EP-0024: Unified Frame Identity And Lifecycle¶

Status: final Type: standards-track

Accepted 2026-06-18: this EP records the operator's rulings on rf2-kz2vfp, rf2-uc6ebw, and rf2-um1jcq; all open issues below are now resolved. The EP defines the post-EP-0023 frame cleanup: one live frame value backed by one registry, one public operation-target grammar, and one explicit UI-owned lifecycle boundary (rf/frame-provider). Normative homes are spec/002-Frames.md, spec/API.md, spec/001-Registration.md, spec/Runtime-Subsystems.md, and spec/Conventions.md.

Abstract¶

EP-0023 made image -> frame -> event stream the public model. The current implementation still realizes a frame through two live concepts: an image-loaded frame object and a separate backing frame record. That split leaks upward as multiple constructors, multiple registries, multiple target spellings, and unclear ownership rules for view-created frames.

This EP proposes one live frame value and one live frame registry. The resolved image generation is stored on the frame itself. Public frame-scoped operations target a frame through the event/query plus opts-map grammar, usually {:frame frame-id} or an established ambient frame context. Lifecycle ownership is separate: create/destroy frames with lifecycle APIs, not with dispatch or subscribe target coercions.

Motivation¶

The frame model is the isolation boundary for app-db, runtime-db, event drains, subscription caches, effect routing, history, and tooling. Users should not need to learn whether a call wants a frame id, a frame object, a handle, a provider context, or a constructor side effect.

The API-review findings show one structural cause:

there are two frame creation paths;
there are two live-frame registries;
the object form points back to the backing record for most stateful work;
Story frames sometimes have to create both halves in a load-bearing order;
teardown walks both structures;
target normalization exists largely to collapse object-shaped inputs back to the id used by the backing record.

This is larger than a facade trim. It changes the framework's public frame contract and the implementation substrate that supports hot reload, Story, tests, SSR, and tooling. It therefore belongs in an EP rather than a bead.

Goals / Non-Goals¶

Goals:

make a live frame one value backed by one registry;
store the resolved image generation on that frame value;
make make-frame the single public constructor for image-loaded frames and ordinary configured frames;
keep frame ids as the public routing address for dispatch, subscribe, reads, providers, and tooling;
keep frame values as lifecycle tokens returned by creation APIs;
make view-owned frame lifetime explicit with one UI boundary, rf/frame-provider, that creates, provides, and destroys;
serve scoping to an existing frame with rf/with-frame (lexical / non-React) and rf/frame-provider-existing (scope an existing frame into a React subtree — a dynamic var cannot cross React's render boundary). The old scope-only frame-provider becomes the new rf/frame-provider-existing; the frame-provider name is repurposed for the owned lifecycle boundary;
keep frame-handle as the callback carry primitive;
remove or retier public spellings that duplicate the target/carry/lifecycle roles;
make registrar generation reads name their data source honestly if they remain public.

Non-goals:

do not revisit the EP-0023 image model itself;
do not reintroduce retired composition vocabulary as current public API;
do not remove frames as the isolation boundary;
do not make every in-render call pass {:frame ...} when a real provider or lexical frame context exists;
do not preserve compatibility shims for off-pattern pre-alpha spellings;
do not decide unrelated facade cleanups from the API-review wave.

Relationships¶

EP-0023 established image-loaded frames and partially superseded the older composition teaching surface. This EP is an amendment-style follow-up that finishes the frame collapse in the implementation and public grammar.
EP-0002 established explicit frame target resolution and the carried-frame invariant. This EP keeps that invariant and narrows the spelling of explicit target operations.
EP-0001 established app-db/runtime-db as frame-owned partitions. The unified frame remains the owner of both partitions.
EP-0007 supplies the one-name-per-fact rule. This EP applies it to frame id, frame value, frame handle, and resolved image generation.
EP-0013 remains historical and partially superseded at the public surface. This EP must not revive its retired public vocabulary.
EP-0009 governs this as a new standards-track EP rather than a silent edit to final EP-0023. If accepted, this EP records the amendment and graduates into the named specs.
API-review findings:
ai/findings/API-review/codex/frame-object-record-unification.md
ai/findings/API-review/codex/frame-targeting-and-lifecycle.md
ai/findings/API-review/codex/registrar-query-addressing.md
ai/findings/API-review/claude/frame-targeting-and-carrying.md — the empirical 0-caller backbone (frame-first (dispatch f ev), frame-bound-fn/frame-bound-fn*, subscribe* have zero real call sites in examples and tools) that grounds the helper-removal slices below.
Related beads already filed:
rf2-ts3fuk fixes unsubscribe target normalization symmetry.
rf2-az1ct6 factors an internal frame-record resolver if not absorbed by this EP.
rf2-ntwwyt moves HTTP test-support helpers out of the core facade.
rf2-10nggz is the home for the registrar query/read address grammar after EP-0023 (ruled: drop :realm, keep :frame). EP-0024 references it as the home and does not re-decide it (see §Registrar and generation reads).
rf2-kgdk03 recorded the spec/002-Frames.md ↔ spec/API.md make-frame contradiction (002-Frames still documented the pre-EP-0023 keyword-returning make-frame). This EP's §One constructor partially resolves it by unifying make-frame; the spec-graduation wave for spec/002-Frames.md subsumes or hands off to the kgdk03 fix.

