What an agent cannot guess

An agent driving the retained browser core knows one thing with certainty: whether its own instruction executed. It clicked, the click happened, succes: true.

That statement is true. It is also compatible with the server having received nothing, the form having been rejected, the session having expired, and the page having been the wrong one all along.

Perception is not only about reading the page. It is about knowing where reading stops.

The gap between “it ran” and “it worked”

Three real cases, all ending in success:

The browser refused the submission. A required field was empty. HTML5 validation blocked it. The click happened — Dinoer is right — and nothing was sent.

The click was synthetic. A JavaScript .click() on a submit button fires the DOM event without guaranteeing the HTTP submission. A configuration update ended green with nothing changed.

The target was not the one expected. A redirect landed on a different page than the scenario assumed; the click landed, on a stranger.

In all three, the agent had no way to know from its own output alone. That is the point of this page.

Time is invisible

A page can be fully loaded and not yet ready.

After a login form is submitted, the network goes quiet and the navigation completes. The agent concludes the session is established and navigates to a protected page — and lands back on the login screen.

The server had not finished. Many applications regenerate the session identifier right after authentication; the browser is done, the server is not, and there is no structural difference between “quiet because finished” and “quiet because between two states.”

{"type": "cliquer", "selecteur": "button[type=\"submit\"]"},
{"type": "pause", "ms": 2000},
{"type": "naviguer", "url": "https://target.local/protected"}

Two seconds of deliberate waiting, because network silence is not the same as readiness. Nothing in the page says so.

What Dinoer does about it

It cannot close the gap — no tool can, from inside the browser. What it can do is stop the agent from reasoning as if the gap did not exist.

Report where you actually are. The compass carries the current URL after redirects, the page title, the last HTTP status. An agent that believed it was elsewhere finds out.

Report drift. session_derive says the session no longer holds. And dernier_code_http disambiguates two failures that look identical — a real expiry redirects with a 302, an application error hidden behind display_errors=0 gives a 500 and the same login page.

Report escalation. repli_js_utilise appears only when a JavaScript click fallback actually ran, never because the flag was set.

None of these are guarantees. They are signals that turn a silent wrong assumption into a readable one.

The habit this argues for

Read the answer, not the verdict. succes: true is the least informative field in the output — it says the instruction ran.

What tells you whether the work happened is everything around it: where you are, what the server said, whether anything had to be forced. An agent that checks only the verdict will be confidently wrong on exactly the cases that matter — and the research pipeline built on top of this same core inherits the same discipline: campagne.py reads succes/statut per source, never assumes a collected page is a usable one.

In short

• Mechanical success and real effect are different claims.
• A quiet network does not mean a ready server.
• The signals exist — compass, HTTP status, drift, JS-fallback flag — and they are worth more than the verdict.
• Knowing where perception stops is part of perceiving.