Runtime Lifecycle

At a glance

A single agent.stream() call involves 3 API layers, 14 callbacks, 4 message chains, and 2 pairs of same-named callbacks. This page pins each one to the timeline.

Key number	Value
Total callbacks	14 (deduplicated)
Same-named callback pairs	2 (`onStepEnd` × 2 / `onEnd` × 2)
stopWhen default (L1)	`isStepCount(20)`
stopWhen default (L2 `streamText`)	`isStepCount(1)`
Timeout tiers	3 (`totalMs` / `stepMs` / `chunkMs`)
Pinned SDK version	`[email protected]`

Three-layer capability matrix

Build the map before reading the timeline — for every parameter/callback, know which layer it belongs to:

	L1: `new ToolLoopAgent({...})`	L2: `agent.stream({...})`	L3: `result.toUIMessageStream({...})`
Role	Static config (define agent)	Single invocation (trigger one run)	Downstream consumer (transform result to UI stream)
Lifecycle	Constructed once, reused	Called once per invocation	Called once per invocation
Structural params	`id`, `model`, `instructions`, `tools`, `runtimeContext`, `toolsContext`, `providerOptions`	`messages` / `prompt`, `abortSignal`, `timeout`	`originalMessages`, `generateMessageId`, `sendReasoning`, `sendSources`, `sendStart`, `sendFinish`
Behavior hooks	`stopWhen`, `prepareStep`, `prepareCall`	`experimental_transform`	—
Callbacks (firing order)	`onStart` → `prepareStep` → `onStepStart` → `onLanguageModelCallStart` → `onLanguageModelCallEnd` → `onToolExecutionStart` → `onToolExecutionEnd` → `onStepEnd` → `onEnd`	Same as L1 (merged with L1 same-named callbacks — both fire in parallel)	`messageMetadata` → `onStepEnd` → `onEnd` → `onError`

L1 and L2 same-named callbacks are merged: if both layers set onStepEnd, they fire in parallel — v7 wraps them with mergeCallbacks, which awaits both via Promise.allSettled (source: agent/tool-loop-agent.ts:285-315, util/merge-callbacks.ts:16; v6 fired L1 then L2 sequentially). L3’s same-named callbacks are entirely independent — they don’t merge with L1/L2 and have different payloads. (v7 renamed the canonical callbacks onStepFinish → onStepEnd and onFinish → onEnd at both the L1/L2 and L3 layers; the v6 names live on as deprecated aliases, resolved via a ?? fallback.)

L3’s other entry point: this page’s L3 column focuses on the toUIMessageStream({ stream }) transform path (passively consuming the agent’s result). v7 canonical is the standalone toUIMessageStream({ stream: result.stream }) helper; the result.toUIMessageStream() method is a deprecated alias. L3 has another entry point — createUIMessageStream({ execute }) — the execute-driven path, which can actively push custom events and merge multiple streams alongside the agent’s. Both share the same underlying handleUIMessageStreamFinish (ui-message-stream/handle-ui-message-stream-finish.ts), so onStepEnd / onEnd / onError fire at the same moments described on this page; but messageMetadata exists only on toUIMessageStream. The execute-driven path is covered in detail in UI Stream Orchestration.

Full timeline diagram

A single N-step agent.stream() call, time flows downward:

Three critical observations:

L1/L2 callbacks and L3 callbacks run concurrently — L2 pushes chunks to stream while L3’s pipe transforms them. So L1/L2 onStepEnd(n) and L3 onStepEnd(n) happen nearly simultaneously, but as independent event-loop tasks.
L3 onEnd always fires later than L1/L2 onEnd — L3 is a downstream transform, it must wait for stream close + consumer drain before flushing. For “after-run” work: use L1/L2 onEnd for engine-side cleanup (token tallying, sandbox close), L3 onEnd for UI-side persistence (saving the assistant message).
messageMetadata runs per chunk — including every text-delta and tool-input-delta. A long multi-tool response can easily emit 1000+ chunks; any synchronous I/O here directly stalls the stream.

Callback firing reference

In firing order. L1 = ToolLoopAgent settings, L2 = streamText (direct pass-through from agent.stream), L3 = toUIMessageStream.

