Playbooks · Coding tasks
Qwen Code 0.27.4: the November 15 build
Written November 15, 2026
Qwen Code 0.27.4 landed November 15: managed sessions reach local files, Auto Mode tightens, and channel replies stop truncating.
What changed
Qwen Code 0.27.4, released November 15, is a release about reach and restraint at the same time. Managed Agent sessions can now work with local directories you grant them and open an interactive terminal, while Auto Mode — the mode that decides on its own whether an action is safe — gets stricter about risky actions, timeouts, and web addresses treated as files. Around those two themes sit channel delivery fixes for Feishu and WeCom, a Goals page that can work across trusted workspaces, and the return of the /memory controls. There are no known breaking changes, which makes this an update most users can take without a migration plan.
Managed sessions reach granted local files
The headline item is that Managed Agent sessions can access granted local directories and open an interactive Managed Web Terminal without persisting credentials or paths. The change list fills in the mechanism: a session can use the browser user's local filesystem through temporary directory grants, and the web terminal uses a short-lived ticket for authenticated WebSocket access while keeping the runtime active for as long as it stays attached. The shape matters for anyone running managed sessions against real project folders. Access is something you grant per directory rather than a standing credential the service stores, and the terminal is a live session rather than a one-off command runner. When the grant or the ticket ends, the access it carried ends with it.
Channel replies stop truncating
Outputs can now go to Feishu and WeCom webhooks, and long replies on managed channels are delivered as ordered segments instead of being truncated. The change list adds the detail that if a reply exceeds even the segment limit, a visible truncation notice appears rather than the text simply ending. Anyone who has pasted a long agent report into a team channel and lost the middle of it will recognise the problem this solves: a segmented reply keeps its order, and a notice at the limit tells readers the output was cut instead of letting them assume the agent stopped mid-thought. The Feishu side is a send-only webhook connection with a secret URL, an optional signing secret, channel control, and scheduled group delivery.
Auto Mode tightens in four places
Auto Mode now requires explicit authorization for risky actions, reviews classifier timeouts, avoids HTTP(S) URL filesystem lookups, and visibly marks Full Access. The timeout item is the subtlest of the four: when the fast classifier times out, the session model can review that timeout once when identifiers differ, allowing or blocking the action under the existing policy instead of leaving the decision unmade. The URL item closes an odd edge where a web address could be treated as a filesystem path to look up. And the Full Access marking is a labelling fix with real value — a session running with full access now says so visibly, so nobody has to guess which permission posture a long-running session is in. Taken together, the four changes push Auto Mode toward asking at the moments that matter and staying quiet the rest of the time.
Goals and memory across workspaces
The Web Shell Goals page can now merge goals from trusted workspaces, show workspace badges, and target a selected workspace when creating a goal. For a user splitting work across several trusted folders, that turns goals from a per-workspace list into a single view with the origin still visible. Memory gets its own cluster: the CLI /memory controls are restored, workspace settings persist, and typed SDK APIs can preview, submit, inspect, cancel, and retry Codex memory imports. That last item is aimed at developers building on the SDK who are moving memory state across tools — the import is no longer a single opaque call but a workflow with states you can inspect and retry.
What to do
Update to 0.27.4. If you run managed sessions, try the directory grant against one project folder first and open the web terminal once before you rely on it in a longer session. If you depend on Auto Mode, expect prompts at the risky-action boundary that previously may not have appeared, and check that any session you intend to run with full access is marked the way you expect. Channel users should send one long output to Feishu or WeCom and confirm the segments arrive in order. No breaking changes are listed, so anything that behaves differently after the update is a bug worth reporting rather than a migration step you missed.
Related playbooks
- Qwen Code 0.27.3: the November 13 build
- Copilot CLI 1.0.110: the November 12 build
- Copilot CLI 1.0.111: the November 13 build
- Copilot CLI 1.0.112: the November 13 build
Sources
Back to: Coding playbooks hub · Also: Qwen Code 0.27.3: the November 13 build · Qwen Code 0.27: agents that ask questions