~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.
Python academiaanthropicarxivbravedeep-researchencryptionhome-automationhomeserverlocallocal-deep-researchlocal-llmmistralollamaopenaipubmedresearchresearch-toolretrieval-augmented-generationsearxngself-hosted
chore: clear changelog fragments for 1.9.2 (#5061)
github-actions[bot]52896f91ee82
diff --git a/changelog.d/+model-dropdown-empty-on-load.bugfix.md b/changelog.d/+model-dropdown-empty-on-load.bugfix.mddeleted file mode 100644index 50549f543a..0000000000--- a/changelog.d/+model-dropdown-empty-on-load.bugfix.md+++ /dev/null@@ -1 +0,0 @@-Fix the Language Model dropdown frequently coming up empty on the New Research page (requiring a manual click of the refresh button). Three root causes are addressed: the page fired two concurrent `/available-models` requests on every load, which contended against a cold Ollama; Ollama's model-list query used a too-tight 2-second timeout (raised to 5s) whose failure was silently turned into an empty list; and a transient empty result could linger in the client cache. The list now loads once per page and self-heals from a transient empty response.diff --git a/changelog.d/+model-required-before-research.bugfix.md b/changelog.d/+model-required-before-research.bugfix.mddeleted file mode 100644index fc7998f37d..0000000000--- a/changelog.d/+model-required-before-research.bugfix.md+++ /dev/null@@ -1 +0,0 @@-Validate that a language model is selected before starting a research from the New Research page. Submitting with an empty model now shows an inline "Please select or enter a model." error and focuses the model field instead of kicking off a run that would fail downstream.diff --git a/changelog.d/+pdf-digits-emoji-font.bugfix.md b/changelog.d/+pdf-digits-emoji-font.bugfix.mddeleted file mode 100644index e5ffd6decb..0000000000--- a/changelog.d/+pdf-digits-emoji-font.bugfix.md+++ /dev/null@@ -1 +0,0 @@-Fixed numbers in exported PDFs rendering as wide, spaced-out emoji glyphs ("2 0 2 6" instead of "2026"). Digits 0-9 carry the Unicode Emoji property, so listing "Noto Color Emoji" in the PDF font stacks (#4730) made Pango draw every digit with the emoji font. Emoji families are removed from the stacks; emoji still render via fontconfig per-character fallback to the installed emoji font.diff --git a/changelog.d/+test-reload-state-restoration.misc.md b/changelog.d/+test-reload-state-restoration.misc.mddeleted file mode 100644index 578e24c9ef..0000000000--- a/changelog.d/+test-reload-state-restoration.misc.md+++ /dev/null@@ -1 +0,0 @@-Fixed cross-test state leaks from tests that reload or evict production modules: the rate-limiter tests restored a fake config, the history-routes fixture leaked auth-bypassed blueprints, the office-format tests could leak a blocked-dependency loader registry, the settings-logger tests leaked the log level, the security import-fallback helper re-created the whole security tree (stacking permanent audit hooks), and four sibling sites (thread-local-session eviction, auth-db reload, env-definitions re-import, loader-registry import guards) now restore the exact pre-test modules.diff --git a/changelog.d/5051.feature.md b/changelog.d/5051.feature.mddeleted file mode 100644index 448b79f9b7..0000000000--- a/changelog.d/5051.feature.md+++ /dev/null@@ -1 +0,0 @@-Research steps in chat mode now expand to show the actual tool output (search results, fetched page content — capped at 4,000 chars) when clicked, and the agent's "selecting next action from …" heartbeat lists every enabled tool instead of sampling 3 with "+N more".diff --git a/docs/release_notes/1.9.2.md b/docs/release_notes/1.9.2.mdnew file mode 100644index 0000000000..34ac65d4ad--- /dev/null+++ b/docs/release_notes/1.9.2.md@@ -0,0 +1,14 @@++### ✨ New Features++- Research steps in chat mode now expand to show the actual tool output (search results, fetched page content — capped at 4,000 chars) when clicked, and the agent's "selecting next action from …" heartbeat lists every enabled tool instead of sampling 3 with "+N more". ([#5051](https://github.com/LearningCircuit/local-deep-research/pull/5051))++### 🐛 Bug Fixes++- Fix the Language Model dropdown frequently coming up empty on the New Research page (requiring a manual click of the refresh button). Three root causes are addressed: the page fired two concurrent `/available-models` requests on every load, which contended against a cold Ollama; Ollama's model-list query used a too-tight 2-second timeout (raised to 5s) whose failure was silently turned into an empty list; and a transient empty result could linger in the client cache. The list now loads once per page and self-heals from a transient empty response.+- Fixed numbers in exported PDFs rendering as wide, spaced-out emoji glyphs ("2 0 2 6" instead of "2026"). Digits 0-9 carry the Unicode Emoji property, so listing "Noto Color Emoji" in the PDF font stacks (#4730) made Pango draw every digit with the emoji font. Emoji families are removed from the stacks; emoji still render via fontconfig per-character fallback to the installed emoji font.+- Validate that a language model is selected before starting a research from the New Research page. Submitting with an empty model now shows an inline "Please select or enter a model." error and focuses the model field instead of kicking off a run that would fail downstream.++### 📝 Other Changes++- Fixed cross-test state leaks from tests that reload or evict production modules: the rate-limiter tests restored a fake config, the history-routes fixture leaked auth-bypassed blueprints, the office-format tests could leak a blocked-dependency loader registry, the settings-logger tests leaked the log level, the security import-fallback helper re-created the whole security tree (stacking permanent audit hooks), and four sibling sites (thread-local-session eviction, auth-db reload, env-definitions re-import, loader-registry import guards) now restore the exact pre-test modules.