mirror of https://github.com/block/buzz.git synced 2026-08-18 06:50:31 +02:00

Files

T

npub1mn7jgtj4w2pd0g0zeuhxsa6jy6p0rewxz4kujt98my82ahfmp72sxjexk7andWill Pfleger 45f199ba5d fix(desktop): resolve P4 pass-2 test-vs-claim defects in workspace-scoped agent store

Four IMPORTANT findings: each assertion now fails if the production guard or
save it claims to prove is removed.

Fix 1 — compensate_drain writer test (genuine store-lock contention):
Add compensate_drain_with_hook seam; on_store_acquired hook fires while BOTH
locks are held. Test writes COMP_SENTINEL under the lock, signals writer,
waits for writer to queue on the store lock, then releases. Final disk state
must contain BOTH COMP_SENTINEL and WRITER_EDIT — fails if the guard is
dropped before compensate_drain_for's save.

Fix 1b — start-contender test (production start-lock seam):
Add start_pair_lazy_for<R: tauri::Runtime> generic seam; production
start_managed_agent_runtime_pair_lazy and start_pair delegate to it.
Contender now calls start_pair_lazy_for (the seam the production adapter
calls) instead of a raw mutex lock — proves compensation blocks production
starts, not just an arbitrary lock acquisition.

Fix 2 — Record-Mesh TOCTOU (async production driver):
Drive restart_local_agent_on_config_change_for (full async driver). Injected
mesh_fn rewrites provider to relay-mesh after the driver has resolved
mesh_model_id=None; epoch re-resolves to Some("auto") != None → TOCTOU guard
fires → Skipped before stop. Removing the in-epoch re-resolve guard would
allow the epoch to proceed, calling stop_fn and failing the assertion.

Fix 3 — Helper-vs-helper serialization (pre-acquisition hook):
Add with_workspace_transition_preflight_with_hook and
install_client_under_workspace_transition_with_hook to mesh_llm_scope.rs.
Both serialization tests use a proper 3-step handshake: holder signals
lock_held BEFORE publishing state; contender fires pre-acquisition hook
signaling at_lock_boundary; only then does the holder publish and release.
Removing the shared lock would let the contender bypass the boundary,
inverting each test's expected outcome.

Fix 4 — Full-tail receipt and disk assertions:
write_receipt_fn captures all receipt fields: key.pubkey, key.relay_url,
pid, desktop_instance_id (asserted equal to app.config().identifier),
started_at. Final disk assertions verify last_started_at.is_some(),
last_stopped_at.is_none(), last_error.is_none() on the matching record.
Removing the post-spawn record save drops last_started_at and fails the
disk assertions.

Co-authored-by: Will Pfleger <pfleger.will@gmail.com>
Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

2026-08-05 01:32:34 -04:00