ur-agent 1.65.10 → 1.65.12

package/docs/VALIDATION.md

@@ -19,7 +19,7 @@ You need:
 
 ```sh
 ur --version
-# expected for this release: "1.65.10 (UR-Nexus)"
+# expected for this release: "1.65.12 (UR-Nexus)"
 ```
 
 ## 0.1 First-workspace model selection (1.45.4)

package/documentation/index.html

@@ -45,7 +45,7 @@
     <main id="content" class="content">
       <header class="topbar">
         <div>
-          <p class="eyebrow">Version 1.65.10</p>
+          <p class="eyebrow">Version 1.65.12</p>
           <h1>UR-Nexus Documentation</h1>
           <p class="lead">A practical, tutorial-style reference for installing, configuring, automating, extending, and operating UR-Nexus.</p>
         </div>

package/extensions/jetbrains-ur/build.gradle.kts

@@ -7,7 +7,7 @@ plugins {
 }
 
 group = "dev.urnexus"
-version = "1.65.10"
+version = "1.65.12"
 
 repositories {
     mavenCentral()

package/extensions/vscode-ur-inline-diffs/package.json

@@ -2,7 +2,7 @@
   "name": "ur-inline-diffs",
   "displayName": "UR Inline Diffs",
   "description": "Review, apply, and reject UR inline diff bundles from .ur/ide/diffs inside VS Code.",
-  "version": "1.65.10",
+  "version": "1.65.12",
   "publisher": "ur-nexus",
   "engines": {
     "vscode": "^1.92.0"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ur-agent",
-  "version": "1.65.10",
+  "version": "1.65.12",
   "description": "UR-Nexus — autonomous engineering workflow engine (plan, execute, test, verify, document, benchmark, reproduce)",
   "type": "module",
   "packageManager": "bun@1.3.14",

package/technical/04-tools.md

@@ -61,6 +61,10 @@ only after `EnterPlanMode` (or `/plan`) has successfully made the active mode
 `plan`. Plan approval may change that mode before permission-edited input is
 revalidated; the executor labels that second validation as post-permission so
 the already-validated exit can finish, while new out-of-mode calls still fail.
+`ExitPlanMode` is exempt from the implementation task-list gate because it is
+the approval/control transition that precedes implementation. Its own plan-mode
+validation remains authoritative, so the exemption does not make a stale
+second exit valid.
 
 For non-trivial work, the task list uses one record per cohesive outcome with
 an observable done check rather than one omnibus record. Genuine single-outcome
@@ -79,7 +83,62 @@ precision-losing numeric IDs are rejected.
 Task-gate recovery names the tracking surface that is actually present:
 interactive Task V2 sessions use `TaskCreate`, while default headless sessions
 use `TodoWrite`. It never instructs a model to recover by calling a tool absent
-from that runtime.
+from that runtime. Runtime inspection tracks actionable and total user tasks
+separately: an all-terminal list is reported truthfully and the model is told
+to reopen or create the cohesive remaining task. Real Edit/Bash mutations stay
+gated. One simple `open <loopback-http(s)-URL>` Bash preview is exempt only
+from the task-list gate; remote/file URLs, flags, shell composition, expansion,
+redirection, backgrounding, sandbox overrides, and permission-time rewrites to
+mutating commands fail closed. The preview command remains a Bash side effect
+and still follows normal permission, sandbox, and plan-worker rules.
+
+Task completion also protects that lifecycle boundary. When the final
+actionable `in_progress` task has a successful `Write`/`Edit`/`MultiEdit`/
+`NotebookEdit` after its recorded start but no later successful inspection,
+runtime, test, shell, or delegated-check result, `TaskUpdate(completed)` is
+soft-deferred: it returns a non-error explanation and leaves the same task
+`in_progress`. The model verifies and retries completion instead of creating a
+duplicate task or discovering an all-terminal dead end on the next corrective
+Edit. The guard is evidence-based and conservative: missing/compacted history,
+non-file work, and intermediate tasks are not guessed into a deferred state.
+
+Live plan mode also treats setup of the exact current session plan artifact as
+planning infrastructure rather than implementation. `Write` creates the plan
+file's parent automatically, but weak models may first emit `mkdir -p` for that
+exact parent or the bounded `ls ... || mkdir -p ... && ls ...` check. Only those
+exact-path shapes bypass the task-list requirement; Bash permission and sandbox
+checks still apply, and a hook rewrite, sibling path, extra command, expansion,
+background launch, or sandbox override fails closed at the final boundary.
+
+`AskUserQuestion` exposes a request-only model schema: one top-level
+`questions` array with 1–4 complete question objects, each containing
+`question`, a header of at most 12 characters, and 2–8 labeled choices.
+Descriptions are optional and are never fabricated from labels. The runtime
+accepts only lossless compatibility forms such as string choices and recognized
+question-text aliases; it does not turn arbitrary prose or flat option rows into
+invented questions. More than four blocking decisions are asked in later
+rounds.
+
+One narrow end-turn recovery exists for weak models that clearly attempted this
+tool but failed to emit a native call. On an interactive main-agent turn with
+no existing tool use, the runtime may recover either one canonical
+`questions` object at the very end of a reasoning block that explicitly says
+to invoke `AskUserQuestion`, or one standalone Markdown decision menu with
+exactly one bold question, 2–8 bold labeled options with descriptions, and a
+terminal instruction to select an option. The recovered object must pass the
+live `AskUserQuestion` schema unchanged before the normal tool executor opens
+the UI. JSON repair, truncation, duplicate/ambiguous candidates, casual “A or
+B?” prose, examples, incomplete menus, background workers, headless sessions,
+and unavailable/disabled tools all fail closed.
+
+Answers and annotations are not model input fields. They are accepted only
+during post-permission validation after the interactive UI has returned one
+non-empty answer for every question; an unchanged generic approval cannot
+produce a successful “user answered” result. The UI uses prototype-safe records,
+provides a real custom `Other` path for both ordinary and preview questions,
+and does not count selecting `Other` itself as an answer. HTML-configured
+previews are escaped into an inert preformatted-text wrapper rather than
+executed as model-provided markup.
 
