Run trace
A React component in Beamline's AI & agents category.
One agent run as a waterfall: every step (model calls, tools, retrieval, checks, sub-agents) nested under the step that started it, each a bar on a shared time axis, with retries as separate attempts on the same bar, tokens and cost per step and for the run; while it runs the bars grow live, and choosing a step shows its timing, attempts, input and output beside the trace.
Use it for
- Debugging or reviewing an agent run: what ran, in what order, what waited on what, what failed and was retried, and what it cost.
- An observability page for model and tool calls (spans from your tracing).
Not for
- A single tool call in a conversation (tool-call).
- The agent's thinking in words (reasoning).
- Background jobs a person follows (task-rows).
- The designed workflow before it runs (flowchart).
Anatomy
- header: the run's status (a glyph and a word), total time, how many steps and how many failed, tokens in and out, cost
- axis: time marks across the lane (0, 2 s, 4 s…), a line at now while running
- row: indent and a chevron for steps with children, the kind's glyph (model, tool, retrieval, check, agent, wait), the name and the model; the bar on the lane from start to end in the kind's colour; failed attempts before it as hatched segments with a gap; the duration at the end
- selected step: the row ringed and a detail panel (beside the trace when wide, under it when narrow): name, kind, start and duration, attempts with their errors, tokens and cost, input and output
- running step: the bar's open end grows with time and carries a soft edge; the duration counts
Variants
- density: compact, default
States
- running (live)
- succeeded
- failed (the failing step and its parents are marked)
- retried
- cancelled
- collapsed
- selected
- empty (no steps yet)
Keyboard
- The trace is a tree grid with one tab stop: ↑/↓ move between visible steps, → opens a step or goes to its first child, ← closes it or goes to its parent
- Home/End go to the first and last step; Enter or Space shows the step's detail
- Escape clears the selection
Motion
New steps grow into the list and the rows below slide down (useFlip). While the run is live the open bars grow on a clock that runs only while something is running and the trace is on screen, then stops. Opening and closing a step folds its children. The detail cross-fades to the chosen step. Reduced motion: bars jump to their length each tick of the clock; nothing slides.
Props
| Prop | Type | Default | Description |
|---|---|---|---|
spans (required) | TraceSpan[] | — | |
now | number | — | ms since the run started, for live bars; omit to let the trace keep its own clock while a span is running. |
title | ReactNode | — | The run's name (the request it answered). |
selected / defaultSelected / onSelectedChange | string | null | — | |
expanded / defaultExpanded / onExpandedChange | string[] | — | Open steps. Default: every step. |
density | "compact" | "default" | — | |
currency | string | — | Cost symbol. Default "$". |
className / classNames / style / ref | root | — |
Built from
- spinner
Dependencies
lucide-react
Import
import { RunTrace } from "@/components/run-trace/run-trace";
Get it
Run trace is part of Beamline: 200+ components and 50+ complete screens for React that your coding agent (Claude Code, Codex, Cursor or any MCP client) installs into your app through Beamline's MCP server, as a ready-built package or as plain React source you can change. $49 launch price, one payment, no subscription, lifetime updates, one licence for your whole team. Get Beamline · Connect your agent
More in AI & agents
- Image generation: The frame an AI image is generated into: a field of brand-hued pixels breathes with the progress while the model works, then the image arrives —…
- Flowchart: A workflow as steps on a dotted canvas: a trigger, conditions that branch into paths side by side, actions and waits, read top to bottom; during a run…
- Answer compare: Two answers to the same prompt side by side, and a person's verdict: A is better, a tie, B is better, or both are bad, with the reasons that decided it…
- Response: An assistant's answer as streamed markdown in the system's reading type: bursty tokens flow in at an even pace, each new word fades in through a short…
- Chat thread: People talking to people: messages grouped by person and day, reactions, delivery and read receipts that glide to the newest message read, who is typing…
- Model picker: Choose the AI model for a request by what it can do and what it costs: a trigger with the model's name opens a searchable list grouped by provider, with…
- Task rows: Background work an agent is doing, one row per task: status as a glyph and a word, what it produced, how long it has run or took, a progress bar while…
- Conversation: The scrolling column of a conversation: follows new content while you are at the bottom, stops when you scroll up to read, and offers a way back; shared…
- Context cards: The knowledge an agent retrieved before it answered: one card per chunk with what it is about, the text it read with the search words marked, how…
- Prompt suggestions: Ready-made prompts a person can start from or follow up with: chips, cards or a list; a picked one lifts off and flies into the composer as its text…
- Plan: An agent's plan: numbered steps with live status, progress, and Approve / Edit / Reject when it needs sign-off; approving shows progress, and a refusal…
- Approval card: The agent stops and asks before it acts: what it wants to do, exactly what that changes, any choices it needs, and Approve or Reject; both run your…
- Tool call: One tool call in an agent's run: status, what it did and a short result on one quiet row that opens to the exact input and output; stops for permission…
- Recommendation card: One suggestion from the agent that the person may take or leave: what it suggests, why, the expected effect and how sure it is (a meter and a word…
- Citations: Inline numbered citations that preview their source on hover or focus, plus a numbered source list; streamed answers flow in at an even pace with the…
- Reasoning: The agent's trace, fed by your stream: a shimmering label and the seconds so far while it works, then "Thought for N seconds"; steps that check off…
- Attachment: A file tile: preview or typed glyph with the type named, name, size, upload progress, error with retry, and a keyboard-reachable remove.
- Prompt bar: The agent composer: a field that grows, @ to add a source, / for a command, a model picker, files and dictation, and one button that sends with progress…
- Selection actions: Select a passage and hand it to the agent: a bar over the selection with quick actions, your own instruction and More; while the agent works it says so…
- Message: One turn in a conversation with an agent: the person's words in a quiet bubble, the agent's answer on the page with its orb mark, trace, tool calls and…