Skip to main content

Agenthood Release Digest: v3.13.0–v3.13.6

Date: August 11, 2026
Author: Agenthood Team

We shipped seven releases between August 5 and August 8, led by v3.13.0 with a new member joining the society, followed by a series of patch releases focused on CI stability, LLM rate-limit handling, RAG performance, and packaging.

ReleaseDateHighlights
v3.13.02026-08-05Builder member joins; runtime triggers and skills-dir watching; marketplace and docs updates
v3.13.12026-08-06Agent analysis: decision-block enforcement and per-run temp isolation
v3.13.22026-08-06Review diff truncation, first-push reviews, agent analysis install fixes
v3.13.32026-08-07LLM rate-limit retry handling; lint gates
v3.13.42026-08-08CI security audit, commitlint, and build gate
v3.13.52026-08-08Tree-sitter parser caching in RAG
v3.13.62026-08-08Packaging fix for postinstall scripts

v3.13.0 — A New Member Joins the Society

The Builder Member

  • New member: the-builder joined Agenthood, brought to full society standards and satisfying the sentinel and librarian checks.
  • Runtime triggers: the-builder was added to the runtime member triggers, so its work routes through the same machinery as other members.

Runtime Improvements

  • Skill file routing: changes to skill files are now routed to the oracle and sentinel triggers.
  • Skills directory watching: the runtime watches the skills directory for operator drift, with the sentinel pattern covered by a test.

Marketplace and Docs

  • Marketplace: the agenthood-all bundle was updated with the new members and member counts were refreshed; marketplace copy now aligns with 19 members.
  • Docs: member skill references now point to the canonical skills home, and member counts and skill links were updated after canonicalization.

CI Baseline

  • Build before analysis: agent analysis now always runs after a build.

Patch Releases: v3.13.1–v3.13.6

CI and Automated Review (v3.13.1, v3.13.2, v3.13.4)

  • Decision-block enforcement (v3.13.1): agent analysis fails when the required decision block is missing or malformed.
  • Run isolation (v3.13.1): agent analysis temp files are now isolated per run, preventing collisions between concurrent runs.
  • Diff truncation (v3.13.2): review diffs are byte-truncated to stay within argument limits and truncated at line boundaries; truncated diffs are marked in the review prompt.
  • First-push reviews (v3.13.2): the first push of a branch is reviewed against the empty tree, so full-diff reviews work correctly from the start.
  • Agent analysis install (v3.13.2): lifecycle scripts are ignored during the analysis install, and the RANGE environment variable is now required.
  • Librarian check (v3.13.2): the member directory glob in the librarian check is guarded against empty or invalid matches.
  • Audit and build gates (v3.13.4): CI now enforces a security audit, declares commitlint, and adds a build gate.

LLM Rate-Limit Handling (v3.13.3)

  • Retry-after clamp: retry-after values are clamped to 300 seconds, so a server-provided value cannot cause an excessive wait.
  • NaN guard: rate-limit errors with a NaN retry-after header are guarded against being interpreted as valid retry delays.
  • Lint gates: lint errors were resolved and CI gates added to keep them from returning.

RAG and Packaging (v3.13.5, v3.13.6)

  • Parser caching (v3.13.5): RAG now caches a tree-sitter parser per language instead of recreating one on each parse.
  • Postinstall fix (v3.13.6): the published package includes scripts/, so postinstall resolves correctly for consumers.

What's Next

The past week was mostly about hardening the automated review pipeline, tightening CI gates, and onboarding a new member. That focus on release stability and society growth is likely to continue in the coming releases.