Agent-Native Clearance
Let coding agents inspect release blockers without raw data access.
Agents should query Clearance the same way reviewers do: through scoped datasets and review primitives, not raw SQL or trace dumps.
Agent Tools
Clearance exposes the agent workflow as SDK calls:
import { explainBlockers, query } from "@policystrata/clearance";
const blockers = await explainBlockers({ limit: 50 });
const heldActions = await query({
dataset: "runtime_events",
decision: "require_approval",
});import clearance
blockers = clearance.explain_blockers(limit=50)
held_actions = clearance.query(
dataset="runtime_events",
decision="require_approval",
)What Agents Can Inspect
| Dataset | Purpose |
|---|---|
runtime_events | Governed actions projected into release impact |
findings | Open blocker, fail, warning, and info findings |
evidence_packs | Reviewable release evidence and current status |
review_decisions | Approval, block, and waiver history |
agents | Observed agent inventory and metadata |
What Agents Should Not Touch
Agents should not request raw prompts, customer rows, credentials, full documents, or trace payloads. Clearance stores metadata, summaries, hashes, artifact refs, trace IDs, span IDs, and review state by default.
Local Skill
This repo includes a Clearance query skill at
.agents/skills/clearance-query/SKILL.md. Install or vendor that skill into
agent workspaces that need to answer:
- Why is this release blocked?
- Which runtime decisions require approval?
- Which evidence pack is waiting for review?
- Who approved, blocked, or waived the release?