Web Authorization Protocol K. McGuinness Internet-Draft Independent Intended status: Standards Track 13 August 2026 Expires: 14 February 2027 Mission-Bound Authorization for OAuth 2.0 draft-mcguinness-oauth-mission-latest Abstract An AI agent is typically given a mission: a task to pursue on a user's behalf. OAuth 2.0 issues access tokens for individual resource requests, but it has no durable, approved artifact that ties those tokens to the one task a user actually authorized. As a result, an agent's authority is a collection of independently obtained tokens with no shared, auditable boundary, and a user's approval is disconnected from what the agent later does. This document defines a Mission: a structured, explicitly approved, integrity-bound authorization artifact for OAuth 2.0. A client submits a Mission Intent through Pushed Authorization Requests; the Authorization Server derives Rich Authorization Requests authorization details from it, binds the approved task and its derived authority to the Approver's consent through integrity anchors, and records a durable Mission. Every access token derived under the Mission carries that authority and a "mission" claim, and issuance is gated on the Mission's lifecycle state. Optional capabilities represent delegation among agents with the OAuth Actor Profile and, as specified by a companion, let a single Mission be honored across trust domains. This is the issuance and governance "mission layer" left unspecified by agent-identity work for OAuth; runtime enforcement of each action is a separate, optional layer. About This Document This note is to be removed before publishing as an RFC. The latest revision of this draft can be found at https://mcguinness.github.io/mission-bound-authorization/draft- mcguinness-oauth-mission.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-mcguinness- oauth-mission/. Source for this draft and an issue tracker can be found at https://github.com/mcguinness/mission-bound-authorization. Status of This Memo This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at https://datatracker.ietf.org/drafts/current/. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." This Internet-Draft will expire on 14 February 2027. Copyright Notice Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License. Table of Contents 1. Introduction 1.1. Why a New Object 1.2. Relationship to Other Authorization Objects 1.3. Relationship to GNAP 1.4. Capability-System Prior Art 1.5. The Mission, the Plan, and Execution 1.6. Relationship to Agent-Identity Specifications 1.7. Applicability 1.8. Scope and Future Work 1.9. Non-Goals 2. Conventions and Terminology 3. Overview 3.1. Principal Model 3.2. Authority Sources 3.3. Protocol Flow 4. Mission Intent 4.1. Submission via PAR 4.2. The Authority Proposal 4.3. Intent Submission Evidence 5. Mission Authority 5.1. Subset Rule 5.2. Common Constraints 5.3. Other Authorization Details Types 5.4. Modeling Tools and Function Calls 6. Mission Approval 6.1. Binding the Mission to the Grant 6.2. Single Accountable Approver 6.3. Integrity Anchors 6.4. Canonicalization Rules 7. Mission Record 7.1. Mission Identifier Format 7.2. Worked Example 8. Mission-Bound Access Tokens 8.1. The Mission Claim 8.2. Resource Server Enforcement 8.3. Remediation Grains 9. Mission Lifecycle and Gating 9.1. Issuance Gating 9.2. Revocation 10. Mission State via Token Introspection 10.1. Caller Authorization and Minimization 10.2. Composite Active State 10.3. Only the Issuer Reports Mission State 10.4. Examples 11. Delegation Within a Mission 11.1. Adopted Model: client_id Names the Requesting Client 11.2. Self-Exchange Down-Scoping 11.3. Delegation Constraints 11.4. Worked Example: Delegated Token 12. Extensibility 13. Authorization Server Metadata 14. Protected Resource Metadata 15. Conformance 16. Security Considerations 16.1. Consent Binding 16.2. Mission Drift 16.3. Downgrade by Omission 16.4. Prompt Injection and the Exfiltration Leg 16.5. Authority Does Not Propagate With Information 16.6. Issuance Scope, Not Runtime Enforcement 16.7. Resource Boundary Canonicalization 16.8. Denial Detail Disclosure 16.9. Token Theft 16.10. Delegation and Chain Compromise 16.11. client_id Conformance and the Approved-Agent Residual 16.12. Signing and Key Rotation 16.13. Compromised or Over-Broad Derivation 16.14. Authority Hash Is Not a Mission Identifier 16.15. Composition and the Effective Ceiling 16.16. The Containment Materialized-Capability Residual 17. Privacy Considerations 17.1. Mission Identifier Correlation 17.2. Token Payload Disclosure 17.3. Intent Retention and Anchor Disclosure 17.4. Third-Party Data Subjects 17.5. Mission Record and Evidence Access 18. IANA Considerations 18.1. OAuth Parameters Registration 18.2. OAuth Extensions Error Registration 18.3. The Mission Resource Access Authorization Details Type 18.4. JSON Web Token Claims Registration 18.5. OAuth Token Introspection Response Registration 18.6. OAuth Authorization Server Metadata Registration 18.7. OAuth Protected Resource Metadata Registration 18.8. Common Constraints Registry 19. References 19.1. Normative References 19.2. Informative References Appendix A. End-to-End Example (Non-Normative) A.1. Stage 0: Agent Identity (by Reference) A.2. Stage 1: Mission Creation A.3. Stage 2: Mission-Bound Token Issuance A.4. Stage 3: The Resource Server Enforces Appendix B. Integrity Anchor Test Vectors Appendix C. Document History Acknowledgments Author's Address 1. Introduction Agent-identity work such as [I-D.draft-klrc-aiagent-auth] establishes how an AI agent authenticates and how a user delegates authority to it: the agent is an OAuth 2.0 [RFC6749] client identified by client_id, the delegating user is the access token sub, and the agent obtains tokens for the resources its task requires. That work deliberately leaves three things out of scope: how an agent's task (its "mission") is translated into authorization, how a user's approval of that task is captured as a durable artifact, and how later token issuance stays bound to what the user approved. Without that layer the gap is invisible to every individual OAuth component. Each token is individually valid and each request individually in scope, yet nothing checks whether the task the user approved is still the one being pursued. A token issued for a task remains usable after the user's approval has lapsed or been withdrawn, because no OAuth object ties the token's validity to the task's authorization: the credential stays secure while the work it authorizes has quietly become unauthorized. This document specifies that missing layer. It defines a *Mission*: a structured, explicitly approved, integrity-bound OAuth authorization artifact. The contribution is a single chain: 1. The client submits a structured *Mission Intent* describing the task (goal, target resources, constraints) instead of requesting raw scopes, optionally proposing concrete authority alongside it as standard authorization_details. 2. The Authorization Server (AS) derives *authorization details* ([RFC9396]), the *Authority Set*: the concrete authority the task needs. 3. At an *approval event*, the Approver consents to that authority, and the AS commits the task as an *intent_hash* and the authority as an *authority_hash* and records a durable *Mission*. 4. Every access token the agent obtains under the Mission carries the derived authorization details and a *mission claim* bearing the authority_hash. A Resource Server enforces statelessly from the token. 5. Token issuance and refresh are *gated on Mission state*, so revoking or expiring the Mission stops the agent from obtaining further authority. The result is that a user approves a task once, and that approval, not a per-request scope grant, bounds and outlives every token the agent derives. Each token carries an authority_hash audit anchor back to the consented authority; a holder of the full Authority Set can verify it, and a holder of only a narrowed subset (a downstream or cross-domain party) treats it as an audit anchor it cannot recompute (Section 16.1). This chain is the first of two deliberate enforcement layers. It gives task-bound issuance, auditability, and a revocation gate over future derivation, which is sufficient for a low-risk workflow whose exposure is bounded by short token lifetimes and narrow authority. It does not evaluate individual actions: an agent taking consequential autonomous actions needs the second layer, the runtime enforcement chokepoint (Section 16.6), specified separately. A deployment chooses its layers deliberately, matching the enforcement it runs to the consequence of what its agents do; the Mission Assurance Levels of [I-D.draft-mcguinness-mission-architecture] name the composed levels informatively. 1.1. Why a New Object OAuth already has objects near this need, but none is the approved task. A scope value or an authorization_details entry ([RFC9396]) expresses authority but neither the task it serves nor a lifecycle of its own. An access token is a short-lived projection; its jti identifies the token, not the task. A refresh token preserves the ability to obtain further tokens but commits no bounded, approved authority. A consent record proves that an approval event happened; it does not govern the resulting work as it continues. The Mission is the durable object these project from: the approved task that bounds and outlives them, and that every derived token refers back to. Rich Authorization Requests express authority; a Mission expresses an approved task with a lifecycle. The Mission is a separate object because that approved-task lifecycle, not a new way to express authority, is what OAuth lacks. A Mission is therefore not another authorization_details type; it is the durable, approval-backed object an Authority Set is derived for and gated by. 1.2. Relationship to Other Authorization Objects A grant, in the sense of FAPI Grant Management [FAPI.GrantManagement], is a durable, queryable, revocable container of consented authorization data. It records consent to authority but carries no task, no integrity commitment, and no derivation gating; a deployment MAY surface Mission revocation through a grant-management- style API. [I-D.draft-klrc-aiagent-auth] names the agent's mission and leaves its translation into authorization out of scope; this document specifies that translation. Decision-layer access-request and approval workflows, such as the OpenID AuthZEN Access Request and Approval Profile [AuthZEN.ARAP], manage approval tasks but define no issuance binding; this document supplies the issuance-bound object such workflows complete into. Nearby individual proposals each carry one Mission property without the others: task-linked Rich Authorization Requests with revocation webhooks carry a task link, intent-digest admission assertions carry an intent commitment, and offline capability attenuation ([I-D.draft-niyikiza-oauth-attenuating-agent-tokens]) carries offline narrowing. None combines the durable approved object, state-gated issuance, and integrity anchors this document defines. 1.3. Relationship to GNAP The Grant Negotiation and Authorization Protocol [RFC9635] occupies much of the same design space as this document: a continuable authorization request, richer client instance identification, and native support for delegation. This document takes a different path for the estates it targets. Rather than introduce a new grant protocol, endpoints, and client machinery, it composes with the OAuth 2.0 surfaces already deployed: Pushed Authorization Requests ([RFC9126]), Rich Authorization Requests ([RFC9396]), DPoP ([RFC9449]), and [RFC9068] access tokens. A Mission derives from and issues through those surfaces rather than replacing them, so an estate that already runs PAR, RAR, and sender-constrained tokens adopts the Mission model without standing up a GNAP grant endpoint or migrating its clients to it. Where a deployment starts from GNAP rather than from OAuth 2.0's authorization code grant, the Mission model, a durable, approval-anchored, integrity-bound task object gating derivation, is the OAuth-native instantiation of a binding- neutral contextual-governance kernel that admits a candidate binding onto that substrate; this document does not define one. 1.4. Capability-System Prior Art The Authority Set's subset rule (Section 5.1) and the offline attenuation it supports continue a lineage of capability systems: macaroons' caveat narrowing, Biscuit's offline attenuation blocks, UCAN's delegation chains, SPKI/SDSI's local-name reduction, and object-capability designs generally, in which a holder narrows what it passes on and the narrowing needs no further contact with an issuer to be enforced. This document's distinction is not the narrowing rule itself but what it narrows: a durable, approval- anchored governance object, centrally revocable by its issuer for the Mission's full lifetime, rather than a bearer credential whose only life is the caveats attached to it. A delegate's or an offline holder's attenuation narrows within that governance object; it does not replace it, and revoking the Mission still reaches everything derived from it that has not already left the issuer's reach (Section 9.2). 1.5. The Mission, the Plan, and Execution The Mission is the durable, AS-held object that commits the approved authority and owns the task's lifecycle. Two related things an agent produces around a task are deliberately not the Mission and carry no authority of their own. The agent's *plan*, how it decomposes the task, chooses tools, and delegates to sub-agents, is the agent's own strategy and is out of scope for this document. It grants nothing: authority a sub-agent exercises is carried by the act chain (Section 11), derived from the Mission and only ever narrowed from it (Section 5.1), never created by the plan. The agent's *execution*, the tokens it derives, the calls it makes, and the decisions taken on them, references the Mission but cannot expand it. Revoking the Mission stops further derivation (Section 9); it does not undo actions already completed. Evaluating each action against the Mission at the point of use is the runtime layer's concern (Section 16.6), not this document's. The invariant across all three: the plan and the execution draw on the Mission's authority; neither enlarges it. 1.6. Relationship to Agent-Identity Specifications This document is complementary to [I-D.draft-klrc-aiagent-auth] and reuses its model: the agent is the OAuth client (client_id); the delegating user or system is the token sub. This document does not define agent identity, credentialing, or posture; it defines the approved-task artifact those identities act within. An agent authenticated and delegated per that specification uses the mechanisms here to obtain Mission-bound tokens. 1.7. Applicability This profile targets OAuth deployments where authority serves a durable, approved task that spans more than one token, request, or audience: an agent pursuing a multi-step objective on a user's behalf, or a workflow whose audit must join activity across hops on a shared task. It is not intended for, and adds cost without benefit to, single-request user flows and short-lived authorizations where the credential's lifetime is the task's lifetime, ordinary machine- to-machine service credentials among them; those use OAuth unchanged. The boundary is that lifetime equality, not the absence of a human: a workload's durable multi-step task is in scope as a service-owned Mission (Section 3.2). A Mission SHOULD be scoped to a concrete task, not to an agent's whole lifetime. A deployment SHOULD prefer narrow, per-task Missions, each separately approved and revocable, over a single broad standing Mission that accumulates authority across unrelated tasks. The durable object is the approved task: keeping it task-scoped is what makes its authority and audit meaningful and bounds the blast radius on compromise to one task. An agent pursuing many tasks holds many Missions, not one broad one. The unit of governance is the action, not the content. A Mission bounds where an agent may act (resources, actions) and how much (constraints and, where metered, cumulative bounds); it does not inspect what content flows within an authorized action, and an approved egress channel carries a status update or an exfiltrated payload with equal authority. Content-level controls (data loss prevention, redaction) are complementary, and mediated execution places their natural insertion point at the mediating enforcement component (Section 16.6, [I-D.draft-mcguinness-mission-runtime]). 1.8. Scope and Future Work This document is a self-contained, minimum-viable profile: it binds Missions to OAuth 2.0 and is implementable on its own, depending only on the OAuth and JOSE specifications it cites. It references the OAuth Actor Profile ([I-D.draft-mcguinness-oauth-actor-profile]), an in-progress individual draft, for the act chain shape the OPTIONAL Delegation capability uses. That reference is informative and confined to Delegation, so the mandatory single-domain core does not depend on it, and this document's RFC path does not wait on that draft's. Cross-domain projection, a single hop that lets an Authorization Server in another trust domain honor a Mission, is specified by the companion Mission Cross-Domain Projection profile [I-D.draft-mcguinness-oauth-mission-cross-domain], which carries the identity-chaining and ID-JAG dependencies with it; the Cross-Domain capability's conformance bar is self-contained in this document (Section 15), so that companion is not a normative dependency. Separate from this document, and not required to implement it, several capabilities are now specified as OPTIONAL companion profiles: * an additional integrity anchor over a structured consent disclosure (consent_rendering_hash, Section 16.1), defined by Mission Consent Evidence [I-D.draft-mcguinness-oauth-mission-consent-evidence]; * mission expansion, defined by Mission Expansion [I-D.draft-mcguinness-oauth-mission-expansion]; and * a cross-domain status or event-distribution mechanism for tighter revocation, defined by Mission Status [I-D.draft-mcguinness-oauth-mission-status] and Mission Lifecycle Signals [I-D.draft-mcguinness-oauth-mission-signals]. A deployment implements this document without any of them. A binding-neutral contextual-governance kernel, of which this document's Mission model is the OAuth-native instantiation, is specified separately and is out of scope for this document. Remaining future work, not yet specified, includes: * the normative carriage of Mission context in Transaction Tokens ([I-D.draft-ietf-oauth-transaction-tokens]), shown only illustratively in the companion's end-to-end example; and * for a community that wants cross-vendor agreement on what a task authorizes within a vertical, an OPTIONAL derivation profile: a registry of standard task types mapped to authority templates, so that two vendors in that profile derive comparable Authority Sets. This document deliberately does not standardize the derivation algorithm itself (Section 5); a vertical profile is the appropriate vehicle where portable derivation is genuinely needed. 1.9. Non-Goals The following are deliberately out of scope. Each is a recurring question for agent authorization; naming it here records that it was considered and where it belongs, not that it was overlooked. * *Semantic / intent verification.* This profile binds a token to an approved authority and task; it does not evaluate whether a given runtime action serves the Mission's purpose beyond matching the approved authorization_details and constraints. Per-action evaluation is the runtime layer's role (Section 16.6). Verifying an agent's declared reasoning against the task is a further attestation problem outside both layers. * *Approval-free authorization upgrade.* The Authority Set is committed at approval; this profile defines no mid-stream widening that bypasses consent. Widening requires a new approval, a successor Mission, as specified by Mission Expansion [I-D.draft-mcguinness-oauth-mission-expansion]; a widening that no consent authorizes remains out of scope. * *Lifecycle event distribution.* A Resource Server learns Mission state from the token lifetime or optional introspection (Section 10); this profile defines no push-based notification of Mission state changes. A Shared Signals ([RFC8935]) / CAEP profile for Mission lifecycle events is specified separately by the Mission Lifecycle Signals profile ([I-D.draft-mcguinness-oauth-mission-signals]), not here. * *Human-in-the-loop suspension.* The base lifecycle here is active, revoked, expired (Section 9). A suspended state with resume/ complete transitions is defined as an OPTIONAL extension by the Mission Status and Lifecycle profile ([I-D.draft-mcguinness-oauth-mission-status]); a pending-human- approval state and a holding-token pause-and-resume protocol remain future lifecycle work. * *Multi-hop cross-domain provenance.* A single cross-domain hop is specified by the companion ([I-D.draft-mcguinness-oauth-mission-cross-domain]); chaining a Mission across more than one trust-domain boundary, and the verifiable provenance that would require, are future work. * *Decentralized agent identity.* Agent identity and credentialing are out of scope ([I-D.draft-klrc-aiagent-auth], and workload identity efforts such as WIMSE); this profile governs the approved-task artifact those identities act within, not the identities themselves. * *Cross-audience unlinkability.* A single canonical Mission Identifier and the authority_hash it shares deliberately let any party correlate a Mission's activity across audiences and resources. This is a design choice, not an omission: a stable, correlatable anchor is what lets a Resource Server, a cross-domain Resource AS, and an auditor bind evidence to one approved Mission, which is a core goal of this document and its companion profiles. Pairwise or unlinkable presentation of Mission-bound authority works against that anchor and is therefore future work, not a v1 property (Section 17.1). 2. Conventions and Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here. All JSON shown in this document is non-normative and illustrative; the member definitions in the surrounding text are authoritative. Agent (Client): The OAuth client acting for the Mission's Subject, identified by client_id. Agent identity is established per [I-D.draft-klrc-aiagent-auth] or ordinary OAuth client authentication. Subject: The user, workload, or organizational principal on whose behalf the Mission is approved (Section 3.2), identified by an (iss, sub) pair and carried in derived tokens' sub claim. Approver: The single accountable principal who approves the Mission at the approval event. Equal to the Subject for self-approval; different for administrator or delegated approval. This document records one accountable Approver; multi-party approval and the provenance of delegated approval authority are deferred (Section 6.2). Mission Issuer (Authorization Server): The OAuth AS that validates a Mission Intent, runs the approval event, records the Mission, and derives tokens. It is the Mission's issuer. "Mission Issuer", "issuer AS", "originating AS", and "AS" are used interchangeably in this document. Resource AS: An Authorization Server in another trust domain that honors a Mission it did not issue, minting its own tokens for its resources; it is never the Mission Issuer. Cross-domain projection is specified by the companion Mission Cross-Domain Projection profile ([I-D.draft-mcguinness-oauth-mission-cross-domain]). Mission Intent: The structured description of the task the client submits (Section 4). Mission Intent Submission (Submission envelope): The object a client submits as the mission_intent parameter value: the Mission Intent under intent, and Intent Submission Evidence under evidence where any is presented (Section 4.1). Intent Submission Evidence: Typed artifacts a client presents in support of claims about a submitted Mission Intent (Section 4.3): authenticated policy input, never authority. The term names inbound, client-presented material; the evidence this document and companion profiles emit and record (a consent-evidence artifact, an audit evidence base) is issuer- or runtime-produced output, not this. Authority Proposal: The authorization_details array a client submits alongside a Mission Intent, a proposal for derivation and never authority (Section 4.2). Authority Set: The set of authorization_details entries the AS derives from a Mission Intent and the Approver approves (Section 5). "Authority Set" names these concrete entries; it does not mean an identity authority, a trust authority, a legal authority, or an issuing authority. Mission: The durable, immutable record created at the approval event (Section 7), identified by a Mission Identifier (Section 7.1). Mission-referenced token: A token that carries a Mission reference (the mission claim or a mission_id) without Mission-derived authority or any gating guarantee. A reference is never authority. Derived token (Mission-derived token): An access token issued under a Mission, carrying its Mission-derived authority (the full Authority Set or a narrowed subset) as authorization_details and a mission claim (Section 8). Mission-bound token: A Mission-derived access token or refresh token whose issuance and refresh are gated on the Mission's active state and bounded by the subset rule (with refresh tokens bound server- side). Only this class carries this profile's gating guarantee, and this document reserves "Mission-bound" for it; a token that merely references or carries Mission data without the gates is not Mission-bound (Section 15). 3. Overview 3.1. Principal Model This document maps principals onto native OAuth constructs: * The *Agent* is the OAuth client, referenced by client_id. Agent identity and credentialing are out of scope (see [I-D.draft-klrc-aiagent-auth]). * The *Subject* and *Approver* are each an (iss, sub) pair, matching the access token sub model of [RFC9068]. The Approver is the accountable principal whose approval created the Mission; it may be a human or an authorized policy authority, and the provenance standing behind a delegated or policy approval is governance state this document does not model (Section 6.2). On a derived token the sub claim is the AS-local sub the AS maps the Subject to under the injective mapping of Section 6, verbatim adoption of the external sub being the common case, and the token iss is the AS; within the issuing AS's namespace this (iss, sub) pair is the AS-local subject principal, authoritative for the Subject, and Resource Servers authorize against it. The Mission separately records the Subject's home issuer and identifier as subject.iss and subject.sub: this pair is the external subject identity, carried as provenance for audit and not on the token, and this document defines no runtime lookup of it (there is no by-Mission status endpoint). The record's (subject.iss, subject.sub) and the token's (AS iss, sub) identify the same Subject in two issuer namespaces. Across trust domains, the companion's cross-domain grant conveys Subject identity to the Resource AS through its own subject-resolution claims ([I-D.draft-mcguinness-oauth-mission-cross-domain]), not through a Mission lookup. Issuer roles obey three invariants: a Mission has exactly one Mission Issuer, its issuer; a Resource AS never creates or alters a Mission; and a local token minted in another domain preserves the mission claim unchanged ([I-D.draft-mcguinness-oauth-mission-cross-domain]). Principals are recorded at the approval event and are immutable. Two principals are equal when their iss and sub are byte-equal, a test that therefore compares principals only within one issuer namespace; the external subject identity and the AS-local subject principal are two representations of one Subject and are never compared to each other under this rule. Dynamic delegation (the actors an agent delegates to during execution) is carried on derived tokens via the act chain (Section 11), not on the immutable Mission record. Richer subject identifier formats (for example, the formats of [RFC9493]) MAY be layered in future versions and are not required here. 3.2. Authority Sources A Mission draws its authority from one of three sources: a delegating person's own authority (*user-delegated*), a workload's own provisioned authority (*service-owned*), or explicitly governed organizational policy with a named accountable owner (*organizational*). The source names whose authority the approval draws on, recorded immutably as the Mission's authority_source; approval_basis records how drawing on it was activated (Section 7), and the two compose: any source may activate through a direct approval event or through a standing-consent basis a companion defines. Approval activates authority the source already holds and manufactures none: the AS establishes the source and verifies the derived Authority Set against it before approval (Section 6). The subject-representation discipline is the same in every source: * The accountable principal is the record's approver, equal to approval_basis.consent_principal, in every source; it is never inferred from the token sub. * sub carries a delegating person only in the user-delegated source. A service-owned or organizational Mission MUST record the workload or organizational principal as subject and MUST NOT record a human principal in its place: some work is not any one person's, and borrowing a human subject for it blurs the actor, the subject, and the accountable principal exactly where the record keeps them distinct. The injective mapping of Section 6 applies unchanged: that principal receives its own AS-local sub, denotes itself, and impersonates nobody. It MUST be an authorization subject the AS recognizes as a resource owner in its own right (the sub model of [RFC9068]), not merely the task's beneficiary. * The actor model does not vary by source: client_id names the Agent, and delegates ride the act chain (Section 11). 