ur-agent 1.78.8 → 1.78.10

package/docs/USAGE.md

@@ -100,11 +100,13 @@ use `ollama.host` in settings if you want plain `ur` to default to a LAN host.
 Models exposed by the chosen Ollama app are valid, including local models and
 Ollama Cloud-backed models.
 
-Ollama Cloud models use a 120-second default bound for both response-header
-waiting and stream consumption. A deliberate stream deadline is returned
-directly instead of being replayed through the non-streaming fallback. Local
-Ollama models keep the five-minute default. Set `API_TIMEOUT_MS` to explicitly
-override either default for the current process.
+Ollama waits up to 15 minutes for response headers so cold loads and long
+prefill are not mistaken for failure. Once streaming begins, local and Cloud
+models use a five-minute *inactivity* watchdog that resets on every chunk; a
+healthy long answer has no total runtime cutoff. Remote/CCR sessions and Cloud
+non-streaming fallback retain a two-minute bound. Set
+`UR_STREAM_IDLE_TIMEOUT_MS` for the stream-silence window or `API_TIMEOUT_MS`
+for an explicit request-wide override.
 
 When project verification requires approval, UR asks once per user turn. The
 same pending compile/test/lint gate is not presented again after the user has
@@ -388,6 +390,16 @@ in dependency order. `ur exec` materializes this graph deterministically; the
 interactive agent follows the same lifecycle and its task/agent tools enforce
 the concurrency boundary.
 
+A new actionable work board is seeded immediately. If the model creates
+explicit subtasks, its first `TaskCreate` atomically replaces that seed instead
+of duplicating it. A successful simple turn completes the seed automatically.
+Replies such as corrections, approvals, and interruptions reuse the current
+seed or explicit unfinished board; the raw reply is not copied into a new task
+row. The agent reconciles and updates the relevant work rather than starting
+empty. A terminal board is archived only when the next prompt is genuinely new
+work; a corrective follow-up reopens the existing automatic task without
+renaming it.
+
 Independent read-only tasks can run in parallel. A task that may write to the
 shared checkout is serialized with other possible writers, even when it comes
 from another top-level prompt or crew worker. Parallel writers require explicit
@@ -398,15 +410,27 @@ on an unrelated remote revision.
 While the agent works, the `◭ Mashoofing…` row stays visible and its
 parenthesized phase changes with activity: `thinking`, `requesting`,
 `responding`, `preparing tool`, or `working`. Elapsed thinking time and token
-activity continue to appear when available.
-
-If a model sends a long internal planning monologue as ordinary answer text,
-UR leaves the useful summary visible and replaces the noisy middle with a slim
-`Reasoning condensed` rail. Press `ctrl+o` to expand the full trace. The
-compaction is presentation-only: session history and the model's next-turn
-context retain the original text. Detection is cached per completed message
-and uses one linear pass; it does not make another model/API call or reduce the
-agent's reasoning budget.
+activity continue to appear when available. The active task follows the
+ellipsis—for example, `◭ Mashoofing… · Fixing timeout handling (thinking)`—and
+is truncated or omitted on narrow terminals instead of wrapping the UI.
+
+The normal screen does not mount live assistant drafts. During tool work it
+shows the persistent `Mashoofing` row and compact tool summaries, then presents
+the stable final answer. Completed "I'll inspect..." text paired with a tool
+call is also omitted from the normal projection. Press `ctrl+o` to inspect the
+complete stored trace; verbose diagnostics can expose live text when needed.
+If self-talk is embedded inside a completed answer, UR replaces only that
+region with a slim `Reasoning condensed` rail. All of this is presentation-only:
+session history, exports, and the model's next-turn context retain the original
+text, and no extra model/API call or reduced reasoning budget is involved.
+
+When a provider reports that the context limit was reached despite the normal
+proactive threshold, UR withholds that transient error, runs one emergency
+compaction, and retries the interrupted turn automatically. Oversized images or
+documents are replaced with markers only in the emergency summary request so a
+large Computer screenshot cannot block recovery; the original transcript is
+not rewritten. If that single bounded recovery fails—or automatic compaction
+was explicitly disabled—UR shows the manual `/compact` or `/clear` action.
 
