ur-agent 1.65.10 → 1.65.12

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
package/CHANGELOG.md CHANGED
@@ -1,5 +1,53 @@
1
1
  # Changelog
2
2
 
3
+ ## 1.65.12
4
+
5
+ - Forced explicit weak-model decision menus through the real
6
+ `AskUserQuestion` UI. A provider-neutral end-turn guard accepts only one
7
+ canonical Ask object from explicit reasoning or one complete rigid Markdown
8
+ choice menu, validates it unchanged against the live tool schema, and rejects
9
+ examples, ambiguity, ordinary prose, workers, and headless sessions.
10
+ - Removed two task-list dead ends without weakening the mutation boundary.
11
+ Live plan mode may run only the exact current plan-directory bootstrap Bash
12
+ shape before tasks exist, and the last actionable task remains `in_progress`
13
+ when a file mutation has no later successful observable check. The same task
14
+ can then be verified or corrected instead of creating a duplicate.
15
+ - Improved exact `Edit` mismatch recovery for stale HTML blocks. Diagnostics
16
+ now select the most distinctive verified line near the real target, such as
17
+ `<script>`, rather than an unrelated generic closing tag; edits remain
18
+ fail-closed and never apply fuzzy replacements.
19
+ - Raised the default Ollama Cloud timeout for the known-slow Kimi K2.7 family
20
+ from 120 to 300 seconds in both streaming and permitted non-streaming paths.
21
+ Explicit timeouts and the stricter remote-session ceiling still take
22
+ precedence.
23
+
24
+ ## 1.65.11
25
+
26
+ - Hardened `AskUserQuestion` for weaker models without inventing user intent.
27
+ Its model schema now exposes only a strict 1–4-question request shape with
28
+ 2–8 nested choices, compact validation explains malformed flat arrays, and
29
+ lossless recovery never fabricates labels, descriptions, or prose-derived
30
+ questions.
31
+ - Made question interaction truthful and safe. Answers are accepted only after
32
+ post-permission validation has collected one real response per question,
33
+ custom `Other` works in both menu layouts, question-keyed UI state is
34
+ prototype-safe, and HTML-configured previews render escaped inert text
35
+ instead of executable model markup.
36
+ - Improved weak-model file-tool recovery. `Write` now explains that
37
+ `file_path` and complete `content` must be supplied together and never treats
38
+ surrounding prose as file data; a narrow deletion-only `Edit` whose desired
39
+ replacement is already uniquely present reports “already up to date” without
40
+ writing, while ambiguous and general stale edits still fail closed.
41
+ - Corrected task-gate control flow and diagnostics. `ExitPlanMode` can reach
42
+ its approval transition without an implementation task, stale out-of-mode
43
+ calls remain invalid, and gate recovery distinguishes a missing plan from an
44
+ all-terminal task list so unfinished Edit/Bash work is reopened rather than
45
+ retried unchanged. A simple loopback browser preview bypasses only the task
46
+ gate while retaining Bash permission and rewrite checks.
47
+ - Kept package validation compatible with both npm 11 and npm 12
48
+ `npm pack --json` output shapes, so upgrading npm does not make a valid
49
+ tarball look missing during tests or release checks.
50
+
3
51
  ## 1.65.10
4
52
 
5
53
  - Reworked approved-plan execution into a capability-aware task graph. Plans