 ## Multi-agent tools
 
@@ -144,8 +203,17 @@ not only a modification timestamp. Full and ranged reads are compared at the
 final write boundary, preventing same-timestamp external replacements from
 being overwritten.
 
+`Write` requires `file_path` and the complete literal `content` in the same
+structured call. Prose outside the call is never treated as file content, and a
+missing-content failure states that no file was written instead of fabricating
+the intended file.
+
 `Edit` remains fail-closed rather than applying a fuzzy replacement to similar
 code. When an exact contiguous `old_string` is absent, its bounded error points
-to a verified matching line when one exists and tells the model to re-read that
-region, use a smaller current 2–4-line anchor, split distant HTML/CSS/JavaScript
-sections, and never retry the unchanged call.
+to the most distinctive verified matching line when one exists, rather than an
+unrelated generic delimiter, and tells the model to re-read that region, use a
+smaller current 2–4-line anchor, split distant HTML/CSS/JavaScript sections, and
+never retry the unchanged call. One narrow idempotent case returns success
+without writing: a non-`replace_all` deletion-only edit whose `new_string` is
+already present uniquely and whose larger `old_string` is absent. General
+stale, fuzzy, empty-replacement, and ambiguous matches still fail.

package/technical/05-providers-and-models.md

@@ -111,15 +111,18 @@ ur --discover-ollama      # scan the LAN for Ollama servers (ollamaDiscovery.ts)
   search, and bounded private cursor state. Compacted context persistence
   requires a 32-byte `UR_OPENAI_RESPONSES_STATE_KEY`.
 - `ollama.ts` selects timeouts from explicit request options, then
-  `API_TIMEOUT_MS`, then runtime defaults. `:cloud` models and remote sessions
-  use 120 seconds; local models use 300 seconds. The same model-aware value is
-  applied while waiting for `/api/chat` response headers and as the absolute
-  deadline in `readOllamaChunks`.
+  `API_TIMEOUT_MS`, then runtime defaults. Remote sessions and ordinary
+  `:cloud` models use 120 seconds; local models and Kimi K2.7 Cloud use 300
+  seconds because large Kimi coding turns can legitimately take longer than
+  two minutes before completing. The same model-aware value is applied while
+  waiting for `/api/chat` response headers and as the absolute deadline in
+  `readOllamaChunks`.
 - `ur.ts` identifies an Ollama Cloud runtime from both the selected provider and
   the `:cloud` suffix. It disables shared automatic request retries for that
-  route, applies the same 120-second bound to any permitted non-streaming
+  route, applies the same model-aware bound to any permitted non-streaming
   fallback, and skips fallback entirely when the Ollama stream deadline itself
-  caused the failure. Explicit `API_TIMEOUT_MS` remains authoritative.
+  caused the failure. Explicit `API_TIMEOUT_MS` remains authoritative, and
+  remote-session safety keeps its 120-second ceiling.
 
