openclaw plugin · MIT licensed

HAWKINS

Everything is connected · by parijat mukherjee

Durable orchestration state (VINES) plus decay-aware shared agent memory (VECNA) for multi-agent OpenClaw swarms. Twelve typed tools. MariaDB-backed. Optional Linear oversight.

Chapter 1 · The Hive-Mind Hierarchy

Three layers, one swarm

OpenClaw gives you isolated specialist agents. openclaw-hawkins gives them a brain, a memory, and a protocol for working together without forgetting what they've already learned.

layer 01 · the nexus

The Nexus

The orchestrator. The operator talks only here. It listens, plans, and dispatches sub-tasks to the specialists — never executing the domain work itself. Drives The Pulse end-to-end.

read more →
layer 02 · the tendrils

The Tendrils

Six isolated specialist agents — system-agent, code-agent, research-agent, data-agent, comm-agent, vision-agent. Each owns one domain, has its own workspace, runs in its own memory space.

read more →
layer 03 · the hive

The Hive

MariaDB-backed persistence. VINES owns state (per-request durability + crash-safe recovery). VECNA owns memory (cross-orchestration knowledge fragments, decay-aware recall).

read more →
Chapter 2 · The Install

Two commands. Sixty seconds.

If you have OpenClaw installed and a reachable MariaDB, the entire swarm comes up in two commands. The setup runner provisions both schemas, creates the six specialist agents from templates, and installs the Nexus protocol doc — all idempotent.

recommended: clawhub install path
openclaw plugins install clawhub:openclaw-hawkins
openclaw hawkins setup
Chapter 3 · The Pulse Protocol

Five named phases

When the Nexus decides a request isn't trivial, it enters The Pulse. Five phases, each with a documented entry condition and exit state. Trivial requests bypass the protocol entirely.

phase 01

Sensitivity Check

Does this request warrant the full protocol? Fires only when estimated work > 30s or spans > 2 specialist domains.

phase 02

Anchoring

Create a Linear parent ticket + a VINES orchestration row. These are the recovery anchors if the host reboots mid-flow.

phase 03

Deep Seeking

Optional research dispatch. Useful for unknown unknowns before committing to a plan.

phase 04

The Connection

Loop dispatch to each Tendril. Each sub-task gets its own Linear sub-ticket, runs in isolation, returns a structured reply.

phase 05

Consolidation

Synthesize the specialist replies into one operator-facing answer. Close all tickets, transition orchestration to success.

→ deeper

read on

The full Pulse contract, including failure / recovery semantics, is in concepts.

Chapter 4 · The Surface

12 typed tools

Every VINES and VECNA operation is exposed as a first-class OpenClaw tool. Any agent in your gateway can call them by name. TypeBox schemas at the runtime boundary — handlers operate on already-validated input.

Chapter 5 · Why Bother

What you get

crash-safe

Crash-safe orchestration

One orchestration_ledger row per request. Recovery scans for unfinished work on boot and cross-references Linear for the resume point. Distinguishes transient API failures from truly orphaned work — a flaky Linear network won't destroy live state.

no more amnesia

The swarm stops forgetting

When system-agent learns something useful at 3 PM, it writes to VECNA. When data-agent hits a similar problem at 9 PM, the Nexus pre-recalls it and injects the prior lesson into the next prompt. Decay-aware so stale entries lose weight automatically.

strict secrets policy

Secrets never in openclaw.json

The plugin's configSchema deliberately rejects mariadb.password and linear.apiKey. Secrets must come from the gateway environment (a 0600 systemd EnvironmentFile) — they can't sit as plaintext in openclaw.json.

linear-anchored

Linear-anchored recovery

Every orchestration carries a linearParentId. vines_recover walks the parent's children to figure out the last completed step and the next pending one — fully automated resume after any restart.

Chapter 6 · Two Audiences

Install path for humans and AI agents

The same plugin, two install paths. Humans get a step-by-step guide with copy-pasteable commands. AI installer agents get a SKILL.md manifest with DETECT → ACT → VERIFY gates so they can install autonomously without supervision.

for humans

/install/

Three install paths: ClawHub plugin (recommended), AI-agent installer, from-source. Each step has DETECT → ACT → VERIFY stanzas. 1Password-first credential handling. Worked examples.

→ open the install guide
for AI agents

/install/llm/

The same content as the SKILL.md manifest, formatted for an installer agent to follow autonomously. Phase 0 host probe, personalisation defaults table, recovery paths for every step that can fail.

→ open the LLM guide