Specification¶

Terms¶

Term	Contract
Frame id	The stable public address of a live frame inside the process. It is data: serializable, comparable, traceable, and suitable for opts maps, providers, tools, and logs.
Frame value	The live lifecycle token returned by `make-frame`. It owns the frame id, durable partitions, runtime subsystem state, queue/drain state, caches, lifecycle hooks, and resolved image generation. Its representation is not an app-facing data contract.
Frame handle	A captured operation bundle for async callbacks. It carries dispatch/subscribe/read functions already targeted at the frame resolved when the handle was created.
Resolved image generation	The sealed registration generation a frame resolves against while it runs. It is a slot on the frame value.
Frame-state value	The serializable EP-0001 projection of a frame. It is not the live frame value.

One live frame registry¶

There is one live frame registry. It maps frame ids to unified frame values.

The frame value owns or reaches every per-frame runtime fact through that one registry entry:

app-db partition;
runtime-db partition;
event queue and drain state;
subscription cache and topology;
epoch/history projection state;
resolved image generation;
lifecycle hooks and teardown bookkeeping;
adapter binding/configuration and host-transient leases where applicable.

The implementation may internally split storage for performance or layering, but there is one public and conceptual owner: the frame value found by frame id. There is no second public "live frame" registry that has to be kept coherent with a backing record registry.

One constructor¶

make-frame is the public constructor for a live frame. It accepts both image-selection options and frame configuration options in one call.

Illustrative shape:

(def frame
  (rf/make-frame {:id :todo/left
                  :images [todo-image]
                  :initial-db {}
                  :fx-overrides {...}}))

The constructor returns the frame value. The frame id is readable from that value through the public accessor chosen during implementation; callers should not depend on the representation.

make-frame must not require a caller to first create a backing frame and then create an image-loaded object for the same id. A Story, test, SSR request, or comparison page creates one frame with one call.

This unified single constructor adopts the previously-deferred option-(a) — extend the object constructor so it honours record-config keys (:fx-overrides, :on-create, :platform, :ssr, :doc, :preset, :tags) alongside the image-selection keys (:images, :id, :initial-db, :capabilities, :adapter) — and thereby reverses the rf2-32siq3.45 option-(b) disposition that currently ships: the fail-loud redirect (:rf.error/make-frame-record-only-key) that rejects record-config keys on the object constructor and points callers at the advanced re-frame.frame/make-frame. Under today's implementation the illustrative shape above throws.

The reversal is legitimate, not incidental. Option-(a) was deferred only because re-frame.live-frame was owned by a concurrent EP-0023 slice — it was not rejected on the merits. This EP's core (one frame value backed by one registry) removes the two-constructor split that motivated the redirect in the first place, so the fail-loud redirect no longer has a reason to exist. The advanced re-frame.frame/make-frame becomes internal or disappears. (Resolves Open Issue #1.)

Duplicate id policy¶

This EP adopts hot-reload-friendly idempotent replacement for make-frame: re-evaluating the same frame declaration updates frame configuration and resolved image generation without destroying durable state unless the caller explicitly asks for reset or destroy. Conflict cases that cannot be reconciled must fail loud.

