ur-agent 1.78.9 → 1.78.11

package/CHANGELOG.md

@@ -1,6 +1,6 @@
 # Changelog
 
-## 1.78.9
+## 1.78.11
 
 - Provider-reported context overflow now triggers one bounded automatic
   compaction-and-retry in the shipped build instead of immediately showing
@@ -9,25 +9,34 @@
   recovery itself; the original transcript remains unchanged. Automatic
   recovery still respects the user's disable setting and never recurses inside
   compact/session-memory workers.
-- Live answer streaming no longer paints partial model self-talk and removes it
-  a moment later. The undecided trailing paragraph stays buffered; once
-  internal planning is identified, one stable `Reasoning in progress` rail is
-  shown until the finalized answer replaces it. Provider-native thinking is
-  also private on the normal screen and remains available only in an explicitly
-  opened transcript or verbose diagnostics. This is presentation-only and uses
-  a monotonic incremental scan, so model context, reasoning depth, and streaming
-  throughput remain intact.
-
-- Every real user work turn now appears in the live task board immediately,
-  even before a model emits `TaskCreate`. The first explicit model task
-  atomically replaces that automatic seed, so parallel task creation does not
-  duplicate or lose work. A simple successful turn completes its seed; an
-  aborted turn leaves it pending for honest recovery.
+- The normal screen no longer paints live answer drafts or completed
+  "I'll inspect..." tool preludes. It shows stable tool progress and the final
+  answer instead; the complete provider text remains untouched in transcript
+  history, exports, verbose diagnostics, and future model context. The
+  `◭ Mashoofing…` activity row is now independent from streamed content and
+  remains mounted throughout active work, so hiding narration never hides
+  liveness. Its primary line now includes the live task-board activity after
+  the ellipsis, width-bounded to stay on one row.
+- Ollama Cloud streams now tolerate five minutes of silence during large-model
+  prefill or tool planning instead of failing after two. The deadline still
+  measures inactivity rather than total runtime and remains configurable with
+  `UR_STREAM_IDLE_TIMEOUT_MS`; remote sessions and cloud non-streaming fallback
+  keep their two-minute bounds, and real dead streams still terminate.
+
+- A new work board appears immediately, even before a model emits `TaskCreate`.
+  Its internal seed now renders as a neutral `Planning tasks…` state and never
+  stores or flashes the user's prompt as a task title or description. The first
+  explicit model task atomically replaces that seed, so parallel creation does
+  not duplicate or lose work. Corrective replies—including `no` / `still`
+  feedback—and interruptions resume the existing seed or explicit unfinished
+  board. A successful simple turn completes its seed; an aborted turn leaves it
+  pending for honest recovery.
 - User interruptions no longer archive pending or in-progress tasks into an
   apparently empty list. The next prompt advances the generation while keeping
-  unfinished work visible, adds the new request, and instructs the agent to
-  reconcile, skip superseded work, and continue updating statuses. Completed
-  or otherwise terminal lists still move to readable history.
+  unfinished work visible and instructs the agent to reconcile, update relevant
+  work, skip superseded work, and continue statuses. Completed terminal lists
+  move to readable history only for genuinely new work; corrective follow-ups
+  reopen the relevant active board.
 - Crew concurrency tests now declare their branches read-only, matching the
   runtime policy that unknown shared-checkout work serializes. The artifact
   HTTP integration test also distinguishes a managed sandbox's blanket socket