The open schema

A relay is one JSON document, relay.context-package.v1. It carries what you’re sharing, why, with whom, and the rules for how their agent may speak for you. The product is Relay’s; the format is open, so anyone can see exactly what is shared and how it may be used.

Your words stay yours

Your intent is quoted, never paraphrased. The model’s reading sits in its own block, and you confirm it.

Every value says where it came from

Stated by you, inferred by a model, or taken from a named source. Nothing unmarked.

Rules travel with the context

The recipient’s agent loads the mandate before the content: what it may say, what it must ask you, what it must never do.

Four questions, 11 blocks

Every package answers the same four questions. Pick one to see its blocks and fields.

Who is it for, and why?

The owner, the one audience it’s written for, and what the owner wants to happen, in their own words.

owneraudienceintent

What did the agent read into it?

The model’s interpretation, kept apart from what the owner said, and reviewed by the owner before anything is shared.

interpretation

What is being shared?

The context itself, and, at deeper levels, how the owner thinks about it.

contextreasoningjudgementprocess

How may their agent speak for you?

The mandate the recipient’s agent loads first: its role, how it opens, what it must never do, and where unanswered questions go.

representationinteractionlifecycle

Around them, the envelope

Identity and approval. The server overwrites the version and approval fields: an agent can draft a package, but only the owner can approve one.

schemaidtitleversiondepthapprovedByapprovedAtproducedBy

Who is it for, and why?

The owner, the one audience it’s written for, and what the owner wants to happen, in their own words.

`owner`

The person whose context this is. Only they can approve it.

namerequired
id
contact
How the audience reaches the human, not the agent.

`audience`

One named recipient (a person, a team or an app), with what they already know and what they care about, so their agent pitches it right.

namerequired
kind
Who receives it. An app (an integration) receives through its own agent, under the same grant, mandate, and receipts as a person.
personteamapp
email
Confirmed by the owner before sharing. Grants are issued to this identity.
relationshiprequired
e.g. co-founder, investor, hiring manager, client
role
organization
knowsAlready[ ]
Shared ground the agent can assume. Prevents over-explaining and wrong framing.
caresAbout[ ]
The owner's model of what matters to them.
likelyQuestions[ ]
register
Tone and level, e.g. 'direct, technical, skip the pitch'

`intent`

What the owner wants to happen. statement is kept verbatim; the outcome, the ask and the priorities sit next to it.

statementrequired
The owner's own words, unedited.
outcomerequired
3 fields inside
- textrequired
- provenancerequired
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- agentReading
  When corrected: what the agent first proposed. Kept so drift between model and owner is visible.
successLooksLike[ ]
3 fields inside
- textrequired
- provenancerequired
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- agentReading
  When corrected: what the agent first proposed. Kept so drift between model and owner is visible.
askrequired
3 fields inside
- typerequired
  decideagreereviewevaluateadviseexploreactinform
- detailrequired
- byWhen
whyNow
3 fields inside
- textrequired
- provenancerequired
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- agentReading
  When corrected: what the agent first proposed. Kept so drift between model and owner is visible.
whyThisAudience
3 fields inside
- textrequired
- provenancerequired
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- agentReading
  When corrected: what the agent first proposed. Kept so drift between model and owner is visible.
interests[ ]
What the owner actually needs underneath the ask (interests, not positions). This is what lets their agent find options the owner did not list.
3 fields inside
- textrequired
- provenancerequired
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- agentReading
  When corrected: what the agent first proposed. Kept so drift between model and owner is visible.
priorities[ ]
Ordered. The first one wins a conflict.
3 fields inside
- textrequired
- provenancerequired
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- agentReading
  When corrected: what the agent first proposed. Kept so drift between model and owner is visible.
nonGoals[ ]
What this is explicitly not about. Stops scope creep in the conversation.
3 fields inside
- textrequired
- provenancerequired
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- agentReading
  When corrected: what the agent first proposed. Kept so drift between model and owner is visible.

What did the agent read into it?

The model’s interpretation, kept apart from what the owner said, and reviewed by the owner before anything is shared.