This is a deliberate behaviour change, not a neutral choice. The current make-frame fails loud on a duplicate live id (the docstring states "a duplicate live id fails loud"). The EP replaces that blanket fail-loud refusal with idempotent replacement, because the fail-loud rule makes hot reload and Story re-evaluation needlessly ceremonial. The change is bounded: re-mount under the same id (hot reload, Story re-evaluation) must preserve durable state, and the fail-loud path is retained for irreconcilable conflicts. The fail-loud-on-every-live-id alternative was considered and rejected. (Resolves Open Issue #5; couples to the UI-owned boundary's idempotent re-mount contract.)

Operation target grammar¶

Frame-scoped operations put the primary datum first and route through an opts map or ambient context.

Canonical explicit forms:

(rf/dispatch [:todo/add "A"] {:frame :todo/left})
(rf/dispatch-sync [:todo/reset] {:frame :todo/left})
(rf/subscribe [:todo/items] {:frame :todo/left})

Canonical ambient forms scope to an existing frame with rf/with-frame:

(rf/with-frame :todo/left
  (rf/dispatch [:todo/add "A"]))

(rf/frame-provider is the UI-owned lifecycle boundary — it creates and destroys — not a pure scoping form; see §Scope, carry, and ownership.)

Frame-first operation arities such as (rf/dispatch target [:event]) are not a second public grammar. If retained internally for macro expansion or advanced implementation use, they are not taught as app API.

Public operation targets should be frame ids. Passing a frame value to dispatch/subscribe is not the canonical app-facing form; callers that own a frame value can read its id and pass the id. Internal normalization may accept a frame value where it is useful for tests or tools, but the API teaches one routing address: the frame id.

Scope, carry, and ownership are separate¶

The public API has three different jobs, and each job gets one spelling.

Job	Public spelling	Contract
Scope descendants to an existing frame (lexical / non-React)	`with-frame`	Does not create or destroy the frame. Establishes context only.
Scope an existing frame into a React subtree	`frame-provider-existing`	Provides an already-created frame id through React context. `:frame` only — a lifecycle opt fails loud. Creates / refreshes / destroys nothing.
Carry a frame across async callback boundaries	`frame-handle`	Captures operations targeted at the current or explicit frame.
Own a frame lifetime	`make-frame` + `destroy-frame!`, `with-new-frame`, and the UI-owned boundary `frame-provider`	Creation and teardown are explicit ownership operations.

Note the naming — a frame-provider name family of two per-adapter React-context components. The old scope-only frame-provider is renamed rf/frame-provider-existing (same :frame-only scope behaviour); the rf/frame-provider name is reused for the UI-owned lifecycle boundary — the component that creates a frame on mount, provides its id to descendants, and destroys it on unmount. Scoping into a React subtree needs a React-context component (frame-provider-existing) because rf/with-frame binds a dynamic var, which cannot cross React's render boundary; with-frame remains for lexical / non-React ambient scope.

Illustrative shape:

[rf/frame-provider {:id :todo/left
                    :images [todo-image]
                    :initial-db {}}
 [todo-root]]

frame-provider creates the frame on mount, provides its frame id to descendants, and destroys the frame on unmount. It is the answer for comparison pages, Story canvases, embedded widgets, and hot-reload-safe view-owned frame lifetimes.

frame-provider is realized per-adapter, against a shared contract. It is not a single component: each substrate (Reagent / UIx / Helix) ships its own frame-provider that hooks the substrate's native lifecycle, exactly as the scope-only provider (now frame-provider-existing) is per-adapter. The EP-0024 spec-graduation wave specifies the shared contract — create-on-mount, provide the frame id to descendants, destroy-on-unmount, and idempotent re-mount (per the duplicate-id policy above) — and each adapter implements its native hook. The per-adapter spec must cover the substrate lifecycle concerns:

React effect-cleanup timing (when destroy-frame! actually runs);
StrictMode double-invoke in dev (mount → unmount → mount): the create/destroy pair must be tolerant of an extra cycle;
destroy-frame!-on-unmount ordering against in-flight subscriptions and event drains, so teardown does not race outstanding per-frame work;
hot-reload-under-the-same-id must not destroy durable state — re-mount is idempotent replacement, not destroy-then-recreate.

This is the EP's only net-new public surface; the rest of the EP trims or unifies. It stays included (it was not split out into a separate EP). (Resolves Open Issue #6.)

