Qwen Code 0.27 Adds Parallel Child-Agent Questions and GitHub Repository Filters

The Qwen team has shipped Qwen Code v0.27.0, an update to its open-source terminal coding agent that lets a managed-agent session keep several questions from its sub-agents open at once. The release was published on GitHub on November 8, 2026, at 2:15 AM IST. Its release notes list no known breaking changes. A TypeScript SDK, version 0.1.30, which bundles the 0.27.0 command-line tool, followed a few hours later.

Qwen Code is published under Apache-2.0 by the QwenLM organization, which describes it as “an open-source AI coding agent that lives in your terminal.” The repository had about 29,100 GitHub stars when this release went out. Beyond the terminal interface, the project now includes a browser-based Web Shell, a daemon that can watch pull requests, and “channels” that connect agent sessions to services such as GitHub.

Several child questions, answered in any order

The headline change is aimed at multi-agent work. In Qwen Code’s managed-agent mode, a parent session can spawn child agents, and those children sometimes need input before they can continue. Under pull request 16484, the release notes say a managed-agent session “can now keep multiple child questions pending, answer child questions in any order, and keep drafts and errors separated per child.”

The documented change separates drafts and errors for each child question and lets users choose which pending question to answer next.

The Web Shell picks up two related features. A task Q&A activity heatmap in the composer shows which days had question-and-answer activity, and clicking a day scrolls the transcript to that day’s first exchange. A new “Ask for details” action on pending Ask User Question panels opens a side task that explains the options on offer without answering the question for the user.

Tighter control over GitHub channels

Qwen Code’s GitHub channel lets agent sessions respond to activity on repositories. Version 0.27 adds a Repository Filter to the channel editor with allowlist and blocklist modes, so operators can limit which repositories a channel reacts to. Skipped notifications now appear as rows in the session timeline rather than vanishing.

Delivery to sessions bound to a pull request is also more reliable. The release notes say the channel now watches all open bound pull requests and keeps polling during public turns, and that plain conversation comments now reach PR-bound sessions. A new daemon view returns a single pull request’s state, mergeability and head commit SHA from its URL. Separately, idle channel sessions are now reclaimed so cached routes no longer use up the maxSessions budget.

Credentials kept out of error messages

Several fixes deal with secrets. Startup errors for unavailable Advisor models now sanitize model selectors, “hiding embedded credentials and NUL bytes while keeping routing unchanged.” Upgrade command refusals now mask passwords, and empty Harness user info is rejected during start and verify. On the project’s own CI, triage agents now run with isolated credentials on a separate pool. A publish job also recreates Git metadata before checkout, so leftover hooks or configuration cannot run in steps that hold credentials.

Fewer stalls in long runs

Much of the remaining change list targets the managed-agent runtime. Storage admission now locks rows through alias keys so concurrent requests queue instead of deadlocking. Snapshot publication waits for an in-progress claim, and session renames wait for an in-flight title update rather than returning a 503 error. Two performance changes remove tenant-level locking from ordinary session binding and tool-execution admission.

Other fixes address everyday editing. The shell tool’s sed simulation now handles CRLF line endings and Unicode in substitution patterns, which the notes say prevents failed edits and file corruption. Commands cancelled with Esc or Ctrl+C are now reported as interrupted, not completed, and keep their partial output. The interactive terminal now prompts for tool approvals requested by background workflow agents, which previously could leave runs stalled. A new FileChanged hook reports add, change and unlink events for watched files in local sessions.

For teams already running Qwen Code’s managed agents or GitHub channels, the parallel-question handling and the repository filter are the changes most likely to affect daily use. The credential-masking fixes are a reason to upgrade even for users who rely only on the command line.