The project
Where Dinoer comes from, what it kept and what it removed, how a measurement on 14 August 2026 narrowed its claim about itself, and who builds it.
Dinoer is a fork of Diwall 1.21.0, made on 25 July 2026 and rebuilt on 9 August 2026 from Diwall 1.23.0. From Diwall it keeps the execution core: navigate, fill a field, click, wait, resolve a credential inside the browser process. It removed the layer that let a model look at a page: screenshots, Set-of-Mark numbering, the vision model.
The premise: to use a page, a model needs to read it, not to see it.
How it grew
Three stages. The third changed what the project says about itself.
Narrowing. Removing the perception layer left one gap: reading the
cleaned text of a page when the accessibility tree is not enough. The
extraire_texte action filled it on 11 August.
The research pipeline. campagne.py chains SearXNG discovery,
lightweight HTTP collection, escalation to the browser core for pages that
need it, and a report written by a delegated model. On a territorial campaign
run on 11 and 12 August, the first version of the report step (concatenate
the corpus, cut it at 60,000 characters) dropped the most relevant pages. A
time-window filter and a semantic ranking fixed it, measured on the same
corpus.
Testing the claim. On 14 August, a campaign caught the delegated model
making twelve web searches of its own. They were forbidden the same day, and
the gap that remains, curl through bash, is documented. Hours later, a
single question to a generalist search assistant surfaced three pages that
Dinoer’s SearXNG discovery had not found. The project’s claim was rewritten
that evening: not “finds more than a web search”, but “a report you can trace
to a fixed set of collected pages”. The measurement →
What did not change, and what did
It runs on your machine. Collection, credentials and the encrypted directory stay local. Search goes through a SearXNG instance you choose, local or remote; the report through a model you configure.
It counts its passage and does not force a door. Every call reports the pages and actions it spent. A page that refuses access is recorded as refused and is not retried by another route.
What changed. In Diwall, the agent does only what a scenario file lists. Dinoer’s research pipeline also hands judgment to a model: whether a page answers the question, whether two pages describe the same event. That model is kept away from the web, with one gap, described on the trust page.
The development log
Dated, newest first, written during the work, the 14 August reversal included.
Who builds it
Three participants.
Ronan Davalan: architect and arbiter. Product vision, security requirements, direction; he validates every architectural decision.
Claude Code (Anthropic): systems engineer and lead developer. The fork of
Diwall’s execution core, the research pipeline (campagne.py,
lib/searxng.py, lib/fetch_leger.py, lib/selection_candidats.py,
lib/extraction.py, lib/tables_reference.py, lib/cache_recherche.py), the
removal of the perception layer. Principal author of the source code.
Gemini (Google): synthesizer and strategic advisor. Independent architectural analysis, cross-checking of technical decisions.
Where it is now
Four programs (shot.py, rpa.py, campagne.py, journal.py), an
encrypted credential directory, scenarios checked against a JSON schema
before a browser starts, documentation in four languages, a .deb package
and a Git-clone installation. Downloads →