Agenthood Release Digest: September 8, 2026
Date: September 8, 2026
Author: Agenthood Team
Here's what's new in Agenthood, covering releases v3.49.0 through v3.56.6. This batch focuses on the skill ecosystem, checkpointing, and CLI ergonomics, alongside a set of security and reliability fixes.
Features
Skills and Registry
- SKILL.md validation (v3.56.0): We now validate SKILL.md files against the agentskills.io spec, catching spec violations before a skill is used in a run.
- SkillRegistryClient (v3.55.0): We've added a registry client with
searchandupgradecommands, making it easier to discover and update skills from the command line. - RemoteSkillSource (v3.54.0): Skills can now be fetched from URLs or git repositories, with caching built in. This simplifies provisioning skills from external sources.
Checkpoints and Trajectories
- Checkpoint persistence (v3.49.0): Runs can now be checkpointed and resumed with
--resume. Long-running work no longer needs to start over after an interruption. - TrajectoryStore (v3.54.0): A new TrajectoryStore records run traces, and
trace visualizeandtrace diffsubcommands make it easier to inspect and compare agent behavior.
CLI and Interactivity
- Interactive confirmations (v3.53.0): The new
--interactiveflag enables human-in-the-loop confirmation before sensitive actions. - Shell completion (v3.51.0):
agenthood completiongenerates completion scripts for bash, zsh, and fish. - Log tailing (v3.52.0): The
logcommand now supports--tailand--followfor streaming and recent log output. - Debug logging (v3.50.0): A
--debugflag enables LLM request and response logging, useful for diagnosing provider issues.
Bug Fixes
Security
- Prompt injection delimiting (v3.55.5): Untrusted task and tool output is now delimited before being passed to the model, reducing the risk of prompt injection.
- SafetyGuard patterns (v3.55.4): SafetyGuard's catastrophic command patterns have been broadened to cover more dangerous commands.
- Secret redaction (v3.55.3): Secrets are now redacted from outbound provider requests, preventing accidental leakage in logs or traces.
Reliability
- Graceful shutdown (v3.55.6): Long-running agent sessions now shut down gracefully on SIGINT, allowing cleanup before exit.
- Concurrent provider imports (v3.56.5): LLM providers are imported concurrently, eliminating cold-start timeouts during startup.
- OpenCode session headers (v3.55.1, v3.55.2): All OpenCode requests, including the Go tier, now include the
x-opencode-sessionheader.
Skills and Tooling
- UTF-8 BOM handling (v3.56.2): Skill frontmatter now parses correctly on Windows-authored files by stripping the UTF-8 BOM.
- Lockfile regeneration (v3.55.7): Lockfile regeneration is now consistent, and
--update-lockworks as intended. - Member-path resolution (v3.56.4): Member-path resolution is unified across verify commands, and CI now enforces lockfile integrity.
- CI whitespace tolerance (v3.56.6): The CI pipeline now tolerates whitespace before the decision marker's closing delimiter, preventing spurious failures.
API and Events
- Re-exposed exports (v3.56.1): The
./dist/llmand./dist/coreentry points are re-exposed for programmatic consumers. - tool.approval event (v3.56.3): The React loop now emits a
tool.approvalevent when a tool awaits approval, instead of a duplicatetool.calledevent.
What's Next
The skill ecosystem is clearly a focus: validation, registry commands, remote sources, and lockfile fixes all landed in this batch. We're continuing to harden that area and expect more tooling around skill discovery and maintenance in upcoming releases.