Glossary (26)
Terms that recur across the corpus, defined in one line each, with see-also chains. Hover any dotted-underlined word elsewhere on the site to peek the definition without leaving the page.
Come here when a word in a concept page made you stop.
-
Adapter
aka provider adapterA class that translates one provider's wire format to a uniform internal API. Hand-rolled adapters are an alternative to LiteLLM.
-
Allowlist
aka whitelistDefault-deny set of permitted hosts, paths, or tool names. Prefer over denylists in security-critical paths.
-
cache_control
aka prompt caching breakpointAnthropic's `cache_control: {type: 'ephemeral'}` marker that designates a content block as cacheable; up to 4 breakpoints per request.
-
Condenser
aka memory condenserOpenHands' name for a memory-compression strategy. Pluggable: LLMSummarizingCondenser, NoopCondenser, etc.
-
Delegate
aka delegationTool-shaped pattern where the parent calls `delegate(task=...)`; the parent pauses, the child runs, the child's result returns as a `tool_result`.
-
Docker
aka containerOS-level virtualization used to bound the agent's filesystem and processes. Most agent sandboxes are a Docker image.
-
Function calling
aka tool_callsOpenAI-style format where the assistant emits a sibling `tool_calls[]` array with arguments as a JSON string.
-
gVisor
User-space kernel from Google; stronger isolation than plain Docker; used when syscall-level isolation matters (cf. Firecracker).
-
LiteLLM
Python library that unifies many provider APIs behind one client interface; used by OpenHands and Strix.
-
Mailbox
aka inter-agent queueIn-process structure (often a dict of lists) that holds pending messages between concurrent agents.
-
MCP
aka Model Context ProtocolOpen protocol from Anthropic for surfacing tools and resources to LLMs over a uniform JSON-RPC interface.
-
Microagent
OpenHands' name for trigger-loaded skill files. Functionally identical to Strix's `skills/`.
-
Partial JSON
aka streaming JSON parserA forgiving JSON parser that returns the partial structure of an incomplete JSON document; needed when tool args stream as deltas.
-
Prompt injection
Attack where adversarial input redirects the model away from its intended instructions. Mitigated by privileged-source system prompts and structured-input boundaries.
-
Queue
aka message queueExternal durable message channel (Redis, SQS). Overkill for in-process; required for distributed agents.
-
ReAct
Reason-then-Act prompting pattern: the model writes a thought, then a tool call, then an observation, then repeats. Predecessor to native tool calling.
-
Reasoning
aka extended thinkingHidden or partially-visible step where the model thinks before answering. Anthropic exposes signed blocks; OpenAI bills opaque tokens.
-
seccomp
aka AppArmorLinux mechanism for restricting which syscalls a process can make. Often layered onto Docker.
-
security_risk
A `Literal["low","medium","high"]` enum field on tool args where the model self-declares risk; the controller routes accordingly.
-
Sentinel string
A literal token (e.g. `__SYSTEM_PROMPT_DYNAMIC_BOUNDARY__`) used as a refactor-safe split point in concatenated prompts.
-
Skill
A markdown file in `skills/` that encodes one methodology, with a trigger and a body. Loaded into the system prompt when matched.
-
SSE
aka Server-Sent EventsStreaming format: server pushes line-delimited events over a persistent HTTP connection. The transport layer for most LLM streaming APIs.
-
Subagent
aka child agent, delegateAn agent spawned by a parent to handle a delegated subtask, usually with a fresh context window.
-
Thinking block
aka reasoning blockAnthropic's structured content block containing a model's internal reasoning. Carries an HMAC `signature` that is model-bound.
-
tool_use
aka tool calling (Anthropic)Anthropic's structured content-block format for function calls; the assistant emits a `tool_use` block, the user replies with a `tool_result` block.
-
Trajectory
aka session, rolloutThe full sequence of an agent's actions and observations in one session. Stored in event-sourced systems for replay/training.