#	Callback	Layer	When	Payload	Use for
1	`prepareCall(baseCallArgs)`	L1	Before each `agent.stream()` starts, after params merge	Full call args, returns overrides	Dynamic model/tools/stopWhen rewrite
2	`onStart()`	L1+L2	After streamText starts, before first step	None	Init logging/timing
3	`prepareStep({...})`	L1+L2	Before each step’s model call	`{ messages, steps, stepNumber, model }`	Compaction, reminder injection, activeTools filtering, model switching
4	`onStepStart()`	L1+L2	Before each step’s model stream (after `prepareStep`)	None	Per-step timing marker
5	`onLanguageModelCallStart()`	L1+L2	Before each provider model call begins	`{ ... }` (call metadata)	Provider-call timing marker, request logging
6	`onLanguageModelCallEnd()`	L1+L2	After each provider model call response is normalized	`{ ... }` (call result)	Provider-call latency, raw-response observability
7	`onToolExecutionStart()`	L1+L2	Before each `tool.execute`	`{ toolCall }`	Permission audit, pre-retry logic
8	`onToolExecutionEnd()`	L1+L2	After each `tool.execute`	`{ toolCall, toolResult }`	Observability, cache writeback
9	`onStepEnd(stepResult)`	L1+L2	Each step end, after `finish-step` chunk emitted	`StepResult`: full step detail	Token tallying, step-level persistence
10	`messageMetadata({ part })`	L3	Each chunk passing through UI transform	`{ part }` (current chunk)	Attach metadata to UI control chunks
11	`onStepEnd`	L3	Each `finish-step` chunk passing through UI transform	`{ responseMessage, messages, isContinuation }`	UI-side step-level persistence
12	`onEnd({...})`	L1+L2	After all steps done, after `finish` chunk emitted	`{ finishReason, usage, steps, ... }`	Engine-side settlement, cleanup
13	`onEnd({...})`	L3	After UI stream drain / cancel	`{ responseMessage, messages, isContinuation, isAborted, finishReason }`	UI-side message persistence
14	`onError(error)`	L3	UI transform error / error chunk / `onStepEnd` throw	`Error` or string	SSE error serialization; the returned string is written into the error chunk’s `errorText` field sent to the client

onEnd throws do NOT route here: the onEnd call (ui-message-stream/handle-ui-message-stream-finish.ts — callOnEnd, ~L120-137) is a bare await with no try/catch. An onEnd throw propagates out through TransformStream’s flush(), rejecting the consumer iterator — it does not invoke onError. Any production onEnd must wrap its own try/catch/finally inside the callback. Full error-capture tiering in UI Stream Orchestration — Error capture, in full.

Same-named callbacks — the biggest trap

These are the v7 canonical names onStepEnd / onEnd (v6: onStepFinish / onFinish, still accepted as deprecated aliases). The “two same-named callbacks fire twice” trap is unchanged — it’s just that the shared name at both layers is now onEnd (and onStepEnd).

`onStepEnd`: L1/L2 vs L3

	L1/L2 (streamText)	L3 (UI stream)
When	Step loop ends, after `finish-step` emit	`finish-step` chunk through UI transform
Payload	`StepResult`: `{ stepNumber, content, text, toolCalls, toolResults, finishReason, usage, response, request, ... }`	`{ responseMessage: UIMessage, messages: UIMessage[], isContinuation }`
What you see	Engine view: raw step output (tool call objects, usage breakdown)	Consumer view: accumulated UI message (assistant message structure)
Use for	Token tallying (billing), step-level logging, driving compaction / context trimming	Incremental UI message persistence, prefetching

`onEnd`: L1/L2 vs L3

	L1/L2 (streamText)	L3 (UI stream)
When	After `finish` chunk emit, before stream close	After stream close + UI transform flush
Payload	`{ finishReason, usage, steps, content, text, reasoningText, toolCalls, toolResults, response, request, warnings, providerMetadata }`	`{ responseMessage, messages, isContinuation, isAborted, finishReason }`
Ordering	Earlier (upstream)	Later (downstream drain)
Use for	Engine-level one-shot settlement: write total usage to DB, close sandbox, commit compaction checkpoint	UI-level one-shot settlement: persist final assistant message, notify client of completion

L1/L2 onEnd closure trap: the L1/L2 onEnd payload carries the entire steps array — every step’s full StepResult (content, toolCalls, toolResults, request, response, all of it). If your callback closure captures this payload and pins it on a long-lived reference (e.g. storing it on an outer session object), the entire large object graph from this invocation is held and never GC’d. Long chat conversations amplify this — 20 steps of cumulative StepResult easily reaches hundreds of MB.

Practical guidance:

Short settlement logic (token counting, step logs) can live in L1/L2 onEnd — closure releases right after the callback returns
Long settlement logic (persistence, background jobs, checkpoint writes) should prefer L3 onEnd; its payload is the folded responseMessage + messages, orders of magnitude smaller
Or: use L1/L2 onStepEnd to incrementally collect only what you need into a small variable (numbers / strings / ids only, never the StepResult itself), then settle on that small variable in L3 onEnd

stopWhen default chain — the second trap

Same parameter name, different defaults at L1 and L2:

L1 new ToolLoopAgent({ stopWhen? })  default: isStepCount(20)   ← agent/tool-loop-agent.ts:132
L2 streamText({ stopWhen? })         default: isStepCount(1)    ← generate-text/stream-text.ts:347

ToolLoopAgent.stream() forwards L1’s stopWhen (default 20) to streamText, so the normal path caps the agent at 20 steps.

