ur-agent 1.65.13 → 1.66.0

package/CHANGELOG.md

@@ -1,5 +1,63 @@
 # Changelog
 
+## 1.66.0
+
+- Unified proactive compaction around one model-aware threshold and one live
+  token estimator. Small context windows now retain positive warning, error,
+  and trigger thresholds; configured 50–95% thresholds are honored; prompt
+  notifications, `/context`, and the SDK expose a clamped live-estimated
+  percentage remaining until the real trigger; reactive/collapse modes do not
+  claim a false
+  countdown; and successful compaction clears stale warnings.
+- Preserved exact task execution state across full, partial, and
+  session-memory compaction. The restored authoritative snapshot keeps Task V2
+  IDs, statuses, owners, and dependency edges (or TodoWrite order/status),
+  prioritizes actionable work under 64-record and estimated 6,000-token
+  bounds, and requires `TaskList` before mutations when records were omitted.
+- Closed a task-gate bypass caused by compacted history. Compact boundaries
+  now carry durable metadata that consumes the initial trivial-call allowance,
+  including SDK serialization, so a model cannot compact and then mutate as
+  though no earlier tool calls occurred.
+- Removed a competing private compaction path from in-process workers. Workers
+  now use the normal agent query loop and therefore inherit the same feature
+  flags, exact trigger, session-memory-first path, reactive/collapse policy,
+  failure circuit breaker, cleanup, and compact-boundary behavior.
+- Brought Kimi/Ollama bare task-call recovery into parity with the live task
+  schemas. Dependency fields, numeric task/dependency IDs, and `failed` or
+  `skipped` terminal updates are normalized safely; invalid IDs, statuses, and
+  unknown fields fail closed.
+- Audited v1.65.0 through v1.65.5 for regression-safe enhancements. Every
+  concrete earlier technique is already present or superseded by stricter
+  current behavior, so no legacy implementation or removed feature was copied
+  back.
+
+## 1.65.14
+
+- Restored the proactive task-first behavior that was present in v1.65.0 and
+  removed a prompt/gate contradiction introduced later. For any non-trivial
+  state change, including a feature-rich single-file build, the model is now
+  told at system, task-tool, approved-plan, Ollama/Kimi, file, notebook,
+  shell, and worker boundaries to complete the available task setup, inspect
+  its result, mark the selected task `in_progress`, inspect that result, and
+  only then mutate state.
+- Made the ordering unambiguous for weaker models: task setup cannot be batched
+  with the mutation it enables, approved-plan handoffs require task calls as
+  the next state-changing calls, terminal task lists must be reopened or
+  extended before new work, and task guidance appears before file-tool
+  selection guidance.
+- Closed planner availability and control-flow gaps across tool modes. Legacy
+  `TodoWrite` now follows the same feature-rich/terminal-list lifecycle and is
+  preferred when Task V2 is incomplete; bare/simple, REPL-simple, coordinator,
+  custom-agent, and override-prompt paths all receive a usable planner and the
+  task-state contract. Coordinator workers must be task-bound before launch;
+  an explicitly planner-less custom tool pool now fails closed with honest
+  configuration recovery instead of naming a missing tool.
+- Prevented task-gate deadlocks and false mutations in supporting tools.
+  Team creation/deletion, shutdown/plan control messages, and emergency task
+  stop are narrow control transitions; skill loading and desktop screenshots
+  are read-only while downstream skill work, desktop input, and arbitrary
+  state-changing tools remain gated.
+
 ## 1.65.13
 
 - Recovered otherwise valid `AskUserQuestion` calls whose UI header exceeds