Agenthood Release Digest: September 18, 2026

Date: September 18, 2026

Author: Agenthood Team

We've shipped several updates over the past two days, including new evaluation capabilities, expanded skill discovery, and a round of reliability and security fixes. Here's what's new.

Evaluation

Cross-provider comparison for evals

The eval CLI now accepts repeatable --provider flags, so you can run the same evaluation against multiple providers in one command and compare results side by side. This makes it easier to benchmark model behavior across providers without maintaining separate runs.

Skills

Packaged tool skills in the runtime catalog

Skills that ship as packaged tools are now included in the runtime catalog. This means they are discoverable and loadable at runtime without requiring an explicit reference, closing a gap where packaged skills were effectively invisible until called directly.

Hardened install against SSRF and path traversal

The agenthood install command has been hardened against server-side request forgery and path traversal attacks. The installer now validates remote skill sources more strictly, addressing the reported security findings.

LLM Provider Handling

OpenAI function shape for chat-completions tools

Tools are now wrapped in the OpenAI function shape when communicating with chat-completions providers. This fixes a compatibility issue where some providers rejected tools that were not formatted according to the OpenAI schema.

Runtime stream guard and structured provider logging

Streaming responses now have a runtime guard to catch and handle stream errors more predictably. Provider interactions also emit structured logs, making it easier to trace requests and diagnose issues.

CI and Release Process

Poll npm registry in release verification

The release verification step now polls the npm registry until the new version appears, instead of sleeping for a fixed duration. This makes the check faster and removes a common source of flaky releases.

Gate ritual matrix entries in a step

Ritual matrix entries are now gated inside a step rather than via a job-level conditional. This simplifies the workflow and avoids the confusion of skipped jobs in the matrix.

Cross-check blocking verdicts against findings

The CI now cross-checks blocking verdicts against the listed findings. A blocking verdict must be backed by an actual finding, preventing false positives and making the review process more reliable.

What's Next

We're continuing to improve provider compatibility and evaluation tooling. Expect further refinements to the skill catalog and additional hardening in the installer in upcoming releases.