Skip to content
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Explore a trace

Give an agent a trace API link, and let it read a Playwright trace in small steps instead of loading the whole file. Get images of the trace for people: filmstrip, animation, frames and the page at th

This page works for traces from CI tests and from sessions an agent recorded. A Playwright trace records everything a test or a browser session did: every action, the page after each action, every network request, and a recording of the screen. A screenshot shows what the page looked like. The trace shows why.

A trace file is a zip of several megabytes, too large for an agent to read at once, and people need the Playwright trace viewer to open it. The Currents trace API solves both:

  • An agent reads the trace one small part at a time. It starts with a summary of the actions and the failure, then asks only for the page, request or screen that the summary points to.

  • People get images from the same API: a filmstrip, an animation, the screen at the failure, or the page as it looked then. Put them in a pull request or an issue. See Post evidence to GitHub, Slack and Linear.

Fix links, Shared Runs, the GitHub run comment, session pages, the MCP tool and the REST API all give a t.crts.sh trace API link. The agent reads its summary first, then only the snapshots, requests or frames it needs.
Where trace API links come from, and how an agent reads one

The trace can come from Playwright Test, Playwright MCP, Playwright CLI, or Vitest browser mode with the Playwright provider (browser.trace: 'on', which writes to __traces__ next to each test file). Attach a local trace to a session to get a link for it. Anyone with the link can open it without a Currents login.

You have
Where the trace API link is
It lasts

A failed test in CI

The Fix link of the test or run: an explore trace link next to each trace

As long as the Fix link

A failed test, and an agent connected to Currents

The failure context the agent gets from the Currents MCP server: the same explore trace link

12 hours

A CI run you shared

The Shared Run page and its Markdown version: a link for each attempt's trace

As long as the Shared Run

A pull request with a failed run

The GitHub run comment: a trace API link for each failed test

As long as the comment's Fix link

A session

The session page: Explore with agent on a trace. The session's Markdown version lists the link too.

As long as the Shared Session

Any test or session trace, from an agent

Ask an agent connected to the Currents MCP server for a trace API link

24 hours, or up to 7 days

Any test or session trace, from a script

The REST API: see Trace API

24 hours, or up to 7 days

For what each kind of link opens and who can open it, see Share links.

On a session page, Explore with agent on a trace copies a prompt with the trace API link, opens the trace summary, or opens the prompt in your coding tool:

A trace on a session page with the Explore with agent menu open: Copy prompt, Open trace index, and Open in Claude, Claude Code, Codex, Conductor, Cursor, GitHub Copilot or Zed
Explore with agent on a trace in a session

Ask the agent

Paste the Fix link of a failed test, then:

For a session where an API response looks wrong:

To show the failure to people:

Example: a checkout that freezes

A session recorded a checkout where the discount code save10 leaves the Apply button stuck. The agent got the trace API link from the session page with Explore with agent. Times are milliseconds from the start of the trace.

  1. The summary, /digest, lists the actions. No action failed, but the page threw a TypeError 33 ms after the click on Apply. About 450 tokens.

  2. The network requests, /requests: all 3 returned 200 before the click, so the server is not the cause. About 300 tokens.

  3. The page text right after the error, /snapshots/2464?format=text: the input holds save10 in lowercase, the button reads Applying…, and the total is still $229.50. About 100 tokens.

  4. The screen at the end, /frames/4024: the button is still stuck. An image of about 750 tokens.

The agent used about 1,600 tokens. Every file in the trace together is about 20,000 tokens.

What the agent reads

The agent asks the trace API for one part at a time: the summary, the page text, the accessibility tree, the HTML, the network requests and their bodies, or the files in the trace. Trace API lists each route, with real output from a failed test.

An agent in a terminal does this with curl. It reads each answer and picks the next call from it, until it can answer. Here it reads the trace of the checkout example above:

Animation. A terminal agent and the trace API exchange four calls in a loop. The agent asks for /digest and learns no action failed and a TypeError came 33 ms after the click on Apply. It asks for /requests and learns all 3 requests returned 200 before the click. It asks for /snapshots/2464 and sees the input holds lowercase save10, the button reads Applying… and the total is $229.50. It asks for /frames/4024 and gets an image of the button still stuck. It answers that save10 is not a key in DISCOUNT_CODES and the error stops the handler before it resets the button. It used about 1,600 tokens, of about 20,000 for the whole trace.
An agent finds the cause of the stuck checkout in four calls to the trace API

Available formats

The trace API returns each part of the trace in more than one format: Markdown and JSON for agents, images and pages for people. Below are three you are likely to post in a pull request or an issue. Add the path in each example to the trace API link to get the same output.

Trace API lists every route and option, with real output for each: the summary, the screen at any time, the page as text, accessibility tree or HTML, network requests and their bodies, the files in the trace, and the trace zip.

Filmstrip

/filmstrip?end=failure&layout=row&width=480 returns the frames up to the failure in one WebP image, and outlines the last one in red. That is the frame /frames/failure returns: the last frame before the failure, which can be seconds earlier while an assertion waits. From a failed test that expected a total of $206.55 and got $183.60:

Three frames of the checkout in one row: save10 typed into the discount field with a total of $229.50, then the same, then outlined in red: SAVE10 applied: 10% off, a discount of −$22.95 and a total of $183.60
/filmstrip?end=failure&layout=row&width=480: the failed test up to the failure

The page at the failure

/snapshots/failure?format=html returns the page when the test failed, with the styles from the trace. The trace API removes the scripts, so buttons and links do nothing. The same failed test, opened in a browser:

The checkout page rebuilt from the DOM snapshot at the failure: save10 in the discount field, SAVE10 applied: 10% off, a discount of −$22.95 and a total of $183.60
/snapshots/failure?format=html: the page when the test failed, opened in a browser

Animation

/animation?format=gif&width=640 returns the recording as a GIF, 640 pixels wide. A test that runs for a few seconds has only a few frames, so this one comes from the checkout session above:

An animation of the checkout: save10 is typed into the discount field, Apply is clicked, and the button stays on Applying… while the total stays at $229.50
/animation?format=gif&width=640: the checkout session as a GIF

GitHub shows these images in Markdown. Linear and Slack do not. See Post evidence to GitHub, Slack and Linear.

Last updated

Was this helpful?