`interpretation`

The restatement the owner confirms or corrects, each inference with its basis and the owner’s verdict, and the ambiguities the owner chose to leave open.

restatementrequired
'Here is what I think you want', written by the agent and shown to the owner before anything else.
restatementVerdict
confirmedcorrectedrejectedunreviewed
inferences[ ]required
6 fields inside
- idrequired
- readingrequired
  What the agent concluded.
- basisrequired
  Why: which words, sources, or patterns.
- confidence
  highmediumlow
- verdictrequired
  confirmedcorrectedrejectedunreviewed
- correction
  The owner's version, when the verdict is corrected.
clarifications[ ]
Questions the agent asked the owner while drafting, with the owner's answers.
2 fields inside
- questionrequired
- answerrequired
openAmbiguities[ ]
What the owner chose to leave open. The recipient agent must not resolve these on its own.

What is being shared?

The context itself, and, at deeper levels, how the owner thinks about it.

`context`

Sections of readable prose, each broken into claims an agent can cite. Only topics the owner named as off-limits are listed as excluded.

subjectrequired
The subject in the owner's context store. Sections are stored once and shared by reference.
sections[ ]required
8 fields inside
- idrequired
  Store item id. The same id across every relay that includes it.
- titlerequired
- purpose
  Why this section is in this relay: how it serves the intent.
- bodyrequired
  Readable prose for humans.
- claims[ ]
  The atomic statements in the prose, so agents can cite and check them.
- versionrequired
- updatedAt
- visibility
  shareable = the recipient may pass it on. recipient-only = for them alone.
  shareablerecipient-only
excluded[ ]
Only topics the owner explicitly named as off-limits, so their agent knows a gap is intentional. Agents never populate this on their own or list what they left out.
2 fields inside
- topicrequired
- reasonrequired

`reasoning`

optional

Decisions and what would reopen them, variables and how far they can move, assumptions, evidence, risks, the strongest case against, and unknowns.

decisions[ ]
12 fields inside
- idrequired
- questionrequired
- statusrequired
  decidedleaningopendelegateddeferred
- decider
  owneraudiencejointother
- byWhen
- options[ ]
- choice
- rationale
- criteria[ ]
- tradeoffsAccepted[ ]
- reversibility
  reversiblecostlyirreversible
- whatWouldChangeIt[ ]
  New facts that would reopen the decision. The most useful thing to give a recipient agent.
variables[ ]
The parameters in play and how much they can move.
7 fields inside
- namerequired
- value
- acceptable
  Range or condition the owner can live with.
- gaterequired
  hard = non-negotiable; elastic = can move with a reason or a trade; soft = a preference.
  hardelasticsoft
- tradesAgainst[ ]
- provenance
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
- visibility
  shareable = the recipient may pass it on. recipient-only = for them alone.
  shareablerecipient-only
assumptions[ ]
6 fields inside
- statementrequired
- confidencerequired
  highmediumlow
- basis
- ifWrong
- howToTest
- provenance
  stated = the owner's words. confirmed = drafted by an agent, accepted unchanged by the owner. corrected = drafted by an agent, changed by the owner. inferred = the agent's reading, not reviewed. sourced = taken from a named source. derived = computed by the system.
  statedconfirmedcorrectedinferredsourcedderived
evidence[ ]
4 fields inside
- supportsrequired
  Claim or decision id.
- kindrequired
  documentdataexperiencethird-partyobservationconversation
- pointer
- strengthrequired
  strongmoderateweak
risks[ ]
4 fields inside
- riskrequired
- likelihood
  lowmediumhigh
- impact
  lowmediumhigh
- mitigation
counterarguments[ ]
The strongest case against the owner's position, with the owner's response. Lets the recipient agent engage critically without inventing the owner's reply.
2 fields inside
- argumentrequired
- ownerResponse
unknowns[ ]
3 fields inside
- questionrequired
- whyItMatters
- whoCanAnswer

`judgement`

optional

The owner’s own principles, preferences and red lines, so their agent can reason the way the owner would.

principles[ ]
2 fields inside
- principlerequired
- example
preferences[ ]
3 fields inside
- preferrequired
- overrequired
- when
redLines[ ]
confidence
How sure the owner is overall, in their words.

`process`

optional

How the package was produced, which sources were consulted with consent, who decides from here, and next steps.

howProduced[ ]
2 fields inside
- steprequired
- byrequired
  owneragentaudience
sourcesConsulted[ ]
4 fields inside
- namerequired
- kindrequired
  conversationdocumentconnectormemoryweb
- consent
  grantednot-needed
- scope
decisionProcess
From here on: who decides what, by when, and how disagreement is resolved.
nextSteps[ ]
3 fields inside
- steprequired
- owner
- due

How may their agent speak for you?

