Guide · Review and trust

What is explainable task assignment?

Explainable task assignment means an owner suggestion arrives with the reasons behind it, not just a name. Relay ranks up to three candidates, shows the signals that put them there and prints a confidence score. A reviewer can question the suggestion, then accept or change it with the evidence in view.

By the Relay maintainers · Last updated 19 November 2026 · Source on GitHub

What does explainable mean here?

A bare assignee is an assertion. An explainable suggestion is a claim with its basis attached: this person has worked on the module, spoke about it in the meeting, has room this week, and has not been overloaded. Relay stores each manual owner change as feedback, so the system can learn from corrections. The goal is not to remove the human decision; it is to give that decision something concrete to examine.

What signals does Relay use?

The documented signals are expertise from past work, recent context from the conversation, availability and a fairness share drawn from workload. Expertise comes from commit history and tracker issues; context comes from who discussed the topic. A confidence score summarises the ranking, but it sits beside the reasons rather than replacing them. Reviewers can request the explanation for a draft and see the source note that fed it.

Signals shown with an owner suggestion
SignalWhere it comes from
ExpertisePast work on the same topic
Recent contextWho discussed it in the source
AvailabilityCapacity and out-of-office status
FairnessCurrent workload across the team

How is a suggestion shown to a reviewer?

Relay returns three candidates in ranked order, each with a short reason and a score, plus the evidence line from the note. The reviewer picks one, or changes the owner entirely. A change is recorded as learning feedback rather than silently discarded. This keeps accountability with the person who made the call and gives future suggestions a better starting point.

In an illustrative draft about a billing webhook, the top candidate is described as owning that module, the second as having reviewed recent changes, and the third as being on the rota this week. The names and scores are demo values.

Illustrative demo data.

Can an AI agent change an owner?

An agent can read suggestions and ask Relay to explain them through the MCP surface. Changing an owner on a draft is a write, so it follows the same confirmation rule as other external writes: the tool call needs confirm: true. A reviewer can also restrict an agent to a viewer key, which permits reads but not approvals. The explainability record is kept in team memory alongside the decision.

What are the limits?

The ranking is a suggestion, not a guarantee of the best owner. Availability data is only as good as what the team records, and fairness depends on workload being up to date. The fairness and dedupe state are in-process, so a multi-instance deployment needs shared state. Live model output can vary between runs, which is another reason to keep a person in the loop before a ticket is pushed.

Next step

Read the triage guide to see where assignment sits in the wider flow, or self-host Relay and inspect a real suggestion against its source note.


Self-host Relay Read the workflow guide