This split keeps the user's question small:

"I already have a frame; how do I scope children?" Use with-frame (lexical / non-React) or frame-provider-existing (into a React subtree).
"This component owns a frame lifetime." Use frame-provider.
"This callback will fire later." Use frame-handle.

Carry primitive¶

frame-handle is the public carry primitive for callbacks that fire after the render or lexical frame context has unwound.

Illustrative shape:

(let [{:keys [dispatch subscribe]} (rf/frame-handle)]
  (set! (.-onclick button)
        #(dispatch [:todo/add "A"])))

frame-bound-fn and frame-bound-fn* are not app-facing carry primitives if frame-handle can express the real use cases. They may move to an internal namespace if implementation code still needs them.

Registrar and generation reads¶

The registrar read grammar (the source-store reads and any frame-generation reads) is owned and settled by rf2-10nggz, not by this EP. That decision is ruled — the read address keeps :frame and drops :realm. EP-0024 references rf2-10nggz as the home and does not re-decide it; there is no fold-in-before-acceptance dependency. This scoping keeps EP-0024 to one decision surface (frame identity and lifecycle).

The source-store reads remain source-store reads, illustrative only:

(rf/registrations :event)
(rf/handler-meta :event :todo/add)
(rf/handler-ids :event)

The standing rule EP-0024 still relies on — argument shape must not smuggle "which data source am I reading?" through a retired composition map — is the substance rf2-10nggz settles. (Open Issue #7 is resolved by scoping this out to rf2-10nggz.)

Teardown¶

Destroying a frame removes one unified frame value from the one live registry and runs teardown for every per-frame subsystem exactly once.

Teardown remains best-effort where individual cleanup hooks are host-transient, but the ownership path is one path. There is no separate public object registry whose cleanup can succeed or fail independently of the backing frame registry.

Vocabulary¶

Use these names consistently:

frame id for the routing address;
frame value for the live lifecycle token;
frame handle for captured callback operations;
frame-state value for the serializable app-db/runtime-db projection;
resolved image generation for the sealed registration generation a frame actually runs.

Do not use retired composition vocabulary as current public API vocabulary.

Rationale¶

The Clojure shape should be small: a value, an id, and data-oriented operations. The current split asks the programmer to remember which functions coerce which shape. That feels flexible, but it hides a two-registry implementation model in every call site.

Putting the event or query vector first preserves the ordinary re-frame reading order: "dispatch this event to that frame." The frame is routing metadata, so it belongs in opts. Lifecycle is not routing metadata, so it belongs in creation and teardown APIs.

Separating scope from ownership also avoids a second ambiguity. A single provider that sometimes creates and sometimes scopes would make cleanup depend on how the frame arrived. So scoping is with-frame only, and frame-provider is the explicitly UI-owned boundary — it says the ownership fact (create on mount, destroy on unmount) out loud, and is realized per-adapter against the shared lifecycle contract.

Backwards Compatibility¶

re-frame2 is pre-alpha. This EP does not require compatibility shims for off-pattern public spellings. The migration is source-level:

replace frame-first operation arities with event/query plus opts-map forms;
replace view-created make-frame plus manual provider lifetimes with the UI-owned rf/frame-provider boundary;
replace scope-only uses of the old frame-provider with rf/with-frame;
replace frame-bound-fn use with frame-handle;
replace any create-twice frame setup with one make-frame call;
migrate registrar target-map reads per rf2-10nggz (the home for the read grammar).

One behaviour change is intentional: make-frame moves from fail-loud on a duplicate live id to idempotent replacement that preserves durable state on re-mount (irreconcilable conflicts still fail loud). See §Duplicate id policy.

Historical EP prose remains historical. Current docs, examples, tools, tests, and API manifests move to the accepted vocabulary.

Bead Plan / Reference Implementation¶

Draft and decision¶

rf2-t0y79n drafts this EP and indexes it.
Operator ruling resolves the open issues below.
If accepted, file the implementation wave as child beads or an epic under this EP. Do not start the structural implementation before acceptance.

Spec graduation wave¶

Update:

spec/002-Frames.md for unified frame identity, the single make-frame constructor (subsuming or handing off the rf2-kgdk03 make-frame contradiction fix), the with-frame scope vs frame-provider owned-lifecycle split, the carry primitive, and teardown;
spec/API.md for the public facade rows and removed/retiered spellings (including the frame-provider re-purpose and the dropped scope-only provider);
spec/Runtime-Subsystems.md for the unified frame ownership of runtime subsystem state;
spec/Conventions.md for the vocabulary distinctions if needed.

