v2026.7.4

Security

• Bumped soupsieve to 2.8.4 (PYSEC-2026-3071, PYSEC-2026-3072). soupsieve is a transitive dependency (via beautifulsoup4), so this is a lockfile-only bump. It clears both advisories from the pip-audit gate.
• Bumped click to 8.4.2 (PYSEC-2026-2132, fixed in 8.3.3). click is a transitive dependency (via typer and uvicorn) with no Dependabot PR of its own; a fresh lockfile resolve pulls 8.4.2, which is past the fixed release. This was the last advisory keeping the pip-audit gate red, which is now green.

Changed

• Upgraded pydantic-ai from 2.0.0 to 2.13.0 (pydantic-ai-slim, pydantic-evals, and pydantic-graph), and raised the pin floors from >=2.0.0 to >=2.13.0. Thirteen minor releases of fixes and features; the full test suite, ty, and a live run pass unchanged. The upgrade pulls openai to 2.46.0, whose stricter typing exposed a latent nullable-response case in the image_gen edit tool (now guarded). Note: InitRunner uses no AG-UI or Vercel adapter, so GHSA-jpr8-2v3g-wgf9 (fixed in 2.5.0) never affected its code paths; this moves off the affected release line regardless.
• Bumped mcp to 1.28.1 (from 1.26.0). A transitive dependency (via fastmcp-slim and pydantic-ai-slim[mcp]); a lockfile-only bump. The MCP gateway and the full test suite pass unchanged.
• Bumped dashboard frontend dependencies: posthog-js to 1.404.1, @xyflow/svelte to 1.6.2, vite to 8.1.5, @sveltejs/kit to 2.70.0, and @sveltejs/vite-plugin-svelte to 7.2.0 (the latter two via the svelte Dependabot group). The production build is unaffected.

Added

• Prompt-cache hit ratio in the run summary. When a provider reports a cache hit, the tokens: … stats line now appends cache hit: NN% (from PydanticAI 2.13's usage.cache_hit_ratio), so the effectiveness of Anthropic/Bedrock prompt caching is visible per run. The suffix is omitted when the ratio is zero or unreported, keeping the line clean for TestModel and non-caching providers.
• The RaiseContentFilterError capability. A role can declare capabilities: [RaiseContentFilterError] to surface a provider content-filter response as a ContentFilterError (which can also be named in model.fallback.on) instead of receiving a silent empty completion. It loads through the existing dynamic capability path with no extra wiring.
• Refreshed the curated model suggestions in initrunner new: Claude Sonnet 5 (direct and via Bedrock), Grok 4.5, and Z.AI GLM-5 and Moonshot Kimi K2.5 via Bedrock. Existing per-provider defaults are unchanged. Any provider:model PydanticAI recognizes (for example moonshotai:kimi-k3) remains usable directly.