Agent runtime persistence adapter

backplane_agent_runtime includes an ephemeral ETS store and a durable store contract. It does not include a database, journal format, or production durable store. A host keeps ownership of its persistence technology and implements Backplane.AgentRuntime.Store around it.

Durable contract

A durable adapter declares all capabilities required by Store.validate_durable_capabilities/1, including incarnation_fencing. Its acknowledgement boundary has these properties:

1. store/3 is the direct commit path used by Execution and Conversation; acknowledge_commit/3 provides the equivalent operation for a host that stages first. Both compare the aggregate revision and atomically write the run, transition events, effect records, and outbox intents. They return only after the backend's documented durable boundary. Implementing only the staged callback is not sufficient. A failed or uncertain commit returns an error and no dependent effect may start.
2. load/3 reconstructs %{run:, revision:, transition:, effects:, outbox:, incarnation:}. Runtime maps use atom keys even if the backend uses another encoding at rest.
3. fence/5 atomically compares both the expected revision and current incarnation, then advances the revision and incarnation. Recovery must fence before dispatching or resolving work. Old callbacks remain stale after the fence.
4. A successful terminal snapshot retains the outcome, finite budget state, settled execution intents, and no outstanding provider/tool/wait ownership. An unknown-outcome terminal intentionally retains unresolved identities as evidence for host reconciliation.
5. Dispatched or unknown mutations remain uncertain across restart. Recovery does not infer success and does not execute them again. Only recorded read-only or idempotent work with :not_dispatched evidence is eligible for automatic resume.

EphemeralStore deliberately provides different guarantees: its context is an ETS table, a restart loses records, and it does not declare durable capabilities or implement incarnation fencing.

Reusable conformance run

Run the package harness from the host adapter's integration tests against the real persistence backend:

assert {:ok, %{mode: :durable, checks: checks}} =
         Backplane.AgentRuntime.StoreConformance.run(MyRuntimeStore, context,
           run_id: "conformance-#{unique_id()}",
           restart: &MyRuntimeStoreTest.restart/1,
           fail_next_commit: &MyRuntimeStoreTest.fail_next_commit/1,
           dependent_effect_count: &MyRuntimeStoreTest.dependent_effect_count/2
         )

assert :terminal_reconstruction in checks

The three callbacks are test controls owned by the consumer, not production Store callbacks. restart reconnects to the same durable namespace. fail_next_commit injects a failure before acknowledgement. dependent_effect_count lets the harness prove that the host did not launch an effect for the rejected run.

The harness checks the direct execution commit and staged acknowledgement paths, atomic snapshots/outbox records, failed direct commits, stale revisions, atomic incarnation fences, restart reconstruction, terminal outcome and budget reconstruction, and conservative handling of uncertain mutations. Passing a fake only proves package handling. A durability claim requires the same run against the real adapter plus a description of its flush, transaction, and power-loss boundary.

Sigma adapter boundary

Current Sigma source keeps session history in Sigma.Session.Writer and Sigma.Session.Log. The writer serializes appends, advances its active leaf only after Sigma.Session.Storage.append/2 succeeds, rebuilds the leaf with Log.snapshot/3, and checks the expected leaf when retrying. Fork publication uses a temporary journal and a create-only hard link. These are useful adapter inputs and remain Sigma-owned.

The current JSONL storage callback is append/read oriented. It does not expose an atomic compare-and-set transaction that writes a runtime transition and outbox together, an incarnation fence, or a documented flush/power-loss acknowledgement. Therefore it must not declare :durable merely by wrapping Sigma.Session.Writer. Sigma still needs a runtime-record adapter or reviewed sidecar that supplies those operations and runs StoreConformance against the actual storage. Existing JSONL replay, retry, fork, Protocol V1, sessions, repositories, and UI stay owned by Sigma.

On restart, the adapter should load the runtime snapshot, atomically call the equivalent of Store.fence/6, reconstruct the budget and outstanding effects, and pass unresolved effect evidence to Recovery.recover/2. It should publish runtime terminal/history projections into Sigma only after the runtime commit; it must not regenerate a mutation from conversational JSONL alone.

Codex resource reconciliation

Run-owned OS commands and Deno cells are cleaned when the run ends, independently of Conversation PID lifetime. Cleanup initiation and Task termination are not confirmation of an external outcome. Registry callback failures, exceptions, timeouts and uncertain backend acknowledgements retain process-local evidence; uncertain run settlement uses the existing unknown_outcome boundary. Storage failure still requires loading/reconciling the store, including when a commit succeeded but its acknowledgement was lost.

Resource registries and collaboration managers remain ephemeral. Hosts needing restart reconciliation must retain serializable resource/run identities and backend-specific evidence in their own recovery system. Never serialize registry values, PIDs, ports, cleanup functions, dispatch closures, credentials or grants.

