Daemon Session Maintenance Writer Lease

Problem

The daemon can delete, archive, or unarchive a persisted transcript after its in-process ACP owner has closed. A different daemon process can still own the same transcript, so the in-process archive coordinator alone does not prevent the daemon from racing an external writer.

The transcript path and writer-lock path must also be resolved from the same workspace runtime. Falling back to the primary daemon runtime can mutate one workspace while checking a lock in another.

Scope

This change covers daemon-owned maintenance:

• REST and ACP delete, archive, and unarchive requests
• disconnect and orphan cleanup
• scheduled-task rollback and keepalive cleanup
• daemon shutdown while maintenance is already running

It does not add lease expiry, heartbeat, hostname-based recovery, automatic steal, force unlock, or a lock-schema migration. Writers that do not participate in the lease protocol still require platform-level single-writer fencing.

Runtime storage binding

Each WorkspaceRuntime resolves one absolute session runtime base directory at creation. Resolution keeps the existing priority:

1. QWEN_RUNTIME_DIR
2. advanced.runtimeOutputDir, resolved relative to the workspace
3. the normal Qwen runtime directory

The resolved directory is stored on the runtime and injected as QWEN_RUNTIME_DIR into every managed ACP child. Environment reload may update other values but preserves this pinned value because changing runtimeOutputDir requires a runtime restart.

Daemon parent operations that list, read, export, organize, or maintain sessions run inside the selected runtime's storage context. Runtime resolution failures do not fall back to the primary runtime.

Lease API

SessionService.acquireSessionWriterLease() derives both the writer-lock root and the active transcript path from the service's fixed Storage instance. Callers provide only the session ID, process kind, version, and reclaim policy. Invalid session IDs are rejected before the lock directory is touched.

Daemon maintenance always uses processKind: 'daemon' and reclaimPolicy: 'never'. The existing lock schema, key, owner record, and acquire/release protocol remain unchanged.

Maintenance protocol

Every session is processed independently:

1. Enter the daemon's per-session exclusive archive coordinator.
2. Close the local owner. Archive requires agent close; delete uses the normal fast close. A missing local owner is allowed.
3. Classify persisted state and preserve existing not-found and idempotent results without creating a lock.
4. Acquire the daemon writer lease.
5. Reclassify while holding the lease.
6. Verify ownership and the transcript fingerprint, then perform one mutation.
7. Release the lease with owner-token verification.

Batch requests may process independent sessions concurrently, but a worker holds at most one cross-process lease and never waits while holding multiple leases.

A failed mutation remains the reported error when release succeeds. A release or ownership failure is the externally safe error even if mutation also failed. Logs record the workspace, session, action, error kind, and whether the transcript mutation reached disk; they never include owner tokens or lock paths. Scheduled-task reconciliation follows the actual transcript mutation, not whether lease release subsequently succeeded.

Orphan cleanup first closes the local owner and respects requireZeroAttaches. A newly attached owner therefore prevents deletion. Late-spawn cleanup awaits close before acquiring the lease and deleting the transcript.

Shutdown

SessionArchiveCoordinator.sealMaintenanceAndWait() synchronously rejects new exclusive maintenance and waits for exclusive operations already admitted. Shared transcript reads are not included, so a long export does not consume the termination budget. REST returns 503 daemon_draining; ACP returns a JSON-RPC server error with data.errorKind = daemon_draining.

Daemon shutdown seals maintenance before child/process teardown and completes only after admitted maintenance leases have been released.

Compatibility and rollout

Batch response shapes and existing archive/delete/unarchive idempotency remain unchanged. Pre-check local session_archiving conflicts (raised by assertNotTransitioning before admission) still surface as a request-level 409. Conflicts raised inside the admission gate are reported per session in the 200 response body (errors[]) for archive, unarchive, and delete alike. Mixed-version writers are unsafe, so deployment and rollback must drain the old daemon and managed ACP processes before starting the new version.

Verification

Tests use real temporary runtime roots for writer contention and root isolation, cover state changes between the initial and locked classifications, and verify close, mutation, release, scheduled-task reconciliation, and shutdown ordering. Unit tests also cover invalid IDs, duplicate IDs, active/archive conflicts, lease release failures, orphan reattachment, and log redaction. Relevant package tests, build, and typecheck are required before merge.