The mandate the recipient’s agent loads first: its role, how it opens, what it must never do, and where unanswered questions go.

`representation`

The rules. Read before the content, every time.

rolerequired
explain = answer from the package; explore = also help the recipient think it through; advise-recipient = may give the recipient its own view, labelled; negotiate-within-bounds = may agree to changes inside elastic and soft variables only.
explainexploreadvise-recipientnegotiate-within-bounds
opening
How the recipient's agent starts. It takes the package in silently (no summary, no echo) and opens with this greeting and these starting points, then follows the recipient's lead.
2 fields inside
- greetingrequired
  One or two sentences: whose context this is, what it's about, what it's for.
- starters[ ]required
  Starting points drawn from the ask. Always offered alongside 'or something else'.
speaksAs
e.g. 'I'm answering from the context Saurabh shared with you, version 3.'
fidelityrequired
4 fields inside
- quoteStatedIntentrequired
  Quote intent.statement verbatim instead of paraphrasing it.
- labelInterpretationsrequired
  Say 'the owner's agent read this as…' for anything with provenance inferred.
- citeSectionsrequired
- extrapolationrequired
  Whether the agent may reason past the package using judgement.principles, and if so, whether it must say so.
  neverlabelledfree
may[ ]
mustAskOwner[ ]required
Topics that are routed back to the owner as a question, never answered.
mustNot[ ]required
whenUnknownrequired
say-unspecified-and-route = tell the recipient and send the question to the owner through the return channel.
say-unspecified-and-routesay-unspecified
passOn
What the recipient may forward. Shareable items only; recipient-only items stay with them.
fidelityChecks[ ]
Questions the audience is likely to ask, with the answer the owner approved during review. The recipient agent uses them as calibration examples, and they double as a regression test for faithful representation.
3 fields inside
- questionrequired
- approvedAnswerrequired
- ownerEdited

`interaction`

Whether the recipient can only read, suggest changes, or rewrite sections, and where their questions go.

permissionrequired
readsuggestedit
returnChannel
Where questions and proposals go. Each targets a section id.
expectedResponse
What the owner hopes to get back, and by when.

`lifecycle`

Live (they see each approved update) or snapshot (frozen at what they accepted), plus expiry and revocation.

moderequired
livesnapshot
expiresAt
refresh
manual24h7d30d
staleSections[ ]
grantId
revocationUrl

Ideas that run through every block

Provenance

Every value carries one of six labels. The recipient’s agent quotes what’s stated, and says “their agent read this as…” for anything inferred. Highlighted: the labels that mean the owner saw it.

stated: the owner's words
confirmed: drafted by an agent, accepted unchanged by the owner
corrected: drafted by an agent, changed by the owner
inferred: the agent's reading, not reviewed
sourced: taken from a named source
derived: computed by the system

Depth

How much was captured. Each level adds to the one before.

briefIntent and context
workingAdds decisions, variables and assumptions
fullAdds judgement, counterarguments, process and rehearsed answers

Gates on variables

How far each parameter can move in a conversation.

hard: non-negotiable
elastic: can move with a reason or a trade
soft: a preference

Their agent’s role

From answering only, to agreeing to changes inside the bounds you set.

explain: answer from the package
explore: also help the recipient think it through
advise-recipient: may give the recipient its own view, labelled
negotiate-within-bounds: may agree to changes inside elastic and soft variables only

Visibility

Per section or claim.

shareable: the recipient may pass it on
recipient-only: for them alone

Use it

Validate packages against the schema, read the worked example, or teach your own agent to prepare relays.

The open schema

Your words stay yours

Every value says where it came from

Rules travel with the context

Four questions, 11 blocks

Who is it for, and why?

What did the agent read into it?

What is being shared?

How may their agent speak for you?

Around them, the envelope

Who is it for, and why?

`owner`

`audience`

`intent`

What did the agent read into it?

`interpretation`

What is being shared?

`context`

`reasoning`

`judgement`

`process`

How may their agent speak for you?

`representation`

`interaction`

`lifecycle`

Ideas that run through every block

Provenance

Depth

Gates on variables

Their agent’s role

Visibility

Use it

JSON Schema

Worked example

Extraction standard

The relay skill