3.3. Protocol Flow Agent (client) Mission Issuer (AS) | | | 1. PAR: intent + proposal --------->| derive authority |<----------- request_uri ------------| (authz_details) | | | 2. authorization request ---------->| Approver consents | | -> authority_hash |<-------------- code ----------------| -> Mission active | | (bound to the grant) | | | 3. token request ------------------>| gate: active? |<----------- access token -----------| + authz_details | | + mission claim v The flow then leaves the AS: (4) the agent calls the Resource Server with the token; the RS enforces the authorization_details statelessly and MAY check the mission claim, with no callback to the AS required. (5) A management revoke, or expires_at passing, moves the Mission to revoked or expired, after which the AS refuses further issuance and refresh; a deployment MAY additionally compose RFC 7009 [RFC7009] refresh-token revocation (Section 9.2). The end-to-end example (Appendix A) walks this flow with concrete messages. 4. Mission Intent Before the approval event (Section 6) only the Mission Intent exists, as untrusted client input (Section 4.1); after it, only the Mission is authoritative. A Mission Intent therefore has no protocol identifier of its own: two submissions of the same Intent produce two distinct pending requests, and a Mission acquires its Mission Identifier (Section 7.1) only at activation. The approved Intent is recorded on the Mission and committed by intent_hash (Section 6.3); it describes the task and carries no authority members. Concrete authority is proposed separately, on the standard authorization_details parameter pushed alongside the Intent (Section 4.2) and committed by proposal_hash when submitted; the granted authority is committed by authority_hash over the derived Authority Set (Section 5). A Mission Intent is a JSON object describing the task. The client submits it as the intent member of the Mission Intent Submission envelope (Section 4.1), in place of scope or alongside a narrowed scope. It has the following members: goal: REQUIRED. A string. A human-readable statement of the task, for rendering to the Approver. Maximum 4096 characters. Prose here persists on the record and can carry personal data about third parties (Section 17.4). resources: REQUIRED. An array of strings. The target resources the task needs, each an absolute URI identifying a protected resource (an OAuth resource per [RFC8707]-style indicators or a Protected Resource per [RFC9728]). constraints: OPTIONAL. An array of strings. Human-readable bounds on the task (for example, "read only invoices from 2026"). They are disclosure and audit context, rendered to the Approver beside the derived Authority Set (Section 6); the AS MUST NOT parse them for machine semantics, and a machine-enforceable bound enters as structure instead (Section 5). success_criteria: OPTIONAL. An array of strings. Human-readable observable outcomes that indicate the task is complete. These are disclosure and audit material only: they are rendered to the Approver and committed by intent_hash (Section 6.3), but carry no machine semantics and MUST NOT be used to derive, widen, or gate authority. purpose: OPTIONAL. A string. A URI identifying the purpose of the task, recorded for disclosure and audit. Its semantics are deployment- or registry-defined and opaque to this document. It is a structured, opaque lookup key permitted only for template- mode selection: the configured mapping MAY key on it (Section 5), and the derived set stays bounded by the Intent and by policy like any derivation. It MUST NOT otherwise be used to derive, widen, or gate authority, and once the Mission is approved it is inert. A deployment MAY consult it for out-of-band policy or logging that does not affect the authority derived here. expires_at: REQUIRED. A string. An RFC 3339 [RFC3339] date-time after which the AS MUST NOT derive tokens under the Mission. The value is the client's proposal like the rest of the Intent: the approved Intent is recorded as submitted, and an AS whose policy cannot accept the proposed expires_at refuses with invalid_authorization_details rather than rewriting it. controls: OPTIONAL. An object of machine-actionable bounds. This document defines the members below; others MAY be added by deployments or defined by companion profiles (an extension member follows the collision-resistant naming and fail-safe rules of Section 12): acr: OPTIONAL. A string. An authentication context class for the approval event: the approval authentication MUST be one the deployment's policy maps as satisfying the named class. No global ordering of acr values exists; the mapping is deployment policy. max_derivations: OPTIONAL. A positive integer (1 or greater). A bound on the number of derivations the issuer AS performs under the Mission, enforced per Section 9. A value of 0 is invalid (it would forbid even the initial issuance); to stop a Mission, revoke it (Section 9.2). An AS MUST reject max_derivations below 1 with invalid_request. This is an issuer-AS derivation cap, not an end-to-end credential cap; what counts as one derivation, including how cross-domain issuances count and what the issuer cannot observe, is defined once in Section 9. A deployment sizes the cap from its derivation cadence (token and grant renewal frequency, times audiences, times the Mission's duration), not as a flat count: steady-state cross-domain work consumes one derivation per short-lived grant per partner audience, so a cap sized for single-domain refresh exhausts in hours under cross-domain renewal. agent_deployment: OPTIONAL. A string. A deployment-defined identifier or digest naming the approved Agent Deployment (the agent's behavioral version: its code, model, system prompt, tool allowlist, data scope, and runtime configuration) the Mission is pinned to. At every derivation under the Mission, the AS MUST refuse issuance when the requesting client's attested or asserted deployment identifier does not match the pinned value, and an AS whose clients present no deployment identifier MUST refuse an Intent carrying this control with invalid_authorization_details. The identifier's format, and what change constitutes a distinct Agent Deployment, are deployment policy; the pin is recorded and committed like every control and never widens or derives authority. Under a class pin, every attested instance of the pinned Deployment derives under the Mission itself, and controls.max_derivations is the explicit fan-out ceiling, whose atomic check and increment (Section 9) already serves exactly this concurrency. This object is the agent's behavioral version, not the deployment's published claims manifest. This document defines no cumulative consumption bounds (for example, a budget, call-count, or activity-duration cap): every bound this document defines is enforced by a party this document names. An experimental companion defines cumulative consumption bounds as controls extension members together with the runtime metering that enforces them ([I-D.draft-mcguinness-mission-metering]). The following table summarizes which party enforces each bound a Mission carries and what holds when that enforcer is absent: +=================+=======================+=========================+ | Bound | Enforced by | When that enforcer is | | | | absent | +=================+=======================+=========================+ | resource and | any Resource Server | a scope-only RS | | actions | that enforces | enforces only the | | | authorization_details | coarse scope | | | (Section 8.2) | projection | | | | (Section 8.2) | +-----------------+-----------------------+-------------------------+ | per-entry | a Resource Server | a Mission-aware RS | | constraints | that understands and | fails closed; a scope- | | | enforces the key | only RS does not | | | (Section 8.2) | evaluate them | +-----------------+-----------------------+-------------------------+ | max_derivations | the issuer AS at each | never absent at the | | | derivation | issuer; it does not | | | (Section 9) | bound another domain's | | | | local minting (see the | | | | cross-domain | | | | companion) | +-----------------+-----------------------+-------------------------+ Table 1 Example Mission Intent: { "goal": "Reconcile Q3 invoices and post adjustments under $500.", "resources": ["https://erp.example.com"], "constraints": [ "Read only invoices issued in 2026-Q3.", "Post journal entries under $500." ], "success_criteria": [ "All Q3 invoices reconciled.", "Each posted adjustment references a source invoice." ], "purpose": "urn:example:purpose:reconcile", "expires_at": "2026-12-31T23:59:59Z", "controls": { "acr": "urn:example:acr:mfa", "max_derivations": 200 } } 4.1. Submission via PAR A client MUST submit a Mission Intent through a Pushed Authorization Request [RFC9126] using the mission_intent request parameter. The parameter value is the UTF-8 JSON [RFC8259] serialization of the *Mission Intent Submission envelope* (the Submission envelope), a JSON object carried as an ordinary OAuth request-parameter value (form-encoded in the application/x-www-form-urlencoded PAR request body, like other OAuth parameters) with exactly these members: intent: REQUIRED. The Mission Intent object (Section 4). evidence: OPTIONAL. A non-empty array of Intent Submission Evidence entries (Section 4.3). When present it MUST be non-empty: omission already expresses absence, and the AS refuses an empty array with invalid_request, preserving one closed canonical syntax for "no evidence". The Submission envelope separates the semantic task from material presented about it: intent_hash commits exactly the intent object, never the Submission envelope or its evidence array (Section 6.3), so an evidence artifact can commit to the intent_hash of the Intent it supports without becoming part of the object that hash covers. Earlier revisions of this document defined the parameter value as the bare Mission Intent object; this document does not (Appendix C), and the bare shape is refused under the closed-envelope rule below. The AS returns a request_uri as usual, which the client uses to start authorization. An AS that cannot parse mission_intent as a JSON object, or that parses it but finds the Submission envelope or the Intent structurally invalid against this document's member definitions, MUST refuse the request with invalid_request. Submission is governed by the following rules: * *Closed top levels.* The Submission envelope is closed: the AS MUST reject with invalid_request an envelope containing a top- level member other than intent and evidence. A bare Mission Intent submitted as the parameter value fails this rule, since its goal and sibling members are unknown envelope members. The Mission Intent's own top level is likewise closed: the AS MUST reject with invalid_request an intent containing a top-level member this document does not define; extension data belongs under controls (Section 4), so deployment-defined semantics are explicit and cannot masquerade as core Intent semantics. Both top levels are closed because their members feed approval rendering, the intent_hash commitment, and evidence dispatch, and the closure also keeps issuer-output members, such as a client-planted authority_hash, out of the Intent. * *Derivation failure is distinct from syntax.* For an Intent that is well-formed but yields no valid Authority Set (an unsupported resource, action, or authorization details type, or a policy that bars the requested authority), the AS SHOULD refuse with invalid_authorization_details ([RFC9396]), whether or not the request carried an authority proposal (Section 4.2), so a client can tell a syntax error from a derivation failure. * *One carriage, through PAR.* The Intent is accepted as the form- encoded mission_intent parameter of the PAR request body, or as a mission_intent claim inside a signed Request Object ([RFC9101]) that is itself pushed through PAR. The AS MUST reject with invalid_request a mission_intent presented on a front-channel authorization request that does not use a PAR-issued request_uri: mission_intent MUST NOT appear as a plaintext front-channel authorization-request parameter, whether or not a Request Object is also present. When the pushed request carries a Request Object (a request parameter value), that object is authoritative: a mission_intent submitted outside the Request Object in the same push MUST be rejected with invalid_request, the [RFC9101] duplication rule applied to this parameter. A single point of precedence keeps the member set the AS validates and bounds in one place; PAR keeps the integrity-sensitive Intent off the untrusted front channel. * *Bounded size.* The AS MUST bound the Submission envelope's total size, the Intent's total size and the lengths of its arrays, and the count and per-entry sizes of evidence entries, refusing a submission that exceeds the deployment-defined limits with invalid_request, so an oversized submission cannot exhaust the AS at rendering, derivation, verification, or hashing. The verification-cost bound of Section 4.3 accompanies these. * *Concrete authority is proposed via authorization_details.* A client proposing concrete authority submits the standard [RFC9396] authorization_details request parameter alongside mission_intent in the same push, as a proposal subject to derivation (Section 4.2); mission_intent itself carries no authority members. A client MAY submit scope and resource ([RFC8707]) values. The AS treats them as a requested subset and MUST NOT grant authority beyond what the Mission Intent yields. * *Pushed parameters are authoritative.* On the front-channel request that redeems the request_uri, the AS MUST ignore any mission_intent, authorization_details, scope, or resource presented, and MUST NOT let such a value widen the authority derived from the pushed parameters. * *A proposal, never authority.* A Mission Intent, and any authority proposal submitted alongside it (Section 4.2), is untrusted client input; trust enters only when the AS validates it and the Approver consents to the rendered result. The AS MUST treat the submission as a proposal and MUST derive and bound authority by policy regardless of what the client submitted. How a client produces the Intent (for example, a "Mission Shaper" deriving it from a natural-language instruction) is out of scope for this document. Example Submission envelope, carrying a compact Intent and one evidence entry of an illustrative, deployment-defined type: { "intent": { "goal": "Reconcile Q3 invoices and post adjustments under $500.", "resources": ["https://erp.example.com"], "expires_at": "2026-12-31T23:59:59Z" }, "evidence": [ { "type": "urn:example:intent-evidence:admission", "assertion": "eyJhbGciOiJFUzI1NiIsImtpZCI6ImFkbS0xIn0..." } ] } 4.2. The Authority Proposal A client MAY propose concrete authority for the task by submitting the standard [RFC9396] authorization_details request parameter, a JSON array of authorization_details objects, alongside mission_intent in the same pushed request (Section 4.1). The Mission Intent carries no authority members: earlier revisions of this document carried the proposal inside the Intent as a proposed_authority member, this document does not (Appendix C), and an Intent carrying that member is refused as an unknown top-level member under the closed-top-level rule of Section 4.1. The submitted authorization_details is a proposal, never authority. The AS MUST treat it as untrusted input, MUST derive and bound the Authority Set by policy regardless of what was submitted (Section 5), and MUST NOT grant authority beyond what the Mission Intent yields. Each submitted entry MUST be of a type the AS advertises and MUST validate against that type's published JSON Schema (Section 5.3, Section 13); where a deployment arranges Mission-bound authorization out of band rather than advertising the metadata endpoint, the supported types and their schemas are established out of band, and this rule applies over that knowledge the same way. This is the [RFC9396] Section 5 validation of the standard parameter: the AS MUST refuse a request carrying an entry of an unadvertised type, or an entry that fails its type's schema, with invalid_authorization_details; a validation failure is never repaired by omitting the entry. Policy narrowing is distinct. During derivation the AS MAY narrow or omit a syntactically valid entry that policy cannot accept (Section 5); it MUST NOT keep such an entry silently, and the granted authorization_details echo (Section 8) MUST reflect every narrowing and omission, so no entry is ever represented as granted when it was not. When a proposal is present, the AS MUST derive each Authority Set entry as a subset (Section 5.1) of some proposed entry of the _same type_: a mission_resource_access entry derives only from a mission_resource_access proposal, under the subset rule of Section 5.1; an entry of another AS-supported type derives only from a same-type proposal, narrowed under that type's own subset rule where it defines one, or carried through unchanged where it defines none (Section 5.3). No entry derives from a proposed entry of a different type. goal and constraints then serve as rendering and bounding context over the proposed authority. Each proposed entry that carries a resource member MUST have it among the Intent's resources; the AS refuses a request violating this with invalid_request. The proposal rides the Intent's carriage rules (Section 4.1): it is accepted only through PAR, inside the Request Object when one is used ([RFC9101]), it is ignored on the front-channel request that redeems the request_uri, and the bounded-size rule applies to it the same way. The AS records the submitted array on the Mission exactly as submitted and commits it by proposal_hash (Section 6.3, Section 7), present iff a proposal was submitted: what the agent asked for is committed separately from the task (intent_hash) and from what was granted (authority_hash), so a narrowed grant can be audited against the proposal that sought it. Submitting authorization_details without mission_intent is an ordinary [RFC9396] request that this document does not govern; the client-side and AS-side duties that keep a governed task from downgrading across that line are stated in Section 13 and Section 16.3. Example authority proposal, submitted alongside the example Intent of Section 4. Derivation narrows invoices.* to invoices.read bounded to a Q3 issuance window, halves the proposed ceiling under the Intent's constraints, and carries the proposed delegation policy through unchanged (the example Authority Set of Section 5): [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.*"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ] 4.3. Intent Submission Evidence The Submission envelope's evidence array carries *Intent Submission Evidence*: typed artifacts the client presents in support of claims about the submitted Intent, such as its originator, an admission or consent decision that applies to it, or the presenter authorized to submit it. Intent Submission Evidence is inbound and client- presented; it is not the evidence this document and its companions emit and record as output (a consent-evidence artifact, a runtime decision record, an audit evidence base, [I-D.draft-mcguinness-mission-audit]). The naming keeps the two apart: evidence in this section is what a client shows the AS about a submission, and emitted Evidence is what the issuer or runtime records about its own decisions. Each entry is a JSON object with a REQUIRED type member: a string naming the evidence type as a collision-resistant name, under the same guidance as anchor typ values (Section 6.3). The specification that owns a type defines the entry's remaining members as a closed schema, the artifact format, the verification procedure, and the verified output facts that verification yields. The generic entry has no other members: as with an [RFC9396] authorization-details type, the selected type owns the entry's exact members and semantics, and this document defines no bag of optional format, issuer, subject, reference, digest, or criticality members. This document defines no evidence types; a companion profile defines the first, and an AS that supports none refuses every presented entry under the dispatch rule below, which is the correct refusal, not a gap. Processing is governed by the following rules: * *Reject, never ignore.* An entry that is not a JSON object, or that lacks type, is structurally invalid and refused with invalid_request. The AS MUST refuse an entry whose type it does not support, and an entry that fails its type's validation or verification, with invalid_mission_intent_evidence (Section 18). Evidence presented for admission MUST NOT be silently ignored, dropped, or demoted to advisory input: a submission is accepted only when every presented entry verified. * *Policy input, never authority.* A verified entry MUST NOT be interpreted as authority, copied into the Authority Set, or treated as the Mission approval event. Verified evidence MAY serve as authenticated input to the AS's admission and derivation policy; the AS remains responsible for deriving and bounding the Authority Set (Section 5), and the approval event (Section 6) remains the sole activation of authority. Verification authenticates the artifact issuer's claims; AS policy decides whether those claims are acceptable for this request. * *Required evidence is resolved before derivation.* Rejecting presented evidence does not defend against evidence a client omits. The AS determines the evidence types its applicable profile, client, resource, or admission policy requires before derivation. When a required type is absent from the submission, the AS MUST refuse with invalid_mission_intent_evidence, and successful processing of a submission without evidence MUST NOT be interpreted as satisfying an evidence requirement. This is the submission-plane form of the downgrade-by-omission duty (Section 16.3). * *Evidence binds one exact Intent.* Evidence bound to an intent_hash applies only to that exact semantic Intent. Any shaping or approval revision that changes intent_hash MUST obtain new evidence, unless the evidence type's specification explicitly authorizes defined transformations and defines how their lineage is verified. Evidence for the predecessor Intent MUST NOT be treated as evidence for the revised Intent. * *The exchange establishes the presenter; the evidence must agree.* The AS establishes the presenter through the containing exchange: client authentication and, where present, proof of possession. An entry that names an authorized presenter (a client_id, a cnf key binding) MUST match the established presenter, and a mismatch fails that entry's verification. Evidence is never an alternative client-authentication mechanism and never selects the presenter. * *Bounded verification.* Beyond the size and count bounds of Section 4.1, the AS MUST bound the verification cost a submission can impose (for example, the number of signature verifications it performs), refusing a submission that exceeds the bound with invalid_request, so presented evidence cannot exhaust the AS any more than an oversized Intent can. The AS processes a submission in this order: 1. Parse the Submission envelope and enforce both closed top levels (Section 4.1). 2. Validate the intent object against the Mission Intent member definitions (Section 4). 3. Compute the provisional intent_hash over the intent object (Section 6.3). 4. Resolve the evidence types policy requires, and refuse a submission missing a required type. 5. Verify every evidence entry under its type's rules, verifying that intent-bound evidence names exactly the provisional intent_hash and is bound to this AS and to the established presenter. 6. Apply admission policy and derive the Authority Set independently (Section 5). 7. Render the Intent, the Authority Set, and the material verified provenance for approval; a change to any of them before the decision is re-rendered and approved over the changed context (Section 6). 8. At activation, record the approved intent, intent_hash, the Authority Set, and the verified evidence facts as submission_evidence (Section 7). The material verified provenance of step 7 is part of the approval surface, not an annotation beside it: where a deployment commits the rendered approval surface, the commitment MUST cover the normalized provenance facts, at least as a digest of their canonical submission_evidence representation (Section 7), so the committed rendering proves which provenance supported the decision. The consent-evidence companion binds this with a submission_provenance_hash inside its committed disclosure ([I-D.draft-mcguinness-oauth-mission-consent-evidence]). Schema validation and the provisional hash precede signature verification, so the AS never verifies artifacts for a submission it would refuse on shape, and so intent-bound evidence has a hash to be checked against. On a surface that carries a Mission-creation idempotency fingerprint (the expansion and child-creation token exchanges, [I-D.draft-mcguinness-oauth-mission-expansion]), presented evidence affects admission, derivation, approval, and side effects and is therefore a member of that fingerprint; the profile that owns the fingerprint lists it. On those surfaces, recovery of a completed operation under the creation-idempotency rules ([I-D.draft-mcguinness-oauth-mission-expansion]) returns the recorded outcome without re-verifying the presented evidence: verification happened when the operation ran, and an artifact whose freshness or status has since lapsed does not invalidate the recovery of an already-completed request. PAR-based creation and surfaces that submit no Mission Intent carry no such fingerprint and retain their own replay and idempotency mechanisms. 