RelayProduct guideOpen Relay

Product guide

How Relay works

Relay keeps work moving when people move on. This guide explains every feature and every number: what it means, exactly how it is calculated, and the code that does it.

The three headline features

Each one works on real records and shows the evidence behind what it says.

Handover Copilot answers#

Answers come only from the evidence behind the brief that you can see, and each one cites it.

How it is calculated

An answer is kept only if every citation points at real evidence in that set. Otherwise the reply is “not in the evidence”, with the closest sources to check. With AI off, you get a labelled search instead.

What it uses

  • The brief's cited evidence and the case's latest analysis, filtered for you
backend/src/relay/intelligence/copilot.py

What-if simulation#

A read-only rehearsal of someone's absence. Nothing is changed: no owner, task or record is written.

What it uses

  • The live catalogue, as you are allowed to see it
  • Open Relay tasks, GitHub pull requests and issues, Linear issues
  • The absence length, to flag deadlines that fall inside it

How to read it

Use it before a planned absence or as a regular drill. Plan a real handover turns the result into an audited case.

backend/src/relay/continuity/simulation.py

Memory Replay#

The history of one responsibility, rebuilt from its sources: owners, pull requests, issues, incidents, decisions and documents.

How it is calculated

Events are facts with their source date. Categories come from the source's labels, then from words in the title, else “implementation”. Chapters and connections are AI inferences that must cite events; anything citing nothing is dropped.

What it uses

  • GitHub, Linear, Notion and Relay records linked to the responsibility
backend/src/relay/intelligence/memory.py

Working with AI, safely

Features that do the busywork and stop before anything changes: a person accepts every result.

Handover Autopilot#

Runs a case's whole analysis in one go and shows each tool it uses as it works: reading the case, gathering evidence, checking continuity rules, four AI reviewers, then the exit interview and handover brief.

How it is calculated

Each reviewer may only search and read the evidence you can see. A point or proposal citing nothing, or citing evidence that doesn't exist, is dropped and counted as “uncited dropped”. Proposals wait for a person; Autopilot never applies anything.

What it uses

  • The case and the responsibilities it covers
  • Relay, GitHub, Linear and Notion records you are allowed to see

How to read it

Review the proposals under the run, accept what is right, and reject the rest with a reason so it isn't suggested again.

backend/src/relay/intelligence/autopilot.py

Ask Relay#

Ask anything about the workspace in plain language. Relay searches the records you can see, then writes an answer that cites each one.

How it is calculated

Search ranks responsibilities, knowledge, cases, people and synced items by text relevance; the top matches become the evidence. Each sentence of the answer must cite them; a reply citing nothing is shown as “not in the evidence”.

What it uses

  • Records you can open yourself; restricted knowledge only for managers

How to read it

Click a source chip to open the record behind it.

backend/src/relay/intelligence/ask.py

Meeting intelligence#

Turns meeting notes or a transcript into decisions, actions, open questions and risks, each linked to the transcript lines it came from.

How it is calculated

With AI on, the model lists items citing line refs (L1, L2…); an item citing no line or a line that doesn't exist is dropped. With AI off, rules find decisions (“decided”, “agreed”), actions (“I'll”, “will send”, “by Friday”), risks (“blocker”, “only she knows”) and unanswered questions.

What it uses

  • The transcript you paste, and nothing else

How to read it

Accepting a decision records it as decision knowledge on the linked responsibility; accepting an action adds a task to the linked case, assigned to the person named if they are a member.

backend/src/relay/intelligence/meetings.py

Decision log#

Every decision in the workspace in one timeline, so a successor can see what is settled, when, and where it was decided.

How it is calculated

Merges decision knowledge (ADRs), decisions from meetings (until accepted ones become knowledge), and decisions reconstructed by case analyses that weren't dismissed. Duplicates with the same wording are shown once.

What it uses

  • Only records you are allowed to see; managers-only decisions stay hidden from members
backend/src/relay/intelligence/meetings.py

30/60/90 continuation plan#

A first-90-days plan for whoever takes over the case's work, built from the case itself.

How it is calculated

For each responsibility in scope: no recipient → “name who takes over” (day 30). Critical or high → the recipient shadows the previous owner (30), and a backup is trained (90). Every item: the recipient leads it (60) and reviews its knowledge (90); with no written knowledge, “document how it works” (60). Unmet mandatory readiness checks go in the first 30 days; open tasks land in the phase their due date falls in.

What it uses

  • The case's scope, recipients, readiness checks and open tasks
  • Which responsibilities have active knowledge

How to read it

Add the steps you agree with to the case's tasks; they get the phase's due date.

backend/src/relay/continuity/plan.py

Handover rehearsal#

Practise the handover before you need it: drills built from this brief's evidence, each answer checked against its source.

