Dinoer v1.0.1 — fixes for the .deb channel and the OpenCode backend

Dinoer 1.0.1 fixes four problems found by auditing the 1.0.0 release, and two more found by running the release cycle for real.

Fixed

• searxng_url on the .deb channel. The key written in /etc/dinoer/dinoer.conf was ignored, and campaigns stopped on "no SearXNG URL configured". lib/searxng.py now reads the file that DINOER_CONF designates, and the dinoer-campaign wrapper exports DINOER_CONF=/etc/dinoer/dinoer.conf, like dinoer-shot and dinoer-rpa.
• OpenCode model. There is no default model any more, because the model list changes and the hard-coded default had disappeared. Without DINOER_OPENCODE_MODEL, the error tells you to pick one from opencode models. A failed call now reports OpenCode's own error and says whether the model is missing from the list.
• OpenCode stays inside the collected corpus. websearch and webfetch are denied whatever the launch directory (through OPENCODE_CONFIG_CONTENT, merged with your own value), so a report only cites the pages that were collected. Standard input is closed on the call. bash stays allowed, as before: a model denied websearch can still reach the web through it.
• Campaign notification. The completion notification no longer carries the local path of the report.
• Installation. postinst and scripts/install.sh now also check for the headless Chromium shell. An interrupted browser download could previously leave the full browser without it and pass for a complete installation, so shot.py could not start.
• Translations. The French, German and Spanish README, guide and manual are back in line with the English source (badges, campaign options table, example paths, Spanish formal register).

Also

• Three model-behaviour entries inherited from before the fork were removed from docs/RADAR_MODELES.md.

Upgrade

• Set DINOER_OPENCODE_MODEL (one of the models listed by opencode models) before running campaigns that call OpenCode.
• Install: sudo apt install ./dinoer_1.0.1-1_all.deb (checksums in SHA256SUMS).

Dinoer v1.0.0 — first public release

Dinoer v1.0.0 — first public release

Dinoer gives language model agents (Claude Code, etc.) the ability to read and drive web interfaces: it runs a local Playwright browser and reports accessible DOM state as text — accessibility tree, extracted content, action results — no screenshots, no vision model in the loop.

Core capabilities

• Text-first navigation and interaction (shot.py, 18 action verbs) with encrypted-directory credential resolution (depuis_secrets) — no plaintext credential ever touches the shell.
• Declarative scenario execution (rpa.py), session persistence, checkpoints for long runs, MFA/TOTP support.
• Research pipeline (campagne.py): SearXNG discovery → respectful collection → escalation to a real browser only when a page needs hands, with a vector search cache, targeted extraction, and per-topic reference-site tables.
• PDF extraction, event fusion, and temporal pre-filtering for research reports.

Install: .deb package (recommended) or git-clone (scripts/install.sh) — see docs/GUIDE.md (also in docs/{fr,de,es}/).

For LLMs: start at docs/GUIDE_LLM.md — mandatory reading before first use.

Website: https://dinoer.davalan.fr — downloads, guides, and full documentation (also in French, German and Spanish).

Full history: CHANGELOG.md.