5. Mission Authority From the Mission Intent, and from the authority proposal where one was submitted (Section 4.2), the AS derives the *Authority Set*: one or more [RFC9396] authorization_details entries of type mission_resource_access (Section 18.3). Derivation is mechanical. It happens once, at the approval event, over the derivation policy then in force, in one of two modes: * *Narrowing mode* (RECOMMENDED): the client submitted an authority proposal (Section 4.2), and the Authority Set is the proposal narrowed to policy. Each derived entry MUST be a subset (Section 5.1) of some proposed entry of the same type (Section 4.2). An entry of an unadvertised type, or one that fails its schema, was refused at validation and never reaches derivation (Section 4.2); a valid proposed entry that policy cannot accept is narrowed or omitted, and the granted echo reflects it (Section 4.2). * *Template mode*: no authority proposal was submitted, and a deployment-configured mapping, keyed on the Intent's purpose or resources, yields the candidate entries, which are then narrowed to policy. The mapping is a lookup, never synthesis; the AS refuses an Intent that matches no configured mapping with invalid_authorization_details, and SHOULD make that refusal distinguishable from a policy refusal in error_description. A deployment publishes its mapping space as deployment documentation; eligibility MAY be scoped per Subject and client, and a mapping keyed on Subject attributes resolves at the approval event, where the Subject is established (Section 6), so a refusal at PAR time is best-effort over what is checkable without the Subject. In both modes the AS MUST bound every derived entry by the Mission Intent: each derived entry's resource MUST be one of the Intent's resources values, and the derived authority MUST NOT exceed any machine-actionable controls bound. In both modes the AS MUST also record the policy version in force as the Mission's policy_version: an opaque audit correlator naming the policy a derivation ran under, not a value whose policy travels. Deriving authority generatively, with model assistance over the structured inputs above, is not one of this profile's modes: a deployment MAY implement it as a local-policy extension, it is the least portable option, and the Intent bounds, the prose boundary below, and the recording rule above all apply to it unchanged. A resources entry the deployment does not recognize either causes refusal with invalid_authorization_details (Section 4.1) or is omitted from the Authority Set, by deployment policy. When an omission or a narrowing leaves the Authority Set short of what was proposed (Section 4.2), derivation is partial. The granted authorization_details echo (Section 8) MUST reflect every omission and narrowing and remains the authoritative statement of what was granted; a client learns of any shortfall by comparing the echo against its proposal, which proposal_hash commits as submitted (Section 6.3). No omitted or narrowed entry is ever represented as granted. The derived Authority Set, not the Mission Intent, is the authority the Approver consents to: the AS renders the Authority Set for approval and commits it as authority_hash (Section 6). The Intent's members describe and bound the task but grant no authority by themselves (Section 4): its structured members constrain what the AS MAY derive mechanically, its prose members bound through disclosure (the Approver refuses authority the words do not support), and none widens. The goal, constraints, and success_criteria members are human- readable disclosure and audit context. The AS MUST NOT parse them for machine semantics or vary the derived Authority Set on an interpretation of their prose; translating a user's words into structure is the shaper's job, before admission and outside the trust boundary ([I-D.draft-mcguinness-mission-shaping]). A client-proposed constraint on an individual Authority Set entry enters through the top-level authorization_details proposal: the Common Constraints (Section 5.2), type-specific constraints, and the collision-resistant deployment extensions the AS understands (Section 12). Authority is further bounded by the Intent's structured members (resources, expires_at, controls), by the template mapping keyed on purpose or resources (a lookup over structured values yielding structured candidate entries, never an interpretation of prose), and by local policy and eligibility. The approval surface renders the prose beside the derived Authority Set (Section 6): the human check that the structure matches the words, never a machine enforcement mechanism. Derivation is governed by local policy and is not a portable algorithm: different Authorization Servers MAY derive different Authority Sets from the same Mission Intent, exactly as different deployments grant different authority for the same [RFC9396] request or the same scope. That locality is intended, not a gap. A Mission Intent has no portable semantics: interoperability begins at the committed result, the derived Authority Set's structure and vocabulary (Section 5, Section 5.2) and its integrity anchors (Section 6.3), which a consumer in any domain interprets, enforces, and audits identically. A consumer enforces the derived Authority Set, never the Intent, and audit establishes what was derived, against intent_hash and policy_version, never whether the derivation was the right reading of the task. A deployment whose partners reason about its derivations SHOULD publish a derivation policy identifier and test fixtures pinning Intent-to-Authority-Set outcomes; the policy itself does not travel. For an open-ended task whose concrete objects cannot be enumerated at approval (for example, "reconcile this customer's ledger," where the individual invoices are not yet known), the AS SHOULD bound the derived authority primarily by constraints that hold as invariants over those objects (the owning customer, the tenant, an amount ceiling, read-only except named write actions, a validity window) rather than by an exhaustive resource enumeration. The runtime layer (Section 16.6) checks each concrete object against the constraint at the point of use. Constraint-bounding lets a Mission cover an open- ended task with tight authority even though the specific objects are unknown at approval, and avoids the over-broad enumeration a deployment would otherwise need in order to anticipate them. A mission_resource_access entry is a [RFC9396] authorization_details object with these members: type: REQUIRED. A string. mission_resource_access. resource: REQUIRED. A string. The single protected resource the entry applies to: an absolute URI identifying an OAuth protected resource ([RFC8707]) or a Protected Resource ([RFC9728]), the same kind of identifier as the [RFC8707] resource value. Carrying it per entry, which [RFC9396] permits, lets one token scope distinct authority to distinct resources. The token aud is derived from the carried entries' resource values and is typically coarser (Section 8). Per [RFC9396] Section 3.2, the resource authorization request parameter does not affect how the AS processes authorization_details, and this member is distinct from the [RFC9396] common locations field. resource_match: OPTIONAL. A string: exact (the default, and the behavior when the member is absent) or prefix. Under exact the entry applies to the resource URI alone. Under prefix the entry authorizes the resource itself and any URI beneath it at a path- segment boundary (the resource followed by / and further path). A resource value under prefix MUST NOT carry a query or fragment component; the AS refuses such an entry with invalid_authorization_details. For prefix purposes, a resource with an empty path and one whose path is / denote the same base: https://a.example and https://a.example/ authorize the same set. These two values are the only ones defined; a consumer MUST treat an entry whose resource_match value it does not recognize as unenforceable and fail closed (Section 8.2). Containment between effective resource sets is compared as defined in Section 5.1. actions: REQUIRED. An array of strings. Permitted action values: each is an action identifier matching [A-Za-z0-9_.:-]+, or an action family (an identifier followed by .*). An action family authorizes every action whose dot-separated identifier extends the family name at a segment boundary (invoices.* authorizes invoices.read and invoices.q3.export, not invoicesx.read). * Like an OAuth scope, an action value carries meaning only at the resource that defines it: a consumer enforces only the actions it recognizes for that resource and honors no others, so an unrecognized action is fail-closed by construction. * An AS SHOULD draw action identifiers from a namespace the serving resource documents, so the set is interpretable cross- vendor rather than ad hoc. * A consent rendering MUST present a family as the breadth it is: all actions under the name, not one action. * An AS SHOULD treat deriving a family as high-risk breadth. constraints: OPTIONAL. An object. Machine-actionable per-resource bounds (for example, max_amount). A member name defined as a Common Constraint (Section 5.2) has shared semantics across deployments; any other name is deployment-defined. * Because a constraints member narrows authority, a Resource Server that cannot enforce one MUST fail closed (Section 8.2). * To avoid that failure mode, the AS SHOULD emit for a given resource only constraints keys that the Resource Server serving it is known (by registration, deployment policy, or the resource's advertised mission_constraints_supported (Section 14)) to understand and enforce. delegation: OPTIONAL. An object. The delegation policy for this entry (Section 11.3). When absent, this entry's authority is non- delegable: it MUST NOT appear in a delegated token. When present, it has these members: max_depth: REQUIRED. An integer. The maximum delegation depth at which this entry's authority may be exercised (Section 11.3). allowed_delegates: RECOMMENDED. An array of objects. The permitted delegates, each a may_act-style matcher (Section 11.3): { "sub": "" } for a specific delegate, or { "sub_profile": "" } for an actor- type class. When absent, eligibility falls to the AS's delegation-authorization policy, which MUST be applied at every exchange (Section 11.3); absence delegates the decision to policy, it never grants blanket eligibility. The member is RECOMMENDED so that eligibility is committed and rendered with the entry rather than left wholly to policy. A companion profile of this document MAY define additional delegation members. Such a member is policy, not authority (Section 11.3): a derived entry's value for it MUST NOT be broader than the parent entry's, and a member the AS does not understand is carried unchanged. Example Authority Set (the read entry is delegable to depth 2 and bounded to a Q3 issuance window by the resource_issued_after and resource_issued_before Common Constraints (Section 5.2); the write entry carries no delegation and so is non-delegable, because delegation is per entry): [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ] 5.1. Subset Rule When the AS narrows the Authority Set for a derived token, a derived mission_resource_access entry A is a subset of a Mission entry B when: 1. A's effective resource set is contained in B's (resource_match, Section 5): when neither entry sets resource_match: "prefix", A.resource equals B.resource; when B is a prefix entry, A (whether exact or prefix) is contained when A.resource equals B.resource or extends its path at a path-segment boundary; a prefix A is never contained in an exact B. 2. Every A.actions value is within some B.actions value: a value is within an equal value; a literal action is within a family whose name it extends at a segment boundary (invoices.read is within invoices.*); a family is within a reference family when its own name extends the reference's name at a segment boundary (invoices.q3.* is within invoices.*). A family is never within a literal action. 3. For every key K in *B*.constraints: * K MUST also be present in A.constraints. A key present in B but absent from A is treated as the broadest possible value and therefore fails this test. * A's value MUST be no broader than B's under K's subset rule: the specification-defined rule when K is a Common Constraint (Section 5.2), the deployment-defined comparison otherwise. The AS MUST refuse to derive an entry that is not a subset of some Mission Authority Set entry. Authority under a Mission MUST NOT widen after the approval event: a request that exceeds the Authority Set on any dimension (a new resource, action, actor, delegation path, longer duration, or constraint relaxation) is refused under this rule, and broader authority requires a fresh approval event, either a new Mission or a successor per the companion [I-D.draft-mcguinness-oauth-mission-expansion]. Resource containment under a prefix reference is compared after RFC 3986 [RFC3986] syntax-based normalization of both URIs: lowercase the scheme and host, remove a default port, uppercase the hexadecimal digits of any percent-encoding ([RFC3986] Section 2.1, so %2f and %2F are equivalent), decode percent-encoded octets of unreserved characters, and remove dot-segments. This normalization applies to comparison only, never to hashing: anchor computation (Section 6.3, Section 6.4) remains byte-exact over the recorded values and is untouched by this rule. The default comparison is deliberately flat: resource matches by exact equality and a literal action by array membership. Hierarchy is opt-in and closed to the two forms above: resource_match: "prefix" for resource containment and .* action families for action containment (Section 5). A deployment that uses neither retains the flat behavior unchanged. The delegation member is policy, not authority, and is not part of this comparison (Section 11.3). A derived entry's delegation, when present, MUST NOT be broader than the parent entry's: * its max_depth MUST be no greater than the parent entry's; * its allowed_delegates MUST be no wider than the parent entry's; and * a derived entry MUST NOT introduce delegation where the parent entry has none. The comparison is representational, not semantic. A candidate that compares as no broader can still permit effects the parent's purpose never contemplated, because narrowing is judged over the entry's members, not over meaning; semantic narrowing is not a property this rule can provide. Where the comparison relation cannot decide (an unrecognized member, an incomparable value), the posture is conservative refusal, as each consuming rule of this document states. 5.2. Common Constraints A constraints member name (Section 5) is either a specification- defined *Common Constraint* or a deployment-defined key. Common Constraints give independently developed deployments one vocabulary they interpret, narrow, and compare identically; further Common Constraints are defined by specification under the naming convention of Section 18.8. A Common Constraint definition fixes: * *Value syntax*: the JSON [RFC8259] value type and any additional rules. * *Subset rule*: how a candidate value is judged no broader than a reference value, used by the subset comparison of Section 5.1. * *Intersection rule*: how two values for the same key combine; the result MUST be no broader than either operand. A constraints member whose name is a specification-defined Common Constraint is interpreted per its definition. Any other member name remains deployment-defined and is interpreted only within the issuing deployment; a consumer that does not recognize it MUST fail closed (Section 8.2). The same duty binds the AS side of narrowing: whenever the AS derives a candidate entry from a ceiling (a client's authority proposal narrowed to the Authority Set (Section 4.2), or an Authority Set entry narrowed to a derived or delegated token, Section 5, Section 11.3), a registered Common Constraint key present in the ceiling entry that the AS does not implement narrowing for MUST NOT be dropped while the entry survives. The AS MUST instead fail closed: refuse the whole derivation, or omit the entry from the result, exactly as an unrecognized resources value already may be; the granted authorization_details echo reflects any such omission (Section 5). Carrying the entry forward with the key dropped would widen effective authority past the ceiling exactly as an unenforced key would at the Resource Server (Section 8.2). This document defines the initial Common Constraints: * max_amount (object): a per-action ceiling on a monetary amount. The value is an object with two members: amount (REQUIRED, a string containing a decimal number) and currency (REQUIRED, an ISO 4217 [ISO4217] currency code). Subset: no broader when the currency values are equal and the candidate amount is less than or equal to the reference amount, compared in decimal value space; differing currencies fail the comparison (no conversion is defined). Intersection: when the currency values are equal, the value with the smaller amount; differing currencies have no intersection and the combination fails. * resource_issued_after (string, an RFC 3339 [RFC3339] date-time): the action applies only to resources issued at or after this instant. Subset: no broader when greater than or equal to the reference. Intersection: the later instant. * resource_issued_before (string, an RFC 3339 [RFC3339] date-time): the action applies only to resources issued at or before this instant. Subset: no broader when less than or equal to the reference. Intersection: the earlier instant. The resource_ qualifier in both names marks that the window bounds resource issuance, not token issuance. * tenant (string): the action applies only to resources of the named tenant. Subset: no broader when equal to the reference value. Intersection: the common value when the two are equal; otherwise there is no intersection and the combination fails. * recipient_domain (string, a DNS name): the action applies only to recipients within the named domain. Subset: no broader when equal to the reference or a DNS subdomain of it. Intersection: the narrower value when one is equal to or a subdomain of the other; otherwise there is no intersection and the combination fails. * time_window (object): the action may be exercised only within the window, evaluated at the point of use (unlike resource_issued_after and resource_issued_before, which bound resource issuance, and unlike token exp, which bounds the credential). The value has two members, not_before and not_after (each an RFC 3339 [RFC3339] date-time); at least one MUST be present, and an absent member is unbounded on that side. Subset: no broader when the candidate window lies within the reference window, an absent candidate bound counting as unbounded and therefore broader than any present reference bound. Intersection: the overlap (the later not_before, the earlier not_after); an empty overlap has no intersection and the combination fails. * data_classification (array of strings): the action applies only to data whose classification label is among the named values. Label semantics are deployment- or registry-defined; the comparison is not. Subset: no broader when the candidate array's members are a subset of the reference array's, compared as exact strings. Intersection: the common members; an empty result has no intersection and the combination fails. * allowed_tools (array of strings): the action may be exercised only through a capability whose identifier is among the named values (a tool or function identity, asserted at the point of use by the enforcing component). Subset: no broader when the candidate array's members are a subset of the reference array's, compared as exact strings. Intersection: the common members; an empty result has no intersection and the combination fails. * requires_action_approval (boolean): when true, each exercise of the action requires a fresh, action-bound approval at the point of use; the enforcing component MUST NOT permit the action on Mission authority alone. A value of false is equivalent to omitting the member. Subset: no broader when the candidate is true or equals the reference (narrowing may add the requirement, never remove it). Intersection: the logical OR of the two values. These comparisons are in value space, not lexical: max_amount amount members are compared as the decimal numbers the strings contain, so "500", "500.0", and "500.00" are equal; resource_issued_after and resource_issued_before values are compared as the instants they denote after normalization to UTC, so two RFC 3339 representations of the same instant that differ only in timezone offset or trailing subsecond zeros are equal; recipient_domain values are compared as DNS names, case-insensitively and on whole labels, so mail.example.com is within example.com and example.net is not. A Common Constraint definition MUST fix its subset and intersection in value-space terms, so that independent deployments compute the same result for the same values and the subset rule of Section 5.1 is reproducible. A decimal-string value in a Common Constraint (max_amount's amount member, and any future Common Constraint with a decimal-valued member) MUST match ^[0-9]+(\.[0-9]{1,18})?$: one or more decimal digits, optionally followed by a single . and one to 18 further decimal digits. A leading - is out of scope for a ceiling value and MUST be rejected. A value of any other form, including scientific notation ("1e300"), a sign, a thousands separator, or a non-numeric token ("NaN", "Infinity"), is malformed, and a consumer MUST reject it rather than attempt to parse it: an authority proposal carrying one is refused at submission (Section 4.2), and a Resource Server treats a malformed decimal value the same as a constraints key it cannot enforce (Section 8.2). Comparison and intersection over two such decimal-string values MUST be computed as exact decimal arithmetic (for example, by scaling both values to integers by their fractional digit count and comparing the integers) and MUST NOT parse either value into an IEEE 754 binary floating-point type: that representation does not hold every value the grammar above admits exactly and can compare or combine two values incorrectly. A numeric constraint value MUST lie within the range JCS [RFC8785] serializes exactly. Monetary amounts avoid that hazard by construction: max_amount carries its amount as a string containing a decimal number, paired with an ISO 4217 [ISO4217] currency code, and a future Common Constraint for a monetary value SHOULD reuse this shape rather than a JSON number. 5.3. Other Authorization Details Types mission_resource_access is the only type this document defines, but the Authority Set MAY include other AS-supported [RFC9396] authorization_details types when an audience consumes them. ("Supported" here means the AS recognizes and documents the type: it appears as a key in the AS's authorization_details_types_metadata_endpoint response, the source of truth for the supported set (Section 13); authorization_details_types_supported, where the AS also advertises it, mirrors that key set. RFC 9396 establishes no IANA registry of type identifiers.) The Mission apparatus is type-agnostic toward such entries: * they are committed by authority_hash and gated on Mission state exactly as mission_resource_access entries are; * narrowing and delegation use the subset semantics the type defines (Section 5.1, Section 11.3). A type whose subset and delegation semantics the AS does not understand MUST NOT be delegated, audience-projected to a Resource AS, or narrowed: the AS cannot prove a transformed copy is still a subset of what was approved. Such an entry MAY be issued only to its original approved audience, carried exactly as approved, and MUST NOT appear in a delegated token or in a cross-domain grant ([I-D.draft-mcguinness-oauth-mission-cross-domain]); * evaluating the entry against a concrete request is the runtime layer's responsibility (Section 16.6), not the AS's. This lets policy-language profiles compose without this document defining them: for example, an entry carrying a Cedar policy set ([I-D.draft-cecchetti-oauth-rar-cedar]), or an analogous AuthZEN policy entry, for an audience that evaluates it. The AS derives such an entry from the Mission Intent and bounds it by the Intent like any other, but treats the carried policy largely opaquely; the Resource Server or Policy Decision Point (PDP) evaluates it at request time. Stated as a limit: the subset rule is fully defined only over mission_resource_access and its Common Constraints. Authority expressed in another type has whatever subset relation that type defines, or none, and where it has none the entry is carried as approved: never narrowed, delegated, or projected. The narrowing guarantee is therefore strongest while authority stays in mission_resource_access entries, and weakens as expressiveness moves into opaque policy-language entries. Example (non-normative): an Authority Set with a Cedar policy entry for a finance audience that consumes Cedar, alongside a mission_resource_access entry for a calendar audience that does not. The Cedar policySet is abbreviated: [ { "type": "account_information", "rarFormat": "cedar", "policySet": "permit(principal, action, resource) when {...};" }, { "type": "mission_resource_access", "resource": "https://calendar.example.com", "actions": ["events.read"], "constraints": { "window_days": 30 } } ] Both entries are committed by the one authority_hash and bound to the Mission. The Cedar entry is evaluated by the finance audience's PDP; the mission_resource_access entry is enforced as in Section 8. Because the Cedar profile defines no subset or delegation rule over policy sets, the AS carries the Cedar entry as approved rather than narrowing it, and it MUST NOT appear in a delegated token or cross- domain grant. Delegation controls on other entries, such as the mission_resource_access entry, apply to those entries only. 5.4. Modeling Tools and Function Calls This section is non-normative guidance. A "tool" an agent invokes, such as a Model Context Protocol (MCP) tool or a function call, is modeled as a mission_resource_access entry. No separate entry type is needed, and the rules above (derivation, subset, delegation, authority_hash) apply unchanged. The mapping is: * resource is the tool provider. For an MCP tool it is the MCP server's URL. The MCP authorization model ([MCP]) makes the server an OAuth 2.0 resource server, so this is the resource identifier a token is audience-bound to. * actions are the tool names the task needs at that provider. Authorizing a tool is authorizing its name as an action, which lines up with MCP filtering its tool list by the caller's granted authority and routing each tool call for authorization. * constraints carry machine-actionable bounds on a tool's arguments, for example an amount ceiling or a recipient domain (the max_amount and recipient_domain Common Constraints, Section 5.2). Like all constraints, they are committed by authority_hash and carried to the point of use, but they are evaluated against the concrete call arguments by a runtime enforcement layer, not at issuance (Section 16.6). For example, a Mission authorized to read invoices and post small adjustments through a finance MCP server, and to send messages through a messaging MCP server, derives: [ { "type": "mission_resource_access", "resource": "https://finance.example.com/mcp", "actions": ["query_invoices", "post_adjustment"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } }, { "type": "mission_resource_access", "resource": "https://mail.example.com/mcp", "actions": ["send_message"], "constraints": { "recipient_domain": "example.com" } } ] Delegation to a sub-agent works unchanged: add a delegation member to a tool entry (Section 11.3). For example, an entry a sub-agent of type ai_agent may invoke at depth 1, narrowed to the read tool only: { "type": "mission_resource_access", "resource": "https://finance.example.com/mcp", "actions": ["query_invoices"], "delegation": { "max_depth": 1, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } } What this profile does not provide for tools is a typed, attenuable per-argument constraint grammar: narrowing one tool's argument schema against another (for example, amount in a range, recipient in a one_of set) as the grant is derived or delegated. Argument bounds here are the same flat, carried constraints used for any resource, evaluated at runtime. Structured per-argument attenuation ([I-D.draft-niyikiza-oauth-attenuating-agent-tokens], with object capability systems such as UCAN as prior art) is a richer primitive deferred to future work; it would extend the delegation and subset model of this document (Section 11.3, Section 5.1) rather than introduce a new entry type. 6. Mission Approval The approval event is the atomic transition at which the Approver consents and the AS creates the Mission. It runs as an OAuth 2.0 [RFC6749] authorization-code flow initiated from the PAR-issued request_uri (Section 4.1). Because the authorization code is the artifact the Mission grant binds to (Section 6.1) and it travels the front channel, the AS MUST bind the code to the requesting client with PKCE ([RFC7636], S256 challenge method) or, equivalently, issue a DPoP-bound authorization code ([RFC9449]). The AS MUST reject a code redemption whose PKCE verifier or DPoP key does not match the binding established for the request. This prevents authorization- code injection from yielding the Mission grant. The AS SHOULD include the iss authorization-response parameter ([RFC9207]) on the authorization response, so the client can detect a mix-up attack on the consent-bearing redirect leg (per the guidance of [RFC9700]). At the approval event the AS MUST, in order: 1. Authenticate the Approver. If the Mission Intent's controls.acr is present, the authentication MUST satisfy it. 2. Establish the Subject: the principal the task is for, recorded as the Mission's subject and mapped to the sub of every derived token (Section 8). When the Approver is the Subject (self- approval), this is the authenticated Approver. When the Approver is a different principal (for example, an administrator or manager approving on a user's behalf), the AS MUST itself establish the Subject's (iss, sub), and MUST authorize the Approver to approve for that Subject under local policy. The AS MUST NOT take the Subject from unauthenticated client input. This document defines no wire parameter for the Subject; how the AS establishes it (administrative selection, a directory, an authenticated reference) is a deployment matter. The Subject may be a workload or organizational principal (Section 3.2); its establishment and the mapping below are unchanged. When the Subject's home issuer (subject.iss) differs from the AS, the AS MUST map the external (subject.iss, subject.sub) pair to an AS- local sub under an injective mapping: one external Subject maps to exactly one AS-local sub, and no two distinct external Subjects map to the same local sub, so a derived token's (iss, sub) pair unambiguously denotes the Subject. Adopting the external sub string verbatim as the local sub is one permitted deployment choice, valid exactly where it collides with no other principal's sub; the injectivity, not the verbatim adoption, is the rule. 3. Establish the authority source: whose authority the approval draws on, recorded as the Mission's authority_source (Section 3.2, Section 7). The AS MUST establish it from trusted configuration or authenticated governance state, never from client assertion. The AS MUST verify the Approver is authorized under local policy to activate the established source, and MUST verify the derived Authority Set lies within that source's authority (for organizational, within the governed policy identified by authority_source.policy). These are distinct checks: an organizational owner may be authorized to activate policy without personally holding every operational permission. The AS MUST refuse when either relationship cannot be established. 4. Render for consent the derived Authority Set in human-meaningful terms, with the goal, constraints, expires_at, and any controls bounds (notably max_derivations) as context: * The object the Approver consents to is the *derived Authority Set*, what the agent may actually do, not the goal or Mission Intent: the authority itself MUST be what is rendered and consented to. An approval surface that renders only the goal, success_criteria, or Mission Intent and not the derived Authority Set does not conform. * When the Approver is not the Subject, the rendering MUST identify the Subject the authority is granted for. * The rendering MUST identify the authority source and, for organizational, the governed policy it draws on (authority_source.policy). A change to the established source before the decision MUST force re-evaluation and re-rendering. * When the client submitted an authority proposal (Section 4.2), the rendering MUST distinguish the entries the client proposed from any narrowing or restructuring the AS applied. The Authority Set, not the Intent, is the consent object because derivation is local policy: nothing commits that the derived authority faithfully reflects the goal the Approver read. 5. Compute the integrity anchors (Section 6.3): authority_hash over the consented Authority Set, intent_hash over the approved Mission Intent, and, when an authority proposal was submitted (Section 4.2), proposal_hash over the submitted authorization_details array. 