Merge reviewed master updates into sleep ingress repair

doc/SPEC-implementation.md

@@ -1729,7 +1729,10 @@ Scheduler must skip invocation when:
 - an existing run is active
 - hard budget limit has been hit
 
-Budget admission and accounting serialize under the company accounting lock.
+Authoritative paid-work admission and accounting mutations serialize under the
+company accounting lock. Preliminary invocation checks use a read-only snapshot
+and take that lock only when a hold, incident, or enforcement transition needs
+reconciliation; reconciliation rereads current state under the lock.
 The dispatch boundary rechecks current company, agent, and project policies and
 reserves capacity before paid work starts. Accounting commits budget blocks and
 durable stop intent without waiting for provider shutdown or reconciling unrelated

doc/execution-semantics.md

@@ -1602,8 +1602,9 @@ agent, workspace, retained session, and termination of prior execution. The
 checkpoint must match the exact Runner, lease, run, session, turn, and item.
 A missing restart claim does not remove these checks or authorize a fresh session.
 
-After a plugin session passes provider recovery and durable run admission, its
-local suspension receipt is removed. The running session then owns authority;
+After a remote sandbox session passes provider recovery and durable run admission,
+its local suspension receipt is removed. This applies to built-in sandbox runners
+and plugin tasks. The running session then owns authority;
 the old receipt must not conflict with a later warm run's journal identity.
 Failed admission retains the receipt. Normal suspension writes a new receipt.
 

doc/observability.md

@@ -45,10 +45,15 @@ Keep credential values out of logs and diagnostic reports.
 ## Budget admission, accounting, and enforcement
 
 The operator-gated `paperclip.budgets` tracer records
+`budget.admission.inspect`, `budget.admission.reconcile`,
 `budget.admission.evaluate`, `budget.admission.reserve`,
 `budget.accounting.transaction`, `budget.accounting.settlement`, `budget.accounting.lock_wait`,
 `budget.exposure.evaluate`, `budget.enforcement.deliver`,
 `budget.enforcement.scope`, `budget.cancellation.deliver`, and `budget.accounting.recover`.
+Inspection spans cover the preliminary read-only snapshot. Reconciliation spans
+exist only when that inspection finds a transition requiring the company lock;
+compare their counts to measure how often preliminary checks bypass the lock.
+Authoritative reservation and transactional evaluation retain their own spans.
 Transaction spans include commit time. Lock-wait spans cover acquisition of the
 company accounting lock. Settlement spans cover the ledger and reservation
 transaction. Enforcement spans measure asynchronous delivery; cancellation