 ## Status bar
 

package/docs/VALIDATION.md

@@ -19,7 +19,7 @@ You need:
 
 ```sh
 ur --version
-# expected for this release: "1.78.8 (UR-Nexus)"
+# expected for this release: "1.78.10 (UR-Nexus)"
 ```
 
 ## 0.1 First-workspace model selection (1.45.4)
@@ -49,10 +49,12 @@ bun test test/ollamaTimeout.test.ts
 ```
 
 Expected: the cloud-model timeout, override-precedence, stream-deadline, and
-fallback-suppression cases pass. A model ending in `:cloud` defaults to 120
-seconds for response headers and streaming; a local model retains 300 seconds.
-`API_TIMEOUT_MS` wins over both. When a cloud stream reaches its deliberate
-deadline, the request fails once instead of starting a non-streaming replay.
+fallback-suppression cases pass. Response headers allow 900 seconds; local and
+`:cloud` streams tolerate 300 seconds of silence, while remote sessions retain
+120 seconds. `UR_STREAM_IDLE_TIMEOUT_MS`, `API_TIMEOUT_MS`, and explicit request
+options follow their tested precedence. When a Cloud stream reaches its
+deliberate inactivity deadline, the request fails once instead of starting a
+non-streaming replay.
 
 ### 0.1.2 Single project-gate approval (1.45.6)
 

package/docs/providers.md

@@ -462,12 +462,14 @@ Local/server providers use their normal endpoints:
 - llama.cpp server mode: `http://localhost:8080/v1`
 - vLLM server mode: `http://localhost:8000/v1`
 
-Ollama models whose names end in `:cloud` use a 120-second default for the
-response-header phase and a 120-second total stream deadline. UR does not
-automatically replay a cloud request after that stream deadline, preventing a
-bounded failure from expanding into the shared non-streaming fallback and
-retry chain. Local Ollama models retain the five-minute default. A positive
-`API_TIMEOUT_MS` or explicit request timeout overrides these defaults.
+Ollama allows up to 15 minutes for response headers so a cold model load or
+large prefill can begin. After headers, local and `:cloud` models use a
+five-minute stream *inactivity* deadline that resets whenever bytes arrive;
+remote/CCR sessions use two minutes. UR does not automatically replay a Cloud
+request after that deadline, preventing a bounded failure from expanding into
+the non-streaming fallback and retry chain; Cloud non-streaming fallback itself
+remains bounded at two minutes. `UR_STREAM_IDLE_TIMEOUT_MS`, `API_TIMEOUT_MS`,
+or an explicit request timeout can override the applicable default.
 
 ## Optional Live Provider Smoke
 

package/documentation/index.html

@@ -45,7 +45,7 @@
     <main id="content" class="content">
       <header class="topbar">
         <div>
-          <p class="eyebrow">Version 1.78.8</p>
+          <p class="eyebrow">Version 1.78.10</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>
@@ -171,8 +171,8 @@ ur config set responses.store false</code></pre>
           </article>
           <article>
             <h3>Ollama Cloud latency bounds</h3>
-            <pre><code>API_TIMEOUT_MS=120000 ur</code></pre>
-            <p>Cloud-tagged Ollama models default to bounded response-header and streaming phases. A stream deadline is surfaced directly instead of starting a non-streaming retry chain. Local models retain their longer default, and <code>API_TIMEOUT_MS</code> remains an explicit override.</p>
+            <pre><code>UR_STREAM_IDLE_TIMEOUT_MS=300000 ur</code></pre>
+            <p>Ollama allows long cold-load headers, then uses a five-minute inactivity watchdog for local and Cloud streams. The timer resets on every chunk, so healthy long answers continue; remote sessions remain shorter, and a Cloud stream timeout is surfaced without starting a duplicate non-streaming retry. <code>UR_STREAM_IDLE_TIMEOUT_MS</code> and <code>API_TIMEOUT_MS</code> remain explicit overrides.</p>
           </article>
         </div>
       </section>

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

@@ -7,7 +7,7 @@ plugins {
 }
 
 group = "dev.urnexus"
-version = "1.78.8"
+version = "1.78.10"
 
 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": "1.78.8",
+  "version": "1.78.10",
   "publisher": "ur-nexus",
   "engines": {
     "vscode": "^1.92.0"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ur-agent",
-  "version": "1.78.8",
+  "version": "1.78.10",
   "description": "UR-Nexus — autonomous engineering workflow engine (plan, execute, test, verify, document, benchmark, reproduce)",
   "type": "module",
   "packageManager": "bun@1.3.14",
@@ -49,6 +49,7 @@
     "bundle": "node scripts/bundle.mjs",
     "build": "node scripts/bundle.mjs",
     "release:check": "node scripts/release-check.mjs",
+    "release:tag": "node scripts/release-tag.mjs",
     "release:create-source-zip": "bun scripts/source-zip-create.mjs",
     "release:check-source-zip": "node scripts/source-zip-check.mjs",
     "typecheck": "bun run tsc --noEmit && bun run typecheck:strict-core",