## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - Agents in a `sandbox` environment run each task in a paid remote
sandbox that an environment lease tracks
> - When a run stops, the control plane must release that lease, and
terminating an agent waits until every lease of the agent is released
> - In a load test, a run stopped after its sandbox was ready but before
its runner started never released its lease, so the sandbox kept running
and the agent stayed in "terminating" forever
> - In the same test, terminating an agent whose failed run had a
scheduled same-run retry waited about 18 minutes, until the stale retry
time
> - This pull request makes lease release finish for every stop window,
records any release that must wait with a retry time, adds a hint-driven
safety net, and makes termination cancel a pending native retry at once
> - The benefit is that a stopped run never leaves a paid sandbox
running, and agent termination finishes in seconds without manual repair
## Linked Issues or Issue Description
**What happened?**
A load test with ten OpenCode agents in a `sandbox` environment stopped
its agents mid-run (pause, then terminate). Two defects appeared.
1. A run cancelled while it selected its provider pack (after the
sandbox was acquired and the workspace was prepared, before the runner
started) kept its environment lease `active` for more than 40 minutes.
The agent stayed in lifecycle `terminating`. The lifecycle worker
retried every pass and wrote `updatedAt` each time with no recorded
reason. Delete returned 409 "Complete termination before deleting the
agent".
2. A failed native run with a scheduled bootstrap retry kept its lease
until the stale retry time. Agent termination took about 18 minutes. The
run still showed "retry_scheduled" after the agent was terminated.
**Expected behavior**
A stopped run releases its sandbox lease within seconds. If release must
wait, the wait has a recorded reason and retry time, and a background
pass finishes it. Terminating an agent cancels a pending retry, releases
its lease, and reaches `terminated` promptly.
**Steps to reproduce**
1. Start a native run in a `sandbox` environment.
2. Stop it after the sandbox is acquired and the workspace is prepared,
before the runner starts.
3. Pause and then terminate the agent.
4. See that the lease stays `active` and the agent stays `terminating`.
**Paperclip version or commit**
Reproduced on `master` at 62033b9994 (load test build 2026.1106.0+2564).
**Relevant logs or output**
Root causes found in the code:
- `releaseEnvironmentLeasesForRun` returned with no record when
workspace retention refused a cancelled run. The refusal was permanent
when the run's native finalization owner had expired but was not
cleared. The owner is cleared only after `agentExecutionsHaveStopped`
sees every lease released, so the two waited on each other.
- `releaseRunLeases` silently skipped a lease whose prepared native
workspace was still "pending export", even for a run that never started
a runner and so never changed those files.
- The destroy path deferred on any busy warm session in the same
environment, also with no record.
- `stopInvocationsForAgents` returned on the first run that was still
live, so it did not release the leases of the other runs of the agent.
- The agent stop cancelled only `queued`, `running` and
`scheduled_retry` runs. A failed native run with a pending same-run
retry kept its lease until that retry fired.
- The lifecycle worker deferred a pending host step by 2 seconds every
pass, with no reason and no backoff.
## What Changed
- `releaseEnvironmentLeasesForRun` (`run-control.ts`):
- A native run with no runner launch evidence (no
`native.process_start_requested`, no runner event, no process id) skips
workspace retention and warm-session closing. No runner process exists
to fence, and no work ran against the prepared files.
- When retention is refused only because the run's native finalization
owner lease has expired, the release clears that expired owner
(compare-and-set on the exact expired owner of a cancelled run with no
result) and retries retention once. A recorded recovery then stops and
retains the sandbox, so the files stay recoverable.
- Every release that cannot finish now records
`metadata.releaseDeferral` (`reason`, `attempts`, `retryAt`,
disposition) on the active lease and publishes an `orphan-active-leases`
hint. The retry time backs off from 30 seconds to 15 minutes, then
retries hourly without end. The record keeps `failingSince` and the last
reason. A new Stop or terminate operation, or a server restart, triggers
an immediate attempt.
- `nativeWorkspaceExportPendingForLease` treats a cancelled run with no
runner launch evidence as owing no export.