doc/plans/2026-11-03-accounting-contention-and-budget-enforcement.md

@@ -0,0 +1,400 @@
+# Accounting contention and best-effort budget enforcement
+
+Date: 2026-11-03
+Status: Investigation complete; first admission optimization implemented locally, remaining stages proposed
+Source revision: `367f29478d14003cc5c0f161b7ce8f3164993634`
+
+Follow-up: [Lock-free accounting design](2026-11-03-lock-free-accounting-design.md)
+defines the requested target architecture. Its append-only receipt path replaces
+the run-lock proposal below; this document retains the investigation evidence
+and the possible behavior-preserving first migration step.
+
+## Recommendation
+
+Keep receipt persistence, deduplication, final settlement, and audit transactional.
+Make budget detection and automatic stopping a measured, short-lag control. Move
+frequent receipt writes to transactions scoped to their run or paid operation.
+Evaluate budget policies in coalesced batches outside those transactions.
+
+First remove avoidable admission transactions and measure the remaining critical
+sections. Do not begin with distributed spending allocations or a new accounting
+service. A short lock at an infrequent boundary can be acceptable; serializing
+every company's receipt stream and admission preflight is the scaling concern.
+
+The first implementation now inspects preliminary eligibility in a repeatable-read,
+read-only snapshot. Healthy checks and checks with no policies bypass the company
+writer lock. Holds, hard blocks, recovering accounting and incidents needing
+resolution fall back to the existing mutating owner, which revalidates current
+state under the lock. Transactional admission, reservations and receipt accounting
+retain their authority. Separate inspection/reconciliation spans expose the bypass
+ratio. This does not implement the later append-only accounting architecture.
+
+Configured-policy tradeoffs: a healthy check adds an incident lookup, and an
+unheld blocking/recovering scope can compute exposure once during inspection
+and again during locked revalidation. Already-held scopes request reconciliation
+before computing exposure. Qualify configured-policy workloads as well as the
+predominant no-policy case; the change is not a claim of lower CPU everywhere.
+
+The first milestone's budget-service suite passes 18 tests against real
+PostgreSQL. A controlled writer holds the company accounting lock while routine
+checks finish; authoritative admission on the same reader instead reaches its
+500 ms lock timeout. Coverage includes healthy/warning policies, manual and
+orphaned pauses, hold release, incident cleanup and cross-company rejection.
+This establishes the lock-bypass contract, not deployed latency or CPU savings.
+Seven additional accounting/enforcement/replay/tracing suites pass 55 tests,
+for 73 passing targeted tests across eight files with no skips.
+The full workspace typecheck passes via `node scripts/typecheck-low-memory.mjs`,
+and `node scripts/check-module-boundaries.mjs` passes.
+The full workspace build passes via `PAPERCLIP_TYPECHECK_LOW_MEMORY=1 pnpm -r
+--workspace-concurrency=1 build`, including the current-source Runner release
+binary, server, UI and CLI. The initial incomplete local Copilot build artifact
+was repaired with the pinned installer before the successful canonical retry.
+
+This proposes a product-contract change for the later stages. It does not change
+the current spec or silently redefine existing configured budgets. The first
+implementation stage below preserves existing enforcement semantics.
+
+## What the investigation established on the source revision above
+
+### The lock protects several different responsibilities
+
+[`withAccountingTransaction`](../../server/src/services/accounting-transaction.ts)
+takes `FOR NO KEY UPDATE` on the company row before invoking its callback. It
+holds that lock until the transaction ends. Different companies can proceed
+independently; writers for one company serialize. Ordinary report reads already
+use a read-only repeatable-read snapshot and do not take this writer lock.
+
+| Path | Work sharing the company lock before this change |
+| --- | --- |
+| Usage receipt | Lock run, validate source/sequence and replay hash, insert immutable receipt, update running usage, evaluate affected budgets if monetary exposure changed |
+| Admission | Read company/agent/project state and policies, calculate exposure, reconcile budget holds/incidents, reserve estimated capacity before dispatch |
+| Final settlement | Recover durable receipt files, validate completion, settle reservation, insert ledger entries, update projections and runtime totals, acknowledge the run, evaluate budgets |
+| Other mutations | Direct cost reports, billing imports/corrections, accounting repairs, paid browser/decision/fast-response/classifier operations, and agent lifecycle operations |
+
+The wrapper was unconditional. Having no configured budget did not bypass the
+lock. The public `getInvocationBlock` method also used this mutating wrapper.
+The implementation above separates its preliminary reads while preserving
+the reconciliation and resume effects that some callers need.
+
+Primary owners are [usage receipts](../../server/src/services/usage-receipts.ts),
+[budgets](../../server/src/services/budgets.ts),
+[reservations](../../server/src/services/budget-reservations.ts), and
+[run settlement](../../server/src/services/run-cost-accounting.ts).
+Queue, dispatch, and lifecycle code should consume their contracts rather than
+implement another budgeting system.
+
+### The current guarantee is already narrower than a dollar cap
+
+Reservations serialize admission against known exposure and estimated holds.
+They do not restrict how much an already-dispatched provider can charge.
+`reservationCents` defaults to zero. A zero reservation still prevents duplicate
+dispatch, but does not reserve a positive allowance for future model spend.
+
+The [implementation spec](../SPEC-implementation.md) explicitly acknowledges that
+mid-run enforcement depends on available provider reports or supported prices.
+Providers reporting only at completion cannot provide exact mid-call detection.
+Browser work can pass a provider cost cap; preserve such supported limits.
+
+The current rules are approximately:
+
+- Admission counts ledger spend plus the maximum of each unfinished operation's
+  reservation and observed spend, plus separate browser exposure.
+- Threshold stopping uses ledger plus observed running spend, rather than
+  treating reservation estimates as money already spent.
+- Pending terminal accounting and configured unknown-price restrictions can
+  independently block work.
+- Final settlement replaces live exposure with the ledger charge atomically.
+
+Accurate recording of received evidence and instantaneous control of external
+spend are separate contracts. Some prices remain estimates or unknown even when
+their receipt is recorded correctly.
+
+### The critical section can do substantial work
+
+When applicable policies exist, [exposure evaluation](../../server/src/services/budget-exposure.ts)
+executes four sequential aggregate statements over ledger entries, held
+reservations joined to runs, pending/live runs, and browser work. Policies are
+batched into those statements; it is not four statements per policy. With no
+billable policies, these four queries are skipped.
+
+A changed monetary receipt can invoke this evaluation synchronously. Duplicate,
+stale, unchanged, and subscription receipts take different branches; not every
+receipt performs the aggregate scans. Native delivery already suppresses
+accepted source-event replays in its receiver, so duplicate suppression is a
+measurement question rather than an assumed large win.
+
+The hot inputs are usage and turn-lifecycle events, not every streamed text
+token. Native capture handles usage reports and turn start/terminal events.
+Legacy capture also durably spools before its database attempt. Recovery can
+still read receipt files while the settlement transaction holds the company
+lock, although it builds an initial file-identity index outside the lock.
+
+[Monthly spend projections](../../server/src/services/costs.ts) update
+`companies.spentMonthlyCents` for each ledger insert, and the applicable agent
+counter. Merely deleting the explicit company lock would leave an implicit
+shared-row write bottleneck and break assumptions around settlement and repair.
+PostgreSQL updates acquire conflicting row locks too. See the
+[PostgreSQL locking contract](https://www.postgresql.org/docs/current/explicit-locking.html#LOCKING-ROWS).
+
+### Live traces justify a focused optimization, not a CPU diagnosis
+
+Read-only observability queries examined one live host. The matched one-hour
+sample is **2026-11-03 16:45–17:45 UTC**, service revision
+`3910352c923afda320b2a5134e523932ee1d0633`.
+The aggregate measurements are reproduced below. Private trace links and host
+identifiers remain outside the repository.
+
+| Observed span | Count | p50 | p95 | p99 | Maximum |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| Company lock acquisition | 9,421 | 13.00 ms | 206.85 ms | 725.14 ms | 4,301.53 ms |
+| Accounting transaction | 9,421 | 87.16 ms | 486.95 ms | 1,147.10 ms | 5,594.19 ms |
+| Admission evaluation | 8,082 | 43.45 ms | 166.84 ms | 342.13 ms | 1,212.92 ms |
+| Final settlement | 216 | 355.90 ms | 1,894.05 ms | 3,149.67 ms | 5,594.33 ms |
+
+These are span counts; nested spans must not be added as independent operations.
+There were 23 `budget.admission.reserve` spans in the same window. The much larger
+admission-evaluation count warrants caller attribution, not an assumption that
+all evaluations represent new paid dispatches.
+
+Of 8,292 exposure evaluations on that host in that hour, **8,280 had zero policies**
+and 12 had one. That is 99.86% with zero policies at this call boundary; it does
+not prove that every company has no budgets. These empty evaluations are cheap,
+but their surrounding admission/accounting work still takes the company lock.
+
+One inspected admission transaction took 4.183 seconds, with 4.090 seconds in
+lock acquisition. Its exposure child had zero policies and took about 0.026 ms.
+A separate settlement transaction took 5.594 seconds, including 4.302 seconds
+in acquisition and a zero-policy exposure evaluation.
+Both inspected traces report mean sample rate 1.
+
+Limits of this evidence:
+
+- The acquisition span measures the complete awaited lock query. It includes
+  database/network round trips and Node scheduling; it is not a PostgreSQL
+  lock-wait measurement. Transaction duration also includes transaction setup
+  and commit. Do not subtract unrelated percentiles to estimate hold time.
+- The spans do not identify the locking company, blocker, receipt frequency,
+  or operation caller. They do not establish one company's throughput ceiling
+  or the cause of the previously observed CPU utilization.
+- The sample identifies one host. It does not establish fleet-wide behavior or
+  independently qualify the public instance serving each request.
+- The local accounting wrapper, receipt, exposure, budget service, and run-cost
+  files match the traced revision in Git. Other deployed components and database
+  workload have not been qualified against this checkout.
+- The existing [budget benchmark](../../scripts/benchmark-budget-enforcement.ts)
+  varies unrelated idle-agent count. It does not test concurrent receipt writers,
+  10,000 active runs, or growing ledger history. It was inspected, not run here.
+
+## Proposed behavior
+
+| Responsibility | Proposed contract |
+| --- | --- |
+| Receipt acceptance | Durable commit before acknowledgement; source ownership, sequence, hash conflicts and company scope remain enforced |
+| Ledger and settlement | Idempotent transaction scoped to the run/invocation; no lost charges, duplicate application, or transient live-plus-final double counting |
+| Budget detection | Coalesced evaluation from committed evidence; a measurable delay is allowed |
+| New work | Check current authoritative pause/Stop and known budget-block state at dispatch; concurrent admission against a not-yet-reflected spend increase can overshoot |
+| Automatic cancellation | Commit durable stop intent, then use the existing scoped cancellation worker; provider shutdown remains a separate obligation |
+| Manual Stop and authorization | Preserve current execution-authority checks; budget approximation does not weaken them |
+| Unknown accounting | Retain unresolved obligations and existing configured unknown-usage restrictions; missing evidence is never a zero-cost receipt |
+
+Suggested initial service targets, to validate before adopting as promises:
+
+- Coalesce dirty budget work over approximately one second while work exists.
+- p99 from committed priced receipt to durable threshold-stop intent at most
+  five seconds under the qualified workload. Measure provider reporting delay
+  and cancellation latency separately.
+- For applicable spend limits, stop admitting additional paid work when the
+  oldest unevaluated accounting change exceeds fifteen seconds. This is a
+  proposed protection against a failed evaluator, not a hard spending cap.
+  An idle company's old evaluation timestamp alone must never trigger it.
+- With no dirty changes, unresolved retries, or actual budget deadlines, the
+  evaluator performs no recurring database scans.
+
+Best effort needs a stated overspend model. As an illustration, 10,000 runs each
+spending $0.001/second consume $10/second; five seconds of extra detection delay
+adds about $50, before unreported and uncancellable in-flight work. Those are
+hypothetical rates, not a measured cost estimate or maximum.
+
+In general, additional spend depends on fleet burn rate multiplied by reporting,
+evaluation, and stop delay, plus already-committed work. Without bounds on those
+quantities, do not advertise a bounded dollar overshoot. A strict end-to-end cap
+would require provider-enforced limits or preauthorized bounded work; distributed
+allocations alone cannot cap a provider that can exceed its allocation.
+
+## Implementation stages
+
+### 1. Attribute the cost and eliminate unnecessary admission transactions
+
+Owners: accounting transaction/tracing services, budget admission API, and its
+queue/dispatch/lifecycle callers.
+
+Add a closed operation-kind label to accounting spans so receipt, finalization,
+preflight, reservation, policy edit, correction, and lifecycle work can be
+distinguished. Capture acquisition-query time, time from acquisition return to
+transaction completion, query count, and replay/outcome categories. Keep the
+operator-gated Observability contract and content-free attributes; this is not
+first-party Telemetry. Label the client-observed held interval honestly because
+it includes commit/return latency. Use sampled PostgreSQL wait/blocker evidence
+to establish actual contention, with SQL values and credentials excluded.
+
+Separate read-only preliminary eligibility from mutating reconciliation.
+Preflights should not take a company writer lock just to discover no applicable
+policy. Keep the existing authoritative reservation and policy check at final
+dispatch in this stage. Budget changes, rollover, settlement and explicit
+reconciliation continue to own their necessary effects. Do not introduce a
+stale cached "no budgets" result that can bypass a newly configured limit.
+
+Measure identical receipt replay and unchanged snapshots before prioritizing
+their optimization. Any early replay return must prove the immutable receipt
+and relevant projection state agree. Native event identity and lifecycle events
+cannot be discarded just because their monetary amount is unchanged.
+
+Deliver a real-PostgreSQL benchmark using the service boundaries below, and the
+small behavior-preserving optimization justified by the caller measurements.
+If this removes today's significant delay, ship it independently; later stages
+remain a deliberate scale/product decision.
+
+### 2. Persist receipts per run and evaluate budgets asynchronously
+
+Owners: usage receipts, native usage delivery, budget exposure, budgets, and the
+existing accounting/enforcement delivery lifecycle.
+
+In one run-scoped transaction:
+
+1. Validate current source/ownership and lock the run.
+2. Insert or verify the receipt with its unique identity and fingerprint.
+3. Replace the run's accepted cumulative usage/exposure projection.
+4. Persist a versioned pending-evaluation marker and register the normal
+   transaction-aware wake.
+5. Commit before acknowledging native delivery or removing legacy spool evidence.
+
+Do not take the company accounting writer lock or evaluate all affected policies
+inside this receipt transaction. Keep final and partial receipts durable; batch
+derived budget evaluation rather than dropping intermediate accounting evidence.
+
+Use the shared delivery coordinator for coalescing, retry and deadlines. Pending
+state must be scoped to the changed run/operation, or an append-only change;
+updating one company dirty counter for every receipt would recreate the lock.
+Consumption must compare revisions or claim durable work so a concurrent change
+cannot be cleared by an older evaluator. A timestamp/sequence high-water mark
+alone cannot safely assume database transactions commit in allocation order.
+
+Read ledger and live contributions in one consistent database snapshot, or use
+transactionally consistent replacement contributions. The present four separate
+read-committed aggregates are not safe to reuse unchanged once other operations
+in the same company can settle concurrently: live-to-ledger
+replacement must not be missed or counted twice across queries.
+
+Evaluate each affected policy once per batch, revalidate its revision before
+committing block/resume state, and retain durable cancellation generations.
+Updates arriving during evaluation remain pending. Recovery reconstructs pending
+work after restart; rollback must not publish a false wake or acknowledgement.
+Use short policy-state transactions where necessary. Preserve manual holds and
+committed stop obligations during policy changes and price corrections.
+
+This is the first stage that permits an additional detection/admission race
+window. Update the product contract together with implementation; removing only
+the wrapper is not a valid intermediate change.
+
+### 3. Remove remaining company writes from frequent settlement
+
+Owners: run-cost accounting, costs, billing reconciliation/import, accounting
+integrity, paid-service invocation owners, and DB/shared contracts.
+
+Settle ledger entries, reservation state, final receipt acknowledgement, and
+run-specific accounting together under the operation's lock. Use database
+uniqueness plus conflict/hash verification for concurrent idempotency; a
+read-before-insert check previously protected by the company lock is insufficient.
+Retain narrow agent coordination for shared runtime totals where needed.
+
+Move company monthly display-counter updates to versioned, coalesced projection
+work. Update integrity inspection/repair and all counter consumers to understand
+projection freshness and checkpoint identity. Reporting from ledger snapshots
+remains available; intentional projection lag must not look like lost money.
+Handle backdated events, credits/corrections, UTC rollover and lifetime budgets.
+
+For receipt-file recovery, stage immutable candidate envelopes outside the broad
+transaction, then validate source/version and closure under the operation lock.
+Retain disk evidence until the final owning transaction commits. A concurrent
+late receipt or replaced source must still prevent premature acknowledgement.
+
+Apply the same boundary to browser runs, decision-model invocations, fast
+responses, prompt-injection classification, and externally reported charges.
+Inventory administrative users of `withAccountingTransaction` separately;
+agent deletion, repair, and lifecycle safety must not depend on a lock removed
+from receipt writers. Specify and test the new lock order, including issue/run
+ownership and foreign-key interactions. Keep supported provider spending caps.
+
+Start with bounded batches and indexed aggregates. If history growth or batch
+time misses the target, maintain incremental scope/window contributions. If
+those counters become the next measured hot row, partition their writes and
+sum a bounded set of partitions. Do not synchronously increment a single new
+company total on every receipt and call the bottleneck solved.
+
+### 4. Qualify workload and roll out
+
+Extend the existing benchmark rather than using idle-agent count as scale proof.
+Use real PostgreSQL with at least two independent controller processes and
+explicitly recorded connection-pool limits. Simulated provider events should
+exercise real receipt ingestion and settlement, followed by a smaller real
+provider qualification. Keep load tests off the active staging instance.
+
+| Dimension | Required cases |
+| --- | --- |
+| Distribution | 100, 1,000, then 10,000 active runs; one company and many companies |
+| Receipt cadence | One update per run per 60 seconds, 10 seconds and 1 second; record actual target-provider cadence before choosing the release load |
+| Policies | None; company; agent/project only; overlapping scopes and windows |
+| History | Increasing ledger history independently of active-run count |
+| Event shape | Cost-changing, unchanged, subscription, duplicate, stale, conflicting and out-of-order reports; reconnect bursts |
+| Races/failures | Simultaneous settlement and receipt, worker crash, failed/ambiguous commit, source replacement, policy change, manual Stop, late final charge, unknown price and UTC rollover |
+
+At 10,000 runs the illustrative receipt cadences imply about 167, 1,000, and
+10,000 receipts/second before admission or settlement. This defines a workload,
+not a claim that current or proposed code supports it. A shared critical section
+held for an average of H seconds has an ideal ceiling near 1/H operations/second
+for that company; do not estimate H from the unrelated percentile summaries.
+
+Acceptance requires durable receipt/ledger equality after quiescence, no duplicate
+effect or acknowledgement loss, stable pending-work depth, correct policy/manual
+hold behavior, and measured detection/cancellation/freshness targets. Record
+p50/p95/p99 ingestion, admission and settlement latency, actual PostgreSQL wait
+events, query count/rows/bytes, CPU/event-loop delay, and reconnect recovery time.
+Prove unrelated companies remain responsive and writers for unrelated runs no
+longer contend on a shared company update in the frequent path.
+
+Initially compute the proposed evaluation in shadow mode without stop/resume
+effects and compare it against the authoritative ledger and existing decisions.
+Then canary the new enforcement contract on an explicitly selected environment,
+with its versioned pending work rebuilt and validated before switching. Do not
+run old and new enforcement authorities simultaneously for the same company.
+Rollback must drain/reconcile accepted receipts and rebuild authoritative state
+before restoring the older path; flipping a flag alone is insufficient.
+
+Relevant existing coverage includes scoped budget enforcement, live attempts,
+native usage delivery, run-cost accounting, budget cancellation, worker recovery,
+and idle-accounting work. Add focused concurrency/failure tests for the changed
+boundaries; run the repository's full checks before a PR-ready implementation
+handoff. Browser qualification is needed if dispatch/Stop or budget UI semantics
+change, not to validate this plan file.
+
+## Contract and delivery boundaries
+
+Before enabling the new behavior, update `doc/SPEC.md`,
+`doc/SPEC-implementation.md` (especially scheduler and budget enforcement), the
+applicable contributor invariants, shared budget contracts, and user-facing
+budget explanations. The UI should say what auto-stop does and disclose possible
+in-flight overspend; it should not expose internal partition mechanics. Decide
+how configured existing reservations retain their estimate/visibility role
+without promising atomic aggregate admission in the new mode.
+
+Distributed bounded spending allocations are deferred. Revisit them only if
+customers need stricter aggregate admission or qualified batch evaluation cannot
+meet latency/overspend needs. They require provider cost bounds as well as
+replenishment, revocation and policy-version coordination to become a monetary
+guarantee. They are not required for the recommended first implementation.
+
+Investigation validation: inspected the source and existing test/benchmark
+contracts, queried live spans and two individual traces, and checked the relevant
+deployed-revision source diff. No service, configuration, database, or deployed
+code was changed. Runtime tests and load tests were not run for this planning
+change; dependency installation and implementation belong to the next stage.