6. Create the Mission record (Section 7) in the active state, atomically with issuance of the authorization code. The atomic coupling of the approval decision to authorization-code issuance is this flow's shape, not the model's: a companion profile ([I-D.draft-mcguinness-oauth-mission-approval]) relocates the approval event for deployments whose approvals are asynchronous or reviewer-narrowed, under the extension seam of Section 12; the steps above, their order, and the atomicity of record creation with the approval decision are what any relocation preserves. Every Mission is rooted in an approved authorization basis (Section 7): the steps above define the direct basis, which MUST be a human approval event. A companion profile MAY define a named standing-consent basis instead (a template ceiling, a drawdown policy), so a policy can approve an instance at machine speed within a bound a human already consented to; such a basis MUST still trace to an accountable human through approval_basis.consent_principal and root_commitment, and creates no fresh approval event per instance. The consent rendering is hardened against client text: * Client-supplied strings (goal, constraints, success_criteria) MUST be rendered as inert text and MUST NOT be interpreted as markup. * The AS SHOULD mitigate Unicode direction-override and confusable- character presentation in them. * The rendering MUST visually distinguish the AS-derived Authority Set from client-supplied text, so crafted client text cannot pass as derived authority. A deployment MUST declare a minimum approval-authentication strength for Missions whose derived Authority Set carries high-risk authority: irreversible, external-commitment, or privileged-administration actions under the deployment's classification, or a consumption bound ([I-D.draft-mcguinness-mission-metering]). The declaration's home is the deployment's published statement of what it enforces (the Mission Deployment Profile of [I-D.draft-mcguinness-mission-architecture]), so the floor is a checkable published fact rather than a floating duty. The approval authentication for such a Mission MUST meet that minimum, and controls.acr can raise the required strength but never lower it below the floor. The material notices of the consent- evidence profile identify these same high-risk classes ([I-D.draft-mcguinness-oauth-mission-consent-evidence]). An Approver who declines, and an Approver whose authentication cannot satisfy controls.acr or the declared approval-strength floor, yield access_denied on the authorization response ([RFC6749]). A token- endpoint resource value outside the Authority Set yields invalid_target ([RFC8707]). Rendering a bound is not the same as enforcing it: a deployment MUST NOT present a rendered bound as enforced when no party enforces it. Which party enforces each bound, and what holds when that enforcer is absent, is summarized in the enforcement table (Section 4). An AS SHOULD make clear to the Approver which rendered bounds its deployment actually enforces, so consent is not given to a limit that binds nowhere. The authority_hash is the *authority commitment*: it commits, by cryptographic digest, exactly the authority the Approver approved. It commits the approved authority, not the way that authority was rendered to the Approver; this profile commits no separate consent disclosure object (see Section 16.1). Every token derived under the Mission carries this value (Section 8). A party holding the full Authority Set can recompute the commitment, confirming the set is the approved one, and then apply the subset rule to the token's carried authority. A party holding only a narrowed subset cannot recompute the commitment and treats it as an audit anchor. For that party, authority_hash alone does not prove that the carried entries belong to the approved set; the relationship is asserted by the AS's signed token and depends on the AS applying the subset rule correctly (see Section 16.1). If the task, the authority proposal, or the derived Authority Set changes between approval rendering and the approval decision, the AS MUST recompute the anchors the Mission records (intent_hash and authority_hash, and proposal_hash where a proposal was submitted) and MUST NOT create the Mission without the Approver's consent to the changed context, each anchor computed over the context actually approved (the approval-event rule of [I-D.draft-mcguinness-mission-substrate], stated here across this document's commitments). Committing the proposal separately keeps the anchors from equivocating: a proposal swapped between rendering and decision changes proposal_hash even where intent_hash is unchanged. The intent_hash commits the *approved Mission Intent*: the task the Approver consented to, as recorded on the Mission. It makes the recorded task tamper-evident: an auditor can verify the Mission's intent against intent_hash and detect any later alteration. intent_hash commits the task; authority_hash commits the authority derived from it. 6.1. Binding the Mission to the Grant At the approval event the AS binds the Mission to the authorization grant it issues: the authorization code, and the refresh token issued from it. The binding is server-side and is what "the referenced Mission" in Section 9 refers to. At each subsequent derivation the AS resolves the Mission from the grant the client presents: the authorization code at the token endpoint (the initial exchange uses grant_type=authorization_code), the refresh token on refresh, or the Mission-bound subject_token on Token Exchange (the actor_token identifies the delegate, Section 11). It then applies the gating of Section 9. A client does not supply mission_id to obtain a derivation; an AS MUST NOT derive Mission-bound authority from a client-supplied mission_id, because the grant, not the identifier, determines the Mission. When the authorization code expires unredeemed, no derivation is possible under the Mission; the AS SHOULD revoke the Mission or allow it to expire. A client learns its mission_id from the mission claim's id on each issued token (Section 8.1) or from the token response. This document defines mission_id as a token-endpoint response parameter: a string carrying the Mission Identifier, returned alongside the issued token. An AS SHOULD return it: the client is not expected to parse the access token, and when the parameter is absent the only path to the identifier is reading the mission claim, which fails for a token that is encrypted or opaque to the client. It is an informational reference only: presenting it authorizes nothing (Section 9), and a client MUST NOT derive authority from it. 6.2. Single Accountable Approver This document records exactly one approver: the accountable principal who approved the Mission. Two richer patterns are deliberately out of scope and deferred: * *Multi-party approval* (M-of-N or dual control), where more than one principal must approve. The number of approvers is orthogonal to the consent commitment: however many principals approve the same rendered Authority Set, the authority_hash is identical, as is every token derived from it. Multi-party approval raises the assurance of _how_ approval was obtained; it does not change the artifacts this document produces. A deployment requiring dual control records one accountable Approver under the same authority_hash; this document does not natively represent the co- approvers. * *Approval-authority provenance* (the deeper standing behind a delegate's authority to approve for another principal, for example, whether an administrator was themselves entitled to approve on a user's behalf). This is governance state about the delegate's own standing, not the named standing-consent basis approval_basis records for a template or drawdown-policy activation (Section 7), and is left to a governance layer. Both remain out of the core. Where a deployment needs them, they are recorded by the Approval Governance Record ([I-D.draft-mcguinness-mission-approval-governance]). Consent Evidence may carry a deliberately partial presentation of that record through co_approvals and its approval-governance members ([I-D.draft-mcguinness-oauth-mission-consent-evidence]). 6.3. Integrity Anchors Every anchor is computed the same way over a domain-separated, issuer-bound envelope: 1. Construct the envelope, where typ selects the committed object and value is that object: { "typ": "", "iss": "", "value": } For intent_hash, typ is mission-intent and value is the approved Mission Intent object: the Submission envelope's intent member, never the envelope or its evidence (Section 4.1). For proposal_hash, typ is mission-proposed-authority and value is the submitted authorization_details array exactly as recorded (Section 4.2); the anchor exists iff a proposal was submitted. For authority_hash, typ is mission-authority-set and value is the Authority Set as a JSON array of entries. 2. Canonicalize the envelope with JCS [RFC8785]. 3. Compute SHA-256 [RFC6234] over the canonical bytes. 4. Encode as sha-256: followed by the base64url, no-padding, encoding of the digest. The typ field domain-separates the anchors so a digest of one object can never be mistaken for another's. The iss binding prevents a committed object from being transplanted across Authorization Servers. The typ value space is an extension point (Section 12): additional committed objects use this same envelope with a new typ and the canonicalization below. This document defines no registry of typ values; each committing specification defines its own and relies on the typ domain separation. To keep that domain separation safe without a registry, a new typ value MUST be a collision-resistant name (for example, a short name prefixed within a namespace the defining profile controls, following the Collision-Resistant Name guidance of [RFC7519] Section 4.2). The mission- prefixed values defined by profiles that extend this document share an author- coordinated namespace for this reason. SHA-256 is the only digest algorithm this document defines; the sha- 256: prefix identifies it. Algorithm agility is future work. A verifier MUST reject an integrity anchor whose algorithm prefix it does not recognize. A verifier MUST NOT treat an unrecognized prefix as sha-256. This ensures that adding an algorithm later cannot be exploited as a downgrade. 6.4. Canonicalization Rules JCS [RFC8785] alone does not make two implementations agree on every byte. The following rules close the remaining gaps; they apply to computing an anchor and to comparing committed values: * The committed value is exactly the object the AS recorded on the Mission: the approved intent for intent_hash, the recorded proposed_authority for proposal_hash, and the authority_set for authority_hash. An auditor reproduces a digest from the record alone. * The AS MUST reject an input object containing duplicate JSON member names before canonicalization; such input is invalid. * JCS does not reorder array elements, and this document defines no element sorting, so array order is significant. The AS MUST emit each array in a fixed, reproducible order; that order is part of the canonical form. * URI-valued members are compared byte-for-byte unless a member's definition specifies a normalization. The only normalization this document defines is the RFC 3986 [RFC3986] comparison of the resource-containment test (Section 5.1), which applies to that comparison alone: the default resource equality test of Section 5.1 remains an exact match, and anchor computation is always byte-exact over the recorded values. Test vectors for the anchors are provided in Appendix B. 7. Mission Record A Mission is the durable record created at the approval event. Its members are immutable after creation except for its state, and it is identified by a Mission Identifier (Section 7.1). Operational issuance bookkeeping, such as the derivation count gated under Section 9, is AS-side state about the Mission, not a member of the immutable record. Naming follows one rule across every surface that carries Mission facts. Record members do not repeat the mission prefix, because the record itself is the Mission; prefixed names belong to surfaces that reference a Mission from outside it, such as the mission_intent request parameter and the mission_id response parameter. Member names are spelled out (issuer, expires_at, created_at) and spelled identically everywhere; the compact JWT names (iss, exp, iat) describe a signed artifact's own envelope, or identify a party in an {iss, sub} object, never the Mission. Like the mission claim (Section 8.1), the record is open (Section 12): a companion profile of this document MAY record additional members set at creation using short names coordinated with it (for example, a lineage member linking the Mission to a predecessor or parent); any other extension MUST use collision- resistant names. The members below are the ones this profile defines: id: REQUIRED. A string. The canonical Mission Identifier (Section 7.1). issuer: REQUIRED. A string. The issuer URL of the Mission Issuer that approved the Mission. Equals the iss of tokens that AS derives; for cross-domain tokens it remains the originating AS even though the issuing iss differs ([I-D.draft-mcguinness-oauth-mission-cross-domain]). state: REQUIRED. A string. The current lifecycle state: active, revoked, or expired in this profile, or an additional state defined by a companion profile, subject to the forward- compatibility rule of Section 9. intent: REQUIRED. An object. The approved Mission Intent. proposed_authority: OPTIONAL. An array. The authorization_details array the client submitted as its authority proposal (Section 4.2), recorded exactly as submitted. Present iff a proposal was submitted; a Mission derived in template mode (Section 5) records none. authority_set: REQUIRED. An array. The consented Authority Set. authority_hash: REQUIRED. A string. The consent commitment over the Authority Set (Section 6.3). intent_hash: REQUIRED. A string. The integrity commitment over the approved Mission Intent (Section 6.3), making the recorded task tamper-evident. proposal_hash: OPTIONAL. A string. The integrity commitment over the recorded proposed_authority (Section 6.3). Present iff proposed_authority is present. It is approval-time provenance, not enforcement input: like approval_basis, it is surfaced on the record and through introspection (Section 10) and is not carried on the mission claim (Section 8.1). submission_evidence: OPTIONAL. An array. The verified Intent Submission Evidence facts (Section 4.3), one element per verified entry, present iff the approved submission carried evidence. Each element carries exactly these members: type, the entry's evidence type; artifact_hash, an integrity anchor (Section 6.3) with typ mission-intent-evidence over the entry exactly as presented; verified_at, an RFC 3339 timestamp of verification; and facts, an object holding the verified output facts the type's specification designates for recording, nested so type-owned facts cannot collide with the common members. Elements preserve the submission's evidence order, so the array has one canonical form (Section 6.4). Like approval_basis, it is provenance, not enforcement input, and it is not carried on the mission claim (Section 8.1). No integrity anchor commits it: its digests are record metadata whose trustworthiness is the trust in this immutable record, not an independently verifiable association between the Mission and the artifacts presented at admission, and an auditor who does not trust the record cannot prove from the anchors which evidence was used. A profile whose threat model requires that association commits normalized provenance under its own anchor typ (Section 6.3, Section 12); this document defines none. subject: REQUIRED. An object. The Subject, an object with iss and sub. approver: REQUIRED. An object. The Approver, an object with iss and sub. MAY equal subject. Equal to approval_basis.consent_principal (below). approval_basis: REQUIRED. An object. The authorization basis this Mission is rooted in: every Mission is rooted in an approved authorization basis, fixed at the approval event and immutable thereafter, like approver and subject. Members: type: REQUIRED. A string: direct, defined in full by this document, or an additional value defined by a companion profile that generalizes approval to a named standing-consent basis (for example, template ([I-D.draft-mcguinness-oauth-mission-template]) or policy_drawdown ([I-D.draft-mcguinness-oauth-mission-child-delegation])), subject to the forward-compatibility rule of Section 9. consent_principal: REQUIRED. An object with iss and sub. The accountable human (or human-accountable principal) who consented. This is the value of approver; this document does not add a second accountable principal. activation: REQUIRED. An object naming what activated this Mission instance, shaped by type. For direct: approval_event_id, mirroring the record's own approval_event_id (below). activation_actor: REQUIRED. An object with iss and sub. Who or what triggered this instance. For direct it equals consent_principal: the Approver triggers their own approval. A standing-consent type names a dispatching or requesting party distinct from the consenting human. root_commitment: REQUIRED. A string. The commitment to the consented root: an integrity anchor where the root is a committed object, otherwise the committed reference that identifies it. For direct, this Mission's own authority_hash. For direct, activation.approval_event_id MUST identify a human approval event (Section 6). A companion profile defining a standing-consent type MUST make its consent_principal and root_commitment trace to an accountable human's approval of the named standing consent, with no fresh approval event per instance. approval_basis is provenance: it is recorded alongside approver and is not folded into intent_hash or authority_hash (Section 6.3). Neither anchor commits it, and it MUST NOT be added to either digest; a profile that commits the Mission Record itself covers it under that profile's own anchor. The mission claim MAY carry approval_basis.type (Section 8.1) as a wire signal; it MUST NOT be relied on to grant or widen authority. authority_source: REQUIRED. An object. The source of the authority the approval draws on (Section 3.2), established at the approval event (Section 6) and immutable thereafter, like approver and approval_basis. Members: type: REQUIRED. A string: user_delegated, service_owned, or organizational, subject to the forward-compatibility rule of Section 9. policy: REQUIRED for organizational, absent otherwise. An object with id, version, and digest: the stable reference to, and commitment over, the governed organizational policy the Mission draws on. authority_source is provenance like approval_basis: recorded alongside it, folded into neither intent_hash nor authority_hash, and not carried on access tokens; Resource Servers enforce authorization_details and do not consult it. client_id: REQUIRED. A string. The Agent (OAuth client) that submitted the Mission Intent. policy_version: REQUIRED. A string. The derivation policy version in effect at the approval event. approval_event_id: REQUIRED. A string. A unique identifier of the approval event, used as the approval idempotency key: a retried or duplicate delivery of the same approval decision (a replayed callback, a double-submitted consent form) MUST NOT create a second Mission, and the AS deduplicates on this identifier. It is otherwise an opaque audit identifier with no wire semantics. created_at: REQUIRED. A string. RFC 3339 timestamp of creation. expires_at: REQUIRED. A string. Mirrors intent.expires_at. The *audit horizon* is the deployment-declared retention window for the Mission record and its evidence: at least the Mission's lifetime plus a declared post-expiry period. After the Mission reaches a terminal state (revoked or expired), the record MUST be retained for the audit horizon. 7.1. Mission Identifier Format A Mission Identifier is an opaque URL-safe ASCII string of [A-Za- z0-9_-] characters, with at least 128 bits of entropy, carrying no semantic content. It MUST NOT be reused. The record and the mission claim carry it as id; a surface that references a Mission from outside carries it as mission_id, as in the token-response parameter (Section 6.1). 7.2. Worked Example { "id": "msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-", "issuer": "https://as.example.com", "state": "active", "intent": { "goal": "Reconcile Q3 invoices ...", "resources": ["https://erp.example.com"], "expires_at": "2026-12-31T23:59:59Z" }, "proposed_authority": [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.*"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ], "authority_set": [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ], "authority_hash": "sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ", "intent_hash": "sha-256:wQ7p4LHnX9Md0LqJ6sZJ8b8mZ3rN2xT5pV4lE6sQqYY", "proposal_hash": "sha-256:kT2mR7vX4qL9nY5pB1sD8fJ6wZ3hC0aGeUoNvSqMrYo", "subject": { "iss": "https://idp.example.com", "sub": "user_3p2q8mN1a0kV7tR" }, "approver": { "iss": "https://idp.example.com", "sub": "user_3p2q8mN1a0kV7tR" }, "approval_basis": { "type": "direct", "consent_principal": { "iss": "https://idp.example.com", "sub": "user_3p2q8mN1a0kV7tR" }, "activation": { "approval_event_id": "ape_8K2nP4qV9rL3tY6sB1z" }, "activation_actor": { "iss": "https://idp.example.com", "sub": "user_3p2q8mN1a0kV7tR" }, "root_commitment": "sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ" }, "authority_source": { "type": "user_delegated" }, "client_id": "s6BhdRkqt3", "policy_version": "deploy-policy:v17", "approval_event_id": "ape_8K2nP4qV9rL3tY6sB1z", "created_at": "2026-10-15T14:32:11Z", "expires_at": "2026-12-31T23:59:59Z" } This recorded intent, proposed_authority, and authority_set, and the anchors above, are this document's canonical worked example; the test vectors (Appendix B) compute over them. A companion that extends this example MUST either reproduce the recorded objects byte-exactly or state explicitly that its example diverges and its anchors differ; an extended example with silently different anchors reads as the same Mission and has repeatedly caused drift. 8. Mission-Bound Access Tokens Access tokens issued under a Mission are JWTs per [RFC9068], which fixes the required claims (including jti) and the at+jwt JOSE header typ ([RFC9068] Sections 2.1 and 2.2); a Resource Server MUST verify the typ per [RFC9068]. In addition to what that profile requires, a derived token: * carries the token's Mission-derived authority as authorization_details ([RFC9396]); this MAY be the full Authority Set or a narrowed subset (Section 5.1); * carries a mission claim (Section 8.1); * sets sub to the AS-local sub the AS maps the Mission's Subject to (Section 6); * carries client_id per its ordinary [RFC8693] Section 4.3 and [RFC9068] Section 2.2 meaning, the client that requested this particular token; the Mission's originally-approved agent is not carried on the token, and remains recorded in the Mission Record (Section 7); * MUST set aud to identify the Resource Server(s) authorized to consume the carried authorization_details, and MUST NOT include an audience unrelated to that carried authority (see below); * MAY carry an act claim when the agent has delegated execution (Section 11); * MAY carry a scope claim, subject to the rule below; * SHOULD be sender-constrained, via a cnf claim [RFC7800]: DPoP [RFC9449] (cnf.jkt) or mTLS [RFC8705] (cnf.x5t#S256). Stated explicitly for estates whose access tokens are opaque reference tokens: this document's token-carried enforcement assumes the JWT above and does not profile opaque tokens. Such an estate has two paths. A Resource Server that validates opaque tokens through introspection receives the mission member and Mission state on the introspection response (Section 10); and an estate whose AS cannot issue Mission-bound tokens at all deploys the standalone Mission Issuer binding, which governs ordinary tokens at the enforcement layer ([I-D.draft-mcguinness-mission-authority-server]). The AS MUST NOT include authorization_details exceeding the Mission's Authority Set. On any issuance that narrows authority (for example, a single-audience token), each emitted entry MUST be a subset of a Mission Authority Set entry under Section 5.1. The aud SHOULD be derived from the resource indicators ([RFC8707]), Protected Resource metadata ([RFC9728]), or the deployment's resource-to-RS mapping. It identifies the Resource Server(s) and need not be byte-equal to the resource values of the authorization_details entries: an aud typically names an RS, API, or security domain, while entries name resources, accounts, tools, or locations beneath it. Bounding aud to the consuming Resource Server(s) prevents a confused-deputy or token-redirection attack, in which a multi-resource Authority Set yields a token an unrelated Resource Server would accept even though it was obtained to act elsewhere. A deployment SHOULD prefer per-RS (single-audience) tokens, narrowed under Section 5.1: the client requests one at the token endpoint with the [RFC8707] resource parameter (and MAY further narrow with scope), and the AS narrows the Authority Set under Section 5.1 to the requested resource(s) and sets aud to the corresponding Resource Server(s). This is the within-domain counterpart of the audience- scoping the Mission Issuer applies when projecting authority to a Resource AS ([I-D.draft-mcguinness-oauth-mission-cross-domain]). Sender-constraining is a SHOULD for this primary token, aligned with [RFC9700]. It is stronger (MUST) for delegated tokens, which face higher replay exposure in the hands of a less-trusted delegate (Section 11); the companion sets the same MUST for the credentials that cross a trust domain ([I-D.draft-mcguinness-oauth-mission-cross-domain]). A deployment SHOULD sender-constrain the primary token as well where its threat model warrants. The token-endpoint response conveys the granted authority to the client. Because what the client submitted was a proposal, never the grant (Section 4.2), and the client is not expected to parse the access token, the AS MUST return the granted authorization_details in the token-endpoint response, per [RFC9396] Section 7, reflecting exactly the (possibly narrowed) set assigned to the issued token; the same applies to refresh and Token Exchange responses. The mission_id response parameter carries the Mission reference beside it (Section 6.1). For example, the agent narrows the canonical ERP Mission (the worked example of Section 7) to a read-only token, presenting the Mission's refresh token with the [RFC8707] resource parameter and narrowing further with scope: POST /token HTTP/1.1 Host: as.example.com Content-Type: application/x-www-form-urlencoded DPoP: eyJ0eXAiOiJkcG9wK2p3dCIsImFsZyI6IkVTMjU2Iiwi... grant_type=refresh_token &refresh_token=rt_4mN8qV2xP7sL1tY9zB3k &resource=https%3A%2F%2Ferp.example.com &scope=invoices.read The issuance is a derivation, gated on the Mission being active (Section 9). The response echoes the narrowed grant and the mission_id reference (Section 6.1); the emitted entry is a subset (Section 5.1) of the Mission's read entry, its constraints carried intact: { "access_token": "eyJhbGciOiJFUzI1NiIsInR5cCI6ImF0K2p3dCJ9...", "token_type": "DPoP", "expires_in": 300, "mission_id": "msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-", "authorization_details": [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } } ] } Mission-bound refresh tokens MUST be sender-constrained or use refresh token rotation. This matters most for a public client, since Mission-state gating bounds a stolen refresh token's usefulness over time but not while the Mission is still active. This strengthens the refresh-token guidance of [RFC9700] Section 2.2.2, whose MUST applies to public clients, to all Mission-bound refresh tokens. The authentication context of the approval event (controls.acr, Section 4) describes the Approver at approval time and is recorded on the Mission, not on derived tokens; a derived token's authority comes from the Mission, not from a fresh authentication, so this document requires no acr or auth_time claim on it. An AS MAY include acr and auth_time per [RFC9068] to convey the approval context, but a consumer MUST NOT treat their absence as an authentication downgrade. authorization_details is the authoritative expression of a Mission- bound token's authority. Any scope the token carries MUST be derived from, and no broader than, the Authority Set. Specifically: * every scope value MUST correspond to authority already present in authorization_details; and * a scope value MUST NOT convey authority, or relaxation of a constraint, that the Authority Set does not grant. Because scope is a coarse string list, it cannot carry the per-entry constraints; it is a compatibility projection, never the authoritative form of the Mission's authority. The AS MUST NOT issue a Mission-bound token whose scope exceeds the Authority Set. A credential the Mission Issuer derives MUST have an exp that does not exceed the Mission's expires_at, so that no credential outlives the approved Mission (not merely that none is issued after expiry). How this bound extends transitively to tokens minted in another trust domain is specified by the companion ([I-D.draft-mcguinness-oauth-mission-cross-domain]). Short-lived access tokens remain this profile's issuance-only recommendation: with no runtime layer, token lifetime is the revocation-latency bound at unmodified Resource Servers. Where a runtime layer covers the high-consequence classes with an active freshness source, the point-of-use decision is the revocation cutoff, and a deployment MAY size lifetimes by action class without losing the kill switch (Section 16.6, [I-D.draft-mcguinness-mission-runtime]). Classes attach to entries while exp attaches to the token: an extended lifetime is appropriate only for a token whose carried entries are all on runtime-gated paths, since a single ungated entry stretches its own revocation latency to the extended lifetime. Narrowed, single-audience tokens (Section 5.1) are the mechanism that keeps gated and ungated authority from sharing one long-lived token. 