hive-coordination

name	hive-coordination
description	Coordinate work across multiple agents and projects in Hive. Covers task claims, blockers, handoffs, campaigns, portfolio management, briefs, and shared memory.

Hive Coordination

This skill covers how agents work together and how work is orchestrated across projects.

Multi-Agent Task Coordination

Agents coordinate on canonical task state in .hive/tasks/*.md, not on prose or checkboxes.

Core tools

hive task ready --json                                      # what is available
hive task claim <id> --owner <name> --ttl-minutes 30 --json # take a lease
hive task release <id> --json                               # give it back
hive task link <src> blocked_by <dst> --json                # declare dependency
hive deps --json                                            # blocker summary
hive context startup --project <id> --task <id> --json      # build context for one task

Pattern: Sequential Handoff

When one task should unlock another:

hive task link <implement-task> blocked_by <research-task> --json

1. Agent A claims and completes the research task
2. Agent A releases it, marks it done
3. The implement task enters the ready queue automatically
4. Agent B claims and continues

Pattern: Parallel Independent Work

When several tasks can run simultaneously:

hive task ready --project-id <project-id> --json

Each agent claims a different task. Avoid overlapping file ownership when possible.

Pattern: Review Loop

When work needs evaluator gates:

hive run start <task-id> --json         # agent implements
hive run eval <run-id> --json           # evaluator checks
hive run accept <run-id> --json         # reviewer approves

One agent implements, another reviews artifacts, a human accepts or escalates.

Common Mistakes

• Flipping owner in AGENCY.md instead of using hive task claim
• Using checkbox text as the dependency graph instead of task links
• Skipping PROGRAM.md when runs or evaluators are involved
• Leaving long-lived claims after the work is done

Campaigns

Campaigns orchestrate multiple tasks across a project with scheduling and lane-based allocation.

Create a campaign

hive campaign create --title "Sprint 1" --goal "Ship MVP features" \
  --project-id <id> --type delivery --driver local --cadence daily --json

Monitor campaigns

hive campaign list --json                       # all active campaigns
hive campaign list --project-id <id> --json     # scoped to project
hive campaign status <campaign-id> --json       # health, runs, lane allocations

Tick a campaign

hive campaign tick <campaign-id> --json

A campaign tick executes one cycle: recommend tasks → start runs → review results.

Campaign lanes

Each campaign allocates work across four lanes:

Lane	Purpose
exploit	Production delivery (default)
explore	Research and experimentation
review	Evaluation queue management
maintenance	Technical debt and ops

Lane quotas are configurable per campaign.

Portfolio Management

Portfolio commands operate across all projects in the workspace.

Portfolio status

hive portfolio status --json

Returns all projects, ready tasks, active runs, and the evaluation queue.

Steer a project

hive portfolio steer <project-id> --pause --json          # pause project
hive portfolio steer <project-id> --resume --json         # resume project
hive portfolio steer <project-id> --focus-task <id> --json  # focus on one task
hive portfolio steer <project-id> --boost 2 --json        # priority boost
hive portfolio steer <project-id> --force-review --json   # force review cycle

Portfolio tick

hive portfolio tick --json                      # recommend next task (default mode)
hive portfolio tick --mode start --json         # claim + start work on recommended task
hive portfolio tick --mode review --json        # finish the next evaluating run
hive portfolio tick --mode cleanup --json       # clean up terminal runs

Each tick runs one bounded phase. The default --mode recommend only returns a recommendation — it does not start work.

Briefs

Generate executive summaries of portfolio activity:

hive brief daily --json     # daily summary of changes
hive brief weekly --json    # weekly portfolio summary

Shared Memory

Memory lets agents persist and share knowledge across sessions.

Record observations

hive memory observe --note "..." --scope project --project <id> --json
hive memory observe --transcript-path <file> --scope global --json

Generate reflections

hive memory reflect --scope project --project <id> --propose --json

Reflections are review-gated. A human must approve before they become canonical:

hive memory accept --scope project --project <id> --json
hive memory reject --scope project --project <id> --json

Search memory

hive memory search "query text" --scope all --limit 8 --json
hive memory search "query text" --scope project --json

Memory scopes: project (per-project knowledge), global (workspace-level knowledge), all.

Rules of Thumb

• Coordinate on task IDs, not prose.
• Use task claims for leases and task links for blockers.
• Treat AGENCY.md as narrative context, not a lock file.
• Refresh projections after coordination changes: hive sync projections --json.
• Campaign ticks respect lane quotas and budget policies — do not bypass them.
• Memory reflections require human approval. Do not assume proposed reflections are canonical.