v1.20.4 — Browser agent reliability, Anthropic Claude 4.x model-id normalization, robust LLM config on model change

[1.20.4] - 2026-05-12

Fixed — Browser agent: model 404, premature step kill, AX-tree starvation

• claude-opus-4.6 (and claude-opus-4.5, claude-sonnet-4.6, claude-haiku-4.5) → 404 not_found_error — the LLM catalogue (llm_models, llm_config_overrides) and domains/llm_config/constants.py stored Anthropic 4.x model ids with a dot (claude-opus-4.6), but the Anthropic API only accepts the dashed form (claude-opus-4-6, …) — core/config/llm.py already used the dashed form. Any LLM type pointed at one of these (e.g. browser_agent after the admin picked "Claude Opus 4.6") crashed at instantiation. Renamed to the dashed form everywhere: pricing/config seeds, mcp_app_react_agent default, unit tests, frontend i18n strings (claude-opus-4-5 template label), docs. New Alembic migration rename_anthropic_model_ids_001 rewrites the existing llm_models.model_name / llm_config_overrides.model rows (idempotent regex ^(claude-[a-z]+-[0-9]+)\.([0-9]+)$ → \1-\2, reversible); DEV migrated, PROD updated via the equivalent in-transaction SQL.
• browser_task_tool step killed mid-task — the parallel executor wraps each tool step in asyncio.wait_for(timeout=step.timeout_seconds or DEFAULT_TOOL_TIMEOUT_SECONDS), and browser_task_tool (a full nested ReAct loop) was absent from _HIGH_LATENCY_TOOLS, so the planner's 30/60/120 s timeout_seconds cancelled the loop before it could finish (observed step durations: exactly 30 009 / 60 020 / 120 009 ms, then asyncio.CancelledError inside browser_task_tool). Added browser_task_tool to _HIGH_LATENCY_TOOLS with a dedicated floor BROWSER_TOOL_TIMEOUT_SECONDS = 300 s and ceiling MAX_BROWSER_TOOL_TIMEOUT_SECONDS = 600 s (the generic MAX_TOOL_TIMEOUT_SECONDS = 120 s no longer applies to it).
• Accessibility-tree truncation hid form controls — BROWSER_AX_TREE_MAX_TOKENS code default raised 5000 → 15000 (new BROWSER_AX_TREE_MAX_TOKENS_DEFAULT constant, referenced from BrowserSettings); .env.example / .env.prod.example updated.

Added — BROWSER_REACT_MAX_ITERATIONS

New env var (default 15, range 1–50) capping the create_react_agent recursion_limit of the browser agent loop run by browser_task_tool — mirrors REACT_AGENT_MAX_ITERATIONS / MCP_REACT_MAX_ITERATIONS. Constant BROWSER_REACT_MAX_ITERATIONS_DEFAULT, field browser_react_max_iterations on BrowserSettings, wired in browser_tools.py (was a hard-coded 15); .env.example / .env.prod.example and BROWSER_CONTROL.md updated.

Changed — Browser agent prompt (prompts/v1/browser_agent_prompt.txt)

Rewritten from web-research best practices (notably the browser-use agent prompt): explicit read-vs-transactional task modes, no fabricated query-param URLs, observe→act→verify loop with a mandatory snapshot after state-changing actions, stale-[ref] handling, autocomplete → click-the-suggestion (not Enter), fill only on real input controls (textbox/searchbox/combobox/spinbutton), cookie-banner/403/login obstacle handling, stuck-loop detection, stop-before-payment + report-only-facts. browser_agent_prompt added to the PromptName Literal (was missing — pre-existing gap).