- Safety net: `sweepOrphanedActiveLeases` (the existing hint- and
deadline-driven orphan-lease work) now retries the run's own release for
a due deferral or for a cancelled native run, after the native owner
lease has expired. The orphan-lease loader uses the stored `retryAt` as
the due time. No new timer or poller.
- Agent stop:
- It cancels a failed native run that has a pending same-run retry (or
an interrupted cancellation of one), through the existing
retry-cancellation path. A pause stamps `pauseStop` on those runs too,
so a budget resume can continue them.
- It releases the leases of every terminal run, and it skips a lease
whose deferral is not yet due.
- `agentInvocationStopBlocker` and `agentExecutionStopBlocker` return
why the stop must still wait.
- Lifecycle worker: a pending host or plugin step records
`pendingReason` and `pendingPasses` on the operation, and its retry
backs off from 2 seconds to a 60 second cap. When a release leaves a run
with no unsettled lease, the stopping agent's saved retry time is
cleared and the lifecycle queue is notified, so Pause or Terminate
finishes on the next pass.
- Cancelled-workspace preservation: a stopped run that never started a
runner has its prepared reference moved to
`nativeWorkspaceSyncReleasedUnchanged` when its lease is released, so
the preservation worker no longer selects it. The worker records the
same settlement itself for older runs, and waits on a live owner's exact
lease expiry.
## Verification
- Red on `master`, green on this branch (embedded PostgreSQL, fake
plugin sandbox provider):
-
`server/src/__tests__/heartbeat-lease-release-before-runner-attach.test.ts`:
agent pause then terminate, and board Stop, for each window between
sandbox acquire and runner attach (artifact discovery, pack select, pack
upload, runner launch, transport connect, session open). Each test
checks that the provider gets a stop or destroy call, the lease is
released, the run is cancelled, and the agent reaches `terminated`
within one extra worker pass. It also checks the orphan-lease safety net
for a leaked lease and a recorded deferral that the background pass
finishes when due.
-
`server/src/__tests__/heartbeat-terminate-pending-native-retry.test.ts`:
terminating an agent whose failed run has a bootstrap retry 17 minutes
away cancels the retry, releases the lease and reaches `terminated` at
once. The test covers an open task and a cancelled task.
-
`server/src/modules/agent-lifecycle/__tests__/pending-backoff.test.ts`:
a host step that stays pending records its reason, and its retry delay
grows to a 60 second cap.
- Related suites pass: run-control, recovery, orphaned active lease
sweep, environment runtime, agent lifecycle, native workspace
preservation and recovery, held comment wake.
- `server` typecheck passes.
## Risks
- A cancelled native run with no launch evidence now gets its prepared
sandbox destroyed instead of kept. The proof is the same launch evidence
that restart recovery uses.
- A launched cancelled run whose retention owes nothing is stopped and
retained, as before. A retention that throws (lock or write error, or a
live owner) now keeps the lease active with a recorded reason and retry
time, instead of a silent skip.
- Release now clears an expired native finalization owner of a cancelled
run. An old controller whose lease already expired then fails its next
owner check and stops, which is the normal lease rule.
- On the first start after deploy, the orphan-lease work releases leases
already leaked by cancelled native runs.
- Task state after pause then terminate: the run-scoped stop keeps the
task status (an open task stays open, a cancelled task stays cancelled),
and a run that held prepared files gets the existing workspace recovery
record on its stopped sandbox. No new state is added.
## Model Used
- Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, used through Claude
Code with tool use (code search, shell, test runs).
## Checklist
- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Feature modules
A feature module under server/src/modules/<name>/ uses three layers. Each
layer has one rule: which layer below it, it may import from.
adapters → application → domain
domain/holds pure business rules. A domain file takes plain data in and returns plain data out. A domain file must not importdrizzle-orm,@paperclipai/db, a service underserver/src/services/, a route underserver/src/routes/, or a Node.js I/O module (node:child_process,node:fs,node:net). A domain function must not read the system clock; the caller passesnowas an explicitDatevalue.application/holds use cases and the ports they need. A use case takes its ports as constructor arguments and calls domain functions for policy decisions. An application file must not importdrizzle-orm, a SQL client, a concrete adapter, or the server's HTTP error helpers. The outer service or route translates application errors into transport responses.adapters/holds the concrete implementations of the ports: Postgres queries, transactions, and process control. An adapter file may importdrizzle-orm,@paperclipai/db, and Node.js I/O modules.
