Structured Context · Draft
Morphz Structured Context Specification v1
Status: Draft specification candidate
Steward: Newvar
Reference implementation: Morphz Runtime
Canonical language: English
Source baseline: Morphz Context Protocol v32 and the 2026-08-15 Runtime status index
Date: 2026-08-21
Chinese translation: zh-CN
1. Scope
This specification defines the normative data model, authority boundaries, transaction behavior, provenance, attention lifecycle, and recovery properties of Morphz Structured Context.
It deliberately separates the public standard from the current serialization and code structure. Protocol v32 is the implementation source used to prepare this candidate; the public specification will receive its own semantic version and MUST NOT inherit every internal field merely because Morphz Runtime currently contains it.
2. Normative language
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, RFC 2119 and RFC 8174, when, and only when, they appear in all capitals.
Examples, rationale, and implementation notes are non-normative unless explicitly identified otherwise.
3. Core entities
3.1 Agent
An Agent is a stable logical cognitive actor. It is not identical to a model process, Provider request, Session, Principal, or operating-system process.
An Agent MUST have a stable identifier. Replacing its model or restarting its Runtime MUST NOT implicitly replace its identity.
3.2 Context
A Context is a first-class, versioned cognitive state owned or used by an Agent. A Context MUST contain or address:
- an immutable Event History;
- a Runtime-owned Kernel projection;
- an Agent-owned Mind projection;
- delivered Inbox or Observation state;
- Session and Attention projections;
- a monotonically ordered sequence of committed Context transactions.
A Context MUST expose a stable identifier and current revision.
3.3 Principal
A Principal is an authenticated or authorized external actor or authority, such as a person, organization, service, or delegated identity. A Principal MUST have a stable identifier within its declared authority domain. Authorization decisions MUST refer to an explicit Principal or an equivalent identity whose scope is observable and auditable.
A Principal MUST NOT be silently substituted for an Agent, Context, or Session identity.
3.4 Session
A Session is a stable interaction connection mounted to a Context. It MUST have its own identity and MUST NOT be used as a synonym for Context, Agent, or Principal.
Multiple Sessions MAY share a Context. Requests for different Sessions MAY execute concurrently, subject to the causal visibility and transaction rules in this specification.
3.5 Event
An Event is an immutable record of an occurrence. It MUST have:
- a stable identifier;
- an authoritative sequence or equivalent ordering coordinate;
- a topic or type;
- an actor or authoritative source;
- enough routing identity to determine its Context and, when applicable, Session;
- direct causal references when the Runtime knows them.
An implementation MUST NOT edit a committed Event in place. Corrections and superseding facts are new Events.
3.6 Observation and Inbox
An Observation is the Agent-visible representation of an Event or Runtime fact. Inbox is the delivery area for Observations that remain available for cognitive processing.
An Observation MUST retain a stable path back to its source Event. Rendering an Observation as a preview, metadata-only entry, or recalled chunk MUST NOT change the original Event.
3.7 Kernel
Kernel is the Runtime-owned projection of authoritative operating facts, including identity, permissions, active execution, budgets, pressure, versions, and control state. Kernel MUST be read-only to the Agent except through explicit Runtime commands.
3.8 Mind
Mind is the Agent-owned cognitive projection. It consists of Frames and Relations and MAY contain arbitrary domain structure inside Frame bodies.
The Runtime MUST understand only the structural metadata needed to validate, version, order, protect, relate, project, and recover Frames. It MUST NOT require a universal business ontology for Frame bodies.
3.9 Frame
A Frame is a stable cognitive unit. A Frame MUST have at least:
- an identifier unique within its Context;
- a revision;
- an Agent-authored body;
- a lifecycle state;
- optional protection state;
- optional source references.
Revising a Frame MUST preserve its identity and increment its revision. Historical Frame bodies MUST remain recoverable from committed transaction facts when a profile claims recovery, even when the active Projection contains only the latest body.
3.10 Relation
A Relation is an explicit edge between stable identifiers. Relation names are open by default. Implementations MUST NOT infer standard business meaning from a Relation unless this specification or an extension profile defines it.
3.11 Projection
A Projection is a derived current-state view over authoritative history. Event History answers what happened; a Projection answers what the current state is.
Projection corruption MUST be recoverable from authoritative records or MUST fail visibly. A Runtime MUST NOT silently invent missing history to repair a Projection.
3.12 Attention
Attention is an explicit Projection or lifecycle state that determines ordering, residency, or visibility for a particular consumer or evaluation. Attention MUST NOT be represented as physical deletion or semantic truth. When an Attention distinction affects available content, the Runtime MUST make that distinction observable.
3.13 Evaluation and Causal Scope
An Evaluation is one bounded Runtime decision or execution cycle for an active Session. A Causal Scope is the explicit set or lineage of local evidence, pending results, and authority available to an Evaluation.
Every Evaluation MUST identify its Session and Causal Scope, or provide observably equivalent behavior. Internal names such as Activation or Thread are implementation details and are not required by this specification.
4. Authority matrix
| State or decision | Agent authority | Runtime authority |
|---|---|---|
| Frame body and cognitive meaning | create and interpret | persist and validate structure |
| Semantic importance and abstraction | decide | MUST NOT assign silently |
| Evidence interpretation | decide | preserve declared source references |
| Event identity, order, and direct cause | interpret | generate and enforce |
| Principal authentication and authorization | request or present credentials | establish, scope, and enforce |
| Permissions and resource limits | observe and obey | define and enforce |
| Context transaction intent | submit | validate, commit, reject, and audit |
| Active Attention | request changes | apply transactionally and project |
| Physical tool result | interpret | record faithfully |
| Current Projection | read and modify through allowed operations | derive from authoritative facts |
5. Context transaction model
5.1 Transaction envelope
A Context transaction MUST declare a base Context revision or an equivalent concurrency token. It MAY include an audit reason and MUST include one when retiring protected or active information, or removing protection.
The Runtime MUST:
- parse the complete transaction before mutation;
- resolve stable references against the authorized Context;
- validate identity, lifecycle, permission, source, and concurrency constraints;
- compute changes against an isolated candidate state;
- commit all authoritative Events and affected Projections atomically;
- return the committed revision and structured result, or a structured rejection;
- leave the previous state intact when the transaction is rejected.
5.2 Core operations
SC-Core defines the following semantic operations:
| Operation | Required semantic effect |
|---|---|
create |
create a new Frame with a stable identifier |
derive |
create a Frame with explicit declared sources |
revise |
replace the active body of an existing Frame while preserving identity |
retire |
remove a Frame or Observation from active semantic Attention without deleting history |
restore |
return a retired Frame or Observation to active semantic Attention |
protect / unprotect |
enable or remove Runtime-enforced retirement protection |
place |
change Attention order without changing Frame meaning |
relate / unrelate |
add or remove an explicit Relation |
retire-session / restore-session |
change Session Attention membership without deleting Session history |
Morphz Runtime also implements checkpoint, rollback, and drop-checkpoint. Their portable
semantics and profile membership are reserved for a later MEP and are not SC-Core requirements.
The concrete wire syntax MAY use the Morphz context_tx S-expression DSL. Independent
implementations MAY expose another API if it produces equivalent observable behavior and passes the
claimed conformance profile.
5.3 Concurrency
An implementation MUST reject a stale transaction when it cannot prove the change is independent of intervening commits.
An implementation MAY safely rebase changes to independent Frames after validating read and write sets. It MUST reject at least:
- concurrent incompatible changes to the same Frame revision;
- creation of the same stable Frame identifier with different content;
- a derivation whose declared source changed in a way that invalidates the submitted read set.
Any extension that defines a Context-wide lifecycle operation MUST also define an adequate Context-level fence. Conflict handling MUST be observable. Silent last-writer-wins behavior is non-conformant.
6. Provenance and reality contract
6.1 Epistemic rationale (non-normative)
Presence in Inbox does not make an Observation true. Recency and usage do not establish semantic authority. A newer physical version does not automatically prove a broader semantic conclusion. An Agent may carry hypotheses or incorrect beliefs while the Runtime continues to preserve what physically occurred.
6.2 Runtime requirements (normative)
The Runtime MUST:
- distinguish Runtime facts from Agent-authored conclusions in its observable representation;
- prevent a source from becoming visible before it physically exists and is authorized for the active Causal Scope;
- preserve declared evidence lineage for
deriveand source-bearingreviseoperations; - preserve the actual source and transaction history even when Agent conclusions are incorrect;
- avoid labeling an Observation semantically authoritative solely because it is recent, frequent, highly used, or the latest physical version.
7. Attention, residency, and recovery
Implementations MAY use full, preview, metadata-only, recalled, resident, swapped-out, or equivalent rendering states. Each state MUST be distinguishable to the consumer when the distinction affects available content.
An implementation MUST NOT:
- represent a preview as the complete original;
- equate exclusion from one model request with semantic retirement;
- equate semantic retirement with physical deletion;
- discard the stable source needed for explicit recall when it claims the item is recoverable.
Resource pressure MAY trigger a Runtime signal or a mechanical Projection decision. It MUST NOT silently author an Agent conclusion or opaque semantic summary.
