ur-agent 1.78.6 → 1.78.8

package/CHANGELOG.md

@@ -1,7 +1,54 @@
 # Changelog
 
-## 1.78.6
-
+## 1.78.8
+
+- Long model self-deliberation emitted through the ordinary answer channel is
+  compacted into a slim `Reasoning condensed` rail. The useful summary and
+  result remain visible, `ctrl+o` expands the complete trace, and the original
+  text still stays in session history and model context; this is a cached,
+  linear-time display change, not reasoning truncation.
+- The `◭ Mashoofing…` activity row remains present for the entire live turn.
+  Its parenthesized detail moves through the real phase (`thinking`,
+  `requesting`, `responding`, `preparing tool`, or `working`) and continues to
+  show elapsed thinking time and token activity instead of disappearing
+  between stream states.
+- Large prompts are decomposed into a bounded dependency graph before work is
+  delegated. Natural lowercase follow-ups and semicolon clauses split
+  correctly, natural phrases such as `agents/subagents` are no longer mistaken
+  for paths, and invalid model-produced graphs (missing targets, self-edges,
+  cycles, duplicates, or excessive fan-out) fall back to the deterministic
+  planner instead of reaching the task store.
+- Parallelism now follows the work, not merely the requested worker count.
+  Independent read-only branches may overlap; unknown or shared-checkout
+  mutations serialize across tasks, crews, and concurrent top-level plans.
+  Mutating subagents may overlap only in explicit isolation, whose worktrees
+  start from the exact clean current revision rather than a possibly stale
+  remote branch.
+- Task outcome learning is atomic across processes and records only work that
+  actually executed. Batched prompt-plan and crew outcomes cannot overwrite one
+  another, malformed stores normalize safely, and model routing waits for five
+  observations and compares confidence-adjusted success rather than trusting a
+  tiny perfect sample.
+
+## 1.78.7
+
+- A short approval turn such as `ok`, `proceed`, or `go ahead` keeps the active
+  task generation. It previously looked like a brand-new request, so the plan
+  was archived just before the agent tried to complete its first task; the next
+  dependent creation then failed too because its blocker had disappeared.
+  Whole-message matching keeps real follow-up requests on the fresh-list path.
+- `AskUserQuestion` recovers a flattened option list even when the model omits
+  the outer question text. Eight `{description, header, label}` choices now
+  render under a neutral prompt with every label and description preserved,
+  instead of producing sixteen missing-field errors and showing no dialog.
+- Batched `TaskCreate` calls keep their emitted order, so the consecutive IDs
+  used in dependency fields cannot be reassigned by lock-acquisition order.
+  Forward dependencies now work with the future task on either side of the
+  edge, adopt both reciprocal relationships when that task is created, and
+  participate in cycle detection while still virtual. A redundant self-edge
+  is ignored without deleting the otherwise-valid task. Together these prevent
+  the cascade of `task_not_found` / `self_dependency` errors that left only one
+  task from a multi-task plan.
 - Delegation is the default for independent work rather than something held in
   reserve. The guidance described subagents as valuable but warned against
   using them "excessively when not needed", which reads as a reason not to —