8.1. The Mission Claim The mission claim is a JSON object: id: REQUIRED. A string. The Mission Identifier (Section 7.1). issuer: REQUIRED. A string. The Mission's issuer (Section 7). A credential's iss names the party that minted it; mission.issuer names the party that approved and serves the Mission, and the two deliberately differ for tokens minted in another trust domain ([I-D.draft-mcguinness-oauth-mission-cross-domain]). authority_hash: REQUIRED. A string. The Mission's authority_hash, the commitment over the complete consented Authority Set (Section 8.2). expires_at: OPTIONAL. A string. The Mission's expires_at (Section 7), in RFC 3339 [RFC3339] date-time form and named identically to the record member it mirrors. It is a bounding and audit commitment with no liveness: a validator can check that the token's exp does not exceed it, and its passing says nothing a state surface does not, since expiry is not revocation and only active permits reliance (Section 9). approval_basis: OPTIONAL. An object carrying type only: the Mission's approval_basis.type (Section 7), as a wire signal of which authorization basis the Mission is rooted in. The record's other approval_basis members are not carried on the token. It MUST NOT be relied on to grant or widen authority; the carried authorization_details and authority_hash remain authoritative. The mission claim is an open object (Section 12): additional members MAY appear alongside the members above. This document defines no registry of mission members. A companion profile of this document MAY use short member names coordinated with it; any other extension member MUST use a collision-resistant name (for example, a name in a namespace the extension controls, per the Collision-Resistant Name guidance of [RFC7519] Section 4.2) and is defined by the profile that introduces it. A consumer MUST ignore members it does not understand and MUST NOT use any additional member to grant or widen authority; the members above remain authoritative. intent_hash and authority_hash are independent commitments to independent objects. That the approved task bounds the derived authority is a governance assertion, made by derivation policy and auditable through policy_version (Section 5), not a cryptographic relation between the anchors: neither anchor proves anything about the other's object. Example decoded token payload: { "iss": "https://as.example.com", "sub": "user_3p2q8mN1a0kV7tR", "aud": "https://erp.example.com", "client_id": "s6BhdRkqt3", "iat": 1797840000, "exp": 1797840300, "jti": "at_9Kp2vN7sR1tY8mZ3qX5b", "authorization_details": [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ], "cnf": { "jkt": "0ZcOCORZNYy-DWpqq30jZyJGHTN0d2HglBV3uiguA4I" }, "mission": { "id": "msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-", "issuer": "https://as.example.com", "authority_hash": "sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ" } } 8.2. Resource Server Enforcement A Resource Server enforces from the token alone; no call to the AS is required. A Resource Server: * MUST validate the JWT per [RFC9068] and verify any sender- constraint binding (cnf). * MUST treat authorization_details as the authoritative expression of authority and enforce the entries whose resource it serves, permitting only the listed actions subject to constraints. Where more than one carried entry names a resource it serves, a request is permitted if at least one carried entry permits it: entries are alternative grants of authority, not conjunctive filters. * MUST, when it matches a concrete request URI against a prefix entry (Section 5), apply the same RFC 3986 [RFC3986] syntax-based normalization the AS uses for containment (Section 5.1), and MUST apply it before the prefix comparison rather than after. That normalization decodes only unreserved octets, so reserved octets stay encoded: an encoded slash (%2F, equivalently %2f since percent-encoding is hex-case-insensitive) remains encoded and MUST NOT be treated as a path-segment separator when matching. An intermediary that percent-decodes %2F, or a downstream component that decodes and removes dot-segments after the Resource Server's check (for example collapsing a decoded %2e%2e), moves the enforced boundary, so a deployment MUST prevent such rewriting or account for it in what it authorizes. * MUST fail closed on any constraints key it does not understand, or understands but cannot enforce, in an entry whose resource it serves: it MUST refuse the request (for example, a 403 with insufficient_scope [RFC6750], or the deployment's usual insufficient-authority error) rather than grant access while ignoring the key. * MUST NOT reduce a constraint to disclosure-only. * MUST NOT, when a token also carries scope, grant on the basis of a scope value any access broader than the corresponding authorization_details entry permits; in particular, scope MUST NOT be used to bypass a constraint carried only in authorization_details. * MUST treat client_id per its ordinary meaning under [RFC8693] Section 4.3 and [RFC9068] Section 2.2: the OAuth client that requested this token. This profile does not redefine it, on a delegated token or otherwise (Section 11.1). The Mission's originally-approved agent is not carried on the token; it is recorded in the Mission Record (Section 7) at the issuer, and a Resource Server MUST NOT infer that identity from client_id. * MAY impose stronger actor-chain requirements when it authorizes or logs the caller on a token that carries an act chain (for example, requiring and recording the chain), but MUST NOT reinterpret client_id to do so. * MAY, for a Mission-governed resource, be configured to require the mission claim, and MUST then reject a token that lacks it with invalid_token. The downgrade this rejection prevents, and the issuance-side duty that pairs with it, are stated once in Section 16.3. A protected resource MAY advertise this requirement through the mission_bound_authorization_required protected resource metadata member (Section 14). * MAY treat the mission claim as audit and correlation context. * SHOULD, when serving Mission-bound requests, log the mission claim's id and the token jti with each served request, so its access logs join to Mission evidence. * MAY recompute authority_hash (Section 6.3): a Resource Server that performs it recomputes the anchor over the full Authority Set it independently holds, compares the result to mission.authority_hash, and MUST reject on mismatch. It additionally verifies that each carried authorization_details entry is a subset (Section 5.1) of that held set. * MUST NOT recompute authority_hash from a token's carried subset: the anchor is computed only over the full Authority Set, and a narrowed token matches by the subset test, never by hashing its carried entries. * MAY, where the AS offers it, introspect the token (Section 10) to observe the Mission's current state per request rather than relying on the token lifetime to bound revocation latency. * MUST, when it introspects, still verify the token's sender- constraint (cnf) locally and MUST NOT treat an active: true result as proof the caller holds the bound key; the AS does not check possession at introspection (Section 10). A deployment MUST NOT route a delegated Mission-bound token to a Mission-unaware Resource Server that authorizes or logs the caller on client_id without processing the act chain. The requirement above binds a Mission-aware RS; a Mission-unaware [RFC9068] RS reads client_id as the immediate client, which is now accurate for that single token, but it still cannot see the delegation lineage carried in the act chain, and it has no way to look up the originally- approved agent from the Mission Record, so it cannot apply actor- chain policy or join a delegate's action back to the Mission's approval in its own audit records. A resource that requires Mission- bound tokens at all advertises that through the mission_bound_authorization_required protected resource metadata member (Section 14), and a Resource Server that serves such a resource is, by that requirement, Mission-aware; a deployment that delegates routes delegated Mission-bound traffic only to a Resource Server it knows to be Mission-aware in this sense. A constraints member narrows authority, so treating an unenforceable key as absent, or reducing it to disclosure-only, would silently widen the grant; that is why an unrecognized or unenforceable key fails closed. A Resource Server that holds only a narrowed token MUST treat mission.authority_hash as an audit correlator, not an enforcement input, and MUST NOT treat it as a cryptographic proof that the carried entries are a subset of the approved set. That subset relationship is an assertion by the AS, authenticated by the token signature, and depends on the AS applying the subset rule correctly. A worked contrast, using the two-entry Authority Set of the test vectors (Appendix B), shows what each party can verify. A single- audience token carries one narrowed entry: journal-entries.write with the approved max_amount of 500.00 USD tightened to 250.00. The Resource Server verifies the token signature and cnf, checks aud, enforces the carried entry, checks the anchor's algorithm prefix (Section 6.3), and logs the mission claim for correlation. It cannot recompute authority_hash: hashing the carried entry digests a one- entry array the anchor never committed, and the tightened max_amount makes the entry a semantic narrowing, not a byte-level member, of the approved set. Whether 250.00 sits within the approved ceiling is the subset test (Section 5.1), and that test needs the approved entry to compare against. A party holding the full Authority Set (a Resource Server provisioned with it, or a policy decision point holding the Mission record) recomputes the commitment over the held two-entry set, matches mission.authority_hash, and verifies the carried entry as a subset of the held journal-entries.write entry: containment verified relative to the authenticated committed set. For that verification to be independent of the issuer, the deployment provisions the set together with its provenance: the full Authority Set, the Mission id and issuer it belongs to, the expected authority_hash, and an authenticated approval-time source for all three (trusted Mission or consent evidence, or an audit commitment recorded before any compromise). Without that pinning, a compromised issuer signs a token carrying a replacement set's hash, and recomputation confirms the replacement. Three denials above are byte-identical 403s to a client, and misrouting them turns a step-up into an authority-widening ceremony or a fail-closed mismatch into a retry loop. A Mission-aware Resource Server SHOULD therefore state which path it denies into, using the mission_denial attribute this document defines for the WWW- Authenticate response header, carried alongside error per [RFC6750], with one of three values: insufficient_authority (the action is outside the token's carried authority, and more requires a new approval or an expansion where that companion is deployed), step_up_required (the authority exists, but the presented token or its binding does not meet the resource's requirements: a step-up, not a widening), or constraint_unrecognized (an applicable entry carries a constraints key the RS cannot enforce, and the request fails closed). A value the client does not recognize is treated as insufficient_authority. The attribute's disclosure considerations are Section 16.8's. A Mission-unaware Resource Server that authorizes only from scope still operates within the Mission at the coarse scope level, because the AS derived and bounded that scope by the Authority Set (Section 8); but it does not enforce the per-entry constraints, which scope cannot carry. A deployment that relies on those constraints MUST route the protected operation through a Resource Server that enforces authorization_details (or the runtime layer that evaluates them). 8.3. Remediation Grains A denial is not the end of the exchange. The family treats "how to ask again" as a graduated challenge assembled from independent grains, each naming a next step without granting anything itself. This document's own grain is mission_denial above: which path a denial leads into. A Resource Server MAY compose a second grain with it: the insufficient_authorization WWW-Authenticate error code and its authorization_remediation parameter, defined by [I-D.draft-zehavi-oauth-rar-metadata]. authorization_remediation is a base64url-encoded JSON object naming the actionable authorization_details the caller lacks, with an OPTIONAL authorization_reference letting the client match a previously issued token to that same gap. It names what mission_denial: insufficient_authority only points at. This document does not fold that grain into mission_denial's carriage, nor redefine either grain's response status: each rides the wire shape its own defining document gives it. A client that decodes authorization_remediation proposes the carried entries back on the standard authorization_details parameter (Section 4.2), where they derive under this document's ordinary rules (Section 5): of an advertised, schema-valid type (Section 13), narrowed same-type (Section 5.1, Section 5.3) like any other proposal. The loop closes natively: the remediation grain's output vocabulary is this document's input carriage, with no re-wrapping between them. A third grain routes the same denial into a governed access request rather than a fresh derivation: the AuthZEN Access Request and Approval Profile's requestable denial over [AuthZEN.ARAP], adopted by the AuthZEN binding companion ([I-D.draft-mcguinness-mission-authzen]). The three grains compose rather than replace one another: a deployment MAY offer any subset, and none widens authority beyond what Section 5 would derive from the same proposal unremediated. 9. Mission Lifecycle and Gating A Mission is in one of three states: * active: tokens MAY be derived. The only state from which issuance proceeds. * revoked: terminated by the Subject, Approver, or policy. Terminal. * expired: expires_at has passed. Terminal. The transitions are: +========+====================+=========+ | From | Event | To | +========+====================+=========+ | (none) | approval event | active | +--------+--------------------+---------+ | active | revoke | revoked | +--------+--------------------+---------+ | active | expires_at reached | expired | +--------+--------------------+---------+ Table 2 These three states are the mandatory core of the Mission lifecycle state space. This profile owns that state space; an OPTIONAL companion profile MAY define an additional state for a lifecycle it introduces (for example, a paused or a superseded state), but only active ever permits issuance. To keep the state space extensible without a registry, a consumer MUST apply this forward-compatibility rule wherever a Mission state is reported, including the Mission record and the introspection mission member: only the exact value active permits derivation or continued reliance, and every other value, including a value the consumer does not recognize, MUST be treated as non-active and non-deriving. A consumer MUST NOT fail open on an unrecognized state. This makes a state added by a companion profile fail safe for a consumer that predates it. 9.1. Issuance Gating The AS MUST refuse to derive a token, at the token endpoint, on refresh, and on Token Exchange ([RFC8693]), unless the referenced Mission is active. Issuance against a revoked or expired Mission MUST fail with invalid_grant. Because derivation is gated on Mission state, revoking or expiring a Mission stops all further authority for the task, including refresh. The active check MUST be evaluated atomically with issuance, as the derivation-count check already is, so a revocation serialized before an issuance is honored by that issuance. When the Mission Intent sets controls.max_derivations, the AS MUST maintain a per-Mission count of *derivations* and MUST refuse with invalid_grant any derivation that would exceed the bound. A derivation is one issuance operation the issuer AS performs for a single request: the initial authorization-code exchange, a refresh, a Token Exchange, or a cross-domain grant issuance ([I-D.draft-mcguinness-oauth-mission-cross-domain]). Each counts as exactly one, regardless of how many artifacts it emits: a code exchange that returns both an access token and a refresh token is one derivation, and a refresh that rotates both is one. The exact rules: * A derivation that fails, including one refused for exceeding the bound, MUST NOT be counted. * The check and increment MUST be atomic with issuance, so concurrent derivations cannot collectively exceed the bound. * The count covers only derivations the issuer AS performs. Tokens another domain mints locally under the Mission are not counted by the issuer, which cannot observe them; the cross-domain issuance that authorized them was counted once, and the local issuer bounds its own minting by its policy ([I-D.draft-mcguinness-oauth-mission-cross-domain]). The AS maintains this count as internal bookkeeping; it is operational state, not part of the immutable Mission record. invalid_grant alone does not tell a client which gate refused. On a refusal under this section the AS SHOULD include, alongside error, the mission_error token-error-response member (Section 18) with one of the values mission_revoked, mission_expired, mission_superseded (where a companion defines supersession), or derivations_exhausted. The member is diagnostic only: it grants nothing, an unrecognized value is ignored, and it is returned only to the authenticated client presenting the Mission's grant. Derived tokens SHOULD be short-lived so that a transition to revoked or expired takes effect promptly without per-request revocation checks. 9.2. Revocation A Mission is revoked when the AS receives an authorized revocation for it. A deployment MUST provide an authenticated means for the Subject, the Approver, or an administrator to revoke a Mission by mission_id, independent of possession of any token (so a Mission can be stopped even when no refresh token is held). This document does not define the wire shape of that operation. Revocation is a management-plane action by a party in the AS's own trust domain, not a cross-party protocol exchange, and the lifecycle gate that makes it effective (Section 9, Section 10) already rides on existing endpoints, so no standardized endpoint is required for interoperability. A standardized Mission management API, with suspend/resume/complete operations, is specified separately by Mission Status [I-D.draft-mcguinness-oauth-mission-status]; this document does not require it. A deployment MAY additionally treat [RFC7009] revocation of a Mission's refresh token as revoking the Mission. A deployment MUST NOT couple routine token revocation to Mission revocation unless it documents that behavior. Already-issued access tokens remain valid until they expire; a deployment requiring lower cutoff latency SHOULD use short token lifetimes. The stateless baseline satisfies the lifecycle-gated capability: a token is a self-contained authorization, verification is stateless, and it needs no status surface. A deployment MAY additionally offer token introspection (Section 10), an OPTIONAL state-observable overlay, so a Resource Server can observe Mission state per request and cut off a revoked Mission without waiting out the token lifetime. A canonical Mission Status surface (keyed by mission_id) and signed status responses are specified separately as another OPTIONAL state- observable overlay by Mission Status [I-D.draft-mcguinness-oauth-mission-status]; this document does not require them. Token validity and Mission validity are distinct: a token can outlive a transition of its Mission, by at most the token lifetime. A deployment whose consumers rely on Mission state beyond a token's lifetime SHOULD offer introspection (Section 10) or the Mission Status companion, so an authorized party can determine the Mission's current state rather than inferring it from token validity. 10. Mission State via Token Introspection This section is OPTIONAL: it is one state-observable overlay on the lifecycle-gated baseline. The stateless baseline (Section 8) needs no introspection; an AS that does not offer it, and a Resource Server that does not use it, are unaffected. It lets a Mission-state-aware Resource Server observe a Mission's current state per request instead of waiting out a token's lifetime. Because it can report Mission state for a token whose Mission is no longer active, this section deviates from the SHOULD NOT of [RFC7662] Sections 2.2 and 4 against including additional information about an inactive token; that deviation is justified and governed by the caller-authorization and minimization rules below (Section 10.1). An AS MAY support OAuth 2.0 Token Introspection [RFC7662] for Mission-bound access tokens. When it does, the response for such a token carries, in addition to the standard members, a mission member: id, issuer, and authority_hash (as in the mission claim, Section 8.1) plus, when the responding AS is the Mission issuer, the current lifecycle state (string); when controls.max_derivations is in force, derivations_remaining (a number): the derivations left under the cap at the time of the response, counting committed issuances (Section 9), so a harness can plan refreshes against the budget; and, when the Mission records an authority proposal, proposal_hash (string): the Mission's proposal_hash (Section 7), surfaced for audit. Like state, only the issuer reports derivations_remaining and proposal_hash (Section 10.3). The core states are active, revoked, and expired (Section 9); a deployment that runs a companion profile defining an additional state reports that state here, and a consumer applies the forward-compatibility rule of Section 9 (only active permits reliance; any other value, recognized or not, is non-active). Only the issuer reports state (Section 10.3). The AS includes the mission member only when it has authenticated the caller, the caller is authorized for the token (Section 10.1), and the presented token resolves to a Mission. For a malformed, unknown, individually expired, or otherwise unresolvable token, the AS responds per [RFC7662] (active: false) with no mission member; it does not reveal Mission state for a token it cannot bind to a Mission. The case below (active: false with mission.state) applies only to a token that is itself valid but whose Mission is no longer active. The composite-active rule (Section 10.2) and the mission member apply equally when a Mission-bound refresh token is introspected. Freshness is per use: this document defines no caching semantics for the mission member, so a Resource Server that relies on introspection for Mission state treats each response as an observation for that decision, not as a cacheable state assertion. A deployment that needs bounded-staleness caching adopts the Mission Status companion, whose signed responses carry explicit freshness ([I-D.draft-mcguinness-oauth-mission-status]). 10.1. Caller Authorization and Minimization The introspection endpoint is protected per [RFC7662]. The AS: * MUST authenticate the calling party. * MUST return Mission data only to a caller authorized to receive it, in particular a Resource Server that is an audience of the token. * MUST audience-filter the response, returning the authorization_details entries and Mission data relevant to the caller's audience and not disclosing entries addressed to other audiences (Section 8). Because this profile returns the mission member and mission.state even when active is false (diverging from the [RFC7662] SHOULD NOT against disclosing detail for inactive tokens), the AS MUST apply this same authorization and minimization to that data and MUST NOT reveal Mission detail to an unauthorized introspection caller. 10.2. Composite Active State The introspection active member reflects the composite authorization, not the token in isolation. The AS MUST return active: true only when the access token is itself valid (valid signature, unexpired, and not individually revoked) AND the Mission is active. The AS does not verify the token's sender-constraint (cnf) at introspection: proof of possession is checked by the Resource Server when the token is presented, not by the AS over an introspection call, so active: true is not by itself evidence the caller holds the bound key. When the token is otherwise valid but the Mission is revoked or expired, the AS MUST return active: false and include mission.state giving the reason, so a Resource Server can distinguish a dead Mission from a bad token. (This deviates from the [RFC7662] SHOULD NOT against including members when active is false.) A Mission transition does not by itself revoke the token as an individual credential; introspection reports the composite authorization as inactive. 10.3. Only the Issuer Reports Mission State An AS MUST NOT include mission.state in an introspection response unless it holds the Mission, that is, unless it is the Mission issuer. Introspection at a non-issuer Resource AS, which returns the claim-shape members only and never state, is specified by the companion ([I-D.draft-mcguinness-oauth-mission-cross-domain]). This is token introspection: it answers "is this token's authorization still good," keyed by the token presented. The canonical Mission Status surface (keyed by mission_id) remains out of scope (Section 9.2). 10.4. Examples While the Mission is active, the response is the standard [RFC7662] body plus the mission member. The canonical ERP token (Section 8.1), introspected at the issuer AS: { "active": true, "iss": "https://as.example.com", "sub": "user_3p2q8mN1a0kV7tR", "client_id": "s6BhdRkqt3", "aud": "https://erp.example.com", "exp": 1797840300, "authorization_details": [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ], "mission": { "id": "msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-", "issuer": "https://as.example.com", "authority_hash": "sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ", "proposal_hash": "sha-256:kT2mR7vX4qL9nY5pB1sD8fJ6wZ3hC0aGeUoNvSqMrYo", "state": "active" } } The same token after the Mission is revoked, reported per the composite-active rule (Section 10.2): { "active": false, "mission": { "id": "msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-", "issuer": "https://as.example.com", "authority_hash": "sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ", "state": "revoked" } } 11. Delegation Within a Mission This section is OPTIONAL. A deployment whose agents never delegate, and a Resource Server that sees no delegated tokens, are unaffected. An agent may delegate execution to downstream actors (a sub-agent, service, or tool that is itself an OAuth client) within a Mission. Delegation is represented with the OAuth Actor Profile [I-D.draft-mcguinness-oauth-actor-profile], which profiles the RFC 8693 [RFC8693] act (actor) claim. A delegate obtains a delegated token by Token Exchange ([RFC8693]). The AS issues the delegated token subject to all of the following: * *The exchange is explicit.* The delegating Mission-bound access token is the subject_token, with subject_token_type of urn:ietf:params:oauth:token-type:access_token. The delegate is identified by an actor_token (with its actor_token_type; for example, an attested client-instance assertion presented as urn:ietf:params:oauth:token-type:client-instance-jwt ([I-D.draft-mcguinness-oauth-client-instance-assertion])) or by its own client authentication, and the AS asserts the actor itself (Section 11.3); a requested_token_type of urn:ietf:params:oauth:token-type:access_token is used. The response carries the matching issued_token_type and a token_type for the issued access token, per [RFC8693] Section 2.2.1. * *Subject is stable.* sub remains the Mission's Subject. The delegate is an actor, not the subject. * *client_id keeps its ordinary meaning.* A delegated token's client_id is the OAuth client that requested it, per [RFC8693] Section 4.3 and [RFC9068] Section 2.2; this profile does not override that definition. The Mission's originally-approved agent is no longer named by client_id on such a token; it remains recorded in the Mission Record (Section 7). * *The act chain identifies the delegates.* The delegated token carries an act claim per the Actor Profile [I-D.draft-mcguinness-oauth-actor-profile] and [RFC8693]: the outermost act is the current delegate, with each earlier delegate nested inside the previous actor's act member (act.act), back through the chain. Each actor object carries the members that profile defines (sub, iss, and the RECOMMENDED sub_profile actor- type classification, e.g. ai_agent). This document does not re- specify the act structure. * *Authority only narrows.* The delegated token's authorization_details MUST be a subset (Section 5.1) of the delegating token's authority, hence of the Mission Authority Set. Delegation MUST NOT add authority. * *The Mission binding rides unchanged.* The delegated token carries the same mission claim (Section 8.1), its id, issuer, and authority_hash, so every actor in the chain operates under the one consented authority. * *Each delegate is bound to its own key.