A module exposes its commands through index.ts, which composes the adapters
and the use cases behind a factory function. Code outside the module imports
that entry point, never a file inside domain/, application/, or
adapters/ directly.
pnpm check:module-boundaries enforces these rules for production source
files. It also checks access to the company deletion entry point below.
Company deletion
The existing company deletion service owns the database transaction and deletion
order. A module can expose a separate company-deletion.ts entry point.
Only services/company-deletion.ts can import this entry point.
The entry point exports an implementation of CompanyDeletionParticipant from
lib/company-deletion.ts. Do not export it through the normal module index.
The service must complete external cleanup before it starts the transaction. It must lock the company before it deletes dependent records or calls a module. Creation commands must use the same company lock. The service removes dependent records before it calls the module. The module checks its deletion conditions and deletes the records it owns. Use the supplied transaction. Do not start another transaction or call external systems. Throw an error if deletion cannot proceed. An error must roll back all database changes in the deletion transaction.
Call each module explicitly in dependency order. No module registry is required.
The agent lifecycle module also has a restricted configuration.ts entry point.
It saves configuration and invalidates verification in the caller's transaction.
The lifecycle guide lists its permitted callers.
Ordinary configuration updates use the root lifecycle command.
The agent lifecycle module owns agent creation, pause, resume, and termination. Use its commands for these changes.
The execution admission module owns project task-execution policy and the
durable admission records that charge project capacity. Today it exposes only
internal policy commands and the deletion guards that project, agent, and
company deletion call. Settlement releases active charges but keeps each
admission's complete project set for later retained-session resolution.
Execution claims still belong to services/heartbeat/queue.ts and native
review dispatch. See execution admission accounting.
Execution admission accounting
The execution-admission module (server/src/modules/execution-admission/) owns
three tables. Nothing writes admissions or caps in production yet, and no claim
path enforces a project maximum.
project_task_execution_policiesstores a project's maximum concurrent tasks apart fromprojects, so project readers never return an unenforced cap. No row means never configured. A row with a null maximum means an explicit "no limit", which a later legacy migration must not overwrite. Maximums are positive. The row is deleted with its project.execution_admissionsstores one admission attempt of a run. A run can be released before dispatch and admitted again, so(run_id, attempt)is unique and a partial unique index allows one unsettled attempt per run. The row keeps the context revision, the nullable primary project as provenance (no foreign key) and the settlement identity, which is set together withsettled_at.admitted_project_idsis the complete admitted project set, sorted, kept after settlement and immutable (a trigger rejects any change). It has no project foreign key, so settled history never blocks project deletion and keeps the deleted project's ID.{}is a known empty set. Null means unknown and is allowed only on settled rows: migration0428filled it from the charges of unsettled rows, but settled0426rows had already lost their set. A retained provider session resumes with the previous set ∪ the current references; an unknown previous set requires a fresh context, never an empty or primary-only guess.execution_admission_project_chargesstores one row per distinct admitted project, including unlimited projects. An admission without projects has no rows. Composite(company_id, …)foreign keys keep admissions and projects in one company. Rows are the only active capacity and exist only while the admission is unsettled: settlement deletes the whole set in one transaction, and(company_id, project_id)is indexed for active counts, so lookups scale with active work, not history.
Archiving a project changes no charge. Deleting a project with a charge returns
409 project_execution_admission_pending; the foreign key is the backstop.
Deleting an agent refuses unsettled admissions on its runs and removes settled
ones, with their sets, together with the runs. Company deletion removes the company's admissions before its runs and
projects.
The proposed global lock order for the admission transaction is: company row
(FOR NO KEY UPDATE), native finalization recovery owner, the whole project set
in ascending ID order (FOR NO KEY UPDATE, one statement), issues, wake
requests, the run, the external-admission advisory lock, the admission row and
its charges, then budget reservations. Policy updates take the same project
lock. Settlement takes no project lock. Only the project-set step is implemented
so far; claim paths keep their current locks until they are integrated.