doc/plans/2026-11-03-lock-free-accounting-design.md

@@ -0,0 +1,432 @@
+# Lock-free accounting and best-effort budgets
+
+Date: 2026-11-03
+Status: Target architecture proposed; preliminary admission optimization implemented locally
+Scope: Accounting ingestion, settlement, budget projections and budget admission
+Evidence: [Accounting contention investigation](2026-11-03-accounting-contention-and-budget-enforcement.md)
+
+## Decision
+
+Make the frequent accounting write path append-only. A receipt writes its own
+immutable fact and commits. It does not lock or update the company, agent, run,
+budget, aggregate counter, or worker checkpoint. Independent runs use independent
+keys, including when they belong to the same company or agent.
+
+Asynchronous workers derive operation state, ledger entries and partitioned spend
+totals. Budget admission reads a recent view without making a synchronized debit
+against a company balance. Threshold detection and cancellation may lag; durable
+evidence, deterministic accounting and execution authorization remain required.
+
+"Lock-free" here means no application mutex, advisory lock, or `SELECT FOR
+UPDATE` in the receipt or budget-check path, and no shared accounting row written
+for each receipt. It is not a formal nonblocking progress guarantee: PostgreSQL
+still uses internal locks, unique-index arbitration, foreign-key checks and
+transaction commit. In particular, two inserts with the same unique key can
+wait for each other. See [PostgreSQL INSERT](https://www.postgresql.org/docs/current/sql-insert.html).
+
+Background batches and infrequent policy changes use optimistic conditional
+writes. Those also take database row locks internally. Moving a compare-and-swap
+onto the same company counter for every receipt would fail this design.
+
+Existing issue ownership, provider-dispatch fencing, manual Stop and resource
+ownership are outside this lock-removal scope. Accounting cannot replace those
+contracts with an eventually consistent spend view.
+
+## Problem fit and first milestone
+
+This design removes the identified company-wide accounting serialization point
+if receipt ingestion, admission, settlement and shared counters all migrate.
+It does not establish that accounting is the current main-loop CPU culprit or
+that the resulting system supports 10,000 active agents. Those remain separate
+claims requiring measurements.
+
+The matched investigation sample contained 8,280 zero-policy exposure evaluations
+out of 8,292. Those exposure evaluations returned almost immediately; the
+surrounding admission path still read state and policies inside the accounting
+transaction. This makes separating eligibility reads from reconciliation the
+most directly supported first change. It does not show that large exposure
+scans or a high receipt rate caused the sampled delays.
+
+The 9,421 accounting transactions in that hour average about 2.6 per second
+across the observed host. Burst rate, company distribution and actual lock hold
+time are missing. The acquisition spans include database/network time and Node
+scheduling, so their long tails alone do not prove PostgreSQL row-lock saturation.
+
+| Problem | What this design changes | Remaining condition |
+| --- | --- | --- |
+| Eligibility checks enter a mutating company transaction | Read-only admission, separate reconciliation | Classify callers and preserve manual holds, resume behavior and execution fencing |
+| Independent receipts serialize by company | Independent immutable inserts | Qualify actual receipt rate, index/FK waits and durable ingestion capacity |
+| Repeated exposure work and shared counters limit growth | Coalesced evaluation and partitioned projections | Show less work per receipt and sufficient projection throughput |
+| Main-loop CPU/event-loop contention | Some synchronous work leaves the request path | Same-process reducers still consume CPU; batching must reduce work or processing must move to another runtime |
+
+The immutable identity model, closure changes and partition publication protocol
+are target-architecture work, not prerequisites for the first latency improvement.
+An asynchronous pipeline also adds backlog and recovery obligations: its stale
+decision gate can become an admission outage if reducers cannot keep up. Merely
+moving identical work behind an ACK would conceal latency without solving load.
+
+Start with an independently shippable milestone: attribute admission calls and
+transaction work, provide a pure eligibility read (including a no-policy fast
+path), and move required reconciliation to its explicit owner. Retain the current
+final admission/reservation authority during this initial behavior-preserving
+step. Having no configured policy must not bypass manual company/project pauses,
+existing holds, unknown-accounting restrictions or duplicate-dispatch protection.
+
+Compare end-to-end dispatch latency, accounting transactions and queries per
+dispatch, database waits, CPU and event-loop delay under the same workload.
+Then qualify receipt coalescing and the remaining single-company bottleneck.
+Use the full migration below when that evidence or the agreed capacity target
+justifies it. The accounting rewrite does not cover the other independently
+identified work in run-event persistence, recovery, snapshots or fanout.
+
+## Architecture
+
+```mermaid
+flowchart LR
+  P["Runner / paid service"] --> I["Validate and append receipt"]
+  I --> J[("Immutable receipt journal")]
+  J --> A["ACK after durable commit"]
+  J --> R["Independent partition reducers"]
+  R --> L[("Ledger + operation views")]
+  R --> T[("Spend totals by scope / window / partition")]
+  T --> B["Coalesced budget evaluator"]
+  B --> V[("Versioned budget decisions")]
+  V --> D["Dispatch: read decision, record own admission"]
+  B --> S[("Durable stop intent")]
+  S --> C["Existing scoped cancellation path"]
+```
+
+The journal, reducers and evaluator can start in the existing server and
+PostgreSQL deployment. Independent scaling is a module/runtime boundary, not a
+requirement for Kafka, Redis, another service, or a separate database today.
+
+## 1. Give every paid operation a stable identity
+
+Use one accounting operation for each heartbeat run, browser invocation,
+decision-model invocation, fast response, classifier invocation, or separately
+reported charge. Record its company and attribution before paid dispatch.
+Retries of the same external operation reuse its identity; a genuinely new paid
+attempt has a distinct attempt identity beneath that operation.
+
+An accounting operation's immutable identity is separate from mutable run state.
+Receipts reference this identity, rather than acquiring foreign-key protection
+on a busy mutable `heartbeat_runs` row. The existing journal directly references
+that row, so removing the explicit lock alone does not eliminate this possible
+wait. Use a company-scoped composite reference to the immutable identity; keep
+company-boundary validation and the existing deletion/retention obligations.
+Separate any nullable live-run pointer from retained original attribution.
+
+Admitted accounting sources have durable identities and provenance. Controller
+reconnection does not manufacture a new paid attempt. Source replacement must
+retain the existing verified handoff/baseline semantics. A higher timestamp,
+random UUID, or newly generated controller source ID is not proof of authority
+over earlier billing evidence.
+
+Source handoff remains an infrequent execution-authority operation. Ordinary
+checkpoints do not mutate its ownership record. Existing authentication and
+company access checks still apply. Authority to report a closing receipt after
+Stop is accounting-only; it never grants authority to continue execution.
+
+## 2. Receipt ingestion only appends evidence
+
+Extend/generalize the existing `run_usage_receipts` contract. The conceptual
+receipt contains:
+
+| Field | Purpose |
+| --- | --- |
+| Company and operation IDs | Tenant boundary and accounting unit |
+| Attempt and admitted source identity | Separate paid attempts from reconnect/replay |
+| Stable source event ID and source sequence | Replay identity and ordering within that source |
+| Receipt kind and normalization version | Partial usage, final usage, source handoff, closure, correction |
+| Cumulative usage, price and provenance | Evidence; unknown price stays unknown |
+| Payload fingerprint | Detect conflicting reuse of an identity |
+| Server-assigned ingestion time and deterministic partition | Recovery, pending-work age and routing |
+
+Provider event time remains separate. Freshness protection never trusts a caller's
+timestamp. External charge callers need a stable request identity to obtain
+idempotent retries; matching dollar amounts or payload similarity cannot prove
+that two otherwise unidentified charges are the same purchase.
+
+Within a small transaction, validate the admitted identity and input, register
+the existing transaction-aware work signal, and insert the immutable receipt.
+Enforce the logical identity with a unique constraint. On conflict, fetch the
+committed existing receipt and compare its fingerprint; equal means replay,
+different means a conflict requiring reconciliation. Do not use `DO UPDATE`
+to replace receipt evidence. At read committed isolation, a conflict check may
+need a subsequent statement to see the winning committed row.
+
+Commit before ACK. An ACK means the evidence is durable, not that the spend view
+has already incorporated it. Preserve legacy durable spooling and native
+redelivery when the database write or its acknowledgement is uncertain.
+
+There is no read-modify-write of `usageJson`, `spentMonthlyCents`, reservation
+capacity, or a company-wide dirty flag here. The new receipt is itself durable
+pending work. Mutable delivery metadata may be attached to it by the consumer;
+its evidence fields are immutable.
+
+Batch inserts with count/byte bounds when useful. Keep lifecycle, final, unknown
+and conflicting evidence intact. Do not drop a receipt solely because its dollar
+amount matches the previous one: it may change completeness or provenance.
+
+## 3. Reduce independent partitions, then publish batches atomically
+
+Route by a stable hash of **operation ID**, not company ID. For example, a
+starting benchmark can use 64 virtual partitions per company. That number is a
+test parameter, not a capacity claim. One company with 10,000 active runs spreads
+over all partitions; one agent with many runs spreads too.
+
+Persist the layout version and partition with the operation's admission. All its
+receipts, closure and corrections follow that assignment. Changing partition
+count requires an explicit migration; do not silently rehash active operations.
+
+The minimal conceptual storage is:
+
+| Record | Write frequency and ownership |
+| --- | --- |
+| Operation identity | Once per admitted operation; immutable attribution |
+| Receipt journal | Append per receipt; unique identity; durable pending metadata |
+| Operation view | Replaced by its partition reducer, at most once per operation per batch |
+| Scope/window/partition totals | Updated in a bounded batch, never by receipt ingestion |
+| Partition publication head | One conditional generation advance per batch |
+| Budget policy and decision | Policy edits and coalesced evaluations; independent of receipt acceptance |
+| Ledger, adjustments, incidents, stop intent | Existing domain records, adapted to operation identities and idempotent writes |
+
+Some records can extend existing tables. Separate immutable identities from
+mutable projections; do not add a table for each timer or delivery stage.
+
+For a reducer batch:
+
+1. In one repeatable-read snapshot, capture partition generation `g`, a bounded
+   set of unprocessed receipt IDs, and their prior operation views.
+2. Reduce those operations and compute replacement contributions outside the
+   publishing transaction. No provider calls or filesystem work occurs during
+   publication.
+3. In one short transaction, conditionally advance the partition head from `g`
+   to `g + 1`, bulk-write operation views and aggregate deltas, insert any
+   qualified ledger settlement, and mark exactly the consumed receipt IDs.
+4. If the conditional advance loses, roll back and retry from a fresh snapshot.
+   Use bounded retries with scheduler backoff under contention.
+
+All writers of those views, including repairs, follow this publication protocol.
+A normally assigned reducer avoids duplicate work; overlapping workers or an
+old worker after failover remain safe because publication requires the captured
+generation. No durable exclusive lease is needed for correctness of this batch.
+
+An input arriving during computation remains pending. Never advance a global
+`last_seen_id` or timestamp past all smaller values: sequence allocation and
+transaction commit order differ. An earlier allocated receipt that commits
+later must still be discoverable. Retain an indexed pending predicate.
+
+This has bounded coordination per partition **per batch**. It cannot stall
+receipt ACKs behind the calculation, and different partitions do not write the
+same aggregate rows. Shared company/agent/project totals are sums of these
+partition contributions, not counters incremented by every producer.
+
+## 4. Define canonical money without depending on arrival order
+
+The reducer preserves current normalization and source-ownership rules as a
+versioned pure function over admitted evidence. Within a source, order by the
+producer sequence, not database arrival time. Across sources, follow verified
+handoff/continuation provenance rather than arbitrarily choosing a winner.
+
+Cumulative reports replace earlier reports for the same accounting component.
+Reports of $2, $3 and $5 represent $5, not $10. Distinct paid attempts contribute
+separately; session-cumulative counters require their established baseline.
+An incomplete/unknown report cannot erase known partial exposure. Final prices
+and audited corrections may legitimately reduce earlier estimates.
+
+Each operation view owns mutually exclusive contribution categories:
+
+```text
+unsettled:   ledger = 0;     live = known observed amount; hold = max(estimate, live)
+settled:     ledger = final; live = 0;                     hold = 0
+```
+
+This is a contribution model, not a declaration that unknown exposure is zero.
+Carry unknown-price, pending-terminal and unresolved-dispatch counts separately
+and apply their configured restrictions. Hosted browser work remains a separate
+paid operation from its parent run.
+
+Settlement and that operation's replacement contribution publish atomically.
+Company/agent/project views sum one contribution per operation. A reader sees
+the old or new contribution, never both the live estimate and its final charge.
+UTC-month and lifetime attribution, unsettled carryover, credits, backdated
+charges and corrections need explicit reducer rules matching the selected
+product contract; do not assign costs to the month in which a worker ran.
+
+### Finalization requires closure, not merely a "complete" checkpoint
+
+The governing execution coordinator supplies immutable closure evidence only
+after the operation cannot start another ordinary paid attempt. It references
+the admitted attempts/sources and the final usage evidence or unresolved debt.
+An adapter's `complete` usage flag alone does not establish that boundary.
+
+A closure record must prove the final cumulative coverage it names. The current
+native `sourceSeq` includes non-accounting events, so gaps between usage receipts
+are not automatically lost usage. Use explicit final references/coverage, or a
+dedicated accounting sequence if the producer contract is extended. Do not
+invent a contiguous usage sequence from today's sparse source sequence.
+
+Publish a settlement certificate and its ledger components in one transaction,
+with one base settlement per `(company_id, operation_id)` and a verified content
+fingerprint. A different closure for an already-settled operation is a conflict
+or an explicit correction, not a second base settlement.
+Only the transaction that inserts that certificate inserts its components; a
+replay verifies the existing certificate. Ledger IDs and component keys are
+stable across retries. Unknown external outcomes remain unresolved.
+
+Late evidence within the declared closure coverage is replay/recovery. Evidence
+beyond it remains durable and creates a reconciliation obligation or an explicit
+correction; it cannot silently reopen the same final charge or disappear because
+the run was marked terminal. Corrections use immutable adjustment identities.
+Existing mutable correction/reporting APIs need an adapter to this ledger model.
+
+## 5. Budget checks read; they do not reserve a shared balance
+
+For dispatch, read current company/agent/project controls and applicable budget
+decisions in one bounded read from the authoritative database, rather than an
+unbounded-lag replica or cache. A policy without a current decision must
+be evaluated or treated as pending; it is not implicitly unlimited. The decision
+records policy revision, budget window, projection generations, status and
+freshness evidence.
+
+If applicable decisions allow work, record this operation's own admission with
+the decision versions and estimated cost. Preserve duplicate-dispatch fencing
+in the execution subsystem. An admission receipt is not proof that an external
+provider call occurred exactly once; existing unknown-outcome recovery remains.
+
+There is no atomic decrement of company remaining budget. Estimate-fit checks
+can use the same snapshot, and admitted estimates can enter the asynchronous
+exposure view. Two dispatches can see the same remaining money and both start.
+That race is accepted by this best-effort contract. Reservations become recorded
+estimates/obligations, not exclusive claims on a shared spending allowance.
+
+New snapshots after a committed budget block deny additional work. Admissions
+already in flight can still race that block. Manual Stop, pause, permissions,
+issue ownership and dispatch authority retain their existing final gates.
+Keep provider-enforced cost caps where supported.
+
+## 6. Coalesce budget decisions and preserve stop authority
+
+The evaluator reads the applicable scope/window partition totals in one
+consistent snapshot, rather than scanning ledger and receipt history on each
+checkpoint. Capture the policy's configuration revision and prior decision
+revision in that same snapshot.
+
+Publish with a conditional update against both captured revisions. A policy
+edit advances its configuration revision and invalidates the previous decision.
+An older evaluator cannot overwrite a newer decision or a changed policy. A
+losing evaluator recomputes; it never retries publication using a fresh version
+number attached to stale calculated totals.
+
+Publish the decision, any incident/approval transition, and durable stop intent
+in the same transaction. Generate idempotency identities from the winning
+transition; losing or repeated evaluations produce no duplicate side effects.
+Only actual state transitions need notification/audit effects. Cancellation
+delivery remains at least once, scoped and recoverable. Already-committed
+per-run Stop obligations survive later policy changes. Resuming a budget hold
+cannot clear a manual hold.
+
+These occasional policy-row updates can contend with one another. Producers
+and admission readers do not wait on that row to record receipts or debit money.
+Do not replace them with per-receipt policy updates.
+
+## 7. Bound lag and recover work without an idle polling loop
+
+Retain the proposed healthy-load targets from the investigation: roughly
+one-second coalescing and a five-second p99 target from durable priced receipt
+to stop intent. Provider reporting and shutdown delay are additional. Targets
+require a workload qualification; none is a guarantee on final billed dollars.
+
+For scopes with spending limits, reject additional paid admissions if an
+unrepresented accounting change has been waiting over fifteen seconds. This
+check considers pending journal inputs **and** reduced contributions not yet
+covered by the decision's projection generations. Carry their oldest pending
+time through publication. Consuming the journal must not hide evaluator lag.
+An old decision with no subsequent relevant changes remains fresh while idle.
+Policy edits and UTC-window changes are explicit new evaluation obligations.
+
+Durable pending rows are truth; notifications are hints. Reuse the existing
+delivery coordinator for local coalescing, finite retries, actual deadlines,
+drain and restart. No unconditional per-company timer scans empty queues.
+
+The current DB work signals are process-local by default, with optional
+commit-bound PostgreSQL notifications. Separate reducers therefore require
+qualified cross-process wake delivery, subscribed before the startup scan and
+reconciled after reconnect. `NOTIFY` alone is not durable storage or a way to
+wake stopped compute. See the [existing signal implementation](../../packages/db/src/work-signals.ts)
+and [PostgreSQL notification behavior](https://www.postgresql.org/docs/current/sql-notify.html).
+Do not claim distributed recovery or host sleep until the actual notification,
+startup, disconnect and external-wake paths are qualified.
+
+Receipt acceptance continues while budgets are blocked. Otherwise stopping spend
+would also stop the evidence needed to settle it.
+
+## Failure cases the design must pass
+
+| Event/race | Required outcome |
+| --- | --- |
+| Same receipt delivered twice | One logical receipt; equal fingerprint accepted as replay |
+| Same identity, different content | Conflict retained/reported; no silent overwrite or second charge |
+| Sequence 8 arrives before 7 | Correct cumulative state; 7 cannot rewind it |
+| Earlier allocated receipt commits after a batch | It remains discoverable as pending work |
+| Producer loses ACK after commit | Safe redelivery with the same identity |
+| Reducer dies before or during publication | All batch effects commit together or remain replayable |
+| Two reducers compute from generation 12 | One publishes generation 13; the other recomputes |
+| Final receipt races partial receipt | Closure and canonical reduction determine one contribution, independent of arrival order |
+| Final bill is lower than live estimate | Correct signed adjustment/replacement; no double counting |
+| Source changes during recovery | Verified handoff retains prior debt; no guessed new baseline |
+| Policy is lowered during evaluation | Old policy revision cannot publish a decision |
+| Manual Stop races budget resume | Manual execution hold remains authoritative |
+| Evaluator stops while receipts continue | Evidence accumulates durably; stale-view admission protection activates |
+| No new accounting work | Reducers and evaluator become quiet |
+
+## Implementation and migration
+
+These steps implement the full target after the first milestone above; they are
+not required as one change to improve today's eligibility path.
+
+1. Add immutable operation identities and the journal acceptance boundary,
+   initially alongside the existing authoritative path. Preserve existing native
+   and legacy receipts; adapt source/closure evidence before removing their
+   current completion fences.
+2. Implement deterministic operation reduction, idempotent settlement and
+   partition publication. Run it in shadow mode with isolated projections and
+   no production ledger, stop, or incident side effects. Compare at a common
+   receipt/closure boundary after draining, not at arbitrary wall times.
+3. Implement versioned budget decisions, freshness, read-only admission and
+   transition-based stop delivery. Replace shared monthly/runtime counters with
+   partitioned projections and update all report, repair and integrity consumers.
+4. At a controlled cutover, fence the old settlement/evaluation authority, drain
+   committed inputs, seed verified ledger/adjustment baselines, and select one
+   authoritative writer path. Resume admission only with current decisions.
+   Continue accepting/replaying durable receipts through this transition.
+5. Rollback similarly requires reconciliation and an authority handoff. Do not
+   let old and new settlement writers charge the same operation independently.
+
+Owning modules are usage receipts/native usage delivery, run-cost accounting,
+costs/reconciliation/integrity, budgets and the existing enforcement worker.
+Dispatch consumes a narrow decision contract; provider normalization remains
+with adapters. Update DB/shared/API/UI contracts and the budget guarantees in
+`doc/SPEC.md`, `doc/SPEC-implementation.md` and contributor guidance before enabling
+the change. The earlier per-run-lock design is not this target's hot path.
+
+Qualification uses real PostgreSQL and multiple independent processes, both one
+large company and many companies, overlapping scope budgets, ledger-history
+growth, duplicate/reconnect bursts and the races above. Verify producer SQL has
+no shared counter/head/policy writes and no company/run `FOR UPDATE`. Audit
+foreign-key, trigger and uniqueness behavior too; string searches cannot prove
+absence of contention.
+
+Measure sustained receipts and bytes per second, ACK/admission/settlement latency,
+projection backlog age, stop latency, database waits and event-loop load. Ten
+thousand agents at one receipt per second would create 864 million receipts per
+day before batching/retention effects. Eliminating serialization does not remove
+database, WAL, index, storage or provider-capacity limits. Start with bounded
+batches and measured retention; introduce a different log substrate only when
+its need is demonstrated.
+
+Design validation in this chat: inspected the existing journal, budget incident,
+enforcement and work-signal contracts and checked PostgreSQL concurrency and
+notification semantics. This is an architecture document, not a prototype or
+load-test result. Implementation details and capacity remain to be qualified.

doc/testing-native-remote-continuity.md

@@ -19,6 +19,10 @@ same Runner and provider history through another healthy follow-up. This is a
 controlled cache fault, not a reconstruction of a historical incident.
 Each recovery case includes another healthy turn after the recovered turn and
 asserts that it does not start another runner or provider process.
+The `warm-suspended` case suspends the first turn, resumes the second turn,
+then keeps that controller warm for the third turn. It must retire the first
+turn's local suspension receipt so the third turn can reuse the resumed Runner
+and provider history without a second process or repeated turn.
 The same-server case also starts preparation lease renewal during recovery and
 holds instruction collection open past a renewal tick after coordinator release.
 The executor must acquire renewal before preparation releases its reference, then

server/src/__tests__/adapter-model-refresh-routes.test.ts

@@ -101,13 +101,15 @@ function registerModuleMocks() {
   }));
 }
 
+registerModuleMocks();
+const [{ agentRoutes }, { errorHandler }] = await Promise.all([
+  vi.importActual<typeof import("../routes/agents.js")>("../routes/agents.js"),
+  vi.importActual<typeof import("../middleware/index.js")>("../middleware/index.js"),
+]);
+
 const refreshableAdapterType = "refreshable_adapter_route_test";
 
 async function createApp() {
-  const [{ agentRoutes }, { errorHandler }] = await Promise.all([
-    vi.importActual<typeof import("../routes/agents.js")>("../routes/agents.js"),
-    vi.importActual<typeof import("../middleware/index.js")>("../middleware/index.js"),
-  ]);
   const app = express();
   app.use(express.json());
   app.use((req, _res, next) => {
@@ -159,11 +161,6 @@ async function unregisterTestAdapter(type: string) {
 
 describe("adapter model refresh route", () => {
   beforeEach(async () => {
-    vi.resetModules();
-    vi.doUnmock("../routes/agents.js");
-    vi.doUnmock("../routes/authz.js");
-    vi.doUnmock("../middleware/index.js");
-    registerModuleMocks();
     vi.clearAllMocks();
     mockCompanySkillService.listRuntimeSkillEntries.mockResolvedValue([]);
     mockCompanySkillService.resolveRequestedSkillKeys.mockResolvedValue([]);

server/src/__tests__/budgets-service.test.ts

@@ -1,7 +1,7 @@
 import { resolveExternalTestDatabaseUrl } from "./helpers/external-test-database.js";
 import { DELIVERY_QUEUES, subscribeDeliveryWork } from "../services/delivery-work-notifications.js";
 import { randomUUID } from "node:crypto";
-import { eq } from "drizzle-orm";
+import { eq, sql } from "drizzle-orm";
 import { afterAll, afterEach, beforeAll, beforeEach, describe, expect, it, vi } from "vitest";
 import {
   agents,
@@ -11,10 +11,14 @@ import {
   companies,
   costEvents,
   createDb,
+  heartbeatRuns,
+  issues,
+  nativeRunFinalizations,
   projects,
 } from "@paperclipai/db";
 import { budgetService, deliverBudgetEnforcement, hasPendingBudgetEnforcement } from "../services/budgets.ts";
 import {
+  EMBEDDED_POSTGRES_TEST_TIMEOUT_MS,
   getEmbeddedPostgresTestSupport,
   startEmbeddedPostgresTestDatabase,
 } from "./helpers/embedded-postgres.js";
@@ -40,13 +44,16 @@ describeEmbeddedPostgres("budgetService release gate enforcement", () => {
     await import("../services/agent-lifecycle.js");
     if (!externalUrl) tempDb = await startEmbeddedPostgresTestDatabase("paperclip-budgets-service-");
     db = createDb(externalUrl ?? tempDb!.connectionString);
-  }, 20_000);
+  }, EMBEDDED_POSTGRES_TEST_TIMEOUT_MS);
 
   afterEach(async () => {
     await db.delete(budgetIncidents);
     await db.delete(approvals);
     await db.delete(budgetPolicies);
     await db.delete(costEvents);
+    await db.delete(nativeRunFinalizations);
+    await db.delete(heartbeatRuns);
+    await db.delete(issues);
     await db.delete(projects);
     await db.delete(agents);
     await db.delete(companies);
@@ -118,6 +125,131 @@ describeEmbeddedPostgres("budgetService release gate enforcement", () => {
     return event!;
   }
 
+  it.each(["no policy", "healthy policy", "warning policy", "manual company pause", "manual project pause", "orphan budget pause", "recovering run"])(
+    "checks %s while another accounting writer holds the company lock",
+    async state => {
+      const { companyId, agentId, projectId } = await createBudgetFixture();
+      if (state === "healthy policy" || state === "warning policy" || state === "recovering run") {
+        const [policy] = await db.insert(budgetPolicies).values({ companyId, scopeType: "company", scopeId: companyId,
+          windowKind: "calendar_month_utc", amount: 100 }).returning();
+        if (state !== "recovering run") await insertCostEvent({ companyId, agentId, costCents: state === "warning policy" ? 80 : 20 });
+        if (state === "warning policy") {
+          const now = new Date();
+          await db.insert(budgetIncidents).values({ companyId, policyId: policy.id, scopeType: "company", scopeId: companyId,
+            metric: "billed_cents", windowKind: "calendar_month_utc",
+            windowStart: new Date(Date.UTC(now.getUTCFullYear(), now.getUTCMonth(), 1)),
+            windowEnd: new Date(Date.UTC(now.getUTCFullYear(), now.getUTCMonth() + 1, 1)),
+            thresholdType: "soft", amountLimit: 100, amountObserved: 80 });
+        }
+      }
+      if (state === "recovering run") {
+        const [issue] = await db.insert(issues).values({ companyId, projectId, title: "Recover accounting" }).returning();
+        const [run] = await db.insert(heartbeatRuns).values({ companyId, agentId, scopeKind: "issue", issueId: issue.id,
+          nativeIssueId: issue.id, runtimeMode: "native", status: "failed", finishedAt: new Date(), costAccountingPending: true }).returning();
+        await db.insert(nativeRunFinalizations).values({ companyId, issueId: issue.id, runId: run.id, phase: "observed" });
+      }
+      if (state === "manual company pause" || state === "orphan budget pause") {
+        await db.update(companies).set({ status: "paused", pauseReason: state === "orphan budget pause" ? "budget" : "manual" })
+          .where(eq(companies.id, companyId));
+      }
+      if (state === "manual project pause") {
+        await db.update(projects).set({ pausedAt: new Date(), pauseReason: "manual" }).where(eq(projects.id, projectId));
+      }
+
+      const reader = createDb(externalUrl ?? tempDb!.connectionString, { maxConnections: 1 });
+      let acquired!: () => void, release!: () => void;
+      const lockAcquired = new Promise<void>(resolve => { acquired = resolve; });
+      const holdLock = new Promise<void>(resolve => { release = resolve; });
+      const writer = db.transaction(async tx => {
+        await tx.select({ id: companies.id }).from(companies).where(eq(companies.id, companyId)).for("no key update");
+        acquired();
+        await holdLock;
+      });
+      try {
+        await Promise.race([lockAcquired, writer]);
+        await reader.execute(sql`set lock_timeout = '500ms'`);
+        if (state === "recovering run") {
+          await expect(budgetService(reader).getInvocationBlock(companyId, agentId, { projectId }))
+            .rejects.toMatchObject({ cause: { code: "55P03" } });
+        } else {
+          const block = await budgetService(reader).getInvocationBlock(companyId, agentId, { projectId });
+          if (state === "manual company pause") {
+            expect(block).toMatchObject({ scopeType: "company", scopeId: companyId,
+              reason: "Company is paused and cannot start new work." });
+          } else if (state === "manual project pause") {
+            expect(block).toMatchObject({ scopeType: "project", scopeId: projectId,
+              reason: "Project is paused and cannot start new work." });
+          } else if (state === "orphan budget pause") {
+            expect(block).toMatchObject({ scopeType: "company", scopeId: companyId,
+              reason: "Budget pause requires a policy or an explicit operator resume." });
+          } else expect(block).toBeNull();
+          if (state === "warning policy") {
+            expect(await reader.select().from(budgetIncidents).where(eq(budgetIncidents.companyId, companyId)))
+              .toMatchObject([{ thresholdType: "soft", status: "open", resolvedAt: null }]);
+          }
+          if (state === "no policy") {
+            await expect(budgetService(reader).evaluateAdmission(companyId, agentId, { projectId }))
+              .rejects.toMatchObject({ cause: { code: "55P03" } });
+          }
+        }
+      } finally {
+        release();
+        try { await writer; }
+        finally { await reader.$client.end({ timeout: 1 }); }
+      }
+      if (state === "recovering run") {
+        const cancelWorkForScope = vi.fn();
+        expect(await budgetService(db, { cancelWorkForScope }).getInvocationBlock(companyId, agentId, { projectId }))
+          .toMatchObject({ scopeType: "company", scopeId: companyId,
+            reason: "New work must wait while a native run recovers its unfinished accounting." });
+        expect(cancelWorkForScope).not.toHaveBeenCalled();
+        expect((await db.select().from(companies).where(eq(companies.id, companyId)))[0])
+          .toMatchObject({ status: "active", pauseReason: null });
+        expect((await db.select().from(agents).where(eq(agents.id, agentId)))[0])
+          .toMatchObject({ status: "active", pauseReason: null });
+        expect(await db.select().from(heartbeatRuns).where(eq(heartbeatRuns.companyId, companyId)))
+          .toMatchObject([{ status: "failed", costAccountingPending: true }]);
+        expect(await db.select().from(nativeRunFinalizations).where(eq(nativeRunFinalizations.companyId, companyId)))
+          .toMatchObject([{ phase: "observed", resultId: null }]);
+      }
+    },
+  );
+
+  it.each([true, false])("releases an agent budget hold when its policy is healthy (active: %s)", async isActive => {
+    const { companyId, agentId } = await createBudgetFixture();
+    await db.insert(budgetPolicies).values({ companyId, scopeType: "agent", scopeId: agentId,
+      windowKind: "calendar_month_utc", amount: 100, isActive });
+    await db.update(agents).set({ status: "paused", lifecycleState: "paused", pauseReason: "budget", pausedAt: new Date(), lifecycleHolds: ["budget"] })
+      .where(eq(agents.id, agentId));
+
+    expect(await budgetService(db).getInvocationBlock(companyId, agentId)).toBeNull();
+    expect((await db.select().from(agents).where(eq(agents.id, agentId)))[0]).toMatchObject({
+      lifecycleState: "resuming", pauseReason: null, lifecycleHolds: [],
+    });
+  });
+
+  it.each(["hard", "soft"])("resolves an obsolete %s incident even when the scope no longer has a budget hold", async thresholdType => {
+    const { companyId, agentId } = await createBudgetFixture();
+    const [policy] = await db.insert(budgetPolicies).values({ companyId, scopeType: "agent", scopeId: agentId,
+      windowKind: "calendar_month_utc", amount: 100 }).returning();
+    const [incident] = await db.insert(budgetIncidents).values({ companyId, policyId: policy.id, scopeType: "agent", scopeId: agentId,
+      metric: "billed_cents", windowKind: "calendar_month_utc", windowStart: new Date(0), windowEnd: new Date(1),
+      thresholdType, amountLimit: 100, amountObserved: 100 }).returning();
+
+    expect(await budgetService(db).getInvocationBlock(companyId, agentId)).toBeNull();
+    expect((await db.select().from(budgetIncidents).where(eq(budgetIncidents.id, incident.id)))[0])
+      .toMatchObject({ status: "resolved", resolvedAt: expect.any(Date) });
+  });
+
+  it("rejects another company's agent or project before allowing a no-policy invocation", async () => {
+    const own = await createBudgetFixture();
+    const other = await createBudgetFixture();
+    const service = budgetService(db);
+    await expect(service.getInvocationBlock(own.companyId, other.agentId)).rejects.toThrow("Budget scope not found");
+    await expect(service.getInvocationBlock(own.companyId, own.agentId, { projectId: other.projectId }))
+      .rejects.toThrow("Budget scope not found");
+  });
+
   it("raises one soft incident per window before hard-stopping and preserving safe local audit details", async () => {
     const { companyId, agentId } = await createBudgetFixture();
     const cancelWorkForScope = vi.fn().mockResolvedValue(undefined);

server/src/__tests__/helpers/native-restart.ts

@@ -16,6 +16,7 @@ import { prepareNativeHeartbeatRun } from "../../services/native-runtime/prepare
 import { createIsolatedTestHome } from "./isolated-test-home.js";
 import { type NativeControllerIdentity } from "../../services/native-runtime/native-restart-recovery.js";
 import { waitForPendingRunFailureReports } from "../../services/run-failure-report.js";
+import { registerLocalRuntimeApiServer } from "../../runtime-api.js";
 
 // An explicitly supplied disposable database also permits qualification on hosts
 // where native Postgres cannot allocate semaphores. Never fall back silently.
@@ -106,6 +107,7 @@ export function nativeRestartFixture() {
         : new Date(),
     };
     server = createServer();
+    registerLocalRuntimeApiServer(server);
     await new Promise<void>((resolveListen) =>
       server!.listen(0, "127.0.0.1", resolveListen),
     );

server/src/services/budget-tracing.ts

@@ -3,7 +3,7 @@ import { withDatabaseOperation } from "@paperclipai/db";
 
 type BudgetOperation = "budget.accounting.lock_wait" | "budget.accounting.recover"
   | "budget.accounting.settlement" | "budget.accounting.transaction"
-  | "budget.admission.evaluate" | "budget.admission.reserve"
+  | "budget.admission.evaluate" | "budget.admission.inspect" | "budget.admission.reconcile" | "budget.admission.reserve"
   | "budget.cancellation.deliver" | "budget.enforcement.deliver"
   | "budget.enforcement.scope" | "budget.exposure.evaluate";
 

server/src/services/budgets.ts

@@ -34,7 +34,7 @@ import type {
 } from "@paperclipai/shared";
 import { HttpError, notFound, unprocessable } from "../errors.js";
 import { logger } from "../middleware/logger.js";
-import { withAccountingTransaction } from "./accounting-transaction.js";
+import { withAccountingReadSnapshot, withAccountingTransaction } from "./accounting-transaction.js";
 import { logActivity, type LogActivityInput, type ActivityPublication } from "./activity-log.js";
 import { computeObservedSpend, computeObservedSpends, type ObservedSpend } from "./budget-exposure.js";
 import { notifyDeliveryWork, DELIVERY_QUEUES } from "./delivery-work-notifications.js";
@@ -198,11 +198,66 @@ export function budgetPoliciesForRun(companyId: string, agentId: string | null,
   );
 }
 
+type AdmissionContext = { issueId?: string | null; projectId?: string | null };
+
+async function readAdmissionScopes(db: Db, companyId: string, agentId: string | null, context?: AdmissionContext) {
+  const scopes: Array<{ scopeType: BudgetScopeType; scopeId: string }> = [
+    { scopeType: "company", scopeId: companyId }, ...(agentId ? [{ scopeType: "agent" as const, scopeId: agentId }] : []),
+    ...(context?.projectId ? [{ scopeType: "project" as const, scopeId: context.projectId }] : []),
+  ];
+  const records = await Promise.all(scopes.map(scope => resolveScopeRecord(db, scope.scopeType, scope.scopeId)));
+  if (records.some(scope => scope.companyId !== companyId)) throw notFound("Budget scope not found");
+  const policies = await db.select().from(budgetPolicies).where(and(eq(budgetPolicies.companyId, companyId),
+    budgetPoliciesForRun(companyId, agentId, context?.projectId ?? null)));
+  return { scopes, records, policies };
+}
+
+function scopePauseBlock(scopeType: BudgetScopeType, scopeId: string, scope: ScopeRecord, hasPolicies: boolean) {
+  if (scopeType !== "agent" && scope.paused && scope.pauseReason !== "budget") {
+    return { scopeType, scopeId, scopeName: scope.name, reason: `${scopeType === "company" ? "Company" : "Project"} is paused and cannot start new work.` };
+  }
+  if (scope.paused && scope.pauseReason === "budget" && !hasPolicies) {
+    return { scopeType, scopeId, scopeName: scope.name, reason: "Budget pause requires a policy or an explicit operator resume." };
+  }
+  return null;
+}
+
+async function inspectInvocationEligibility(db: Db, companyId: string, agentId: string | null, context?: AdmissionContext) {
+  const { scopes, records, policies } = await readAdmissionScopes(db, companyId, agentId, context);
+  if (policies.length > 0) {
+    if (records.some(scope => scope.budgetHeld)) return { needsReconciliation: true } as const;
+    const observations = await computeObservedSpends(db, policies);
+    if (policies.some((policy, index) => observedBlocks(policy, observations[index])
+      || (policy.isActive && policy.hardStopEnabled && policy.amount > 0 && observations[index].recoveringRunCount > 0))) {
+      return { needsReconciliation: true } as const;
+    }
+    const incidents = await db.select({ policyId: budgetIncidents.policyId, thresholdType: budgetIncidents.thresholdType })
+      .from(budgetIncidents).where(and(eq(budgetIncidents.companyId, companyId),
+        inArray(budgetIncidents.policyId, policies.map(policy => policy.id)), eq(budgetIncidents.status, "open")));
+    const byId = new Map(policies.map((policy, index) => [policy.id, { policy, observed: observations[index] }]));
+    if (incidents.some(incident => {
+      const { policy, observed } = byId.get(incident.policyId)!;
+      return incident.thresholdType === "hard" || shouldResolveSoftIncident(policy, observed);
+    })) return { needsReconciliation: true } as const;
+  }
+  for (const [index, { scopeType, scopeId }] of scopes.entries()) {
+    const block = scopePauseBlock(scopeType, scopeId, records[index],
+      policies.some(policy => policy.scopeType === scopeType && policy.scopeId === scopeId));
+    if (block) return { needsReconciliation: false, block } as const;
+  }
+  return { needsReconciliation: false, block: null } as const;
+}
+
 function unpricedUsageBlocks(policy: PolicyRow, observed: ObservedSpend) {
   return policy.unpricedUsagePolicy !== "allow"
     && (observed.unpricedEventCount > 0 || observed.unknownBrowserRunCount > 0);
 }
 
+function shouldResolveSoftIncident(policy: PolicyRow, observed: ObservedSpend) {
+  return !policy.isActive || !policy.notifyEnabled
+    || compareCents(observed.totalExact, Math.ceil(policy.amount * policy.warnPercent / 100)) < 0;
+}
+
 function observedBlocks(policy: PolicyRow, observed: Awaited<ReturnType<typeof computeObservedSpend>>) {
   if (!policy.isActive || !policy.hardStopEnabled || policy.amount <= 0) return false;
   return compareCents(observed.totalExact, policy.amount) >= 0 || (observed.pendingRunCount - observed.recoveringRunCount) > 0 || unpricedUsageBlocks(policy, observed);
@@ -654,18 +709,12 @@ export function budgetServiceInTransaction(db: Db, publications: ActivityPublica
       // A closed UTC window, inactive policy, or budget raise releases only
       // budget-owned pauses. Manual pauses remain operator decisions.
       await resolveOpenIncidentsForPolicy(policy.id, policy.companyId, "approved", "budget_service", "hard");
-      if (!policy.isActive || !policy.notifyEnabled || compareCents(observed.totalExact, Math.ceil(policy.amount * policy.warnPercent / 100)) < 0) {
+      if (shouldResolveSoftIncident(policy, observed)) {
         await resolveOpenIncidentsForPolicy(policy.id, policy.companyId, "approved", "budget_service", "soft");
       }
     }
-    // Agent status has its own admission gate and error code. Preserve a
-    // manual pause without misclassifying it as a budget failure.
-    if (scopeType !== "agent" && scope.paused && scope.pauseReason !== "budget") {
-      return { scopeType, scopeId, scopeName: scope.name, reason: `${scopeType === "company" ? "Company" : "Project"} is paused and cannot start new work.` };
-    }
-    if (scope.paused && scope.pauseReason === "budget" && !policies.length) {
-      return { scopeType, scopeId, scopeName: scope.name, reason: "Budget pause requires a policy or an explicit operator resume." };
-    }
+    const pauseBlock = scopePauseBlock(scopeType, scopeId, scope, policies.length > 0);
+    if (pauseBlock) return pauseBlock;
     if (scope.budgetHeld && policies[0]) await resumeScopeFromBudget(policies[0]);
     if (recoveringPolicy) return { scopeType, scopeId, scopeName: scope.name,
       reason: "New work must wait while a native run recovers its unfinished accounting." };
@@ -748,14 +797,7 @@ export function budgetServiceInTransaction(db: Db, publications: ActivityPublica
 
   async function evaluateAdmissionInTransaction(companyId: string, agentId: string | null,
     context?: { issueId?: string | null; projectId?: string | null }) {
-    const scopes: Array<{ scopeType: BudgetScopeType; scopeId: string }> = [
-      { scopeType: "company", scopeId: companyId }, ...(agentId ? [{ scopeType: "agent" as const, scopeId: agentId }] : []),
-      ...(context?.projectId ? [{ scopeType: "project" as const, scopeId: context.projectId }] : []),
-    ];
-    const records = await Promise.all(scopes.map(scope => resolveScopeRecord(db, scope.scopeType, scope.scopeId)));
-    if (records.some(scope => scope.companyId !== companyId)) throw notFound("Budget scope not found");
-    const policies = await db.select().from(budgetPolicies).where(and(eq(budgetPolicies.companyId, companyId),
-      budgetPoliciesForRun(companyId, agentId, context?.projectId ?? null)));
+    const { scopes, records, policies } = await readAdmissionScopes(db, companyId, agentId, context);
     const observations = await computeObservedSpends(db, policies);
     const reservationIndexes = policies.map((policy, i) => policy.isActive && policy.hardStopEnabled && policy.metric === "billed_cents" && policy.amount > 0 ? i : -1).filter(i => i >= 0);
     const admission = (block: Awaited<ReturnType<typeof reconcileScope>>) => ({ block,
@@ -1188,8 +1230,13 @@ export function budgetService(db: Db, _hooks: BudgetServiceHooks = {}) {
     upsertPolicy: (companyId: string, input: BudgetPolicyUpsertInput, actorUserId: string | null) =>
       mutate(companyId, service => service.upsertPolicy(companyId, input, actorUserId)),
     evaluateCostEvent: (event: typeof costEvents.$inferSelect) => mutate(event.companyId, service => service.evaluateCostEvent(event)),
-    getInvocationBlock: (companyId: string, agentId: string | null, context?: { issueId?: string | null; projectId?: string | null }) =>
-      mutate(companyId, service => service.getInvocationBlock(companyId, agentId, context)),
+    getInvocationBlock: async (companyId: string, agentId: string | null, context?: AdmissionContext) => {
+      const eligibility = await withBudgetSpan("budget.admission.inspect", () =>
+        withAccountingReadSnapshot(db, companyId, tx => inspectInvocationEligibility(tx, companyId, agentId, context)));
+      if (!eligibility.needsReconciliation) return eligibility.block;
+      return withBudgetSpan("budget.admission.reconcile", () =>
+        mutate(companyId, service => service.getInvocationBlock(companyId, agentId, context)));
+    },
     resolveIncident: (companyId: string, incidentId: string, input: BudgetIncidentResolutionInput, actorUserId: string) =>
       mutate(companyId, service => service.resolveIncident(companyId, incidentId, input, actorUserId)),
   };

server/src/services/native-runtime/native-runner-restart-recovery.integration.test.ts

@@ -150,12 +150,13 @@ describeEmbeddedPostgres("native runner restart recovery with real processes", (
     expect(next.kind).toBe("bootstrap_incomplete");
   });
 
-  it.skipIf(!process.env.PAPERCLIP_TEST_RUNNER_CONTAINER).each(["warm", "hard", "stale-projection", "legacy", "changed", "same-server", "diskless", "diskless-per-turn", "diskless-per-turn-read-retry", "diskless-warm-suspended", "diskless-warm-suspended-read-retry", "diskless-live-inspection"])("preserves remote warm continuity across completed turns: %s", async (scenario) => {
-    const hardRestart = !["warm", "same-server"].includes(scenario);
+  it.skipIf(!process.env.PAPERCLIP_TEST_RUNNER_CONTAINER).each(["warm", "hard", "warm-suspended", "stale-projection", "legacy", "changed", "same-server", "diskless", "diskless-per-turn", "diskless-per-turn-read-retry", "diskless-warm-suspended", "diskless-warm-suspended-read-retry", "diskless-live-inspection"])("preserves remote warm continuity across completed turns: %s", async (scenario) => {
+    const hardRestart = !["warm", "same-server", "warm-suspended"].includes(scenario);
     const diskless = scenario.startsWith("diskless");
     const perTurn = scenario.startsWith("diskless-per-turn");
-    const warmSuspended = scenario.startsWith("diskless-warm-suspended");
+    const warmSuspended = scenario === "warm-suspended" || scenario.startsWith("diskless-warm-suspended");
     const resumedAfterSuspension = perTurn || warmSuspended;
+    const suspendedBeforeThird = perTurn || (diskless && warmSuspended);
     const container = process.env.PAPERCLIP_TEST_RUNNER_CONTAINER!;
     const tunnel = await createRunnerTestTlsTunnel({ port: Number(process.env.PAPERCLIP_TEST_RUNNER_PORT),
       key: process.env.PAPERCLIP_TEST_RUNNER_TLS_KEY!, cert: process.env.NODE_EXTRA_CA_CERTS! });
@@ -310,6 +311,8 @@ describeEmbeddedPostgres("native runner restart recovery with real processes", (
       };
       const first = hardRestart ? await executeInController(firstOptions) : await executeFirst();
       expect(first.exitCode).toBe(0);
+      if (scenario === "warm-suspended") expect(await closeIdleWarmNativeSessionsForRestart())
+        .toEqual({ closed: 1, busy: 0, failed: 0 });
       await finalizeNativeRun({ db: fixture.db, runId: fixture.runId, workspaceFinalizeStatus: "succeeded", projectRunStatus: true });
       const snapshot = await runner.execute({ command: "sh", args: ["-c", `cat '${remoteCwd}'/.paperclip-runtime/paperclip-runner/sessions/*/runner/runner-state.json`] });
       expect(snapshot.exitCode).toBe(0);
@@ -476,9 +479,9 @@ describeEmbeddedPostgres("native runner restart recovery with real processes", (
       const finalCalls = await runner.execute({ command: "cat", args: [`${remoteCwd}/provider-calls.log`] });
       expect(finalCalls.stdout.split("\n").filter(line => line === "turn/start")).toHaveLength(3);
       expect(finalCalls.stdout.split("\n").filter(line => line === "thread/start")).toHaveLength(1);
-      expect(finalCalls.stdout.split("\n").filter(line => line === "process-start")).toHaveLength(providerStarts + (resumedAfterSuspension ? 1 : 0));
+      expect(finalCalls.stdout.split("\n").filter(line => line === "process-start")).toHaveLength(providerStarts + (suspendedBeforeThird ? 1 : 0));
       const finalProcess = await runner.execute({ command: "sh", args: ["-c", `cat '${remoteCwd}'/.paperclip-runtime/paperclip-runner/sessions/*/runner/runner-process.identity`] });
-      if (resumedAfterSuspension) expect(finalProcess.stdout).not.toBe(processAfter.stdout);
+      if (suspendedBeforeThird) expect(finalProcess.stdout).not.toBe(processAfter.stdout);
       else expect(finalProcess.stdout).toBe(processAfter.stdout);
       expect(await closeIdleWarmNativeSessionsForRestart()).toEqual({ closed: perTurn ? 0 : 1, busy: 0, failed: 0 });
       const suspended = await runner.execute({ command: "sh", args: ["-c", `cat '${remoteCwd}'/.paperclip-runtime/paperclip-runner/sessions/*/runner/runner-state.json`] });

server/src/services/native-runtime/native-session-executor.test.ts

@@ -12967,9 +12967,10 @@ describe("runnerd provider runtime wiring", () => {
     },
   );
 
-  it.each((["codex", "claude"] as const).flatMap(provider => (["unchanged", "external-checkpoint", "changed", "runtime-context", "retirement-failed", "stale-owner", "active-owner", "missing-owner", "foreign-environment", "plugin-replaced", "plugin-managed-github"] as const).filter(scenario => provider === "codex" || scenario.startsWith("plugin-")).map(scenario => [provider, scenario] as const)))("rotates an owned warm %s remote runner before checking changed configuration: %s", async (provider, scenario) => {
+  it.each((["codex", "claude"] as const).flatMap(provider => (["unchanged", "external-checkpoint", "suspension-receipt", "suspension-active-owner", "changed", "runtime-context", "retirement-failed", "stale-owner", "active-owner", "missing-owner", "foreign-environment", "plugin-replaced", "plugin-managed-github"] as const).filter(scenario => provider === "codex" || scenario.startsWith("plugin-")).map(scenario => [provider, scenario] as const)))("rotates an owned warm %s remote runner before checking changed configuration: %s", async (provider, scenario) => {
     const plugin = scenario === "plugin-replaced" || scenario === "plugin-managed-github";
-    const reusable = ["unchanged", "external-checkpoint", "plugin-managed-github"].includes(scenario);
+    const suspensionReceipt = scenario.startsWith("suspension-");
+    const reusable = suspensionReceipt || ["unchanged", "external-checkpoint", "plugin-managed-github"].includes(scenario);
     const { permissionPolicy: _permissionPolicy, ...claudeProfile } = QUALIFIED_ACPX_PROFILES.claude;
     const first = { ...scopedNativeExecution(`warm-admission-${provider}-${scenario}`),
       schema: "paperclip.native-execution-input.v5" as const,
@@ -12978,7 +12979,7 @@ describe("runnerd provider runtime wiring", () => {
     if (provider === "claude") first.session.driverKind = "acpx_runtime";
     first.session.lifecyclePolicy = { mode: "warm", idleTimeoutMs: 60_000 };
     const next = { ...first, binding: { ...first.binding, runId: `${first.binding.runId}-next` },
-      session: { ...first.session, lifecyclePolicy: { mode: "warm" as const, idleTimeoutMs: ["unchanged", "external-checkpoint", "runtime-context", "stale-owner", "missing-owner", "plugin-replaced", "plugin-managed-github"].includes(scenario) ? 60_000 : 90_000 } } };
+      session: { ...first.session, lifecyclePolicy: { mode: "warm" as const, idleTimeoutMs: suspensionReceipt || ["unchanged", "external-checkpoint", "runtime-context", "stale-owner", "missing-owner", "plugin-replaced", "plugin-managed-github"].includes(scenario) ? 60_000 : 90_000 } } };
     if (scenario === "runtime-context") {
       next.runtimeContext = { ...first.runtimeContext,
         mcp: { ...first.runtimeContext.mcp, digest: "1".repeat(64) } };
@@ -13084,11 +13085,17 @@ describe("runnerd provider runtime wiring", () => {
     const ordinarySelect = nextDb.select.bind(nextDb);
     nextDb.select = ((fields?: Record<string, unknown>) => fields?.status === heartbeatRuns.status
       && fields?.runnerProfileJson === heartbeatRuns.runnerProfileJson
-      ? priorNativeRunDb(scenario === "stale-owner" ? older : first, scenario === "active-owner" ? "running" : "succeeded").select(fields as never)
+      ? priorNativeRunDb(scenario === "stale-owner" ? older : first, ["active-owner", "suspension-active-owner"].includes(scenario) ? "running" : "succeeded").select(fields as never)
       : ordinarySelect(fields as never)) as Db["select"];
     try {
       await executePaperclipNativeSession({ db: firstDb, execution: first,
         runnerInstanceId: identity.runnerInstanceId, useRunnerd: true, runnerExecutionTarget: pluginTarget as never,...pluginOptions });
+      if (suspensionReceipt) {
+        // Remote suspension leaves a local identity receipt, while the Runner
+        // owns its full state in the sandbox. A resumed turn must retire it.
+        await writeFile(join(root, "plugin-runner-state.json"), JSON.stringify(durableRunnerState(identity, "suspended")));
+        await rm(join(root, "runner/runner-state.json"));
+      }
       if (scenario === "stale-owner") await writeRunnerAuthority(root, { ...identity, runId: older.binding.runId }, "ready");
       if (scenario === "missing-owner") {
         await closeWarmNativeSessionsForEnvironment({ environmentId, reason: "fixture owner loss" });
@@ -13101,12 +13108,14 @@ describe("runnerd provider runtime wiring", () => {
         runnerInstanceId: `runner-next-${scenario}`, useRunnerd: true,
         runnerExecutionTarget: (scenario === "foreign-environment" ? { ...target, environmentId: `${environmentId}-other` }
           : plugin ? {...pluginTarget,leaseId:"next-core-lease"} : target) as never,...pluginOptions });
-      if (["stale-owner", "active-owner", "missing-owner", "foreign-environment"].includes(scenario)) {
-        await expect(continuation).rejects.toThrow(scenario === "active-owner"
+      if (["stale-owner", "active-owner", "suspension-active-owner", "missing-owner", "foreign-environment"].includes(scenario)) {
+        await expect(continuation).rejects.toThrow(["active-owner", "suspension-active-owner"].includes(scenario)
           ? "prior_owner_active" : "prior_owner_terminal_state_indeterminate");
         expect(state.execute).toHaveBeenCalledOnce();
         expect(firstSession.close).not.toHaveBeenCalled();
         expect(order).toEqual([]);
+        if (suspensionReceipt) expect(JSON.parse(await readFile(join(root, "plugin-runner-state.json"), "utf8")))
+          .toMatchObject({ ...identity, lifecycle: "suspended" });
       } else if (scenario === "retirement-failed") {
         await expect(continuation).rejects.toThrow("retirement unconfirmed");
         expect(state.execute).toHaveBeenCalledOnce();
@@ -13120,8 +13129,9 @@ describe("runnerd provider runtime wiring", () => {
         }
         if (scenario === "plugin-managed-github") expect(rebind).toHaveBeenCalledOnce();
         expect(order).toEqual(reusable ? ["execute-next"] : ["retire", "execute-next"]);
-        if (["unchanged", "external-checkpoint"].includes(scenario)) {
+        if (["unchanged", "external-checkpoint", "suspension-receipt"].includes(scenario)) {
           if (!persistedController) await writeRunnerAuthority(root, { ...identity, runId: next.binding.runId }, "ready");
+          if (suspensionReceipt) await rm(join(root, "runner/runner-state.json"));
           const third = { ...next, binding: { ...next.binding, runId: `${next.binding.runId}-third` } };
           const thirdDb = leaseDb(third);
           const thirdSelect = thirdDb.select.bind(thirdDb);
@@ -13139,6 +13149,11 @@ describe("runnerd provider runtime wiring", () => {
             runnerExecutionTarget: pluginTarget as never,...pluginOptions })).resolves.toMatchObject({ exitCode: 0 });
           expect(firstSession.close).not.toHaveBeenCalled();
           expect(state.execute).toHaveBeenCalledTimes(3);
+          if (suspensionReceipt) {
+            expect(existsSync(join(root, "plugin-runner-state.json"))).toBe(false);
+            expect(JSON.parse(await readFile(join(root, "control-plane/control-plane-state.json"), "utf8")).identity)
+              .toMatchObject({ ...identity, runId: next.binding.runId });
+          }
           if (restore) {
             expect(restore.mock.calls.map(([, binding]) => binding.runId)).toEqual([first.binding.runId, next.binding.runId]);
             expect(existsSync(root)).toBe(true);

server/src/services/native-runtime/native-session-executor.ts

@@ -9844,7 +9844,8 @@ async function executePaperclipNativeSessionWithinScope(
               );
             },
             onSession: async (session) => {
-              if (session && usesPluginTask) {
+              if (session && input.runnerExecutionTarget?.kind === "remote"
+                && input.runnerExecutionTarget.transport === "sandbox") {
                 rmSync(resolve(scopedRunnerdStateRoot(input.execution), "plugin-runner-state.json"), { force: true });
               }
               if (session && session === existingWarmSession) remoteWarmRecoveryPending = false;