v0.5.0

Breaking: leaner public response (F11)

The payload goes straight into an LLM context, so redundant fields cost real tokens:

• answer is omitted when a provider produces no prose (keyword providers no longer send an empty string)
• result rows no longer repeat the provider name that is already the enclosing dictionary key

Reliability

• web_search arguments are now validated against the declared inputSchema before dispatch; stdio and HTTP return a structured invalid_arguments tool error (empty query, wrong types, unknown providers, out-of-range values, unknown fields)
• provider misconfiguration no longer breaks startup: only enabled providers are constructed, and a bad AGENT_WEB_SEARCH_TIMEOUT raises a readable ValueError
• tests are isolated from developer-machine provider credentials

Fixes & cleanup

• Grok: correct tool-name matching direction (fixes mode=both), dead code removed
• Parallel: User-Agent version no longer falls back to a stale 0.1.0
• ddgs: removed a dead ImportError branch with a misleading install hint
• ARK: continuation threshold is now a named constant documenting the second billed request
• Exa: removed a dead min(100, ...) bound
• Dropped redundant starlette/uvicorn hard dependencies (mcp already requires them)
• Docs: corrected timeout semantics and Parallel result-cap behavior (en/zh)

Tooling

• Stricter ruff rule set (BLE/I/UP/B/SIM/RUF/C4/PT/E501), all findings fixed
• ruff format applied and enforced in CI

97 tests pass on Python 3.10-3.13.