How it is calculated

Drills come from the brief's procedures, decisions, open work and readiness checks you may see, one of each kind first, up to six. Each source's key terms are its title words, then its most frequent meaningful words (up to 8). Score = key terms your answer uses ÷ key terms × 100. Ready ≥ 70, partly ≥ 35. No AI is involved, so the grade is the same every time.

What it uses

  • The handover brief's evidence, filtered to what you are allowed to see

How to read it

Missed terms point at what to re-read; the source excerpt is shown under the grade. Only the score is recorded, never your answer.

backend/src/relay/intelligence/rehearsal.py

Numbers and how they are calculated

Every number in Relay can be traced to the records behind it. The ⓘ next to a number in the app opens the same definition you see here.

Readiness score#

How well the work would carry on if this person were unavailable, out of 100. 100 means nothing they hold would be left exposed.

How it is calculated

For each affected responsibility, weighted by criticality (critical 4, high 3, standard 2, low 1): someone still covers it 60%, it has usable (non-stale) knowledge 25%, and it keeps a second holder 15%. Score = earned ÷ possible × 100, rounded.

What it uses

  • Current primary, co-owner and backup assignments, including team holders
  • Active knowledge and whether it is past its review window
  • Only responsibilities this person holds; everything else is unaffected

How to read it

Below 60 usually means critical work would have nobody. Add a backup in the simulator to see which change moves the score most.

backend/src/relay/continuity/simulation.py (_score)

Overload flag#

Marks a cover candidate who would end up holding far more than is usual in this workspace.

How it is calculated

Overloaded when responsibilities held after the change ≥ max(2 × typical, typical + 3). Typical = the average number held by people who hold any.

What it uses

  • Active ownership assignments across the workspace

How to read it

Prefer a candidate without the flag, or split the work between two people.

backend/src/relay/continuity/simulation.py (candidates)

Continuity Debt#

A backlog, not a grade: the unresolved continuity problems in scope, added up so you can see whether it is shrinking.

How it is calculated

Open ownership risks (high 5, medium 2, low 1) + responsibilities with no knowledge (critical 3, high 2, standard 1) + responsibilities whose knowledge is all overdue for review (critical 2, high 1) + overdue tasks (1 each) + cases past their end date (2 each).

What it uses

  • Accepted risks are excluded and listed separately, so the number can't be lowered silently
  • Every component lists the items behind it

How to read it

Open a component to see exactly which items add points, then fix the largest first.

backend/src/relay/insights/service.py

Continuity risks#

Rule-based checks on the current catalogue. No AI and no hidden score: each risk names the rule that raised it.

How it is calculated

No primary owner (high for critical/high work) · No backup (high for critical work) · Critical work with fewer than 2 owners (high) · An owner is archived (high for critical/high work) · One person is the only owner of 2 or more responsibilities (high if any is critical/high).

What it uses

  • Current assignments, criticality, and whether holders are archived

How to read it

A risk resolves by itself when the gap is fixed. You can acknowledge it, or accept it until a date with a reason.

backend/src/relay/risks/service.py

Ownership coverage#

How much of the work has someone accountable for it, and someone to step in.

How it is calculated

With a primary = responsibilities with a current primary owner ÷ all in scope. With a backup = responsibilities with a current backup ÷ all in scope. Critical coverage counts critical work only.

What it uses

  • Active (non-archived) responsibilities in the selected scope
backend/src/relay/insights/service.py (CoverageInsight)

Key-person exposure#

People who are the only holder of work. It describes where coverage is thin, not how anyone performs.

How it is calculated

Responsibilities where this person is the only current owner, critical first.

What it uses

  • Active ownership assignments; team-held work counts the team, not one member

How to read it

Add a backup for their critical work, or run the simulator for them.

backend/src/relay/insights/service.py

Knowledge freshness#

Whether a piece of knowledge has been confirmed recently enough to rely on.

How it is calculated

Fresh when reviewed (or written by a person) within the workspace's review window, 90 days by default (7 to 730). Imported content stays unreviewed until a person confirms it.

What it uses

  • Last review date, origin (person or import), expiry date, and disputed or draft state

How to read it

Stale knowledge still shows, labelled, so you can decide whether to trust it.

backend/src/relay/catalog/knowledge.py (_freshness)

Handover follow-through#

After a case completes, whether each transferred responsibility actually took hold with its new owner.

How it is calculated

Settled: the new holder is recorded, has reviewed or worked on it, and its tasks are closed. Settling in: still within 14 days. Check in: 14 days without a review or any recorded activity, or open tasks.

What it uses

  • Ownership records, reviews, linked activity and case tasks after completion

How to read it

It measures the work, not the person. A check in is a prompt for a conversation.

backend/src/relay/continuity/follow_through.py

Expertise map#

Who holds each responsibility and who has written down how it works, so you can see where one absence would stop the work.