* The delegated token MUST be sender-constrained (Section 8) to the *delegate's own* key: its cnf is the delegate's DPoP or mTLS key, not the delegating party's. The delegate proves possession of that key in the Token Exchange. A compromised delegate key therefore cannot be replayed as the agent or as another actor in the chain, and each actor's credential is independently revocable by key. * *Each delegation is gated.* Issuing a delegated token is a derivation event; the AS MUST refuse it unless the Mission is active (Section 9). Where a deployment authenticates client instances ([I-D.draft-mcguinness-oauth-client-instance-assertion]; for AI agents, its agent profile [I-D.draft-mcguinness-oauth-ai-agent-instance]), the delegate identified by the outermost act is the concrete instance: act.sub is the instance identifier and act.cnf is the instance-specific key. This profile's requirement that a delegated token be sender- constrained to the delegate's own key then lands on an instance- possessed key by construction. An allowed_delegates matcher can select instance-grade actors (Section 11.3), for example { "sub_profile": "client_instance" }. The sub_profile values used here (ai_agent, client_instance) are drawn from the entity-profiles vocabulary those instance profiles use; the Actor Profile [I-D.draft-mcguinness-oauth-actor-profile] remains the structural reference for the actor object. 11.1. Adopted Model: client_id Names the Requesting Client This profile keeps client_id's ordinary meaning under [RFC8693] Section 4.3, which [RFC9068] Section 2.2 incorporates: the OAuth client that requested the token, the immediate client, on every issued or derived token, including a delegated one. It does not override that definition for delegated tokens. The Mission's originally-approved agent is no longer carried in a token claim; it remains recorded in the Mission Record (Section 7), without redefining a registered claim. Downstream delegates are carried in the act chain (Section 11), whose outermost entry is the current delegate; earlier delegates nest inward through act.act. A design considered and rejected: freezing client_id to the Mission's approved agent on every derived token, overriding its meaning under [RFC8693] and [RFC9068]. That design kept the approved party at a fixed claim name, but only by having client_id silently mean something other than what [RFC9068] Section 2.2 registers it to mean: a generic [RFC9068] Resource Server or logging pipeline, which reads client_id as the immediate client because that is its registered meaning, would attribute a delegate's action to the approved agent instead of to the actor that performed it, with no error to surface the mismatch. It also bought no operational relaxation: it is safe only where every Resource Server processes the act chain anyway, since that is where the executing party's identity then lives, which is exactly the discipline this profile already requires of a Mission- aware Resource Server (Section 8.2). The Mission Record (Section 7) gives the same stable approved-agent binding without that conformance cost. The operational rule of Section 8.2 holds under the adopted model too, for a different reason: a Resource Server that authorizes or logs the caller on a token carrying an act chain MUST process that chain to see the delegation lineage, and a deployment MUST NOT route a delegated Mission-bound token to a Mission-unaware Resource Server that authorizes or logs on client_id alone (Section 8.2). Under the adopted model such a Resource Server attributes the immediate action correctly, since client_id is accurate, but it still cannot see the delegation lineage or the originally-approved agent, and so cannot apply actor-chain policy or join an actor's execution back to the Mission's approval. It is the mission claim's presence, not the binding choice, that signals a token may carry an act chain a consumer needs to process; a consumer that is not Mission-aware has no way to opt into or out of that need, which is why routing a delegated token to one remains forbidden. 11.2. Self-Exchange Down-Scoping An agent MAY present its own Mission-bound access token as the subject_token of a Token Exchange ([RFC8693]) with no actor, to obtain a narrowed token (for example, a single-audience one). The AS MUST verify that the client authenticated at a no-actor exchange is the Mission's approved agent, per the Mission Record's client_id (Section 7); any other party's no-actor exchange is refused, since a delegate narrows only through a delegated exchange that names it in the act chain. The result MUST be a subset (Section 5.1) of the presented token's authority, carries the same mission claim (Section 8.1), and adds no act chain. It is a derivation and is gated on the Mission being active (Section 9). Because it names no actor, it does not delegate: it re-scopes the agent's own authority downward. 11.3. Delegation Constraints What may be delegated, how far, and to whom is governed per Authority Set entry by the entry's optional delegation member (Section 5). Because the policy lives in the entry, it is committed by authority_hash with the rest of the Authority Set and travels with the entries wherever they are carried, including across a cross- domain projection ([I-D.draft-mcguinness-oauth-mission-cross-domain]), needing no separate mechanism. *Delegation depth.* The delegation depth of a token is the number of actors in its act chain (the nesting depth of the act claim), counted from the approved agent: the agent's own non-delegated token is depth 0, the first delegate is depth 1, and each further delegate adds 1. The depth checked against max_depth is that of the token being issued, computed after appending the new outermost actor, not the depth of the delegating token. A credential projected across a trust domain carries no act chain and enters the target domain at depth 0 ([I-D.draft-mcguinness-oauth-mission-cross-domain]). *Per-entry enforcement.* When the AS issues a token to a delegate (the actor that becomes the outermost act) at delegation depth d, it includes a Mission Authority Set entry in the delegated token's authorization_details only if all of the following hold: 1. the entry carries a delegation member (otherwise it is non- delegable, which is the default); 2. d is less than or equal to the entry's delegation.max_depth; and 3. the delegate is permitted by delegation.allowed_delegates or, when that member is absent, by the AS's delegation-authorization policy, which the AS MUST apply at every exchange: an absent matcher list is a decision deferred to policy, never a blanket grant of eligibility. An entry failing any of these narrows out of the delegated token, consistent with the subset rule (Section 5.1). The delegation member is policy, not authority, and is not part of the subset comparison; surviving entries are carried with their delegation member intact so the next hop is evaluated the same way. *Matching allowed_delegates.* Each entry is a matcher object modeled on the RFC 8693 may_act actor object ([RFC8693] Section 4.4): where may_act names a single party eligible to act on a token, allowed_delegates is a per-Authority-Set-entry _list_ of such matchers, generalized to actor-type classes. A { "sub": ... } matcher permits a specific delegate by client identifier; a { "sub_profile": ... } matcher permits any actor of that type (for example, ai_agent). An actor's sub_profile MAY carry multiple space- separated values; a { "sub_profile": ... } matcher is satisfied when its value is among the actor's values. A deployment can thus permit a specific client, a class of actors, or both, and a delegate is permitted if it matches any entry. The AS MUST authenticate the delegate at the Token Exchange and assert the actor's sub and sub_profile itself. A self-asserted sub_profile MUST NOT satisfy a matcher; otherwise a client could claim any actor type to bypass the constraint. A { "sub": ... } matcher is a client identifier in the issuing AS's namespace and is not portable across a trust domain; how a Resource AS evaluates conveyed matchers, failing closed and narrowing out any sub matcher it cannot resolve, is specified by the companion ([I-D.draft-mcguinness-oauth-mission-cross-domain]). *Empty result.* If narrowing leaves no entries for the delegate, the AS MUST refuse with invalid_target ([RFC8693] Section 2.2.2) rather than issue a token with empty authority: the requested delegation has no authority to carry, while the subject grant itself remains valid for other exchanges. *The Resource Server enforces none of this.* Delegation constraints are applied by the AS at issuance; a Resource Server sees only the already-narrowed authorization_details and enforces those as usual (Section 8). 11.4. Worked Example: Delegated Token Suppose the Mission's Authority Set has two entries on the ERP: invoices.read, delegable to ai_agent actors through depth 2; and journal-entries.write, which carries no delegation member and is therefore non-delegable. The approved agent s6BhdRkqt3 delegates to sub-agent tool-runner-7, an ai_agent, at depth 1. The read entry is permitted (depth 1 <= 2, ai_agent allowed) and the write entry narrows out. The decoded delegated access token: { "iss": "https://as.example.com", "sub": "user_3p2q8mN1a0kV7tR", "aud": "https://erp.example.com", "client_id": "tool-runner-7", "iat": 1797840600, "exp": 1797840900, "jti": "at_3qX5bN7sR1tY8mZ9Kp2v", "authorization_details": [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } } ], "act": { "sub": "tool-runner-7", "iss": "https://as.example.com", "sub_profile": "ai_agent" }, "cnf": { "jkt": "qVx7y2N0p4Lq9Md3sZJ8b8mZ3rN2xT5pV4lE6sQqYY" }, "mission": { "id": "msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-", "issuer": "https://as.example.com", "authority_hash": "sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ" } } sub is still the user. client_id is now tool-runner-7, the delegate that authenticated the Token Exchange and requested this token (Section 11.1); client_id and the outermost act coincide here because tool-runner-7 authenticated the exchange itself, not because this profile requires the two to match. A deployment where a different, already-authenticated client requests the exchange on a delegate's behalf, asserting the delegate only through an actor_token, would instead show that client's own identifier in client_id. client_id no longer names s6BhdRkqt3, the originally-approved agent; it remains recoverable from the Mission Record via mission_id (Section 7). The cnf is tool-runner-7's own key, not the agent's, so this token cannot be replayed as the agent. The non-delegable write entry was dropped; the read entry survives, carrying its delegation member so a further hop can be evaluated: a depth-3 delegate, or a non-ai_agent one, would narrow it out too. The rest of the mission claim is unchanged. 12. Extensibility This profile is a base layer that other agent-authorization work is expected to extend. Extensions build alongside the stable interface below; they MUST NOT redefine it. An extension MAY rely on these remaining stable across revisions of this profile: * the mission claim members id, issuer, and authority_hash (Section 8.1); * the mission_resource_access authorization details shape (Section 5); and * the act delegation chain (Section 11). The profile's extension points are each a declared seam rather than new machinery: * *Authority types.* The Authority Set is open to other AS-supported authorization_details types (Section 5.3); the Mission apparatus (commitment, gating, delegation) is type-agnostic toward them, subject to the delegation and projection limits in Section 5.3. * *Intent Submission Evidence types.* The evidence array of the Submission envelope is open to evidence types defined by companion profiles (Section 4.3): each type is a collision-resistant name whose owning specification defines the entry's closed schema, verification, and verified output facts, and an AS refuses an entry of a type it does not support rather than ignoring it. * *Integrity anchors.* Additional committed objects use the same domain-separated, issuer-bound envelope with a new typ (Section 6.3). A consent-disclosure commitment, an instruction- text attestation, or a delegation receipt can be committed this way without changing this profile. A profile that commits an evidence or disclosure object MUST commit it with this envelope and a collision-resistant typ, not by hashing the bare object, so the domain separation and issuer binding hold uniformly. A mission descriptor embedded in such an object uses the mission claim shape (Section 8.1), optionally extended with collision- resistantly named members (for example, an intent_hash for audit), and is never authority-bearing on its own. * *The mission claim.* It is an open object (Section 8.1): additional, collision-resistantly named members ride the mission binding (for example, a runtime decision reference, a delegation receipt, or an attestation reference), and consumers ignore unknown members and never derive authority from them. * *Lifecycle state.* The lifecycle state space (Section 9) is open to additional states defined by companion profiles for lifecycles they introduce. The forward-compatibility rule in Section 9 keeps this safe without a registry: only active permits issuance, and a consumer treats every other state, recognized or not, as non- active. * *Approval-event sequencing.* The approval-event steps, their order, and the atomicity of record creation with the approval decision are the model's (Section 6); the coupling of that decision to authorization-code issuance is this flow's. A companion profile MAY relocate the approval event relative to code issuance (for example, deferring the decision beyond the authorization response), provided the steps and their atomicity hold unchanged and no Mission reference exists before the record is active; the Mission Deferred Approval companion is such a profile ([I-D.draft-mcguinness-oauth-mission-approval]). This document defines no extension registry, capability-negotiation mechanism, or profile-version field; an extension declares its own identifiers and, where it needs discovery, its own metadata. The extensibility of the typ value space, the mission claim, and the lifecycle state space rests on collision-resistant naming and the fail-safe rules above rather than on central registration. 13. Authorization Server Metadata An AS MAY advertise support for this specification in its authorization server metadata [RFC8414]: mission_bound_authorization_supported: OPTIONAL boolean. When true, the AS supports the core Mission Issuer surfaces of this profile (Section 15): the mission_intent authorization request parameter through PAR (Section 4), derivation of mission_resource_access authorization details (Section 5), Mission-bound access tokens (Section 8), and the mission JWT claim (Section 8.1). It asserts Mission Issuer support only; it makes no claim about any Resource Server, nor about the OPTIONAL capabilities (delegation, introspection, cross-domain projection), which are discovered out of band or by attempt (Section 15). An AS that advertises this profile SHOULD also include mission_resource_access in its authorization_details_types_supported metadata ([RFC9396]), so that RFC 9396-aware clients discover the authorization details type through the standard mechanism. A client MAY use the RFC 9396 client metadata authorization_details_types at registration to declare the types it understands. An advertised type, mission_resource_access included, appears in authorization requests only as a proposal subject to derivation (Section 4.2): a client submits entries of advertised types on the authorization_details parameter alongside mission_intent, and the AS derives, narrows, or refuses under Section 5. The granted entries on issued tokens and echoes are issuer-derived, never the submission carried through by right. Discovery is OPTIONAL: a deployment MAY arrange Mission-bound authorization out of band, and this member only lets an AS advertise it. When the member is absent or false, a client MUST NOT infer that the AS supports this specification. A client holding a Mission Intent MUST NOT silently downgrade the task to an ungoverned authorization request against an AS whose support is not advertised and not otherwise established: submitting the same authority as bare scope or authorization_details obtains tokens no Mission governs, the client-side face of downgrade by omission (Section 16.3). The client surfaces the inability instead; where the estate's AS cannot change, the standalone Mission Issuer binding is the governed alternative ([I-D.draft-mcguinness-mission-authority-server]). An AS that advertises mission_bound_authorization_supported: true MUST also publish pushed_authorization_request_endpoint ([RFC9126]), since a Mission Intent is accepted only through PAR (Section 4.1). An AS that advertises mission_bound_authorization_supported: true MUST also advertise authorization_details_types_metadata_endpoint [I-D.draft-zehavi-oauth-rar-metadata]. That endpoint's response is a JSON object keyed by authorization_details type identifier, each value carrying, per [I-D.draft-zehavi-oauth-rar-metadata], a JSON Schema for exactly one authorization_details object of that type (schema or schema_uri) and optionally version, description, documentation_uri, and examples. That response's key set is the source of truth for which types the AS supports: authorization_details_types_supported ([RFC9396]), where the AS also advertises it, mirrors those keys and MUST NOT list a type absent from them. The AS MUST publish, within that response, an entry for mission_resource_access whose schema validates the object shape Section 18.3 defines, including the Common Constraints structure (Section 5.2); Section 18.3 remains the normative definition, and the published schema is its machine-readable form. Where a deployment arranges Mission-bound authorization out of band rather than advertising mission_bound_authorization_supported, the supported types and their schemas are likewise established out of band. This member and the mission_bound_authorization_required member of Section 14 are unauthenticated discovery data: their integrity rests on the metadata retrieval protections of [RFC8414] and [RFC9728], whose security considerations apply. 14. Protected Resource Metadata A protected resource MAY advertise, in its protected resource metadata [RFC9728]: mission_bound_authorization_required: OPTIONAL boolean. When true, the protected resource accepts only Mission-bound tokens: a token that lacks the mission claim (Section 8.1) is rejected (Section 8.2). When absent or false, the resource makes no such requirement. mission_constraints_supported: OPTIONAL. An array of strings: the constraints keys (Common Constraints and deployment-defined names) the resource understands and enforces (Section 8.2). It gives the AS's duty to emit only keys the serving resource is known to understand (Section 5) a discovery surface, and lets a client predict a fail-closed constraint mismatch before making the request. When absent, that knowledge is established out of band. 15. Conformance An implementation conforms in one of three roles. A *Mission Issuer* (the Authorization Server) implements the core issuance surfaces: * submission of a Mission Intent, in the Submission envelope, via PAR (Section 4.1), with the Intent Submission Evidence dispatch and refusal rules (Section 4.3); * derivation of mission_resource_access authorization details (Section 5); * the approval event with its integrity anchors and recorded approval_basis (Section 6, Section 7); * issuance of Mission-bound access tokens carrying the mission claim (Section 8); * the subset rule (Section 5.1); and * gating of issuance on Mission state (Section 9). A *Mission-aware Resource Server* implements Resource Server enforcement (Section 8.2). A *Mission Client* implements the client surfaces: * submission of the Mission Intent via PAR only (Section 4.1), proposing concrete authority, where it does, on the authorization_details parameter pushed alongside mission_intent (Section 4.2); * reading its granted authority from the token-response authorization_details echo (Section 8); and * obtaining mission_id from the mission_id token-response parameter or the mission claim's id (Section 6.1), treating it as a reference, never a credential. Beyond these mandatory roles, an implementation MAY additionally claim three OPTIONAL capabilities. Each is independent, and an implementation that supports none of them is still conformant: * *Delegation* (Section 11): issuing and consuming derived tokens that carry the act delegation chain. * *Introspection* (Section 10): reporting Mission state through the mission token introspection response member. * *Cross-Domain*: projecting a Mission so it is honored by an Authorization Server in another trust domain. An implementation claiming this capability preserves, across the hop: the Mission reference (mission.id, mission.issuer, authority_hash) carried intact; authority that only narrows (Section 5.1); projection performed only by, or under the authorization of, the Mission issuer, gated on the Mission's active state (Section 9); and projected credential lifetimes capped by the Mission's expires_at (Section 8). This bar is self-contained in this document; the companion Mission Cross-Domain Projection profile ([I-D.draft-mcguinness-oauth-mission-cross-domain]) specifies the interoperable mechanism that satisfies it, and implementations that interoperate across the hop implement that companion. A conforming implementation names the optional capabilities it supports (for example, "Mission Issuer with Delegation and Cross- Domain"); each capability's defining section or document states its detailed requirements. A token carrying a mission claim is not, by itself, Mission-bound authorization. Conformance as a Mission Issuer requires the gates: authority derived from the approved Intent and committed by the anchors (Section 6), issuance bounded by the subset rule (Section 5.1), and derivation gated on Mission state (Section 9). An implementation that carries Mission metadata without these gates conforms to no role in this document and does not implement it: in particular, it MUST NOT advertise mission_bound_authorization_supported as true (Section 13), the machine-checkable form of that claim. The mission_bound_authorization_supported metadata (Section 13) advertises Mission Issuer support only. It makes no assertion about any Resource Server, which does not advertise through Authorization Server metadata. The OPTIONAL capabilities are discovered first through existing OAuth metadata ([RFC8414]): introspection_endpoint for introspection, and grant_types_supported containing urn:ietf:params:oauth:grant-type:token-exchange for delegation and for the companion's cross-domain grant issuance. Absent such a signal, a capability is discovered out of band or by attempt: a Token Exchange, a cross-domain grant issuance, or an introspection request fails if the issuer does not support it. The smallest useful conforming deployment, noted here informatively, is a Mission Issuer that derives in narrowing mode from the client's authority proposal (Section 5), emits only the Common Constraints of Section 5.2, and implements none of the OPTIONAL capabilities; a scope-only Resource Server still operates at the coarse scope level (Section 8.2). This note names a starting point and creates no new conformance class. 16. Security Considerations 16.1. Consent Binding The security goal of this document is that a user's approval of a task bounds every token derived for it. The authority_hash commits the exact Authority Set the Approver consented to, and every derived token carries it in the mission claim. The complete mechanism also requires the AS to issue only subsets of that set and the Resource Server to verify the AS's token signature and enforce the carried authority. The hash alone does not prove containment of a narrowed token's authority. The requirements that uphold the commitment live at the approval event (Section 6): the AS computes authority_hash over the same Authority Set it rendered for consent, and re-renders and re-consents if that set changes. authority_hash commits the full Authority Set, while a derived token may carry a narrowed subset, so a Resource Server cannot in general recompute it from the token alone. Recomputation is optional, and its rules live with the Resource Server's other duties (Section 8.2); a Resource Server that does not hold the full set treats authority_hash as a whole-Mission audit and correlation anchor ([I-D.draft-mcguinness-oauth-mission-cross-domain]) while enforcing the token's authorization_details directly (Section 8). It relies on the signed token as the AS's assertion that the carried authority was correctly projected from the approved set; authority_hash supplies no independent subset proof. This document's flat Authority Set commitment defines no selective inclusion-proof mechanism: authority_hash digests the complete set as a single array (Section 6.3), recomputation therefore needs the full set, and the baseline defines neither a retrieval surface for that set nor a proof format for anything less. Both are undefined here, not impossible. Retrieval carries a real privacy and authorization burden rather than an architectural prohibition: the minimization rules of Section 10.1 keep other audiences' entries out of token introspection, so a distinct control-plane surface serving a policy decision point, auditor, or privileged Resource Server needs authorization at least that strong. A selective proof composes with the existing subset rules: a profile could commit the approved entries to a structure that supports inclusion proofs, prove the approved parent entry against that approval-time root, and apply the type-specific subset test (Section 5.1) between the carried narrowed entry and the disclosed parent; the cryptography stays generic, and only the semantic comparison is type-owned, the division this document uses throughout. Either mechanism is a future profile with privacy, commitment, and trust-bootstrap requirements to meet. Under the baseline, a deployment that needs assurance independent of the token signature provisions the verifying party out of band, the Resource Server itself or a policy decision point holding a materialized view of the Mission record, and uses the recomputation path of Section 8.2. What such verification buys depends on when the issuer is compromised. A proof or retrieval response issued under the same trust root as the token adds nothing against an issuer malicious at approval time: that issuer can approve and commit arbitrary authority, and no containment mechanism changes that. The same checks do defend against projection implementation errors, against corruption of the record after an independently anchored approval commitment, and against post-approval signing-key compromise where the original commitment is pinned outside the issuer. The pinning is what makes the difference; the worked example in Section 8.2 lists what a deployment provisions to get it. intent_hash extends the same protection to the task itself: it commits the approved Mission Intent, so an auditor can detect any later alteration of the recorded task, independently of the authority derived from it. proposal_hash, present when the client submitted an authority proposal (Section 4.2), commits what the agent asked for: approval-time provenance that lets an auditor compare the narrowed grant against the request that sought it. The anchors are domain- separated (Section 6.3); none is a substitute for another. The task and the authority are committed separately, rather than folded into one hash over the whole Mission, because they are distinct objects with distinct uses. authority_hash commits what a Resource Server enforces and what a cross-domain projection carries, so it MUST be verifiable from a token that conveys only the authority, without the Intent. intent_hash commits the task as audit material, tamper-evident even where the authority is projected without the Intent traveling with it. One combined hash could not serve both a token that carries authority alone and an auditor that holds the task alone. Neither anchor proves the Approver understood the rendered task, nor that the AS rendered it faithfully; they commit what the AS recorded, and make post-hoc tampering of those records detectable. This profile commits the task (intent_hash) and the authority (authority_hash) the Approver consented to, but deliberately does not commit the *rendered consent disclosure* itself: the locale, disclosure-template version, and material notices the Approver was shown are not bound by any anchor here. Because of this gap, a buggy or malicious rendering layer could mislead the Approver, showing a narrower or different task than the Authority Set actually committed, without leaving any committed trace. A deployment whose Missions carry high-risk authority SHOULD therefore record presentation-level audit evidence: for example, a hash over the exact consent disclosure rendered to the Approver, retained so the disclosure shown can be reconstructed and audited after the fact. Binding this on the wire (a consent_rendering_hash over a structured consent-disclosure object) is specified as an OPTIONAL companion profile by Mission Consent Evidence [I-D.draft-mcguinness-oauth-mission-consent-evidence]; an AS that does not implement it MAY record equivalent evidence out of band. As that profile makes explicit, such a commitment binds the structured disclosure the AS records, not the pixels actually presented; it narrows this gap for audit but does not close it. 16.2. Mission Drift Because issuance is gated on Mission state and bounded by the Authority Set, an agent cannot escalate beyond the approved task by acquiring additional tokens: every derived token is a subset of the approved authority (Section 5.1), and no token can be derived once the Mission is revoked or expired. Deployments SHOULD keep derived token lifetimes short so state transitions take effect promptly. 