Glossary

One line per term, in alphabetical order. Each links the page that explains it.

Term	Meaning
Approval	A reviewer's yes or no on one tool call before it runs. See Approval.
Approval id	The id of a pending approval request; you resume with it. See Human in the loop.
Approval port	The object that decides approvals: AutoApprove, AutoDeny, CallbackApproval, DeferredApproval, or your own ApprovalPort. See Approval.
Approval store	Where pending approvals and decisions live: memory, SQLite or Redis. See Approval stores.
Bridge	Integration code that maps a framework's own approval or tool protocol onto the executor. See Integrations.
Cache scope	How widely a cached tool result is shared: run, thread or global. See Cache.
Call id	The id of one tool call (ToolContext.call_id). See Context.
Confirm rule	A tool's confirm setting: True, False or a function that decides per call whether approval is needed. See Approval.
Context	The ToolContext passed to a tool: run, thread and call ids, state, ledger, cache and event sink. See Context, events and ports.
Decision	An ApprovalDecision: approved or not, who decided, and why. Decision is an alias. See Approval.
Event	A structured record of something that happened (call started, call failed, ...). See Tracing.
Event sink	Where events go: logs, a callback, OpenTelemetry. See Tracing.
Executor	Runs tool calls: validation, hooks, approval, retries, timeouts, cache and limits. See Executor.
Flow	A graph of nodes the FlowRunner runs, or compiles for a framework. See Flows.
Flow pause	A flow stops at a node that waits for approval and resumes later from its checkpoint. See Flow pause and resume.
Hook	Code that runs before, after or on error of a tool, model or node call. See Hooks.
Importer	A function that turns foreign tools into ToolSpecs: import_mcp, import_openapi, import_langchain. See Integrations.
Integration	A module in agent_adapter.integrations that exports tools to one model API or framework. See Integrations.
Kind (error)	The stable id of an error, such as tool-not-found. See the error catalogue.
Ledger	Records the cost of every model and tool call as LedgerEntry rows. See Cost.
One-use token	An approval the adapter issues after its own port approved, bound to one run and call and spent once. A framework's own approval never counts alone. See Approval.
Port	A protocol you implement to plug in a backend: LLMPort, CachePort, LedgerPort, EventSink, ApprovalPort, EmbeddingPort. See Context, events and ports.
Portable schema	A JSON schema every supported provider accepts; SchemaNotPortable says what breaks it. See Schemas.
Price	The per-token cost of a model, from Prices. See Cost.
Principal	An authenticated HTTP caller; its id scopes the runs, calls and approvals it can see. See HTTP server.
Problem	An RFC 9457 error body the HTTP server returns, with the error's kind in type. See HTTP server.
Registry	Holds the tools you define and exports them to every target. See Registry.
Run id	The id shared by every call in one agent run. Run limits count per run. See Context.
ShortCircuit	Raised by a hook to return a result without calling the tool or model. See Hooks.
Spec	A ToolSpec: a tool's name, description, argument model and settings. See Tools.
Strict mode	OpenAI schema mode: every property is required, so an optional argument becomes nullable and null means "not given". See Strict mode and nulls.
Target	A model API or framework a tool is exported to; targets limits a tool to some of them. See Registry.
Thread id	The id of a conversation that spans several runs. See Context.
Tool	A Python function registered with @tool or wrap, callable by any model or framework. See Tools.
Tool loop	Model call, tool calls, results back to the model, repeat: run_tool_loop. See Tool loop.