LocalCommand separates active/output records, unresolved session obligations, and recent confirmed-release receipts. An obligation's confirmed evidence stays pinned until the ResourceRegistry consumes it; only then may it enter the bounded receipt cache. Receipt acknowledgement runs as a supervised, bounded effect after registry settlement. A failed acknowledgement retains the backend obligation and can cause admission backpressure; it does not undo already confirmed OS cleanup. Missing or evicted receipts never prove successful release and must not cause blind cleanup retries. Hosts must reconcile unresolved identities explicitly. Owner-wide cancellation fences existing never-launched reservations in the backend before acknowledgement; a consumed or evicted receipt cannot make an old launch identity valid again. A verified LocalCommand reconciliation updates current cleanup status in retained output for the exact session, owner, and incarnation without shortening the output cache lifetime or changing the command's execution outcome. The earlier cleanup_error remains historical diagnostic evidence. A still-active ResourceRegistry session can release through its existing callback after host backend reconciliation; an already failed or uncertain registry entry has no public retry/reset and remains unresolved.

Conversation timer generations and catalog callback tokens are ephemeral fences, not persistent execution contexts. A failed nested producer loses staging before settlement acknowledgement, while a successful producer becomes eligible only after that acknowledgement. A lost Store reply still requires loading and reconciling the store; neither staging nor receipt eviction permits replay. An already-dispatched mutating direct or nested tool with an untrustworthy outcome retains its committed active invocation and execution intent under the existing unknown_outcome boundary. Outer error handling, Task cleanup and late results do not consume that evidence. This adds no persistence fields or Store callbacks; the admitted catalog used for classification is process-local. After acknowledgement loss, the host must inspect the stored run before any continuation or replacement. The current R22 implementation still fails command lifecycle regressions for trusted confirmed non-start refusals; see the follow-up validation record. Do not treat the new classification as a fully validated host upgrade boundary.

A replacement child run uses a fresh storage aggregate and committed history; it never automatically replays a restored nonterminal run or a completed tool mutation. The old run's revision history remains distinct from the replacement.

Explicit disabled deadlines

The durable run deadline is integer() | :infinity | nil. Only :infinity disables the root deadline. nil keeps the legacy unset meaning and receives a finite default on admission. A finite admitted record remains finite even when new options request infinity. An explicit saved infinity remains infinite when options are omitted.

JSON adapters encode the sentinel as "infinity" and decode it with an explicit case into :infinity. Do not create atoms from stored strings. Effect, commit, cleanup, work and native command bounds remain independently finite. Restoring an unfinished infinite run still requires reconciliation and fencing before admission; it never authorizes replay of its external effects.

Goal records and accounting recovery

Conversation revokes its Goal binding in a supervised task before emitting the terminal run notification. The acknowledgement is bounded by commit and cleanup limits. Goal death or missing acknowledgement leaves a controlled storage_failed recovery boundary; hosts must reconcile it before new admission, even if the journal already contains a terminal run.

Goal.Store is a separate host-owned thread record contract:

@callback load(term(), binary()) :: {:ok, map()} | {:error, Error.t()}
@callback compare_and_swap(term(), binary(), non_neg_integer(),
  binary() | nil, map()) :: {:ok, map()} | {:error, Error.t()}

For an absent thread, return the empty record from Goal.State.new(thread_id, []). CAS must compare both revision and current goal_id (nil for no Goal), persist the entire next record atomically, advance revision exactly once and return that committed record. A conflict returns Error class :resource_conflict. Clearing retains a tombstone revision, so delete/recreate cannot reuse an old receipt or mutation fence. Adapter acknowledgements must represent their documented durable transaction/flush boundary.

Persist the full private snapshot, including receipt digests, event revisions, cumulative usage baselines, mutation invocation responses, audit counters, budget-steering delivery state and continuation deferral. The public Goal.get/1 response is insufficient for recovery. Never store PIDs, references, functions, ports, authority capabilities or BEAM monotonic timestamps. Elapsed whole seconds are persisted; a restored owner establishes a fresh local clock. The file-backed test fixture proves package restart handling, not production adapter durability.

Conversation checkpoints the Goal receipt outbox with the acknowledged provider/tool/turn transition. Goal CAS records the receipt and response before Conversation clears the outbox in another checkpoint. A lost reply may therefore leave either an undelivered receipt or an already-applied receipt. Redeliver the same receipt identity and payload; matching duplicates return the saved response, while changed duplicates conflict. Recover accounting without replaying the provider or external tool. Goal tool mutation and invocation response are also stored together so a repeated committed invocation returns its original result.

After acknowledgement failure, load and fence/reconcile the ordinary run first. Goal.restore/1 reloads its record and invalidates live bindings/reservations; it does not scan run journals or launch a continuation. The host reconstructs fresh authorized bindings, then uses GoalBridge.redeliver(goal, binding, run) for each retained committed outbox. Resolve changed Goal identity and uncertain effects explicitly before fresh admission. Only after delivery and ordinary Session readiness may the host signal Goal.idle/2. Restarting clears process-local pending-receipt tracking, so persisted run outboxes remain the host's recovery authority.

Descendant bindings capture root thread and Goal identity at admission. Their capabilities authorize root usage accounting only; they cannot read or mutate the root Goal through its tools. Child threads that need their own Goal tools must receive their own host-owned Goal. Plan mode is inherited at delegation, and its lifecycle does not change root Goal status or failure audits. Their reports debit that root once, including while the root run is idle; completing a child never adds its aggregate usage a second time. Replacing or clearing the root fences old descendant receipts from the new Goal. fork_snapshot/2 copies the Goal snapshot into an empty target thread, retaining its Goal identity, status, budget, totals and timestamps. Run receipts, live bindings, reservations and continuation ownership are fresh in the target.