16.3. Downgrade by Omission A token bearing equivalent authorization_details but no mission claim is governed by no Mission state, revocation, or consent commitment. A deployment that designates a resource Mission-governed MUST NOT issue tokens for that resource outside a Mission, except under documented policy exceptions. The same rule has a per-client form: a deployment MAY register a client as Mission-governed, and an AS MUST reject a bare authorization_details request, one carrying no mission_intent, from a client so registered, so a governed client cannot strip the Intent from its submission to obtain ungoverned tokens; the client-side face of this duty is stated in Section 13. On the enforcement side, a Resource Server for such a resource rejects a token lacking the mission claim (Section 8.2), and MAY advertise the requirement through mission_bound_authorization_required (Section 14). 16.4. Prompt Injection and the Exfiltration Leg An agent that reads attacker-influenceable content can be prompt- injected; this profile assumes that and does not try to make the agent immune. Injection is dangerous when one agent combines access to private data, exposure to untrusted content, and the ability to communicate externally; the robust defense is architectural, constraining one of those, not making the model resistant. This profile constrains the data-access leg: a Mission narrows authority from everything the agent's standing credentials allow to the resources the approved task needs, and per-task Missions (Section 1.7) shrink the blast radius further. It contributes one thing against the untrusted-content leg: success_criteria is inert, granting, widening, and gating no authority, and goal and purpose shape authority only through the pre-approval derivation whose result the Approver reads and consents to (Section 4, Section 5); authority is fixed at the approval event, so injected text cannot talk an approved Mission into expanding itself. This profile does not constrain the external-communication leg and provides no information-flow control. It models authority over resources and actions, not how an agent uses authority it holds: within an approved Authority Set, an injected agent can read what the Mission permits and write to a sink the Mission permits, and the flat subset and constraint model cannot express "may read secrets, may write documents, but not write secrets into documents." Constraining exfiltration by a compromised agent is the runtime enforcement layer's role (Section 16.6), and even there it is bounded, not closed (see that profile's security considerations). Closing within-scope data laundering needs a separate taint or information-flow layer this profile does not define. 16.5. Authority Does Not Propagate With Information The Mission Drift consideration above bounds escalation by token acquisition: an agent cannot exceed the approved task by acquiring additional tokens. The same bound holds for information. No authority may be acquired by information propagation alone. An agent may inherit another agent's knowledge; it never inherits another agent's authority. Information may cross a boundary without authority crossing with it. A work product produced under one Mission, such as a file, message, memory entry, queue event, or other durable shared artifact, is input when an agent operating under another Mission reads it, not authority. The receiving Mission determines what may be done with the information under its own Authority Set (Section 5.1), and the producing Mission's authority does not transfer through the artifact by copying, referencing, embedding, or communicating it. An agent that needs authority to act on what it read acquires it only through an authorized derivation or delegation bounded by the Mission (Section 11), never from the artifact. This extends a possession-independence the profile already requires. Revocation acts on the mission_id independent of possession of any token (Section 9.2); here, authority is likewise independent of possession of any information. It is not information-flow control, which this profile does not provide (Section 16.4): the profile does not constrain what agents communicate, only what that communication can confer. Coordination between agents remains possible and cannot route around Mission authority. The threat is emergent authority through coordination. Multiple agents executing independently bounded work communicate through shared state, so discoveries, credentials, techniques, or intermediate results persist across runtimes and Missions, and individually acceptable actions compose into behavior that no single Mission authorized. This differs from a compromised or multiplied agent acting within one Mission's Authority Set: the composing units are independent Missions coordinating through a carrier outside any Mission's gate. The mechanism that upholds the invariant across such a carrier, work-product provenance and a non-transitive Mission-to- Mission handoff, is specified by Mission Work Products [I-D.draft-mcguinness-oauth-mission-work-products]; this document takes no normative dependency on it. 16.6. Issuance Scope, Not Runtime Enforcement This profile governs the issuance and derivation of authority: it bounds what authority a Mission yields, binds it to the Approver's consent, and gates derivation on Mission state. It does not evaluate individual runtime actions. In particular, it does not: * evaluate a request's parameters against the Mission at the point of use; * produce runtime enforcement evidence for each consequential action; * bind tool or function identities to the Mission; or * re-evaluate at execution time to close the approval-to-execution (time-of-check to time-of-use) gap. Mission governance is necessary but not sufficient. An active Mission still bounds a set of authority an agent may exercise freely within a token's lifetime, so an active Mission can become ambient authority for individual consequential actions. Preventing that requires a runtime enforcement layer that evaluates each consequential action against the Mission and records evidence; such a layer composes with this profile and is out of scope here. Which party enforces each Mission-carried bound is summarized in the enforcement table (Section 4). Short token lifetimes and narrow authority bound, but do not eliminate, this exposure. Where the Resource Server or a composing runtime layer does match a concrete request URI against a prefix entry, that match MUST apply the same RFC 3986 [RFC3986] normalization the AS uses for containment (Section 5.1, Section 8.2), so the boundary enforced at the point of use is the one the AS derived, not a divergent one. 16.7. Resource Boundary Canonicalization A prefix entry (Section 5) draws an authority boundary in URI space. The AS's containment test (Section 5.1) and the Resource Server's request matching (Section 8.2) MUST apply the single RFC 3986 [RFC3986] normalization defined in Section 5.1, so issuance and enforcement draw the same boundary. A matcher that normalizes differently from the party that derived the entry can admit a request the Approver never authorized, or deny one it did. The normalization MUST run before the prefix comparison, never after, and the boundary MUST be evaluated on the same normalized form a downstream component will act on. The following vectors each turn on such a difference: * *Encoded slash (%2F).* A reserved octet, left encoded by this normalization (which decodes only unreserved octets) and never a path-segment separator when matching. An intermediary that decodes it before the Resource Server enforces shifts the boundary; a deployment MUST prevent that rewriting or account for it. * *Encoded dot (%2e, %2e%2e).* The . octet is unreserved, so this normalization decodes it and then removes dot-segments. The hazard is order: normalizing before the comparison collapses a decoded %2e%2e out of the path and draws the boundary correctly, while matching the raw string first and letting a downstream component decode and collapse afterward can escape the prefix. * *Double or empty path segments.* An empty path and / denote the same base (Section 5); a matcher MUST NOT coalesce // or trailing- slash differences of its own, since the Section 5.1 normalization does not. * *Default-port presence.* https://a.example:443 and https://a.example are one authority only after the default-port removal Section 5.1 performs; a matcher that skips it splits one boundary into two. * *Unicode and IDNA host.* The Section 5.1 normalization lowercases the host but does not define Unicode to A-label (IDNA) conversion, so a deployment that accepts non-ASCII hosts MUST ensure both sides compare the same form. * *Percent-decoding order.* This normalization decodes only unreserved octets and runs before comparison; decoding reserved octets, or decoding after the comparison, changes which characters count as separators and can move the boundary. These vectors make the rule concrete. For a prefix entry whose resource is https://api.example/orders, a concrete request URI matches as follows: * https://api.example/orders and https://api.example/orders/123 ALLOW: the base itself, and an extension at a path-segment boundary. * https://API.EXAMPLE/orders/123 ALLOW: the scheme and host are case-folded, so the host matches. * https://api.example:443/orders/123 ALLOW: the default port is removed, so the authority matches. * https://api.example/Orders/123 DENY: the path is case-sensitive, so Orders is not orders. * https://api.example/orders%2F123 (equivalently %2f) DENY: the hex digits normalize to %2F, which stays encoded and is not a path- segment separator, so this is the single segment orders%2F123, not an extension of /orders. * https://api.example/ordersX DENY: the match extends only at a path-segment boundary, not within a segment. * https://api.example/orders/%2e%2e/admin DENY: normalization decodes %2e%2e to .. and removes the dot-segment, yielding https://api.example/admin, outside the prefix. A deployment MAY agree out of band on the canonicalization profile its AS and Resource Servers apply; this document defines one rule (Section 5.1), so no profile identifier is required for interoperation. 16.8. Denial Detail Disclosure The mission_denial attribute (Section 8.2) tells a caller which path a denial leads into, and thereby reveals authorization shape: step_up_required confirms to the presenting party that the authority exists, where insufficient_authority denies its existence. Introspection guards the same class of fact behind caller authorization (Section 10.1); a Resource Server applies the same care here. It SHOULD return the attribute only on a response to a validly signed, audience-correct token whose holder its deployment accepts learning the distinction, and SHOULD omit the attribute otherwise; when in doubt, insufficient_authority is the value that reveals least, and omission reveals nothing. 16.9. Token Theft Derived tokens are sender-constrained (DPoP [RFC9449] or mTLS [RFC8705]) at the levels set in Section 8 and Section 11: SHOULD for the primary access token, MUST for delegated tokens; the companion sets the same MUST for its cross-domain credentials ([I-D.draft-mcguinness-oauth-mission-cross-domain]). A stolen token is bounded by the Authority Set and the Mission lifetime regardless, but sender-constraint prevents replay by a different party. 16.10. Delegation and Chain Compromise Delegation (Section 11) widens the set of parties holding Mission- derived authority. Because authority only narrows down the chain, a compromised actor can act only within the authority it was delegated, for the lifetime of the token it holds: the exposure of any actor is bounded by its narrowed authorization_details times its token lifetime. The per-entry delegation constraints (Section 11.3) bound this exposure at approval time: non-delegable entries never reach a delegate at all, max_depth caps how far an entry can propagate, and allowed_delegates restricts who may receive it. Note that max_depth bounds the _length_ of a delegation chain, not its _breadth_: fan-out to many distinct depth-1 delegates is bounded only by allowed_delegates, so a deployment that needs to limit breadth MUST constrain allowed_delegates (and MAY use max_derivations to cap total derivations). Because each delegated token is bound to the delegate's own key (Section 11), a compromised delegate is confined to its own narrowed credential: it cannot replay as the agent or another actor, and its credential can be revoked by key without revoking the rest of the chain. Deployments SHOULD keep delegated token lifetimes short and SHOULD make only the entries that need delegation delegable. A distinct path deserves its name: audience replay into the exchange. A Mission-bound token obtained by or issued to one party, presented as a subject_token, would launder into a fresh delegated credential bound to the presenter. The gates above bound it: the exchange is a derivation gated on Mission state, the AS applies delegation- authorization policy at every exchange (Section 11.3), the result is bound to the authenticated delegate's own key and narrowed by the subset rule, and a no-actor exchange is accepted only from the Mission's approved agent (Section 11.2). Sender-constraining the primary token (Section 8) closes the remaining gap, since a token stolen from an audience then fails presentation at the token endpoint. 16.11. client_id Conformance and the Approved-Agent Residual Because this profile keeps client_id's ordinary [RFC9068] meaning (Section 11.1), a generic [RFC9068] Resource Server, or a logging, SIEM, or audit pipeline built for ordinary OAuth tokens, that keys identity and attribution on client_id attributes a Mission-derived token, delegated or not, to the correct requesting client without needing to understand this profile. What such a component cannot see is different from misattribution: it has no visibility into the delegation lineage carried in the act chain (Section 11), and it has no way to look up the Mission's originally-approved agent, which this profile does not carry in client_id and instead leaves recorded in the Mission Record (Section 7) at the issuer. A component that authorizes or logs on the approved agent's identity, or that must join a delegate's action back to the Mission's approval, MUST NOT assume client_id carries it. A deployment that delegates MUST route delegated Mission-bound traffic, including to logging and audit infrastructure, only through components that process the act chain (Section 8.2), and SHOULD review any existing component that authorizes or logs solely from client_id for this gap before exposing it to delegated Mission-bound tokens. 16.12. Signing and Key Rotation The mission claim and authorization_details are carried inside the [RFC9068] JWT and are covered by the AS's token signature; their integrity reduces to the AS's signing key. An AS MUST publish its verification keys (for example, via [RFC8414] jwks_uri), and the retired-key rule is: rotation retires a key from signing, never from resolvability within its retention bound. For this document's artifacts the floor is token lifetime: a verification key stays resolvable while tokens signed under it remain valid. Verification for audit outlives validity, so the key SHOULD remain resolvable at least as long as the audit horizon (Section 7) of any Mission whose tokens were signed under it. A companion that anchors a longer-lived artifact to the same keys (a status assertion, a Mandate, registered evidence) states its own retention bound as an extension of this same rule, not a new one. Revocation for a known or suspected compromise is distinct from routine retirement: the issuer publishes the compromised key as revoked, or marks it with a compromise time, rather than silently rotating it out. Key custody carries the model: a compromised issuer signing key voids every guarantee the signature carries. Issuer signing keys SHOULD be held in non-exportable or HSM/KMS-grade custody with dual-controlled generation, and SHOULD be segmented by artifact class under distinct kid values within the one jwks_uri, so that high-value, low-volume signing (long-lived evidence and portable artifacts) can sit under stricter custody than high-volume online token signing; verification is already kid-indexed, so the segmentation needs no wire change. Recovery from a signing-key compromise is deployment-owned and belongs in the deployment's documented procedures. 16.13. Compromised or Over-Broad Derivation The AS is trusted to derive authority no broader than the Mission Intent. Both derivation modes (Section 5) are mechanical, a proposal narrowed to policy or a configured mapping, rather than free-form inference, and the recorded policy_version names the policy a derivation ran under so the derivation can be audited. General OAuth security guidance [RFC9700] applies. 16.14. Authority Hash Is Not a Mission Identifier authority_hash commits the approved Authority Set, not the Mission. Two distinct Missions that approve byte-identical authority carry the same authority_hash: a successor Mission that re-approves the same Authority Set, or an unrelated Mission with the same derived authority, differs in its intent_hash, approver, and id while sharing the authority_hash. It is therefore not globally unique to a Mission and MUST NOT be used as a Mission Identifier or as a replay or idempotency key for a Mission. The canonical Mission Identifier names the Mission; authority_hash names the authority the Mission approved. A consumer that needs to bind to or correlate a specific Mission uses the Mission Identifier, and intent_hash and approver distinguish Missions that share an Authority Set. 16.15. Composition and the Effective Ceiling Delegation depth (Section 11.3) resets to 0 at each cross-domain hop ([I-D.draft-mcguinness-oauth-mission-cross-domain]) and, where a deployment runs the child-delegation profile, at each child generation ([I-D.draft-mcguinness-oauth-mission-child-delegation]). max_derivations (Section 4) is a per-Mission bound the issuer AS enforces for that Mission alone; a Child Mission's own max_derivations is independent of its parent's, and the parent's cap does not bound the child subtree by default. The aggregate surface a Mission's descendants can reach, the product of delegation depth, the number of trust domains projected into, and the number of child generations, together with the derivations summed across an entire child subtree, can therefore exceed what a single approval appears to bound at consent time. This is a composition property of independently-bounded mechanisms, not a defect in any one of them. A deployment SHOULD disclose the composed bound, not only the immediate Mission's, at the consent surface, and MAY impose a global cap out of band where a single approval's apparent bound must hold in practice. Bounding aggregate consumption (calls, spend, or activity over the life of a Mission and its descendants) is the metering profile's role ([I-D.draft-mcguinness-mission-metering]), not a property this document or its composition partners provide by themselves. 16.16. The Containment Materialized-Capability Residual Where a deployment runs the Mission Containment profile ([I-D.draft-mcguinness-oauth-mission-containment]), containment narrows an Authority Set entry's authorization to derive going forward and propagates to Child Missions justified by that entry, per that profile. It does not reach back into authority already materialized before the containment transition: a cross-domain grant already redeemed at a Resource AS ([I-D.draft-mcguinness-oauth-mission-cross-domain]), or an offline attenuation root already minted and now attenuating on its own outside the issuer's reach, continues to operate for its own remaining lifetime. This is the same residual bound that revocation carries (Section 9.2): a party that already holds a materialized credential is bounded by that credential's own lifetime, not by a state change at the issuer it can no longer observe in time. A deployment that needs containment to take effect quickly against already-materialized authority SHOULD keep cross-domain grant and offline attenuation root lifetimes short, so the residual window is one the next lease or re-mint closes. 17. Privacy Considerations A Mission Identifier is a correlation handle: a deployment limits exposure by giving stable Mission Identifiers only to parties that enforce, audit, or observe that Mission, preferring audience-scoped projections of authority where possible, and minimizing status and introspection disclosures to authorized callers. The subsections that follow and the introspection minimization rules (Section 10.1) give the specific rules. An Intent Submission Evidence artifact can carry personal data (an originator identity, a consent reference); PAR keeps it off the front channel, and the record retains the designated verified facts under the same access governance as the rest of the Mission's evidence (Section 17.5). 17.1. Mission Identifier Correlation This document carries a single canonical Mission Identifier on every derived token; the companion's cross-domain projection carries it across trust domains unchanged ([I-D.draft-mcguinness-oauth-mission-cross-domain]). Any party that observes credentials for the same Mission, whether a Resource Server, a Resource AS, or an auditor spanning audiences, can correlate that activity by the Mission Identifier, and mission.issuer further identifies the issuing AS. This is intentional: a stable, correlatable Mission anchor is what lets a Resource Server, a cross- domain Resource AS, and an auditor bind credentials and evidence to one approved Mission, which the governance and audit properties of this document and its companion profiles depend on. The cost is that this profile does not provide cross-audience unlinkability, and that is a deliberate non-goal for this version (Section 1.9), not an unfinished feature. Audience-pairwise (or request-pairwise) Mission references, in which the issuer projects a distinct opaque identifier per audience and resolves them server-side, are the fuller mechanism for unlinkability; because they work against the stable anchor, they are future work rather than a v1 property. A deployment that carries the canonical Mission Identifier on the wire accepts this correlation as part of its privacy posture; the operative control is limiting who receives the stable identifier, per the guidance above, and a deployment that publishes its posture records the property there rather than treating documentation as a separate duty. 17.2. Token Payload Disclosure The carried constraints and a multi-resource Authority Set disclose the shape of the task and its business bounds (for example, an amount ceiling) to every holder and every audience of a derived token. Per- RS single-audience tokens are the minimization measure: they carry only the entries the consuming Resource Server needs, and this document recommends them (Section 8). The privacy considerations of [RFC9396] apply to the carried authorization_details. 17.3. Intent Retention and Anchor Disclosure The Mission record's Intent members (goal, constraints) are personal- data sinks: they carry whatever task description the user supplied. Record retention and erasure follow the deployment's retention policy, with the audit profile's erasure record as the transparency- side mechanism ([I-D.draft-mcguinness-mission-audit]). The integrity anchors are unsalted commitments: a party holding a candidate Intent can confirm it against intent_hash, so over low-entropy or guessable content the anchor is a disclosure channel, and deployments treat it as one when the Intent itself is sensitive; the same confirmation channel exists for a candidate proposal against proposal_hash. 17.4. Third-Party Data Subjects A task can be about a person who holds no Mission role: in a background check, the employer's agent queries a registrar about a candidate who is neither Subject nor Approver nor resource owner. Mission approval records the accountable Approver's authorization of the undertaking (Section 6); it is not, by itself, evidence of such a person's consent to disclosure or of any other legal basis a disclosure requires, and whether a basis is required and what satisfies it is deployment and legal policy outside this protocol. Where the resource domain requires data-subject consent or another basis, that disclosure policy MUST be satisfied in that domain's own lane, wherever the deployment evaluates it (the Resource Server, a gateway, a policy decision point, or an authorization server acting for the domain) and through that lane's mechanisms (claims gathering, a resource-domain consent artifact), and access MUST be refused while required evidence is absent or invalid. A conforming deployment treats that refusal as the resource's answer, never as a Mission gap to route around. Mission approval and Mission authority MUST NOT be treated as the data subject's consent; a Mission record MAY retain a verified consent reference or facts as submission_evidence (Section 7), and those facts are provenance and policy input only: the resource domain validates them independently under its current disclosure policy, and that policy remains authoritative. Third-party personal data can enter through any Intent, proposal, authority, or recorded-evidence member: the prose members (goal, constraints, success_criteria), purpose, resources, deployment- defined controls, any type-owned member of a proposed entry, the derived Authority Set entries that reach tokens (Section 8), and submission_evidence facts, including a consent reference. Whatever the member, it persists on the Mission record for its audit horizon (Section 7) and concentrates at the AS with the record (Section 17.5), and the committed members are confirmable through the unsalted anchors by any party holding a candidate value (the retention consideration above). Clients SHOULD reference a third party through resource-scoped or pseudonymous identifiers rather than identifying prose; an opaque identifier is minimization, not anonymity, and personal-data obligations follow it. Retention and erasure follow the deployment's policy. Where approval-event evidence was registered under the audit transparency profile, its erasure record and data-subject-request basis are the transparency- side mechanism ([I-D.draft-mcguinness-mission-audit]): it records an erasure, it neither performs one nor overrides retention law, and the operational Mission record and its audit-horizon retention floor (Section 7) are untouched by it. 17.5. Mission Record and Evidence Access The Mission record concentrates the task, its authority, and its principals at the AS, and every evidence artifact joins on the Mission Identifier, so the join is a correlation surface equal to the identifier itself. Tokens carry references and authority, never the record: nothing in this profile puts goal, constraints, or other Intent content in a credential. Access to the record and to Mission evidence is policy-governed and auditable: reading a Mission's evidence is a privileged operation, not a byproduct of holding a Mission reference, and retention is deployment policy bounded below by the audit horizon (Section 7). 18. IANA Considerations 18.1. OAuth Parameters Registration This document registers the following in the "OAuth Parameters" registry: * Name: mission_intent * Parameter Usage Location: authorization request * Change Controller: IESG * Specification Document(s): this document, Section 4.1 * Name: mission_id * Parameter Usage Location: token response * Change Controller: IESG * Specification Document(s): this document, Section 6.1 * Name: mission_error * Parameter Usage Location: token response * Change Controller: IESG * Specification Document(s): this document, Section 9 PAR [RFC9126] carries authorization-request parameters without a distinct usage location, so the pushed submission of mission_intent needs no separate registration. The mission_error member rides the token-endpoint error response; "token response" is the registry's applicable usage location, and the member relies on the error response's JSON extensibility rather than defining a new error code, so generic [RFC6749] error handling is undisturbed. The mission_denial attribute rides the WWW-Authenticate scheme's extensible auth-param space ([RFC6750], Section 8.2), for which no IANA registry exists; no action is required for it. 18.2. OAuth Extensions Error Registration This document registers the following in the "OAuth Extensions Error" registry [RFC6749]: * Name: invalid_mission_intent_evidence * Usage Location: authorization endpoint, token endpoint * Protocol Extension: Intent Submission Evidence (Section 4.3) * Change Controller: IESG * Specification Document(s): this document, Section 4.3 The error is returned where the containing exchange returns its errors: on a PAR submission, in the PAR error response; on a token- endpoint carriage defined by a companion profile, in the token error response. 