Fixed — reasoning_effort ↔ model coherence (robustness on model/provider change)

Switching a model/provider on an LLM type could leave a reasoning_effort whose shape no longer matched the new model's reasoning_widget (e.g. a DeepSeek {"effort": "off"} enum value left on a config whose effective model became the Qwen toggle_budget default — the admin UI rendered it as "thinking disabled", so it looked fine), which then crashed the typed reasoning builder at get_llm() time (RuntimeError: ... must be ReasoningEffortToggleBudget, got ReasoningEffortEnum. Validation upstream is broken.). Three layers now prevent this:

1. Frontend — the admin dialog normalizes reasoning_effort to the newly selected model on model/provider change: kept only if its shape fits the new reasoning_widget (and, for enum, the value is allowed), otherwise reset to null. New coerceReasoningEffortForModel / reasoningEffortMatchesModel in components/settings/llm-config/reasoningHelpers.ts.
2. Write path — LLMConfigService.update_config validates reasoning_effort against the effective model (update.model, or LLM_DEFAULTS[llm_type].model when the model override is null), rejecting an incompatible combination with 422 + structured ctx.
3. Merge runtime — core/llm_config_helper.py::merge_config → new _reconcile_reasoning_effort drops any still-incompatible reasoning_effort at merge time (stale override row, outdated seed, manual edit, a past bug), falling back to the model's default and logging llm_config_reasoning_effort_dropped — get_llm() degrades gracefully regardless of the drift's origin.

Shared non-raising predicate reasoning_effort_matches_widget(caps, value) (twin of validate_reasoning_effort) in domains/llm_config/reasoning_validation.py — one source of truth for "is this value valid for this model?", reused by layers 1 and 3.

Fixed — Production build: stale GeoIP database date

apps/api/Dockerfile.prod hard-coded ARG DBIP_DATE=2026-03 for the DB-IP City Lite GeoIP download; db-ip.com keeps only the last ~2 months online, so the URL started returning 404 ("not in gzip format" after curl -L wrote the error page) and gunzip failed the production image build. The geoip-downloader stage now resolves the latest available month at build time — current month, then the previous one — using curl -fsSL (so a 404 cleanly falls through to the next candidate); still overridable with --build-arg DBIP_DATE=YYYY-MM. Build-only change, no runtime impact.

Tests

• tests/unit/domains/llm_config/test_reasoning_validation.py::TestReasoningEffortMatchesWidget — none/enum/budget_int/toggle_budget cases incl. the {"effort": "off"}-on-Qwen regression.
• tests/unit/domains/llm_config/test_config_helper.py::TestEffectiveConfigReasoningReconciliation — incompatible reasoning_effort override dropped, compatible kept, unknown model left untouched.
• apps/web/src/components/settings/llm-config/__tests__/reasoningHelpers.test.ts — 10 cases for reasoningEffortShape / reasoningEffortMatchesModel / coerceReasoningEffortForModel.
• test_reasoning_validation.py / test_constants.py / test_llm_defaults_compliance.py / test_reasoning_builders.py updated for the claude-opus-4-6 etc. rename.

Documentation

• docs/technical/LLM_CONFIG_ADMIN.md — new "Cohérence reasoning_effort ↔ modèle" section (the 3-layer guarantee), reasoning_validation.py / ReasoningWidget.tsx / reasoningHelpers.ts added to the file map, PUT-semantics note.
• docs/technical/BROWSER_CONTROL.md — BROWSER_REACT_MAX_ITERATIONS + browser step-timeout + BROWSER_AX_TREE_MAX_TOKENS notes; new "Agent prompt" section summarising the prompt's behavioural contract.
• In-app FAQ (apps/web/locales/*/translation.json, 6 languages): new faq.changelog.versions.v1_20_4 entry (3 user/admin items); faq.intro.features.browserControl blurb refined (multi-step flow, stops at login/payment, reports a summary); faq.sections.tool_examples_external q18/q19 refined (multi-step example, "good to know" note, corrected timing). docs/knowledge/06_external_services.md propagated from the updated FAQ. docs/GETTING_STARTED.md compatibility line bumped to v1.20.4 (with v1.20.1–v1.20.4 highlights). README.md top banner updated to v1.20.4 (v1.20.3 demoted to a <details> block).
• README.md / docs/technical/LLM_PRICING_TEMPLATES.md / docs/guides/GUIDE_CONFIG_ARCHITECTURE.md / the 6 apps/web/locales/*/translation.json — claude-opus-4.5 → claude-opus-4-5 in prose/examples.
• Version bumped to 1.20.4 across apps/api/pyproject.toml, apps/web/package.json, package.json, and apps/web/src/lib/version.ts (LAST_UPDATED = 2026-05-12T22:48:00, shown on the landing page).