The registry.

In-thread approval boundary for agent actions. Three semantic variants (search, review, write). Idle, expanded, confirmed and dismissed, and every resolution can be undone.

$ npx shadcn@latest add https://www.hitlkit.dev/r/hitl-card.json

Single-row agent status with icon, label, detail, and state in words. Use to wrap any in-progress agentic task.

$ npx shadcn@latest add https://www.hitlkit.dev/r/subagent-status-card.json

Collapsible step-by-step renderer for agent reasoning. Thought, action, result. Makes supporting facts visible per §3.3 of the paper.

$ npx shadcn@latest add https://www.hitlkit.dev/r/mini-trace.json

Five-segment ordinal scale indicating AI involvement in a piece of work. Embodies the scaffolding and expertise-matching principle from Dhillon et al. (2024).

$ npx shadcn@latest add https://www.hitlkit.dev/r/ai-generation-scale.json

The compact drag form of the AI generation scale. One header row and one track, no wrapping at 320px. Pointer drag snaps to five stops; arrow keys step and Home/End jump. Full slider ARIA.

$ npx shadcn@latest add https://www.hitlkit.dev/r/ai-generation-slider.json

At-a-glance, read-only provenance for a list row or a header. Five thin segments filled to the current level plus the level name, as a single role=img element with no focusable children.

$ npx shadcn@latest add https://www.hitlkit.dev/r/ai-generation-meter.json

The densest form of the AI generation scale: one pill with a five-dot micro-indicator and the level name. Static by default; given onAction it grows stepper buttons with 24px hit areas.

$ npx shadcn@latest add https://www.hitlkit.dev/r/ai-generation-badge.json

Pill chips representing context attached to an agent run. Notes, files, URLs. Removable with overflow truncation.

$ npx shadcn@latest add https://www.hitlkit.dev/r/context-chips.json

Multi-question approval card. Single-choice, multi-select, and freeform text. Submits to a resolved line that keeps the answers visible and can be reopened.

$ npx shadcn@latest add https://www.hitlkit.dev/r/qa-flow.json

Compound widget for a draft-in-progress document. Title, target section, word range, evidence notes, and six status states.

$ npx shadcn@latest add https://www.hitlkit.dev/r/writing-agent.json

Three-mode config surface for a long-running research task. Create, follow-up, read URL.

$ npx shadcn@latest add https://www.hitlkit.dev/r/research-agent.json

Sequential approve and reject flow across mixed agent items. Auto-advances, can step back one item, and resolves to a summary state.

$ npx shadcn@latest add https://www.hitlkit.dev/r/batch-queue.json

Ranked search result card with rank badge, metadata, snippet, and relevance bar.

$ npx shadcn@latest add https://www.hitlkit.dev/r/search-result-card.json

The canonical decision row. Approve, reject, or can't tell; pending, approved, rejected, abstained; undo back to pending.

$ npx shadcn@latest add https://www.hitlkit.dev/r/approve-reject-row.json

The atomic palette. Accent swatches, the four approval badges, and the decision row that sets them.

$ npx shadcn@latest add https://www.hitlkit.dev/r/shared-primitives.json

Before/after diff card for proposed text or code edits. Per-hunk red/green strips; apply, keep original, or can't tell; undo. The primitive for any in-place edit the human should approve before it lands.

$ npx shadcn@latest add https://www.hitlkit.dev/r/diff-result.json

Single source-backed citation card. Claim on top, source attribution below, optional supporting quote and confidence. Verify, reject, or can't tell; undo.

$ npx shadcn@latest add https://www.hitlkit.dev/r/citation-result.json

Multi-step plan the human can reorder, rename, add to, or delete from before the agent executes. Locked steps cannot be removed. Submit or cancel; undo.

$ npx shadcn@latest add https://www.hitlkit.dev/r/editable-plan.json

Preview a tool call (name, args, optional rationale and signals) so the human can approve, reject, or say they can't tell before execution. External scopes are flagged. Pairs with the gates layer.

$ npx shadcn@latest add https://www.hitlkit.dev/r/tool-call-preview.json

Where a claim is grounded, not merely that it is. One row per pointer with source, locator in human units, and excerpt; sources consulted but not drawn from are listed so silence never reads as safety. Grounded, not supported, or can't tell; undo.

$ npx shadcn@latest add https://www.hitlkit.dev/r/evidence-pointer.json

Tailwind classname merge helper used by every HITL Kit primitive.

$ npx shadcn@latest add https://www.hitlkit.dev/r/hitl-utils.json

The props every decision surface shares (busy, error, help, allowAbstain, allowUndo, autoFocus) and the interrupt card's local state. Event and action types come from @hitl-kit/core.

$ npx shadcn@latest add https://www.hitlkit.dev/r/hitl-types.json

The card frame, decision bar, resolved line, inline error, help line and chip that every decision surface is built from. One vocabulary, one set of button styles.

$ npx shadcn@latest add https://www.hitlkit.dev/r/hitl-shared.json

Shared icon, color and plain label for the six agent execution states (idle, running, completed, error, skipped, cancelled).

$ npx shadcn@latest add https://www.hitlkit.dev/r/hitl-subagent-meta.json

The five-point AI-generation ordinal and its accent progression, shared by every scale variant so a level cannot mean one thing in the slider and another in the badge.

$ npx shadcn@latest add https://www.hitlkit.dev/r/hitl-ai-levels.json

Valid core events for the decision surfaces, for demos and tests. Spread one onto its primitive or feed it to HitlEventRenderer.

$ npx shadcn@latest add https://www.hitlkit.dev/r/hitl-fixtures.json