[HyeockJinKim]

Release 26.10.0a3 from main.

• Base: main at f8a77d8fdedd591eba4eb69d817f37dc9a44fdd5 (chore: merge diverged alembic heads #15896), stacked on fix: declare the dotfile concern its field group is wired under #15899 (fixes test_ops.py failing on main)
• WebUI bundle: unchanged (26.10.0-alpha.0)
• External tool binaries: all-smi updated

Changelog

Features

• Write the AppProxy worker ingress state from the existing circuit created, route updated and removed events (#15650)
• Let a user see and manage all own sessions regardless of which keypair created them, in session rename, the legacy GraphQL session queries and the session event stream (#15761)
• Add the session owner user id to session event payloads and session enqueue plugin hooks, and deprecate access key fields of v2 session, kernel and login session types (#15762)
• Add an async storage status check interface and a probe task that runs each async check as an asyncio task, alongside the thread pool probe for checks that block in the kernel (#15763)
• Null the owner of sessions and kernels when the owning user is purged instead of keeping a dangling user id (#15830)
• Replace the thread pool probe with an async probe task, and run the vfs mount read in an executor the check is given; the marker file is parsed as a volume id, and a marker that is missing, empty, or not a volume id is recorded as a check error (#15831)
• Add the user-owned ssh_keypairs table that will hold users' SSH keypairs (#15832)
• Store the agents a deployment designates in a table referencing agents.uuid (#15844)
• Add a Sentinel monitor that tracks Valkey master changes through +switch-master subscriptions and a periodic query (#15866)
• Add purgedId with entityId/fieldId to v2 GraphQL purge responses, give each purge mutation its own Purge*Payload, and return the target Node from soft delete and restore responses (#15887)
• Add a soft delete mixin with shared soft delete / restore updaters, and move role presets from deleted to deleted_at, exposing deletedAt on the role preset node and filter and deprecating deleted. (#15888)
• Add the AppProxy worker ingress.delivery_mode config (api or redis) and accept ingress pushes on /v2/ingress with the api delivery mode (#15890)

Improvements

• Drop the CUDA toolkit (libcudart) requirement of the open-source CUDA plugin by querying devices through the CUDA driver API (libcuda) and NVML, enabling toolkit-free containerized agent deployments (#13599)
• Share user dotfiles across keypairs and create new files as non-default variants excluded from default-file APIs and sessions (#15666)

Fixes

• Fix persistent service containers such as the socket relay staying on the old container at the first agent start after an image version bump and staying stopped at the next start (#15846)
• Fix Jupyter login cookies piling up with every App Proxy session until the browser can no longer open apps (#15856)
• Fix manager /metrics blocking other requests on the same worker and growing without bound, and report system metrics per worker instead of summing stale values (#15862)
• Reconnect Valkey clients to the new master right after a Redis Sentinel failover instead of after about 30 seconds of failed requests (#15867)
• Fix mgr ops list --concern dotfile rejecting the area the dotfile operations are wired under (#15899)

Documentation Updates

• Settle how relations are identified (by their end UUIDs) and how write mutations respond in BEP-1075, BEP-1069 and the API guides (#15873)

External Dependency Updates

• Remove the redis-py dependency by reimplementing the redlock distributed lock on Valkey Glide, which now follows Sentinel master switches (#15871)

Miscellaneous

• Fix scripts/agent/build-socket-relay.sh writing null as the socket relay image version and saving an archive name the agent does not load (#15853)
• Merge the two alembic heads left by two revisions written against the same parent, so alembic upgrade head resolves again. (#15896)

Live test

Run on a local stack at the base commit. Overall: PASS

#	Check	Result	Evidence
0	Profiles	✅ PASS	local-admin and local-user on http://127.0.0.1:8090, logged in
1	Service health	✅ PASS	manager 10/10 HEALTHY; agent and storage pass
2	DB schema	✅ PASS	current = head e234d84a162f (mergepoint, #15896)
3	Superadmin	✅ PASS	admin domain search --limit 1 succeeds
4	Basic read	✅ PASS	domain get default
5	Agent	✅ PASS	1 agent, ALIVE and schedulable
6	Images	✅ PASS	155 images
7	Session	✅ PASS	as a regular user, stable/python:3.13-ubuntu24.04-arm64 → RUNNING → logs → TERMINATED
7-1	Deployment	✅ PASS	as a regular user, created with 0 replicas → found in my deployment search → STOPPED after delete
8	VFolder	✅ PASS	as a regular user, project vfolder created → found in project search → delete → purge, delete-complete
9	Regular user	✅ PASS	my session search succeeds
10	Permission	✅ PASS	admin domain search as a regular user → 403
11	Error logs	✅ PASS	no manager ERROR or traceback; one agent WARNING dropbearmulti execution failed (Linux binary on macOS, session unaffected)

🤖 Generated with Claude Code

---

📚 Documentation preview 📚: https://sorna--15898.org.readthedocs.build/en/15898/

---

📚 Documentation preview 📚: https://sorna-ko--15898.org.readthedocs.build/ko/15898/

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.