| name | tracy-debugging |
| description | Invoke when fetching web pages from localhost, debugging PHP errors, or interpreting Tracy output (BlueScreen, Tracy Bar, dump). Read BEFORE running curl or Chrome to any local development PHP URL – with Tracy >= 2.12 and a detected agent, Tracy mirrors BlueScreen, Tracy Bar and dumps as markdown into the JS console for easy machine reading. For Chrome MCP, call list_console_messages() to read Tracy output. Essential when: 500 error, blank page, PHP exception, slow page, N+1 queries, or inspecting variables with dump().
|
Tracy Debugging
Tracy is enabled automatically in debug mode. In debug mode, Tracy displays errors in the browser; in production mode, errors are logged to the log/ directory and the user sees a generic error page.
Requires Tracy ≥ 2.12 for the markdown-to-console output described below. With older versions, only the visual HTML BlueScreen and Tracy Bar are available — you will need to scrape HTML instead.
Enabling Debug Mode
Debug mode is controlled in app/Bootstrap.php:
$this->configurator->setDebugMode('secret@23.75.345.200');
$this->configurator->setDebugMode(true);
$this->configurator->enableTracy($this->rootDir . '/log');
In production, Tracy logs exceptions to log/exception-*.html files (full BlueScreen snapshots) and errors to log/error.log. Tracy ≥ 2.12 additionally writes a .md sibling next to every .html exception — see Batch reading production exceptions below.
How Tracy detects an agent
The markdown-to-console output is conditional — it activates only when Tracy thinks an AI/automation agent is reading the page. Detection works in two stages:
- JavaScript stage. The Tracy Bar's own JavaScript checks
navigator.webdriver. When true (which Chrome and Firefox set automatically under WebDriver, Chrome DevTools Protocol, Playwright, Puppeteer), it sets a cookie tracy-webdriver=1;path=/;SameSite=Lax.
- PHP stage. On every request, PHP calls
Tracy\Helpers::isAgent(), which checks $_COOKIE['tracy-webdriver'] === '1'. If yes, markdown outputs are emitted alongside the normal HTML rendering.
Implication for curl / API testing: curl does not run JavaScript, so it never sets the cookie itself. To get markdown outputs in a curl response, set the cookie manually:
curl --cookie 'tracy-webdriver=1' https://example.l/admin/products
Without that cookie, Tracy still works normally for humans (red BlueScreen, visual Tracy Bar in the corner), but the markdown JSON inside <script>console.log(...)</script> / <script>console.error(...)</script> is not emitted.
Chrome MCP — reading Tracy output
When using Chrome MCP (or any browser-automation MCP server: Playwright MCP, mcp-chrome, Puppeteer-based, Browser Use), navigator.webdriver is true, the cookie is set automatically after the first page load, and from then on every page response includes markdown for the agent. Read it all with one call:
list_console_messages()
What you get:
[error] — BlueScreen error report (full markdown with stack trace, source snippets, arguments, environment). Only present when an exception occurred.
[log] — Tracy Bar summary (execution time, memory, plus any panel info like warnings, SQL queries, dumps). Present on every page rendered while the cookie is set.
[log] for individual dump() calls — each dump($var) in PHP also pushes a plain-text version through Helpers::consoleLog() for the agent.
[error] on a production 500 page — Tracy ≥ 2.12 prints Tracy: Server Error 500. Details have been logged on the server. so the agent knows where to look even when the visible page is the generic placeholder.
The page title on a BlueScreen is set via JavaScript to Tracy: <ExceptionType>: <message>[ #<code>] where <ExceptionType> is the actual exception class or PHP error severity (RuntimeException, Notice, Error, etc.). Useful for grep / wait_for in scripts.
No screenshots or snapshots needed. The visual HTML BlueScreen and Tracy Bar are still rendered for the human user — the markdown is added on top.
Tracy Bar
A floating panel in the bottom-right of every successful page. For humans it is rendered as HTML; for a detected agent the same content is mirrored as a markdown summary into console.log(). Typical contents:
- Execution time and memory usage
- SQL queries (count, total time, individual queries with parameters)
- Request and response details
- Logged messages, warnings, recent dumps
Custom panels can expose their own markdown rendition by implementing Tracy\IBarPanel::getAgentInfo(): ?string. Built-in panels (SQL, Errors, Dumps, etc.) already do this.
BlueScreen (exceptions & fatal errors)
When an exception is thrown or a fatal error occurs, Tracy renders a full HTML BlueScreen for the human. For a detected agent the same exception is also emitted as markdown via console.error() with:
- Exception class, message, and code
- Stack trace with file paths and line numbers
- Source code context around the error
- Arguments at each stack frame
- Request parameters, headers, environment
The BlueScreen provides enough information to diagnose most issues without looking at logs.
Using dump() for debugging
Insert dump($variable) anywhere in PHP code to inspect values:
dump($product);
dump($query->getSql());
dump($form->getValues());
The visual HTML dump is rendered inline as usual; for a detected agent, the same value is also written through Helpers::consoleLog(Dumper::toText($var)) into the browser console, so it shows up in list_console_messages() as a [log] entry.
Batch reading production exceptions
In production, Tracy logs exceptions into log/exception-<hash>.html. Tracy ≥ 2.12 writes a parallel log/exception-<hash>.md next to each one with the same content in markdown — created at the moment the HTML is first written, never overwritten afterwards.
This makes it possible to hand the whole log folder to an agent for batch triage without parsing hundreds of HTML files:
You have a folder of .md exception reports in log/.
For each report:
1. Read the exception (file/line/message/stack).
2. Compare it with the current state of the codebase — the bug may already be fixed.
3. Group related reports thematically.
4. Propose a fix plan and wait for approval before changing code.
The .html siblings stay around for human inspection; ignore them unless you specifically need a rendered view.
Debugging workflow
With Chrome MCP (preferred for web pages):
- Navigate — open the page in Chrome via
navigate_page.
- Check console —
list_console_messages() to read Tracy errors ([error]) and Bar info ([log]).
- Inspect — read further console messages or use
take_snapshot if you need to see the visible HTML structure.
- Fix — make the code change.
- Verify — reload and check the console again.
With curl (API endpoints, CLI):
- Add the cookie —
--cookie 'tracy-webdriver=1', otherwise you only get raw HTML.
- Reproduce — fetch the page to see current behavior.
- Inspect — read the markdown blocks emitted via
<script>console.log(...)</script> / <script>console.error(...)</script> in the response body. They are JSON-encoded; decode the first argument to get the markdown.
- Add dump() — insert
dump() calls in PHP to inspect specific values; refetch to read them from the same blocks.
- Fix — make the code change based on findings.
- Verify — fetch once more to confirm the fix.
Common patterns
Check what SQL queries a page generates (Chrome MCP):
navigate_page to the URL.
list_console_messages().
- Find the
[log] entry with the markdown Tracy Bar; the SQL panel section lists queries with timings.
Inspect a variable at a specific point:
dump($this->getParameter('id'));
Refetch (or reload in Chrome) and read the new [log] entry that appears.
Read production exception batch:
Open log/ and process every exception-*.md file in order.
Online documentation
For detailed information, use WebFetch on these URLs: