Review queues
Agents do most of their work on their own, but some decisions need a human. Review queues is the hub where those pending items live — approval requests, newly installed skills awaiting sign-off, and extracted learnings waiting for review before they're reinforced.
What this page is for
Autonomous agents are fast and cheap, but they make the wrong call sometimes. For the decisions that matter most — sending an external email, installing an untrusted skill, reinforcing an extracted lesson into the agent’s future behavior — Exolvra holds the action in a queue until a human decides. This hub is where those three queues live, all in one place, with a live count of what’s pending.
Opening /admin/review shows three cards. Each card shows the current pending count — if anything is waiting, the count is highlighted so you can see at a glance what needs your attention.
The three queues
| Queue | What it holds | When a human must act |
|---|---|---|
| Approvals | Agent-proposed actions gated by the approval rule system — sends, deletes, destructive shell commands | Every request, before the agent will execute |
| Skill approvals | Skills newly installed from external sources that haven’t been signed off by an admin | Before the skill becomes available to any agent |
| Learning review | Lessons the learning system extracted from past successful runs — claims like “prefer web_fetch over web_search for PDFs” | Before the learning is reinforced into future agent context |
How the queues differ
Approvals are synchronous — the agent is waiting on your decision right now. A pending approval means an agent is paused mid-turn. Resolve these first. An approval rejected without a useful reason teaches the agent nothing; a specific rejection sharpens it for next time. See Approvals for the full workflow.
Skill approvals are asynchronous — the skill was installed, but it can’t be used until you sign it off. Nothing is paused. Review these on your own schedule. A skill is a bundle of prompts, schemas, and sometimes executables; signing it off means you’ve read the SKILL.md and believe the skill is safe. If you have skill signing enforced (under Settings → Security), unsigned skills are not loaded at all — the approval becomes a prerequisite for use.
Learning review is asynchronous — extracted lessons are harmless until approved. Rejecting a learning means the system forgets it. Approving one means the lesson goes into future agent context when relevant. Learning extraction is opt-in; you enable it from Settings → Agents. Disable it if you don’t want the review burden.
Common tasks
Clear the approvals inbox
Work down the list oldest-first. Read the agent’s reasoning. Click Approve if it looks right, Reject with a specific reason if it doesn’t. Use Discuss to ask follow-up questions before deciding — the agent sees your message on its next turn. Unanswered requests expire after 24 hours by default and count as rejections.
Review a newly installed skill
Open Skill approvals. Each pending entry shows the skill name, source (ForgeHub, local, GitHub), author, and the manifest contents. Open the skill detail and read through the manifest — the point of Skills 2.0 is that the manifest tells you what the skill does, which tools it uses, and what its prompts look like, in one place. If you’re comfortable, click Approve. If not, click Reject or Delete.
Approve a learning
Open Learning review. Each entry shows the claim (“prefer web_fetch over web_search for PDFs”), the confidence the extractor had, and the runs that contributed evidence. If you agree, click Approve — it becomes active in the agent’s context. If the claim is generic, harmful, or simply wrong, reject it.
Turning queues off
Two of the three queues can be disabled:
- Skill approvals are required only when skill signing is enforced — toggle that off under Settings → Security and skills go live immediately on install.
- Learning review can be turned off entirely from Settings → Agents, or the require-approval toggle can be lifted so learnings auto-reinforce without human review.
Approvals cannot be globally disabled — they only exist because an approval rule was defined. If you have no rules, no requests land in the queue. See Approvals for the rule configuration page.
Common pitfalls
Letting the queues back up. A stale approval blocks an agent. A stale skill approval means a user installed a skill and can’t use it. Clear the three queues daily if you have rules or signing enforced.
Approving skills without reading the manifest. The point of Skills 2.0 is that the SKILL.md describes the skill’s behavior, tools, and permissions in a single file. If you’re signing a skill without reading it, you’re not doing a skill review — you’re rubber-stamping.
Reinforcing generic learnings. Learnings like “be careful with destructive operations” are too broad to be useful. Approve only specific, actionable claims.
Where to go next
- Approvals — the full approvals reference
- Admin overview — all six hubs at a glance
- Settings → Security — toggle skill signing
- Settings → Agents — configure the learning system