Agenthood Release Digest: September 28, 2026
Date: September 28, 2026
Author: Agenthood Team
We've shipped a set of releases over the past two days, from v3.69.0 through v3.70.2. This digest covers the new features and the fixes that keep our release pipeline and test suite healthy.
Release Process and CI Hardening
Several fixes landed to make our release pipeline more reliable, from cleaning generated artifacts to catching stale output before it gets tagged.
v3.70.1: Clean Release Notes Before Tagging
- CI publish job: The publish job now cleans release notes before tagging, preventing stale artifacts from being included in the release.
v3.70.0: Herald and Bare-Ref Gate Fixes
- Herald: Generated release notes are now cleaned before gating and publishing, closing a gap where leftover content could slip through (#969).
- Herald: The release gate now correctly recognizes the linkified closing keyword, so issue references are validated as intended (#968).
- CI: The bare-ref gate now rejects
cleanLine's own output, avoiding a false pass caused by the cleanup utility itself (#971).
v3.70.2: Sync-Skills Validation
- CI: Added validation for
sync-skillsto prevent stale generated artifacts from being committed or published (#961).
Test Typecheck Coverage
v3.70.2: Full Typecheck Coverage for Tests
We closed the remaining typecheck gaps in the tests/ directory, so the whole test suite now typechecks cleanly.
- Tests: Achieved full typecheck coverage for the
tests/directory (#963). - Tests: Addressed remaining typecheck errors in test files, including aligning 7 test files with current source API types (#963).
- Tests: Fixed shared helpers and cascade mocks to match the updated types (#963).
- Tests: Resolved
opencode-pluginChildProcesstype mismatches (#963). - Tests: Restored spy mocks and the
vitestimport, clearing the typecheck gate (#976). - Tests: Reverted some
opencode-plugintype casts toas anywhere the external types could not be safely narrowed (#963).
New Features
v3.70.0: Confidence Patterns for Mediator and Steward
- Members: Added Jev-inspired confidence patterns to the mediator and steward agents, giving them a structured way to express and act on confidence in their decisions (#935).
v3.69.0: Routing and Skills Improvements
- Routing: The confidence-gated cascade is now validated as a record, ensuring the routing configuration stays well-formed (#958).
- Skills: The context window is now budgeted at skill activation, so skills can reserve space up front and avoid overrunning the available context (#966).
What's Next
The recent releases point toward continued investment in two areas: keeping our release and test infrastructure airtight, and making confidence and context management more explicit in agent behavior. We expect further refinements to the confidence patterns and context budgeting in upcoming versions.