 ## Capability-aware routing
 

package/technical/09-multi-agent.md

@@ -149,6 +149,11 @@ use legacy `TodoWrite` by default, or Task V2 when
 - Dependencies block transition or claim until prerequisites are complete.
 - Actionable `pending` or `in_progress` entries open the gate; completed and
   internal entries do not.
+- The last actionable `in_progress` task cannot be terminalized immediately
+  after a recorded file mutation with no later successful observable check.
+  `TaskUpdate` soft-defers that completion and keeps the same task actionable;
+  it does not infer or auto-create a replacement task. A later successful
+  inspection/runtime/test/delegated check allows the explicit completion retry.
 - Reads remain unrestricted. Ordinary mutations have a default allowance of
   three preceding tool calls, counted by tool call rather than message.
   Delegation and child mutations always require an actionable parent task.
@@ -159,10 +164,13 @@ use legacy `TodoWrite` by default, or Task V2 when
   exempt so the agent can repair the plan.
 - Creating or updating the exact current-session plan-mode Markdown file is
   also exempt: that file is the planning artifact, not an ordinary workspace
-  change. The exemption and live actionable-task state are re-evaluated at the
-  final execution boundary after permission-hook input rewrites; sibling files
-  and the plans directory are not exempt, and the filename alone is not exempt
-  outside live plan mode.
+  change. A bounded Bash bootstrap may only create/check that file's exact
+  parent (`mkdir -p`, optionally guarded by the known `ls` pattern); this
+  compatibility path remains subject to Bash permission and sandbox checks.
+  The exemption and live actionable-task state are re-evaluated at the final
+  execution boundary after permission-hook input rewrites. Sibling paths,
+  general plans-directory commands, added shell operations, and the filename
+  alone outside live plan mode are not exempt.
 - Configure the behavior at
   `tasks.requireBeforeChanges.{enabled,freeReads}`.
 

package/technical/12-security-sandbox-stability.md

@@ -38,6 +38,23 @@ feature. Its AI-classified `auto` mode is therefore source-only. The accepted
 `permissions.classifierPermissionsEnabled` schema field is not read by a
 shipped enforcement path and must not be treated as an active control.
 
+Tools that require user interaction are semantically validated again after the
+permission decision even when their input signature did not change. This keeps
+an unchanged generic approval or hook allow-result from being mistaken for the
+interaction result.
+For `AskUserQuestion`, model requests cannot supply `answers` or `annotations`;
+post-permission execution requires an exact, complete, non-empty answer map.
+Question-keyed UI state uses null-prototype records and own-property checks, so
+model-controlled names such as `constructor` or `__proto__` cannot become
+inherited answers or mutate the record prototype. The tool rejects those
+reserved question names at its outer schema as an additional boundary.
+
+Model-provided Ask previews are never trusted as executable HTML. When an SDK
+selects the HTML preview format, the runtime escapes raw input into one inert
+`<pre data-ur-preview="text">…</pre>` wrapper and validates that wrapper before
+interaction. Scripts, event handlers, URL attributes, style injection, and
+closing-tag escapes remain text rather than active markup.
+
 Bash adds command parsing, command-injection checks, dangerous-pattern
 classification, project safety policy, path validation, and sandbox-aware
 permission decisions. `UR_CODE_DISABLE_COMMAND_INJECTION_CHECK` weakens one of

package/technical/README.md

@@ -1,6 +1,6 @@
 # UR-Nexus — Technical Specifications
 
-> Audited against the executable source and tests for `ur-agent` v1.65.10.
+> Audited against the executable source and tests for `ur-agent` v1.65.12.
 > Command, tool, flag, provider, and setting claims are checked against the
 > implementation rather than copied from product prose. Release validation
 > keeps this version synchronized and packages the complete `technical/`