Six bugs that combined to break the Visualize capability on Gemini 2.5
Flash (and similar thinking-by-default models). Each is independently
useful, but the user-visible symptom — Visualize "kind of works but
output is randomly truncated at ~370 chars" — needs all of them.

1. Gemini 2.5/3.x reasoning-tokens default (root cause)

   Gemini 2.5+ models burn most of `max_tokens` on internal "thinking"
   tokens by default. With `max_tokens=4096`, ~3900 went to reasoning
   and only ~160 came out as actual content, causing finish_reason=length
   on every multi-step pipeline (Visualize codegen + review, Deep Solve,
   anything that asks for a structured output beyond a sentence).

   Default `reasoning_effort="none"` for Gemini 2.5/3.x models when the
   caller doesn't specify, in all three execution paths:
   - provider_core/openai_compat_provider.py:_build_kwargs (live path)
   - executors.py:sdk_complete / sdk_stream (legacy SDK path)
   - cloud_provider.py:_openai_complete / _openai_stream (aiohttp fallback)

2. visualize capability had no agents.yaml entry

   `get_agent_params("visualize")` silently fell through to the 4096
   default because there was no section_map entry and no
   DEFAULT_AGENTS_SETTINGS entry. Added both, with a 16384-token budget
   appropriate for full HTML pages.

3. Review stage crashed hard on JSON parse failure

   `ReviewAgent.process` does `ReviewResult.model_validate(extract_json_object(response))`.
   When the model returned prose instead of JSON (common with large SVGs
   that the model can't escape into a JSON string), the parse raised and
   killed the entire turn. Wrapped pipeline.run_review() in try/except
   so review failure falls back to the unreviewed draft and the user
   still gets a rendered result.

4. Codegen output not trimmed to the root tag

   Models often wrap SVG/HTML in prose ("Here you go: <svg>…</svg>
   Enjoy!") or emit a closing code fence on the same line as `</html>`,
   which `extract_code_block`'s regex (requiring a leading \n before
   the fence) doesn't strip. Added defensive root-tag trimming for
   render_type=="svg" and render_type=="html".

Verified end-to-end on Gemini 2.5 Flash via the CLI and headless
Playwright: full 22 KB long-division HTML page, no truncation, all
interactive elements present, multi-step walkthrough completes
correctly (7852 ÷ 6 → 1308 R 4).

… endpoint (#485)

Fixes #481. Makes `require_auth` and `require_admin` `async def` so the user ContextVar set inside the dep is visible to async endpoints; sync deps run via `anyio.to_thread.run_sync` under `copy_context()` and the set is discarded on return.

…en Visualize pipeline

Squash-merge of #490 (skinred78) with two non-trivial conflict resolutions:

* `openai_compat_provider.py`: PR inlined reasoning_effort logic that has since
  been refactored into `build_openai_compatible_reasoning_kwargs`. Keep the
  helper call; the Gemini default-down will be folded into the helper in a
  follow-up commit instead of duplicating it inline.

* `capabilities/visualize.py`: i18n was applied to the review-stage messages
  after the PR forked. Kept the new try/except review-fallback structure but
  routed all three messages through `i18n.t(...)`; added `review_skipped_error`
  key to en/zh visualize.yaml.

Other PR changes applied as-is:
* `executors.py` / `cloud_provider.py`: inline Gemini 2.5/3.x default-down
* `loader.py`: section_map entry for `visualize`
* `init.py`: `capabilities.visualize` default with max_tokens=16384
* `code_generator_agent.py`: defensive root-tag trim for svg/html

Closes #489.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Follow-up to #490. The PR inlined the "disable thinking for Gemini 2.5/3"
gate in 5 places across 3 files. This commit collapses them to one
registry and three thin call sites.

Changes:

* `services/llm/reasoning_params.py`: new `_PROVIDER_DEFAULT_OFF_PATTERNS`
  registry + `default_reasoning_effort_for(provider, model)` public helper.
  `build_openai_compatible_reasoning_kwargs` now consults the registry, so
  the openai-compat path (which lost its inline gate during the #490 merge
  conflict resolution) is restored via the helper.

* `services/llm/executors.py` (sdk_complete + sdk_stream) and
  `services/llm/cloud_provider.py` (_openai_complete + _openai_stream) now
  call `default_reasoning_effort_for(...)` instead of inlining the
  ('gemini-2.5', 'gemini-3') startswith check.

* Use substring (not startswith) match so `models/gemini-2.5-flash` is
  also covered — some OpenAI-compat clients prefix model ids with `models/`.

* `services/config/loader.py:get_agent_params`: when a module's section is
  missing from the user's stale `agents.yaml`, fall back through
  `DEFAULT_AGENTS_SETTINGS` before the global `(0.5, 4096)` default. This
  lets the `capabilities.visualize` default (`max_tokens=16384` from #490)
  reach existing installs, not just fresh ones.

* `capabilities/visualize.py`: hoist the duplicated lazy
  `from deeptutor.agents.visualize.models import ReviewResult` import
  from two branches into the top of `run()`.

Tests:

* `tests/services/llm/test_reasoning_params.py` — 17 new cases covering
  Gemini 2.5/3 + `models/` prefix + case-insensitivity + the legacy
  Gemini 1.5/2.0 / other-provider untouched paths + the explicit-override
  takes-precedence rule.
* All 107 tests in `tests/services/llm/` still pass; 5 pre-existing
  `test_chat_params_config` failures (8192 vs 8000 drift) are unrelated.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…streaming everywhere

- Centralize Gemini 2.5/3 reasoning_effort=none in
  reasoning_params.default_reasoning_effort_for so Visualize / Chat /
  Solve / agentic loop stop returning empty bodies on those models.
- Visualize: per-capability max_tokens default (16k) seeded from
  DEFAULT_AGENTS_SETTINGS, defensive root-tag trim on SVG/HTML output,
  graceful fallback when JSON-mode review step crashes.
- Fix #485: require_auth / require_admin are async so the
  set_current_user ContextVar reaches the endpoint instead of being
  discarded by anyio.to_thread.run_sync's worker-thread context copy.
  Adds _install_current_user helper shared by HTTP + WebSocket.
- Reasoning + native-tools chat protocol: formal content stream must
  still start with FINISH/TOOL/THINK/PAUSE; tool-call deltas no longer
  force-resolve labels and implicit_think_label is ignored, so
  protocol-repair catches missing labels instead of mis-routing turns.
- Smooth streaming on every chat surface: useSmoothStreamText through
  AssistantResponse, pin-to-bottom (useLayoutEffect) on book chat +
  quiz follow-up, data-chat-scroll-root opt-in for overflow-anchor:none.
- Sidebar: collapsible Recents region with own scroll viewport,
  deterministic Lucide icon per session via SessionAvatar, Docs link
  next to GitHub footer, "New Chat" button removed (nav handles it).
- Add Lemonade local provider (port 13305) — registry entry, README
  Docker host-gateway row, providers.md docs.
- Context-window models-endpoint probe honors DISABLE_SSL_VERIFY via
  TCPConnector(ssl=False).
- README: insert v1.4.2 release row, push v1.3.10 inside "Past
  releases" fold, bump install opt-in note to v1.4.2.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>