Events

A run is a flat stream of events, each addressed by run_id and span_id and ordered by seq. The tree is rebuilt from parent_span_id. Every SDK, the OpenTelemetry endpoint and a hand-written curl send the same five event types.

Does ingest accept this batch?

Drop run_idseq: -1type: "chat"schema_version: 2Reset

A batch as the SDK sends it (first three events of anthropic-py.json). Edit anything.

{ "events": [ { "error": null, "input": "for [redacted:email]", "metadata": {}, "name": "py-anthropic", "output": null, "parent_span_id": null, "phase": "start", "run_id": "01M4G4WC0QX80S2SNE1RJH1KG5", "schema_version": 1, "seq": 0, "span_id": "01M4G4WC0QX80S2SNE1RJH1KG6", "status": "running", "ts_end": null, "ts_start": "2026-10-09T10:55:57.463Z", "type": "run" }, { "error": null, "latency_ms": 55, "metadata": { "framework": "anthropic" }, "model": { "model_id": "claude-mock-1", "provider": "anthropic" }, "parent_span_id": "01M4G4WC0QX80S2SNE1RJH1KG6", "request": { "messages": [ { "content": "Weather in Paris? cc [redacted:email]", "role": "user" } ], "params": { "max_output_tokens": 256 }, "system": "You are a weather bot.", "tools": [ { "description": "Weather for a city", "name": "get_weather", "parameters": { "properties": { "city": { "type": "string" } }, "required": [ "city" ], "type": "object" } } ] }, "response": { "finish_reason": "tool_use", "reasoning": null, "text": null, "tool_calls": [ { "input": { "city": "Paris" }, "tool_call_id": "toolu_1", "tool_name": "get_weather" } ] }, "run_id": "01M4G4WC0QX80S2SNE1RJH1KG5", "schema_version": 1, "seq": 1, "span_id": "01M4G4WECN5KHERP8QJ5AEWW7D", "ts_end": "2026-10-09T10:55:59.893Z", "ts_start": "2026-10-09T10:55:59.838Z", "type": "llm", "usage": { "input_tokens": 12, "output_tokens": 5, "total_tokens": 17 } }, { "error": null, "input": { "city": "Paris" }, "latency_ms": null, "output": { "forecast": "sunny" }, "parent_span_id": "01M4G4WECN5KHERP8QJ5AEWW7D", "run_id": "01M4G4WC0QX80S2SNE1RJH1KG5", "schema_version": 1, "seq": 2, "span_id": "01M4G4WEDHWTQG91S7PBGMJDSR", "tool_call_id": "toolu_1", "tool_name": "get_weather", "ts_end": "2026-10-09T10:55:59.921Z", "ts_start": "2026-10-09T10:55:59.920Z", "type": "tool" } ] }

POST /api/ingest would

Accept.

3 events pass the schema ingest validates with.

#Fields

Generated from Runback's event schema, so it always matches what ingest accepts. The record format an auditor verifies, with canonicalisation and the six checks, is the open spec at /spec.

Common to every event

Field	Type	What it is
schema_version required	1	The wire-format version. Always 1.
run_id required	string	The run this event belongs to: one agent invocation.
span_id required	string	This event's own id: unique, and sortable by time (a ULID).
parent_span_id required	string | null	The event this one sits under, or null for the top of the run.
seq required	integer	This event's position in the run. Sort by this, not by timestamp: it survives clock skew.
ts_start required	string	When the event started, in ISO 8601.
ts_end required	string | null	When it ended, in ISO 8601, or null while it is still running.
actor .type "user" | "api_key" | "system" .id string .label string, optional	object	Optional: who or what caused this event, a person, an API key or a schedule. Reported by your code, once per collector or per call.

type: "llm"

Field	Type	What it is
type required	"llm"	Which kind of event this is.
model required .provider string .model_id string	object	The provider and model id that was called.
request required .system string | null .messages object[] .tools object[] .params object	object	Exactly what the model was sent (system prompt, messages, tools and parameters), captured as sent rather than rebuilt from the reply.
response required .text string | null .reasoning string | null .finish_reason string | null .tool_calls object[]	object	What the model returned: its text, reasoning, finish reason and any tool calls it asked for.
usage required	object | null	Input, output and total tokens, or null if the provider did not report them.
latency_ms required	number | null	How long it took, in milliseconds, or null if not measured.
error required	object | null	The error the call raised (name, message, stack), or null.
key_projection .keep string[], optional .drop string[], optional	object	Optional: which request fields count when this call is matched across runs. Fields under drop, such as request ids and timestamps, are ignored; keep limits matching to the fields it names.
metadata	object	Optional free-form context from your code, such as the graph node a call ran inside.

type: "tool"

Field	Type	What it is
type required	"tool"	Which kind of event this is.
tool_name required	string	The tool that was called.
tool_call_id required	string	Links this call to the model's tool-call request that asked for it.
input	any	The arguments the tool was called with.
output	any | null	What the tool returned, or null if it failed or is still running.
latency_ms required	number | null	How long it took, in milliseconds, or null if not measured.
error required	object | null	The error the tool raised (name, message, stack), or null.
key_projection .keep string[], optional .drop string[], optional	object	Optional: which input fields count when this call is matched across runs. Fields under drop are ignored; keep limits matching to the fields it names.
policy_block .rule string .detail string	object	Set when a rule stopped this call before it ran: the tool never ran and output is null. Names the rule and why.
policy_evaluated .passed boolean	object	Set whenever a rule check ran on this call before it executed. passed is false when the call was stopped.
metadata	object	Optional free-form context from your code, such as the graph node a call ran inside.

type: "reasoning"

Field	Type	What it is
type required	"reasoning"	Which kind of event this is.
text required	string	The reasoning or log text.
label required	string | null	A short label for the step, or null.
graph_node .name string .step number, optional .graph_name string, optional	object	Optional: the graph node this step belongs to, when your framework reports one.
routed_from	string[]	Optional: the node or nodes that routed here.

type: "run"

Field	Type	What it is
type required	"run"	Which kind of event this is.
phase required	"start" | "end"	Either start, when the run begins, or end, when it finishes.
name required	string	The run's name.
input	any | null	What the run was asked to do, or null.
output	any | null	What the run produced, or null.
status required	"running" | "success" | "error" | null	One of running, success or error, or null when not reported.
error required	object | null	The error the run ended with (name, message, stack), or null.
metadata required	object	Free-form context from your code, such as your agent's version.

type: "env"

Field	Type	What it is
type required	"env"	Which kind of event this is.
kind required	"now" | "date" | "random" | "uuid" | "fetch"	What was read: the clock (now, date), a random number, a generated UUID, or a network fetch.
key required	string	What identifies the read: the method and URL for a fetch, otherwise the kind.
output	any	The value your code received, recorded so a replay hands back the same one.