How it is calculated

People per responsibility = distinct people who hold it (primary, backup or other role) or wrote its active knowledge. A named backup is assigned but may never have done the work; a capable backup has also written its knowledge. Rows run from fewest people and most critical down.

What it uses

  • Current ownership assignments (teams count as shared cover, shown separately)
  • Active knowledge items and who wrote them

How to read it

A red 1 on critical work is a single point of failure: name a backup, then ask them to document one procedure so they become capable.

backend/src/relay/search/service.py

Trust in what you read

How Relay separates what it knows from what it suggests.

Fact, Rule and AI inference#

Every statement in a brief says where it comes from, so you know how far to trust it.

How it is calculated

Fact: read directly from a record. Rule: produced by a deterministic Relay rule. AI inference: proposed by a model from cited evidence, never stored as truth until a person accepts it.

What it uses

  • Only evidence the reader is allowed to see; restricted items are removed first
backend/src/relay/intelligence/validate.py

Confidence#

How strongly the cited evidence supports an AI statement: high, medium or low.

How it is calculated

Set by the model from the evidence it cites, and lowered where it infers a reason the evidence doesn't state. Facts and rules carry no confidence: they are read, not judged.

What it uses

  • Cited evidence only

How to read it

Treat low confidence as a question to ask the previous owner, not an answer.

backend/src/relay/intelligence/capabilities.py

Principles

What AI does, and what it never does

  • AI drafts: handover briefs, answers to questions, chapters of a project's history, interview questions and coordination messages.
  • Every AI statement cites the evidence it rests on. A statement citing nothing, or citing something that doesn't exist, is dropped.
  • AI never changes anything. Models only hold read-only tools; owners, cases and knowledge change only when a person accepts a proposal.
  • With AI switched off, Relay still works: rule-based briefs, labelled search and every deterministic number stay available.

Permissions

  • Relay shows only what your role allows: owner, admin, continuity manager, knowledge steward, team manager, member, auditor or guest. Unknown roles and permissions are denied.
  • Restricted and confidential knowledge is limited to managers by a database rule, not just by the interface.
  • AI receives only the evidence the person asking may see, so a citation never reveals something you couldn't open yourself.
  • Every workspace is isolated at the database level, and every access decision is written to an append-only audit log.

When sources disagree

  • A conflict is shown only when two distinct sources say different things, and both are cited side by side.
  • Relay never picks a winner silently: the person reviewing the brief decides, and the decision is recorded.

When data is stale or missing

  • Each connected source shows when it last synced. A failed or partial sync is labelled wherever its data is used.
  • Knowledge past its review window stays visible with a stale label, so you can judge it yourself.
  • Briefs list what they could not see, so the reader knows what is missing as well as what is there.

Roles: who does what in a handover

Eight roles, each for a real job. Permissions come from one registry and are enforced inside every service and by the database. In a sample workspace you can try Relay as any role from the banner at the top.

owner
Accountable for the workspace itself: who administers it and whether it exists.
admin
Runs the workspace day to day: members, integrations, AI settings and policies.
continuity manager
Runs handovers: opens cases, approves plans and moves ownership, without needing admin rights.
knowledge steward
Keeps knowledge trustworthy: verifies it, settles disputes and retires what is out of date.
team manager
Leads a team: manages the responsibilities their team holds, nothing else.
member
Does the work: writes knowledge, completes tasks and receives handovers.
auditor
Checks compliance: reads everything, including the audit log, and changes nothing.
guest
An outside contractor or consultant: can read what they need, never change it.
What each role can do
Can they…owneradmincontinuity managerknowledge stewardteam managermemberauditorguest
See the workspace, people, responsibilities and casesYesYesYesYesYesYesYesYes
Read AI output: briefs, Autopilot results, Ask RelayYesYesYesYesYesYesYesNo
Comment on casesYesYesYesYesYesYesNoNo
Write knowledge: runbooks and decisionsYesYesYesYesYesYesNoNo
Verify knowledge and settle disputesYesYesNoYesNoNoNoNo
Manage responsibilitiesYesYesYesNoOwn teams onlyNoNoNo
Open and run handover casesYesYesYesNoNoNoNoNo
Approve handovers and sign off readinessYesYesYesNoNoNoNoNo
Run AI: Autopilot and case analysisYesYesYesNoNoNoNoNo
Accept or reject AI proposalsYesYesYesNoNoNoNoNo
Run the What If? simulatorYesYesYesNoNoNoNoNo
See insights: health and continuity debtYesYesYesNoNoNoYesNo
Read the audit logYesYesNoNoNoNoYesNo
Manage members and invitesYesYesNoNoNoNoNoNo
Connect GitHub, Linear, Notion and SlackYesYesNoNoNoNoNoNo
Change policies and automationsYesYesNoNoNoNoNoNo
Delete the workspaceYesNoNoNoNoNoNoNo

◐ own teams only. Members see AI output built only from evidence the whole workspace may read.