18.3. The Mission Resource Access Authorization Details Type mission_resource_access is an authorization_details type per [RFC9396] Section 2, defined by this document in Section 5. RFC 9396 does not establish an IANA registry of authorization details types (type identifiers are interpreted by the Authorization Server), so this document creates no registry entry for it and requires no IANA action here. If a registry of authorization details types is established in the future, this type SHOULD be registered in it. 18.4. JSON Web Token Claims Registration This document registers the following in the "JSON Web Token Claims" registry: * Claim Name: mission * Claim Description: Reference to the Mission a token was derived under, with the consent-commitment authority_hash. An open object; additional members may be present and are ignored if unknown. * Change Controller: IESG * Specification Document(s): this document, Section 8.1 18.5. OAuth Token Introspection Response Registration This document registers the following in the "OAuth Token Introspection Response" registry ([RFC7662]): * Name: mission * Description: The Mission a token was derived under. Same object shape as the mission JWT claim (Section 8.1); a response from the Mission's issuer additionally carries a state member giving the current lifecycle state, where a derivation cap is in force a derivations_remaining member, and where the Mission records an authority proposal a proposal_hash member (Section 10). * Change Controller: IESG * Specification Document(s): this document, Section 10 18.6. OAuth Authorization Server Metadata Registration This document registers the following in the "OAuth Authorization Server Metadata" registry ([RFC8414]): * Metadata Name: mission_bound_authorization_supported * Metadata Description: Boolean indicating that the Authorization Server supports the Mission Issuer core surfaces of this document. * Change Controller: IESG * Specification Document(s): this document, Section 13 18.7. OAuth Protected Resource Metadata Registration This document registers the following in the "OAuth Protected Resource Metadata" registry ([RFC9728]): * Metadata Name: mission_bound_authorization_required * Metadata Description: Boolean indicating that the protected resource accepts only Mission-bound tokens. * Change Controller: IESG * Specification Document(s): this document, Section 14 * Metadata Name: mission_constraints_supported * Metadata Description: JSON array of the constraints keys the protected resource understands and enforces. * Change Controller: IESG * Specification Document(s): this document, Section 14 18.8. Common Constraints Registry This document establishes the "Mission Common Constraints" registry. The registration policy is Specification Required [RFC8126]. A Designated Expert reviews a submission for the discipline Section 5.2 requires: a name matching ^[A-Za-z0-9_.:-]+$ not already registered, a JSON [RFC8259] value syntax precise enough for independent implementations to agree on, and a subset rule and an intersection rule both stated in value-space terms, with the intersection of any two valid values never broader than either operand. Registration does not require IETF review or a Standards Track document; a Specification Required reference that a Designated Expert can review against these criteria suffices. Each registration records: * *Key Name*: the constraints member name. * *Value Space*: the JSON value type and any additional syntax rules. * *Subset Rule*: how a candidate value is judged no broader than a reference value. * *Intersection Rule*: how two values for the key combine into a result no broader than either operand. * *Change Controller*: IETF, or the registrant for any other registration. * *Reference*: the specification defining the key. This document populates the registry with the Common Constraints it defines (Section 5.2); the "Ref" column below is this document, at the section shown, for every row, and the full Subset Rule and Intersection Rule text is there, not restated in the table: +========================+===========+============+==========+=======+ |Key Name |Value Space|Subset / |Controller|Ref | | | |Intersection| | | +========================+===========+============+==========+=======+ |max_amount |Object: |Same |IETF |Section| | |amount |currency, | |5.2 | | |(decimal |candidate <=| | | | |string), |reference; | | | | |currency |intersection| | | | |(ISO 4217) |is the | | | | | |smaller | | | | | |amount | | | +------------------------+-----------+------------+----------+-------+ |resource_issued_after |String, RFC|Candidate >=|IETF |Section| | |3339 date- |reference; | |5.2 | | |time |intersection| | | | | |is the later| | | | | |instant | | | +------------------------+-----------+------------+----------+-------+ |resource_issued_before |String, RFC|Candidate <=|IETF |Section| | |3339 date- |reference; | |5.2 | | |time |intersection| | | | | |is the | | | | | |earlier | | | | | |instant | | | +------------------------+-----------+------------+----------+-------+ |tenant |String |Equal; |IETF |Section| | | |intersection| |5.2 | | | |is the | | | | | |shared value| | | +------------------------+-----------+------------+----------+-------+ |recipient_domain |String, DNS|Equal or a |IETF |Section| | |name |subdomain; | |5.2 | | | |intersection| | | | | |is the | | | | | |narrower | | | | | |value | | | +------------------------+-----------+------------+----------+-------+ |time_window |Object: |Candidate |IETF |Section| | |not_before,|window | |5.2 | | |not_after |within | | | | |(RFC 3339 |reference | | | | |date-time) |window; | | | | | |intersection| | | | | |is the | | | | | |overlap | | | +------------------------+-----------+------------+----------+-------+ |data_classification |Array of |Candidate |IETF |Section| | |strings |array a | |5.2 | | | |subset of | | | | | |reference | | | | | |array; | | | | | |intersection| | | | | |is the | | | | | |common | | | | | |members | | | +------------------------+-----------+------------+----------+-------+ |allowed_tools |Array of |Candidate |IETF |Section| | |strings |array a | |5.2 | | | |subset of | | | | | |reference | | | | | |array; | | | | | |intersection| | | | | |is the | | | | | |common | | | | | |members | | | +------------------------+-----------+------------+----------+-------+ |requires_action_approval|Boolean |Candidate |IETF |Section| | | |true or | |5.2 | | | |equal to | | | | | |reference; | | | | | |intersection| | | | | |is the | | | | | |logical OR | | | +------------------------+-----------+------------+----------+-------+ Table 3 Names are kept collision-free by the convention Section 5.2 already uses: a specification-defined name is coordinated through this registry, and any other name is either collision-resistant or remains deployment-defined and outside the registry. 19. References 19.1. Normative References [I-D.draft-zehavi-oauth-rar-metadata] Zehavi, Y., "OAuth 2.0 RAR Metadata and Error Remediation", Work in Progress, Internet-Draft, draft- zehavi-oauth-rar-metadata-06, 9 August 2026, . [ISO4217] International Organization for Standardization, "ISO 4217:2015, Codes for the representation of currencies and funds", ISO 4217:2015, August 2015. [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, March 1997, . [RFC3339] Klyne, G. and C. Newman, "Date and Time on the Internet: Timestamps", RFC 3339, DOI 10.17487/RFC3339, July 2002, . [RFC3986] Berners-Lee, T., Fielding, R., and L. Masinter, "Uniform Resource Identifier (URI): Generic Syntax", STD 66, RFC 3986, DOI 10.17487/RFC3986, January 2005, . [RFC6234] Eastlake 3rd, D. and T. Hansen, "US Secure Hash Algorithms (SHA and SHA-based HMAC and HKDF)", RFC 6234, DOI 10.17487/RFC6234, May 2011, . [RFC6749] Hardt, D., Ed., "The OAuth 2.0 Authorization Framework", RFC 6749, DOI 10.17487/RFC6749, October 2012, . [RFC6750] Jones, M. and D. Hardt, "The OAuth 2.0 Authorization Framework: Bearer Token Usage", RFC 6750, DOI 10.17487/RFC6750, October 2012, . [RFC7519] Jones, M., Bradley, J., and N. Sakimura, "JSON Web Token (JWT)", RFC 7519, DOI 10.17487/RFC7519, May 2015, . [RFC7636] Sakimura, N., Ed., Bradley, J., and N. Agarwal, "Proof Key for Code Exchange by OAuth Public Clients", RFC 7636, DOI 10.17487/RFC7636, September 2015, . [RFC7662] Richer, J., Ed., "OAuth 2.0 Token Introspection", RFC 7662, DOI 10.17487/RFC7662, October 2015, . [RFC7800] Jones, M., Bradley, J., and H. Tschofenig, "Proof-of- Possession Key Semantics for JSON Web Tokens (JWTs)", RFC 7800, DOI 10.17487/RFC7800, April 2016, . [RFC8174] Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, May 2017, . [RFC8259] Bray, T., Ed., "The JavaScript Object Notation (JSON) Data Interchange Format", STD 90, RFC 8259, DOI 10.17487/RFC8259, December 2017, . [RFC8414] Jones, M., Sakimura, N., and J. Bradley, "OAuth 2.0 Authorization Server Metadata", RFC 8414, DOI 10.17487/RFC8414, June 2018, . [RFC8693] Jones, M., Nadalin, A., Campbell, B., Ed., Bradley, J., and C. Mortimore, "OAuth 2.0 Token Exchange", RFC 8693, DOI 10.17487/RFC8693, January 2020, . [RFC8705] Campbell, B., Bradley, J., Sakimura, N., and T. Lodderstedt, "OAuth 2.0 Mutual-TLS Client Authentication and Certificate-Bound Access Tokens", RFC 8705, DOI 10.17487/RFC8705, February 2020, . [RFC8707] Campbell, B., Bradley, J., and H. Tschofenig, "Resource Indicators for OAuth 2.0", RFC 8707, DOI 10.17487/RFC8707, February 2020, . [RFC8785] Rundgren, A., Jordan, B., and S. Erdtman, "JSON Canonicalization Scheme (JCS)", RFC 8785, DOI 10.17487/RFC8785, June 2020, . [RFC9068] Bertocci, V., "JSON Web Token (JWT) Profile for OAuth 2.0 Access Tokens", RFC 9068, DOI 10.17487/RFC9068, October 2021, . [RFC9101] Sakimura, N., Bradley, J., and M. Jones, "The OAuth 2.0 Authorization Framework: JWT-Secured Authorization Request (JAR)", RFC 9101, DOI 10.17487/RFC9101, August 2021, . [RFC9126] Lodderstedt, T., Campbell, B., Sakimura, N., Tonge, D., and F. Skokan, "OAuth 2.0 Pushed Authorization Requests", RFC 9126, DOI 10.17487/RFC9126, September 2021, . [RFC9207] Meyer zu Selhausen, K. and D. Fett, "OAuth 2.0 Authorization Server Issuer Identification", RFC 9207, DOI 10.17487/RFC9207, March 2022, . [RFC9396] Lodderstedt, T., Richer, J., and B. Campbell, "OAuth 2.0 Rich Authorization Requests", RFC 9396, DOI 10.17487/RFC9396, May 2023, . [RFC9449] Fett, D., Campbell, B., Bradley, J., Lodderstedt, T., Jones, M., and D. Waite, "OAuth 2.0 Demonstrating Proof of Possession (DPoP)", RFC 9449, DOI 10.17487/RFC9449, September 2023, . [RFC9700] Lodderstedt, T., Bradley, J., Labunets, A., and D. Fett, "Best Current Practice for OAuth 2.0 Security", BCP 240, RFC 9700, DOI 10.17487/RFC9700, January 2025, . [RFC9728] Jones, M.B., Hunt, P., and A. Parecki, "OAuth 2.0 Protected Resource Metadata", RFC 9728, DOI 10.17487/RFC9728, April 2025, . 19.2. Informative References [AuthZEN.ARAP] OpenID Foundation, "OpenID AuthZEN Access Request and Approval Profile 1.0", 2025, . [FAPI.GrantManagement] OpenID Foundation, "Grant Management for OAuth 2.0", 2022, . [I-D.draft-cecchetti-oauth-rar-cedar] Cecchetti, S., "Cedar Profile for OAuth 2.0 Rich Authorization Requests", Work in Progress, Internet-Draft, draft-cecchetti-oauth-rar-cedar-02, 21 February 2024, . [I-D.draft-ietf-oauth-transaction-tokens] Tulshibagwale, A., Fletcher, G., and P. Kasselman, "Transaction Tokens", Work in Progress, Internet-Draft, draft-ietf-oauth-transaction-tokens-11, 30 July 2026, . [I-D.draft-klrc-aiagent-auth] Kasselman, P., Lombardo, J., Rosomakho, Y., Campbell, B., Steele, N., and A. Parecki, "AI Agent Authentication and Authorization", Work in Progress, Internet-Draft, draft- klrc-aiagent-auth-03, 6 July 2026, . [I-D.draft-mcguinness-mission-approval-governance] McGuinness, K., "Mission Approval Governance", 2026, . [I-D.draft-mcguinness-mission-architecture] McGuinness, K., "An Architecture for Mission-Bound Authorization", 2026, . [I-D.draft-mcguinness-mission-audit] McGuinness, K., "Mission Audit Transparency", 2026, . [I-D.draft-mcguinness-mission-authority-server] McGuinness, K., "Mission Authority Server", 2026, . [I-D.draft-mcguinness-mission-authzen] McGuinness, K., "Mission-Bound Runtime Enforcement: AuthZEN Profile", 2026, . [I-D.draft-mcguinness-mission-metering] McGuinness, K., "Mission Consumption Metering", 2026, . [I-D.draft-mcguinness-mission-runtime] McGuinness, K., "Mission-Bound Runtime Enforcement", 2026, . [I-D.draft-mcguinness-mission-shaping] McGuinness, K., "Mission Intent Shaping", 2026, . [I-D.draft-mcguinness-mission-substrate] McGuinness, K., "Mission Substrate Requirements", 2026, . [I-D.draft-mcguinness-oauth-actor-profile] McGuinness, K., "OAuth Actor Profile for Delegation", Work in Progress, Internet-Draft, draft-mcguinness-oauth-actor- profile-00, 30 April 2026, . [I-D.draft-mcguinness-oauth-ai-agent-instance] McGuinness, K., "OAuth 2.0 AI Agent Instance Profile", Work in Progress, Internet-Draft, draft-mcguinness-oauth- ai-agent-instance-00, 4 July 2026, . [I-D.draft-mcguinness-oauth-client-instance-assertion] McGuinness, K., "OAuth 2.0 Client Instance Assertion", Work in Progress, Internet-Draft, draft-mcguinness-oauth- client-instance-assertion-01, 24 June 2026, . [I-D.draft-mcguinness-oauth-mission-approval] McGuinness, K., "Mission Deferred Approval for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-child-delegation] McGuinness, K., "Mission Child Delegation for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-consent-evidence] McGuinness, K., "Mission Consent Evidence for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-containment] McGuinness, K., "Mission Containment for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-cross-domain] McGuinness, K., "Mission Cross-Domain Projection for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-expansion] McGuinness, K., "Mission Expansion for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-signals] McGuinness, K., "Mission Lifecycle Signals for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-status] McGuinness, K., "Mission Status and Lifecycle for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-template] McGuinness, K., "Mission Template for OAuth 2.0", 2026, . [I-D.draft-mcguinness-oauth-mission-work-products] McGuinness, K., "Mission Work Products", 2026, . [I-D.draft-niyikiza-oauth-attenuating-agent-tokens] Aimable, N., "Attenuating Authorization Tokens for Agentic Delegation Chains", Work in Progress, Internet-Draft, draft-niyikiza-oauth-attenuating-agent-tokens-01, 15 June 2026, . [MCP] Model Context Protocol Project, "Model Context Protocol: Authorization", 2025, . [RFC7009] Lodderstedt, T., Ed., Dronia, S., and M. Scurtescu, "OAuth 2.0 Token Revocation", RFC 7009, DOI 10.17487/RFC7009, August 2013, . [RFC8126] Cotton, M., Leiba, B., and T. Narten, "Guidelines for Writing an IANA Considerations Section in RFCs", BCP 26, RFC 8126, DOI 10.17487/RFC8126, June 2017, . [RFC8935] Backman, A., Ed., Jones, M., Ed., Scurtescu, M., Ansari, M., and A. Nadalin, "Push-Based Security Event Token (SET) Delivery Using HTTP", RFC 8935, DOI 10.17487/RFC8935, November 2020, . [RFC9493] Backman, A., Ed., Scurtescu, M., and P. Jain, "Subject Identifiers for Security Event Tokens", RFC 9493, DOI 10.17487/RFC9493, December 2023, . [RFC9635] Richer, J., Ed. and F. Imbault, "Grant Negotiation and Authorization Protocol (GNAP)", RFC 9635, DOI 10.17487/RFC9635, October 2024, . Appendix A. End-to-End Example (Non-Normative) This appendix walks one Mission from an agent through Mission creation, token issuance, and Resource Server enforcement in a single trust domain. It is illustrative and adds no normative requirements. The OAuth pieces use the rules in this document; the identity setup is by reference to [I-D.draft-klrc-aiagent-auth]. Identifiers and hash values are illustrative and are not computed from the displayed JSON. This walkthrough is the baseline issuance path: stateless enforcement bounded only by token lifetime. No stage calls back to the AS for Mission state; each party enforces from the credential it holds. Stage 3 notes where the OPTIONAL runtime layer adds a point-of-use check. Scenario: agent s6BhdRkqt3, acting for alice (user_3p2q8mN1a0kV7tR), reconciles Q3 invoices in the home ERP under Mission msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-. A.1. Stage 0: Agent Identity (by Reference) The agent is an OAuth client with a workload identity (for example a WIMSE or SPIFFE identity), and alice has delegated to it through an ordinary authorization-code flow, per [I-D.draft-klrc-aiagent-auth]: client_id is the agent and the token sub is alice. This document adds the Mission layer on top of that identity; Stage 0 is otherwise unchanged from that specification. A.2. Stage 1: Mission Creation The agent submits this Submission envelope through PAR (Section 4.1), carrying the Mission Intent and no evidence, and proposing concrete authority alongside it on the authorization_details parameter (Section 4.2): { "intent": { "goal": "Reconcile Q3 invoices and post adjustments under $500.", "resources": ["https://erp.example.com"], "constraints": [ "Read only invoices issued in 2026-Q3.", "Post journal entries under $500." ], "success_criteria": [ "All Q3 invoices reconciled.", "Each posted adjustment references a source invoice." ], "purpose": "urn:example:purpose:reconcile", "expires_at": "2026-12-31T23:59:59Z", "controls": { "acr": "urn:example:acr:mfa", "max_derivations": 200 } } } The submitted authority proposal, on authorization_details in the same push: [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.*"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ] The AS (as.example.com) validates both, derives this Authority Set (each entry a same-type subset of a proposed entry, Section 4.2), and renders it for alice's consent: [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ] After approval, the AS records Mission msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9- in the active state with authority_hash sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ, intent_hash sha-256:wQ7p4LHnX9Md0LqJ6sZJ8b8mZ3rN2xT5pV4lE6sQqYY, and proposal_hash sha-256:kT2mR7vX4qL9nY5pB1sD8fJ6wZ3hC0aGeUoNvSqMrYo. A.3. Stage 2: Mission-Bound Token Issuance The agent redeems the authorization code at the token endpoint. The AS resolves the Mission from the grant (Section 6.1), gates on it being active (Section 9), and issues a Mission-bound access token for the ERP, echoing the granted authorization_details in the token response (Section 8). The decoded token: { "iss": "https://as.example.com", "sub": "user_3p2q8mN1a0kV7tR", "aud": "https://erp.example.com", "client_id": "s6BhdRkqt3", "iat": 1797840000, "exp": 1797840300, "jti": "at_9Kp2vN7sR1tY8mZ3qX5b", "authorization_details": [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "constraints": { "resource_issued_after": "2026-07-01T00:00:00Z", "resource_issued_before": "2026-09-30T23:59:59Z" }, "delegation": { "max_depth": 2, "allowed_delegates": [{ "sub_profile": "ai_agent" }] } }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ], "cnf": { "jkt": "0ZcOCORZNYy-DWpqq30jZyJGHTN0d2HglBV3uiguA4I" }, "mission": { "id": "msn_8RfX2Lqv9TqMv4z7sA2bN1k0YpEdHc9-", "issuer": "https://as.example.com", "authority_hash": "sha-256:l3KvZ4mP5x0wQrR6tY2nD9bM7sX1cF8gH2vJ4kE5pNQ" } } Everything enforcement needs is in the token: the audience, the sender-constraint (cnf), the authority with its constraints, and the mission claim carrying the authority_hash consent anchor. The token is short-lived (300 s) and its exp is far below expires_at; revoking the Mission stops further derivation, and this token dies at its own expiry (Section 9.2). A.4. Stage 3: The Resource Server Enforces The agent calls the ERP Resource Server (erp.example.com) with that token. The Resource Server validates the JWT and the cnf binding and enforces the authorization_details whose resource it serves, permitting invoices.read within the Q3 issuance window and journal- entries.write up to the max_amount ceiling of 500.00 USD (Section 8.2). It treats the mission claim as audit and correlation context and makes no call to the AS. This is stateless enforcement from the token alone. journal- entries.write is a consequential write, so where the deployment runs the runtime profile ([I-D.draft-mcguinness-mission-runtime]) it also obtains a point-of-use PDP permit against current Mission state before executing. The baseline bounds the write only by token lifetime and the carried constraints. The cross-domain continuation of this same Mission, projected to a partner ERP in another trust domain and enforced there, is walked through in the companion's end-to-end example ([I-D.draft-mcguinness-oauth-mission-cross-domain]). Appendix B. Integrity Anchor Test Vectors These non-normative vectors let an implementation verify its anchor computation (Section 6.3, Section 6.4) byte for byte. All use the issuer https://as.example.com. Each canonical-bytes block is the exact JCS [RFC8785] output: a single line, UTF-8, with no whitespace outside string values. It is shown here wrapped only for layout; remove the layout line breaks, adding no characters, to recover the canonical form. Note that JCS sorts object member names (so iss precedes typ precedes value, within an entry actions precedes constraints precedes resource precedes type, and within max_amount amount precedes currency) and preserves array order. intent_hash, over this Mission Intent as the envelope value with typ mission-intent: { "goal": "Reconcile Q3 invoices", "resources": ["https://erp.example.com"], "expires_at": "2026-12-31T23:59:59Z" } Canonical bytes of the envelope: {"iss":"https://as.example.com","typ":"mission-intent","value":{"e xpires_at":"2026-12-31T23:59:59Z","goal":"Reconcile Q3 invoices"," resources":["https://erp.example.com"]}} intent_hash = sha-256:6mIFoCz79uCHNzKLfBpBwqFjoFXdpmpuc65486IqimQ authority_hash, over this Authority Set as the envelope value with typ mission-authority-set: [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"] }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "500.00", "currency": "USD" } } } ] Canonical bytes of the envelope: {"iss":"https://as.example.com","typ":"mission-authority-set","val ue":[{"actions":["invoices.read"],"resource":"https://erp.example. com","type":"mission_resource_access"},{"actions":["journal-entrie s.write"],"constraints":{"max_amount":{"amount":"500.00","currency ":"USD"}},"resource":"https://erp.example.com","type":"mission_res ource_access"}]} authority_hash = sha-256:vUCCfjGulit9u0qJ0Z6pQSNerZtXMqRlfJNCr4PzLro The third pair exercises the two nesting shapes the pair above does not: an Intent controls object with nested members, and an Authority Set entry whose delegation.allowed_delegates is an array of matcher objects, where JCS sorts each object's members but preserves the array's order (the sub_profile matcher stays before the sub matcher). intent_hash, over this Mission Intent as the envelope value with typ mission-intent: { "goal": "Reconcile Q3 invoices", "resources": ["https://erp.example.com"], "expires_at": "2026-12-31T23:59:59Z", "controls": { "acr": "urn:example:acr:mfa", "max_derivations": 20 } } Canonical bytes of the envelope: {"iss":"https://as.example.com","typ":"mission-intent","value":{"c ontrols":{"acr":"urn:example:acr:mfa","max_derivations":20},"expir es_at":"2026-12-31T23:59:59Z","goal":"Reconcile Q3 invoices","reso urces":["https://erp.example.com"]}} intent_hash = sha-256:DHUg4zS3HHnWtXlO6hu9sTN_jX4LyjZ4tOJiTDAvWAI authority_hash, over this Authority Set as the envelope value with typ mission-authority-set: [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.read"], "delegation": { "max_depth": 2, "allowed_delegates": [ { "sub_profile": "ai_agent" }, { "sub": "s6BhdRkqt3" } ] } } ] Canonical bytes of the envelope: {"iss":"https://as.example.com","typ":"mission-authority-set","val ue":[{"actions":["invoices.read"],"delegation":{"allowed_delegates ":[{"sub_profile":"ai_agent"},{"sub":"s6BhdRkqt3"}],"max_depth":2} ,"resource":"https://erp.example.com","type":"mission_resource_acc ess"}]} authority_hash = sha-256:notrA9wZaP3I5Gx8UzN0mfzUjHYPeX4Ri_B3ilh7BbA The last vector exercises the third anchor. proposal_hash, over this submitted authorization_details proposal as the envelope value with typ mission-proposed-authority (Section 4.2): [ { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["invoices.*"] }, { "type": "mission_resource_access", "resource": "https://erp.example.com", "actions": ["journal-entries.write"], "constraints": { "max_amount": { "amount": "1000.00", "currency": "USD" } } } ] Canonical bytes of the envelope: {"iss":"https://as.example.com","typ":"mission-proposed-authority" ,"value":[{"actions":["invoices.*"],"resource":"https://erp.exampl e.com","type":"mission_resource_access"},{"actions":["journal-entr ies.write"],"constraints":{"max_amount":{"amount":"1000.00","curre ncy":"USD"}},"resource":"https://erp.example.com","type":"mission_ resource_access"}]} proposal_hash = sha-256:udzftXYQy0pvYNxz4KgtmyL_EV8ry4DhIbBFfwILEBA An implementation that canonicalizes the same value under the same typ and iss, computes SHA-256, and encodes as sha-256: followed by base64url with no padding (Section 6.3) reproduces these anchors exactly. A divergence indicates a JCS or encoding difference to resolve before interoperating. Appendix C. Document History [[ To be removed from the final specification ]] -01 * Breaking change to the Intent carriage shape: the mission_intent parameter value is the Mission Intent Submission envelope (Section 4.1), intent plus an OPTIONAL typed evidence array, and the bare-Intent value is refused under the closed envelope. Anchors are stable: intent_hash commits exactly the intent object, so committed values, recorded anchors, and the test vectors are unchanged. The Intent Submission Evidence hook (Section 4.3) adds type-dispatched verification that rejects unknown or failing entries, the policy-input-never-authority rule, the required- evidence anti-downgrade rule, evidence invalidation across shaping and revision, the presenter conjunction, verification-cost bounds, the invalid_mission_intent_evidence error registration, and the submission_evidence record member as record-trusted provenance metadata. * Breaking change to the authority-proposal carriage: the proposal moves from the Intent's proposed_authority member, which is removed, to the standard top-level authorization_details parameter pushed alongside mission_intent (Section 4.2), and the old prohibition on submitting the two together inverts. Anchor inputs changed: intent_hash no longer covers the authority proposal, and the new proposal_hash (typ mission-proposed-authority) commits the submitted proposal, recorded on the Mission and surfaced through introspection, never on the mission claim. Worked examples and test vectors are recomputed; an approval-event rule recomputing every recorded anchor and a Mission-governed-client bare-request rejection accompany the change. * Derivation is mechanical, in two modes: narrowing (RECOMMENDED) and template. The deterministic-reproducibility and policy- inspectability rules are retired, policy_version stays as an opaque audit correlator, generative derivation is demoted to a local-policy extension, and the derivation trust boundary (no portable Intent semantics) is stated. Template mode publishes its mapping space as deployment documentation and distinguishes no- mapping from policy refusals. * New wire surface: the mission_error and derivations_remaining introspection members, the mission_denial WWW-Authenticate attribute, and the mission_constraints_supported protected- resource metadata member. The mission claim gains an OPTIONAL expires_at; the mission_id token-response parameter is promoted to SHOULD. * Authority vocabulary: four new Common Constraints (time_window, data_classification, allowed_tools, requires_action_approval) with prefix-match fixes; the OPTIONAL controls.agent_deployment class pin; max_derivations sizing guidance. * Delegation hardening: delegation-authorization policy applies at every exchange, allowed_delegates is RECOMMENDED and its absence is never a blanket grant, self-exchange is accepted only from the Mission's client_id, and the delegated-token routing guardrail is raised to MUST NOT. * Model precision: the Mission-referenced, Mission-derived, and Mission-bound token-class taxonomy; anchors as independent commitments with authority_hash an audit correlator at a narrowed- token Resource Server; the subset rule named representational, not semantic; approval-event sequencing named as an extensibility seam; Intent carriage closed to the form-encoded PAR parameter. * Operational and privacy: verification-key retention anchored to the audit horizon, issuer-key custody and per-artifact-class kid guidance, a Mission-record and evidence-access privacy section, and extended token lifetimes scoped to fully runtime-gated tokens with single-audience narrowing as the splitting mechanism. * Editorial throughout, no normative change: requirement lists for the PAR submission and rendering duties, twice-stated rules reduced to one home, a third integrity-anchor test vector, and registry hygiene. * External review resolutions: GNAP and capability-system prior art positioned in Relationship to Other Authorization Objects; the mission claim named as the actor-freezing wire signal, with the Resource Server routing guardrail and a new Security Considerations subsection on client_id misattribution; RFC 6750 and RFC 9700 promoted to normative and the OAuth Actor Profile reference demoted to informative; an actual Common Constraints IANA registry replacing the future-revision deferral; and new Security Considerations on composition's effective ceiling and the containment materialized-capability residual. -00 * Initial individual draft. Acknowledgments This work builds on the OAuth 2.0 Rich Authorization Requests, Pushed Authorization Requests, and JWT access token specifications, and is intended to complement agent-identity work including [I-D.draft-klrc-aiagent-auth]. Author's Address Karl McGuinness Independent Email: public@karlmcguinness.com