with one click
hive-essentials
// Hive mental model and orientation. Read this first before using any other Hive skill. Covers the entity hierarchy, observe-and-steer pattern, drivers, sandboxes, console vs CLI, and workspace conventions.
// Hive mental model and orientation. Read this first before using any other Hive skill. Covers the entity hierarchy, observe-and-steer pattern, drivers, sandboxes, console vs CLI, and workspace conventions.
| name | hive-essentials |
| description | Hive mental model and orientation. Read this first before using any other Hive skill. Covers the entity hierarchy, observe-and-steer pattern, drivers, sandboxes, console vs CLI, and workspace conventions. |
Read this before the other Hive skills. It gives you the mental model; the other skills give you the workflows.
Hive is an agentic workflow orchestrator. It manages parallel task execution across multiple agents (human or LLM) with governance, evaluation, and observability.
The console is the primary interface for humans. The CLI is the primary interface for agents. Both read and write the same underlying state.
Workspace
āāā Projects (goals with policy)
ā āāā AGENCY.md (narrative context for humans)
ā āāā PROGRAM.md (execution policy: budgets, allowed paths, evaluators)
āāā Tasks (units of work, canonical in .hive/tasks/*.md)
ā āāā status (proposed ā ready ā claimed ā in_progress ā review ā done)
ā āāā claims (time-limited work leases)
ā āāā links (blocks, parent_of, relates_to, duplicates, supersedes)
āāā Runs (governed execution attempts on a task)
ā āāā worktree (isolated git branch)
ā āāā transcript (event stream)
ā āāā approvals (driver ā human round-trip)
ā āāā eval_results (evaluator decision)
āāā Campaigns (multi-task orchestration with lanes and cadence)
ā āāā Lanes (exploit, explore, review, maintenance)
āāā Memory (observations and reflections, review-gated)
Hive separates execution from judgment:
As an agent, your job is to execute within policy and surface decisions that need human judgment. You do not self-approve governed runs unless the evaluator policy allows it.
Drivers are the execution backends that actually run agent work. Each driver has different capabilities.
| Driver | What it does | Key capabilities |
|---|---|---|
local | Runs commands in a subprocess | Always available, no scheduling |
manual | Human-in-the-loop placeholder | Interactive, no automation |
codex | OpenAI Codex integration | Async, stateful, supports approvals |
claude | Claude SDK adapter | Async, stateful, supports approvals |
pi | Staged/honest driver | Available, RPC depth deferred |
Check what is available:
hive drivers list --json
hive driver doctor <driver-name> --json
Sandboxes provide execution isolation for runs.
| Backend | Isolation | When to use |
|---|---|---|
podman | Container (rootless) | Default for safe local runs |
docker-rootless | Container | Alternative to podman |
e2b | Managed cloud | Ephemeral, upload-only sandboxes |
daytona | Remote | Team/self-hosted environments |
Check what is available:
hive sandbox doctor <backend-name> --json
The console (hive console open) is a React dashboard for humans. It shows:
The CLI (hive <command> --json) is for agents and power users. It accesses the same state. Use --json on every command for machine-readable output.
As an agent, use the CLI. But be aware that a human may be watching your work in the console and may steer runs or approve requests through it.
.hive/
tasks/ canonical task state (markdown with YAML frontmatter)
runs/ run metadata, transcripts, artifacts, eval results
memory/ observations and reflections (per-project and global)
cache/ SQLite index (rebuildable)
campaigns/ campaign state and decision logs
projects/
<project-id>/
AGENCY.md narrative project context (human-readable, not machine state)
PROGRAM.md execution policy (budgets, paths, evaluators, promotion rules)
--json everywhere. Every CLI command supports --json as a global flag. Always use it for structured output..hive/tasks/*.md is the source of truth, not AGENCY.md checkboxes.hive task claim with --ttl-minutes to hold work. Release when done.hive sync projections --json.When you arrive in a Hive workspace:
hive doctor --json # workspace health check
hive task ready --json # what work is available
hive deps --json # dependency/blocker summary
hive console home --json # dashboard state as JSON
hive drivers list --json # available execution backends
hive onboard <project-slug> --json # full setup: project, tasks, evaluator template
This creates a project with AGENCY.md, PROGRAM.md, starter tasks with dependency chains, and a local-smoke evaluator template. The defaults are forgiving: auto_close_task: true and allow_accept_without_changes: true so first runs succeed cleanly.
hive quickstart is a legacy alias for hive onboard.
Once you understand the mental model:
hive-work-loop ā the core agent work cycle (claim ā work ā finish ā promote)hive-project-setup ā creating and configuring projects, tasks, and evaluatorshive-coordination ā multi-agent patterns, campaigns, and portfolio managementhive-mcp ā MCP server integration for host applicationshive-maintainer ā for developing Hive itself (PR discipline, releases)Discipline for developing Hive itself ā PR sizing, review handling, merge discipline, release workflow, delegation, and cleanup hygiene. Use this skill when working on Hive repo changes that span multiple PRs or review cycles.
Coordinate work across multiple agents and projects in Hive. Covers task claims, blockers, handoffs, campaigns, portfolio management, briefs, and shared memory.
Use the Hive MCP server for the thin v2 search and execute tool surface. Use this skill when configuring MCP access to Hive or when an agent needs workspace search and bounded local execution.
Set up and configure Hive workspaces, projects, tasks, and evaluator policy. Use this skill when bootstrapping, creating projects, managing tasks, or configuring PROGRAM.md.
The core agent work cycle in Hive ā from finding a task through claiming, launching a run, handling approvals, finishing, and promoting. Use this skill for task-first project work, governed runs, and clean handoff.