The registrar read grammar is NOT graduated by this wave — it is owned by rf2-10nggz (which updates spec/001-Registration.md as needed).

Implementation wave¶

Expected slices:

Add resolved generation to the unified frame value and route all per-frame reads through one registry.
Collapse make-frame to one constructor over image options plus frame configuration options — adopting option-(a) and removing the rf2-32siq3.45 option-(b) fail-loud redirect; folds into rf2-tu2vr7 (the make-frame backing collapse, which already owns re-frame.live-frame). The advanced re-frame.frame/make-frame becomes internal or disappears.
Migrate Story, tests, SSR, and examples off create-twice setup.
Implement the idempotent-replacement duplicate-id/hot-reload policy (behaviour change from fail-loud; preserve durable state on re-mount).
Add the UI-owned frame-provider boundary as a per-adapter affordance (Reagent / UIx / Helix) against the shared create/provide/destroy/idempotent contract.
Retier or remove frame-first operation arities, frame-bound-fn, frame-bound-fn*, subscribe*, and direct make-frame-handle exposure (move to internal namespaces; keep frame-handle public) — grounded by the 0-caller backbone in ai/findings/API-review/claude/frame-targeting-and-carrying.md.
Remove the second live-frame registry and any teardown hook whose only job was keeping it coherent.
Run the docs guide-impact tail and final correctness/completeness review, per the EP wave-end standing rule.

The registrar frame-generation reads are out of scope for this wave — they are owned by rf2-10nggz.

Existing beads that can land independently or be absorbed:

rf2-ts3fuk can fix unsubscribe target normalization before this EP lands.
rf2-ntwwyt can move HTTP test-support facade helpers independently.
rf2-az1ct6 should be absorbed if the unified frame value lands first.

Open Issues¶

All open issues are resolved as of the 2026-06-18 acceptance (operator rulings on rf2-kz2vfp, rf2-uc6ebw, rf2-um1jcq).

Was the two-layer implementation deliberately transitional? Resolved: treat it as unrealized collapse debt and converge now. The single unified make-frame adopts the deferred option-(a) and reverses the rf2-32siq3.45 option-(b) fail-loud redirect (see §One constructor).
What is the exact public accessor from frame value to frame id? Resolved: provide one accessor frame-value → id; do not expose the representation.
What is the final live frame representation? Resolved: choose the smallest single-registry representation that lets the frame value own the resolved generation and lifecycle without reintroducing a second registry. It may be the existing record, a frozen handle over it, or an object wrapper only if the wrapper is the single registry value.
Should public operations accept frame values, or only frame ids plus ambient context? Resolved: teach frame ids as the routing address; frame values stay internal/tests-and-harnesses only and must not create a second public spelling.
What duplicate-id policy should make-frame use? Resolved: hot-reload-friendly idempotent replacement (preserving durable state on re-mount) with fail-loud irreconcilable conflicts. Flagged as a deliberate behaviour change from current fail-loud-on-every-live-id (see §Duplicate id policy).
What is the final name for the UI-owned lifecycle boundary? Resolved: rf/frame-provider, realized per-adapter (Reagent / UIx / Helix) against the shared create/provide/destroy/idempotent contract. The old scope-only frame-provider is dropped; scoping is rf/with-frame (see §Scope, carry, and ownership).
Should frame-generation registrar reads be public, and what are their names? Resolved by scoping out: the read grammar is owned and settled by rf2-10nggz (drop :realm, keep :frame). EP-0024 references it as the home and does not re-decide it; the prior fold-in-before-acceptance dependency is removed (see §Registrar and generation reads).
Which helper spellings are removed vs retiered? Resolved: keep frame-handle public; move make-frame-handle, frame-bound-fn, frame-bound-fn*, subscribe*, and the frame-first operation arities to internal namespaces. Grounded by the 0-caller backbone in ai/findings/API-review/claude/frame-targeting-and-carrying.md.

Recommendation¶

Accepted 2026-06-18. This is the smallest durable decision surface that matches the findings: one frame value, one registry, one target grammar, and one explicit UI-owned lifecycle boundary (rf/frame-provider). That is the post-EP-0023 frame model the public API already wants to teach. The implementation wave is filed under this EP and gated on this acceptance.