But if you call streamText(...) directly without setting stopWhen, your agent runs for exactly one step — one tool call and it halts. Classic beginner trap.

Built-in stop conditions (combinable: stopWhen: [isStepCount(N), hasToolCall('complete')]):

Factory	Semantics
`isStepCount(N)`	Stop after reaching step N
`hasToolCall(name)`	Stop after a tool call with the given name

Typical production combo: [isStepCount(N), hasToolCall('complete')] — the number is a hard ceiling (prevents the agent from spinning in a loop), hasToolCall('complete') is the “task finished” signal (lets the agent declare its own end). Pick N based on task complexity: simple Q&A 10-20, general assistant 30-50, deep research / multi-step editing 50-100.

Three-tier timeout — the third trap

timeout is an object with three granularities:

agent.stream({
  timeout: {
    totalMs: 600_000, // Entire invocation: 10 min
    stepMs: 300_000, // Single step: 5 min (model + tools combined)
    chunkMs: 120_000, // Gap between two chunks: 2 min
  },
});

All three are independent timers, but v7 folds them into one abort signal via mergeAbortSignals(...) (generate-text/stream-text.ts:732-737) — any one tripping aborts the whole stream, and the merged signal is also handed to the running tool. The tool-execution consequences of chunkMs specifically (why it can abort a long silent tool, and the heartbeat fix) are covered in Stream Timeouts & Tool Heartbeat.

Dimension	Watches for	Typical scenario
`totalMs`	Absolute call duration	Long-task total ceiling
`stepMs`	Single step from prepareStep to finish-step	Model response stalls
`chunkMs`	Gap between two adjacent chunks	Mid-stream hang (TCP half-open, slow provider thinking phase)

Tutorials usually only mention totalMs — but chunkMs is the lifesaver in production. A model stream can emit two lines then hang (TCP half-open from the provider, a thinking phase taking too long, etc.); totalMs is nowhere near, but chunkMs terminates it immediately. A solid default in production is chunkMs: 120_000 (2 minutes) — enough to cover long reasoning-model thinking gaps, but not so long that a real disconnect quietly hangs forever.

`prepareCall` — the little-known upstream hook

Besides prepareStep (per-step), L1 also has prepareCall (per-invocation):

new ToolLoopAgent({
  model,
  instructions,
  tools,
  prepareCall: async (baseCallArgs) => {
    // baseCallArgs = all merged args (settings + method options)
    // Return overrides — or return nothing to use baseCallArgs as-is
    return {
      ...baseCallArgs,
      tools: dynamicallyDecideTools(baseCallArgs.messages),
      stopWhen: isStepCount(deriveStepLimit(user)),
    };
  },
});

When to use:

Dynamically swap tool sets (user tier / A-B test) without rebuilding the agent instance
Pick a model based on the call’s input
Set stopWhen per call

prepareCall vs prepareStep:

	`prepareCall`	`prepareStep`
Layer	L1	L1 or L2
Invocation count	Once per `agent.stream()`	Once per step (N times per invocation)
Can override	All call args (tools, stopWhen, instructions, messages, model, …)	Per-step args (messages, system, model, toolChoice, activeTools)
Use for	Static config dynamization	Runtime context adaptation (compaction, reminders, dynamic tool sets)

Picking between them: most projects only need prepareStep — runtime compaction, per-step reminder injection, swapping tool sets based on conversation length, these are all per-step scenarios. prepareCall makes more sense for deployments where the same agent instance is reused across requests (the agent is constructed once at module top level, and each HTTP request rewrites call args based on user tier / AB experiments). If your agent is reconstructed per request (the orchestrator layer already does config dynamization), prepareCall is redundant.

The four step/tool lifecycle hooks

In v7 these are canonical — no experimental_ prefix. The v6 names (experimental_onStart, experimental_onStepStart, experimental_onToolCallStart, experimental_onToolCallFinish) remain as deprecated aliases, resolved via a ?? fallback in generate-text/stream-text.ts:718-724. Note the rename of the tool pair: onToolCall{Start,Finish} → onToolExecution{Start,End}. (v7 also promoted onLanguageModelCallStart / onLanguageModelCallEnd out of experimental_ — the provider-call pair added to the timeline above.) They’re core building blocks for observable agents:

Hook	Use for	Typical landing scenarios
`onStart`	Call-level “begin” marker	Emit a UI init event, start the whole-call timer, log run start
`onStepStart`	Step-level “begin” marker	Reset step counter, clear per-step buffer, start per-step timer
`onToolExecutionStart`	Tool-level “begin” marker	Audit log, permission check, blocking validation
`onToolExecutionEnd`	Tool-level “done” marker	Result caching, metric reporting, tool-level error routing

Difference from onStepEnd: onStepEnd is the aggregation callback at step end (full StepResult); onStepStart is the transition point at step start (no payload, pure signal). For “inter-step cleanup” use the former; for “step initialization” use the latter.