Task Handoff guide
One claimed task, carried through one supervised run.
The runner copies exact Wagglet instructions into an agent, stays responsible for the human checks, then returns to the same session with a separate delivery prompt.
When to use it
Use Task Handoff for a bounded unit of work
- Good fit
- A clear outcome, named scope, permitted context, credible checks, and a runner qualified to supervise the work.
- Poor fit
- Ambiguous goals, sensitive data the runner may not access, high-risk judgment without a qualified reviewer, or work intended to stay Human-only.
- Agent responsibility
- Read the current task, stay inside scope, perform the work and checks, stop on missing authority, and report what actually happened.
- Human responsibility
- Start the right environment, provide authorized access and judgment, perform the named checks, stop unsafe work, and route the result to review.
Runner workflow
From claim to review
- 01
Read before you claim
Check the outcome, LLM instructions, Human instructions, repository, branch, attachments, and acceptance evidence. Claim only work you may access and can supervise with your own agent account.
- 02
Copy the LLM instructions
While you are the current claimer, use Copy LLM instructions and paste the complete result into a fresh agent session. Wagglet includes the official skill URL, a ready-to-run current-context command, and a fallback task snapshot.
- 03
Let the agent read current bounded context
The starting credential is read-only. It can refresh the claimed task and, when that task belongs to a Story, inspect the permitted Story overview and relevant published linked tasks. It cannot browse the board, comment, edit, change status, or deliver.
- 04
Supervise the run
Follow the Human instructions. Supply only authorized access, answer decisions you own, inspect the real environment, and perform the required human checks. Stop when the task leaves scope or a specialist decision is needed.
- 05
Copy the separate delivery prompt
When the work is ready, use Deliver and paste its report prompt into the same agent session. That later prompt carries the separate one-time capability used to file the branch, outcome, changes, warnings, and useful failure information.
- 06
Leave the verdict to review
Delivery records what the runner and agent produced. An authorized reviewer still accepts it as Done, sends it back to the same runner, or reopens it for a new claim. Acceptance, Git merge state, and deployment remain separate facts.
- Prompt 1ReadClaim-rooted, refreshable context
- TogetherRunAgent work, human access and checks
- Prompt 2DeliverSeparate one-time report capability
- ReviewerAccept or returnA second decision closes work
Capability boundary
What the copied credential can—and cannot—reach
- The read credential is rooted in the exact active claim, stored only as a hash, sent in an authorization header, and expires after at most seven days.
- Ending the starting claim invalidates it. Moving or unlinking the starting Story item also invalidates the traversal boundary.
- Story scope is live: a newly linked published task may enter scope, while an unlinked or moved task leaves scope immediately.
- Another person's Private Draft never enters agent context. Missing or truncated context is named explicitly so the agent should ask instead of guess.
- Attachment URLs returned in Task Handoff are independent bearer links. Keep them private; revoking the read credential does not revoke a URL already returned.
Canonical procedure
Give the agent the raw skill
Read or link the canonical Wagglet Task Handoff Skill. If you need ongoing search and actions across the records already authorized for you, use the separate MCP Workspace Workflow instead.