ur-agent 1.76.7 → 1.77.1

package/CHANGELOG.md

@@ -1,5 +1,103 @@
 # Changelog
 
+## 1.77.1
+
+- Finishing a task no longer produces a long write-up. Both output-efficiency
+  sections said "be concise" but neither addressed what actually ran long, so
+  the rules are now specific: never paste code or file contents already written
+  to disk (cite `file_path:line`), report an audit or review as its findings
+  one line each, no closing recap of the conversation, and long explanations
+  only when the user asks for one. Subagent reports carry the same rule, since
+  their output is relayed verbatim.
+
+## 1.77.0
+
+- Edit no longer demands a prior Read. A matching `old_string` is checked
+  against the bytes on disk, which is exactly what "has this been read?" and
+  "has it changed since?" were asking — and stronger, since a stale snapshot
+  cannot survive a match against fresh content. The verified content is
+  recorded so the write path and later staleness checks work from it. This
+  removes a full model round trip from the most common mutating call. A
+  genuinely absent `old_string` is still refused, and now says the file has not
+  been read so the model knows to read it.
+- Write no longer refuses an unread existing file. It reads and records the
+  file itself — one local read instead of a round trip spent asking for content
+  nobody needed to see — and the "modified since read" check is unchanged, now
+  working from that recorded baseline.
+- Bash no longer issues a `mkdir` syscall on the critical path of every
+  command. The task output directory is process-wide and cannot change after
+  the first one, so it is created once; a failure is not cached, so the next
+  command retries.
+
+## 1.76.11
+
+- AskUserQuestion recovers the payload shape where a model flattens one
+  question's choices straight into `questions`, so six options arrived as six
+  question objects carrying a label and a description and no question text.
+  Every entry reported "question must be a non-empty string" and "options must
+  be an array". The array is folded back into the options of a single question
+  using the question text the payload already carries. A genuine
+  multi-question payload is never retargeted, and with no question text
+  anywhere the payload is still reported rather than given an invented
+  question.
+
+## 1.76.10
+
+- `npm publish` builds before it validates. `prepack` ran `release:check`
+  against whatever `dist/` happened to be on disk, so publishing straight after
+  a version bump always failed on a bundle that predated it — which is exactly
+  how 1.76.9 failed to publish. It now runs `build && release:check`, so the
+  check validates the artifact that actually ships.
+- A backgrounded subagent that exhausts its turn budget says so in its
+  completion notification instead of presenting a truncated result as a
+  finished one. The notification status has no value for "incomplete", so the
+  notice is carried in the message, matching how the handoff warning is
+  surfaced on the same path.
+
+## 1.76.9
+
+- Crew fan-out no longer abandons sibling workers when one throws. Both paths
+  passed raw worker promises to `Promise.race`/`Promise.all`, so the first
+  throw propagated while every other worker kept running unawaited, leaking
+  worktrees and child processes and turning a second failure into an unhandled
+  rejection. Failures are collected and rethrown once all workers finish.
+- `bunfig.toml` and the JetBrains `build.gradle.kts` are bumped with every
+  other surface; both were left behind by an earlier hand-edited bump.
+- `packageSmoke` no longer declares a 30s per-test budget that silently
+  overrode the release gate's own timeout, and the OpenAI Responses usage test
+  no longer asserts the double-counted `input_tokens` that the usage-accounting
+  invariants forbid.
+
+## 1.76.8
+
+- Streaming requests no longer time out while the provider is still working.
+  SSE comment keepalives (OpenRouter's `: OPENROUTER PROCESSING`) and provider
+  `ping` events were dropped by the SSE readers, so the inactivity watchdog saw
+  a frozen stream while bytes were arriving and aborted it as `Request timed
+  out`. They now surface as pings that rearm the watchdog.
+- Streaming requests get their own header-wait ceiling
+  (`provider.streamTimeoutMs` / `UR_STREAM_REQUEST_TIMEOUT_MS`, default 15
+  minutes) instead of sharing the 120s non-streaming timeout, which cut off long
+  prompts during queueing and prefill. Idle thresholds raised to 300s.
+- The Ollama stream deadline is rearmed per chunk, so it bounds silence rather
+  than total runtime; a local model answering a long prompt is no longer killed
+  mid-answer.
+- Edit resolves `old_string` against blocks the model re-indented, `Read`
+  line-number gutters copied into the string, and non-breaking/zero-width
+  character differences. `new_string` is re-indented to the file's depth. A
+  failed match now names the line where the block diverges instead of echoing
+  the string back.
+- AskUserQuestion no longer replaces an unrepairable payload with `null`, which
+  made every failure report a missing `questions` array regardless of the real
+  defect.
+- Concurrent tool and subagent fan-out closes every generator it started when a
+  consumer aborts or a sibling throws; previously their `finally` blocks never
+  ran, leaving subprocesses and streams alive after the batch reported done.
+- Both tool executors read one concurrency resolver, so
+  `UR_CODE_MAX_TOOL_USE_CONCURRENCY` and `UR_MAX_CONCURRENT_TOOLS` apply on
+  either path instead of whichever executor happened to be active.
+- A subagent that exhausts its turn budget reports `partial` with the reason
+  rather than `completed` with a truncated result.
 ## 1.76.7
 
 - Reverted to the 1.76.4 behavior and removed the intermediate version surface transitions.