feat(ai): modernize the Anthropic path and fix conversation replay

[yupanzi]

Summary

Modernize the Anthropic path and fix conversation replay:

Anthropic request surface

• Move to the Beta Messages API with adaptive thinking, context management, and output_config.effort. Feature use is gated per model (anthropicModelSupportsModernFeatures) so models that reject the modern surface keep the classic request shape.
• New Reasoning Effort setting (low/medium/high/xhigh/max, default xhigh), exposed in general settings next to Max Tokens, with hints explaining both knobs.
• AIMaxTokens is sent as configured (never clamped) and the default is provider-aware: 64000 for Anthropic, 8192 for OpenAI. Existing installs with the old 4096 default are migrated (covered by general_setting_upgrade_test.go). Default Anthropic model bumped to claude-opus-5.

Conversation replay

• The frontend now sends tool round-trips structurally (role: tool with id/name/args/result) and the backend rebuilds real tool_use/tool_result blocks. Tool messages without a result (denied/cancelled) are skipped so no dangling tool_use is produced. Covered by conversation_history_test.go.

Robustness

• Secret data in tool results is masked before reaching the model.
• SSE streams emit a keepalive comment every 20s; the chart's gateway request timeout is raised to 900s and the matching ingress-nginx annotations are documented in values.yaml — these are two halves of the same fix.
• Resource tool schemas now state their caps (list items, log tail lines/bytes) so the model narrows queries instead of truncating blindly.
• AI chat markdown tables no longer overflow on long unbreakable tokens.

Why

The Anthropic integration predates the current Claude API surface: budget_tokens now returns 400 on current models, the 4096-token default truncates answers (thinking and answer share max_tokens), and tool turns were replayed to the model as flattened [Tool: ...] text, which poisons later turns into emitting tool calls as plain text/XML instead of native tool use. Long agent turns also died behind ingress-nginx, which closes a connection after 60s of backend silence while a tool is legitimately still running.

Related issue

Closes #668

Validation

• go test ./pkg/ai/... ./pkg/model/... ./pkg/settings/... (new: conversation history, handler streaming, settings migration tests)
• pnpm run type-check && pnpm run lint && pnpm run test
• helm lint + helm template for the chart change

Checklist

• I reviewed this PR myself before requesting review.
• I understand the changes, including AI-generated parts (if any).
• For new features, a feature request issue is linked.
• I cleaned up AI noise (unnecessary comments, dead code, and unrelated changes).
• This PR is reasonably scoped (or split into smaller PRs).

🤖 Generated with Claude Code