8. Session and causal visibility
Each Evaluation MUST identify its active Session and Causal Scope. Context-wide committed Mind state MAY be visible across authorized Sessions, while local Session evidence MUST respect its Causal Scope.
A late tool result MUST resume or signal the Causal Scope that created it. It MUST NOT be delivered as if it belonged to an unrelated Session merely because that Session is currently active.
Cross-Session messages or signals MUST identify source and destination explicitly. Referencing a Session MUST NOT implicitly import its transcript, activate it, or copy its private evidence.
9. Durability and recovery
For a durable conformance profile:
- committed Context transactions MUST survive a clean restart;
- a crash before commit MUST leave no partial authoritative mutation;
- a crash after commit MUST allow the same current state to be reconstructed;
- rebuilding Projections MUST reproduce the same observable Context revision and active state;
- retries MUST be idempotent where the protocol exposes a stable request or transaction identity.
10. Canonical representation (reserved)
This Draft does not define a canonical wire representation. Byte-for-byte fixture equality is therefore not an active conformance requirement.
Before Candidate status, v1 MUST define ordering, escaping, stable identifiers, optional fields, and version negotiation for its canonical representation. The current Morphz Runtime renderer is an implementation source, not the frozen public wire standard. Fields that exist only for current scheduling or Provider optimization SHOULD remain implementation extensions unless portable consumers require them.
11. Conformance profiles
The v1 candidate reserves these profiles:
- SC-Core: object model, authority boundaries, transaction semantics, provenance, and Attention;
- SC-Durable: SC-Core plus restart, replay, idempotent stable-identity retries, and Projection recovery;
- SC-Concurrent: SC-Durable plus conflict detection, causal routing, and concurrent Session work;
- SC-Distributed: SC-Concurrent plus multi-Runtime fencing, leases, and cross-process recovery.
The exact required cases are defined by the matching Conformance Suite release. Morphz Runtime does not claim public certification for these profiles until the standalone suite is extracted and a signed report is published.
12. Versioning and extensions
The public specification uses semantic versions independently of internal Context Protocol numbers.
- Patch releases clarify text or add tests without changing conformant behavior.
- Minor releases add backwards-compatible optional behavior or profiles.
- Major releases may change required observable behavior and MUST provide a migration statement.
An extension MUST use a namespaced identifier, declare its required base version, and fail visibly when required semantics are unavailable. An extension MUST NOT redefine a core term while claiming compatibility with the unchanged core version.
13. Open decisions before Candidate status
The following decisions remain intentionally open:
- canonical wire representation and negotiation fields;
- the minimal portable Event and Projection schemas;
- the profile, if any, for checkpoint and rollback operations;
- the precise boundary between retirement and future residency or swap semantics;
- compatibility rules for Frame-level rebase across independent implementations;
- normative privacy and Principal visibility profiles;
- final intellectual-property, contribution, and compatibility-mark policies.
Each decision requires an MEP or an explicitly recorded specification review before v1 Candidate.
14. Security considerations
Conforming implementations operate across trust boundaries. They MUST treat Agent-authored Context transactions, external Observations, tool results, recalled content, and cross-Session messages as potentially untrusted input.
At minimum, an implementation MUST:
- authenticate or otherwise establish the Principal for protected operations and authorize every Context, Session, source, recall, and cross-Session access;
- scope stable references so that forged or guessed identifiers cannot bypass authorization;
- preserve Event and transaction integrity, detect invalid replay where a stable request identity exists, and prevent concurrency from widening a Causal Scope;
- enforce explicit resource limits so denial-of-service pressure fails visibly without silently rewriting semantic state;
- prevent secrets from entering Agent-visible Context unless that disclosure is explicitly authorized;
- redact credentials and secrets from conformance reports, diagnostics, and exported evidence.
Implementations SHOULD document their trust boundaries, credential lifecycle, retention policy, and recovery assumptions. Conformance to this specification is not a certification of an implementation's overall security.
15. Intellectual-property status
This Draft is licensed and governed by the active IPR Status Notice. Apache-2.0 provides its stated copyright and patent grants; trademark, compatibility-mark, and certification rights remain separate. Implementers MUST NOT infer a broader standards-essential patent commitment than the rights actually granted by Apache-2.0.
16. Errata and interpretations
Suspected errors or ambiguities MUST be recorded through the public issue and MEP process described in MEP-0001. Editorial errata MAY clarify language without changing conformant behavior. Any interpretation that changes required observable behavior, profile membership, or compatibility results requires a Standards Track MEP and a versioned Specification or Suite update.