managed service that controls cryptographic keys and performs authorized operations such as checkpoint signing without exposing private key material to Heartwood.
Launch capability
The process-lifetime secret that every Heartwood gateway API request, event stream, and WebSocket must carry, delivered through a one-time launch link or a platform proxy header.
Large language model (LLM)
The model used by the agent to interpret requests, reason, produce text, and propose tools.
Model Context Protocol (MCP)
An open protocol through which an agent discovers and calls tools that an external service provides.
Model credential boundary
The technical separation, or documented lack of separation, between provider authentication used for model calls and the operating-system identity that runs agent tools.
Model transfer bundle
A reproducible Heartwood ZIP64 artifact containing one model's immutable source metadata, runtime settings, selected files, and per-file SHA-256 digests for verified import into another project.
OAuth
A standard authorization protocol that lets a user grant account access without giving Heartwood the account password.
One-time code
The short-lived code displayed during device-code sign-in and entered on the provider's verification page.
OpenAI-compatible service
A model endpoint implementing the relevant OpenAI API request and response shapes; compatibility does not imply operation by OpenAI.
OpenHands
The upstream agent SDK and coding-tool platform used by Heartwood for conversations and tool execution.
Pilot evaluation
A bounded trial in which a small group of researchers uses one standard research environment so the deployment can record installation, workflow, reproducibility, and support evidence before a wider rollout.
Protected health information (PHI)
Individually identifiable health information protected under applicable policy or law.
Project
The process current directory and its descendants that Heartwood treats as the agent workspace.
Project state
Private non-secret configuration, sessions, models, Skills, logs, caches, runtime files, and audit artifacts under .heartwood/.
Quantization
A lower-precision model representation, such as AWQ, GPTQ, or FP8, used to reduce memory and storage requirements.
Research Skill
A versioned instruction package with declared tools, metadata, and workflow guidance available to the OpenHands agent.
Research specialist
A bounded OpenHands child agent that performs one focused planning or review task for the parent agent within inherited or narrower project, credential, model, network, policy, and confirmation boundaries.
Responses API
OpenAI's current API shape for model responses, tool calls, and related agent interactions.
safetensors
A tensor-storage format designed to load model weights without executing serialized code.
SHA-256
A cryptographic hash function Heartwood uses to verify that model files and other artifacts have not changed.
Slurm
A scheduler that allocates compute resources for jobs on platforms such as Stanford Carina.
Standard research environment
One recorded combination of Heartwood release, platform base image, compute shape, model routes, storage layout, and update policy that a deployment supports for a group of researchers.
Tensor parallelism
A runtime layout that divides one model across a fixed number of GPUs.
The Update Framework (TUF)
A signed metadata system that protects software distribution against repository compromise, rollback, freeze, and artifact-substitution attacks.
Terra
A cloud platform for biomedical research workspaces, data, workflows, and interactive Jupyter compute.
uv
A Python package and environment manager used by Heartwood to install locked analysis dependencies in a separate environment.
vLLM
A GPU-oriented inference server used by supported Heartwood NVIDIA deployments.
Wheel
A prebuilt Python package file (.whl) that can be installed without running a source-build backend.
ZIP64
An extension of the ZIP archive format that supports files and archives larger than the original ZIP size limits.