[Benoît Cortier (CBenoit)]

An identity directory owned by another account stops all Agent Identity processing until it is fixed, because the account that created it can keep a handle that still changes it after a reset. Start-up already refused such a directory, but while the agent ran, a directory that changed owner only made each state change fail: runners kept renewing and checking in, and their agent channel streams stayed open.

The pending-directory poll now provisions the identity directories, as every state change does before it writes, then checks the stored identities, without taking the state lock. A directory whose owner the agent doesn't accept stops every runner and its streams, as an untrusted authority directory or identity.json already does, and the agent resumes without a restart once the owner is fixed. A directory the agent owns with wrong permissions is still repaired and stops nothing; the poll now resets such permissions on each interval, which every state change already did.

Issue: DGW-478
Security: yes
Co-Authored-By: Claude Opus 5.5 [email protected]

🤖 Generated with Claude Code

[Benoît Cortier (CBenoit)]

Evidence and notes

Lines changed (git diff --numstat b43cbd68e e71f43664):

Area	Added	Removed
Production (agent-identity state.rs, private_file.rs; devolutions-agent identity poll)	45	21
Tests (devolutions-agent/src/identity/mod.rs test module)	132	9
Other	0	0

How it works: the poll calls the new Store::check_state, which runs provision_directories (the same Existing::RepairAccess provisioning start-up and every state change run) and then check_stored_identities. A ForeignIdentityDirectory in the error chain returns the existing Stopped::Untrusted, so run_identity drops every runner and its streams and goes back to start-up, which refuses until the owner is fixed. runner.rs is unchanged.

Tests at e71f43664:

Platform	agent-identity	devolutions-agent lib	devolutions-agent main	Owner test
Windows, not elevated	227 passed	201 passed	6 passed	skipped (5 cases, needs elevation)
Linux (WSL), normal user	210 passed	135 passed	6 passed	skipped (needs root)
Linux (WSL), root, identity::tests only	—	12 passed	—	ran: identity, authorities, keys, rejected, pending

The owner test (an_identity_directory_given_to_another_account_after_start_up_stops_identity_processing) failed on the code before the fix (pending poll kept running) and passes after it, as root on Linux. The CI runners run it only where they are root or elevated; the Windows icacls /setowner path has not run locally.

cargo fmt --check and cargo clippy -p agent-identity -p devolutions-agent --all-targets -- -D warnings are clean on Windows and Linux.

Decisions and side effects worth a look:

• The poll now resets wrong permissions on the identity directories each interval (default 5 s). Every state change already did this; crates/agent-identity/INTENT.md §Local state names start-up, so the reset is now periodic rather than also per write.
• The check takes no state lock, so an outside handle on <data-dir>\identity (Explorer, a backup) can't stall the poll. On Windows, while a runner holds the state lock, the poll's provisioning can fail with a sharing violation; that is an ordinary error, retried at the next poll and logged as a warning at most once a minute.
• At start-up a foreign directory is still logged through the generic "state is unavailable" error, whose text names the directory.
• The start-up half of the test asserts only that start-up is refused: for identity and keys, the device-key directory check refuses before provisioning, so the error type differs by case.

Note

LLM-assisted content (no human feedback).

[Benoît Cortier (CBenoit)]

UNTESTED properties

No new UNTESTED: comment. The change relies on one that already exists:

• devolutions-agent/src/identity/mod.rs, run_identity: dropping the runners aborts every runner task and, with it, its agent channel streams. Tests check that the pending poll returns Stopped::Untrusted, not that a live runner stops.

Note

LLM-assisted content (no human feedback).