Campaigns

Campaigns turn a feature spec into a reviewed, dependency-aware multi-agent workflow. AgentCraft plans the work, dispatches tasks into Git worktrees, integrates them on a Campaign branch, verifies the combined result, and keeps the final merge under your control.

Campaigns are enabled by default in AgentCraft 0.5.0. Set AGENTCRAFT_CAMPAIGN=0 before starting the server only when you need to disable the pipeline.

Start a Campaign

1. Open the Quest Board and select the Campaign tab.
2. Click + Submit Feature.
3. Describe the outcome, constraints, and acceptance criteria. Attach reference images for visual work.
4. Wait while a planning agent decomposes the spec.
5. Review the generated plan:
   • Refine opens a planning hero for conversational changes.
   • Review lets you inspect and edit tasks.
   • Approve Campaign begins execution.
   • Reject leaves the plan unstarted.

AgentCraft creates the Campaign branch when the first task dispatches. Each task runs from the latest Campaign state in its own worktree, then merges back before dependent work begins.

Monitor and Control It

An active Campaign card shows completed and running task counts, verification attempts, automatic repair status, and PR/CI state when configured.

• Chat with the Campaign orchestrator for status or changes.
• Add a task with the + control.
• Pause to stop new dispatches while current tasks finish.
• Resume after resolving a blocker.
• Cancel to stop the Campaign and its active heroes.
• Retry, edit, cancel, or manually review individual tasks when intervention is required.

Campaign heroes are grouped with task context in the roster. On the map, rings, dependency lines, completion flags, and failure markers make the execution graph visible.

Verification and Evidence

When the spec includes end-to-end or visual acceptance criteria, AgentCraft can capture a baseline before dispatch. A blocked baseline pauses the Campaign instead of starting from an unverified assumption.

After task execution, AgentCraft reviews the combined Campaign and runs final verification. Screenshots, recordings, the verification report, and the Campaign workpad persist under ~/.agentcraft/plan-memory/<planId>/, outside disposable worktrees.

The first failed final verification can create one critical automatic repair task and rerun the full verification once. A remaining failure moves to human review.

Review and Merge

Open Review & Merge to inspect:

• the combined diff;
• task results and evidence;
• baseline and final verification status;
• screenshots and walkthrough recordings;
• the Campaign workpad and verification summary.

The final action adapts to the workflow: Approve & Merge for a local branch, Create PR when the Campaign is ready for GitHub review, or Merge PR after PR feedback and checks are resolved.

Loop Delegation

Recurring loops receive linked Campaign status and recent loop history on every wake. If a loop discovers a concrete feature or fix that deserves planned delegation, its agent can submit a Campaign spec instead of implementing it during the loop turn.

AgentCraft links the plan to the loop, shows the relationship in the Active Loops panel, and deduplicates repeated submissions of the same spec from the same loop. Normal loop iterations do not create Campaigns unless the agent explicitly submits an actionable spec.

See Scheduled Tasks for loop setup and controls.

Recovery

Campaign state persists across server restarts. Task evidence and plan evidence live under ~/.agentcraft/task-memory/ and ~/.agentcraft/plan-memory/, so hero dismissal and worktree cleanup do not erase the review record.