Two coding tasks can work on separate copies of a repository and still act through the same operating-system account. One can wait for quota while the other changes the interfaces it depended on. A worktree is useful here: it can keep their uncommitted files apart. The harder question is what each task may do when it wakes, and whether its starting assumptions still hold.
Z.ai’s November 13, 2026 ZCode 3.18.2 release adds worktree support for parallel task development, a session-management tool, and changes to permission and sharing controls. It follows October 20’s documented persistent remote processes and the November 3 update that says workflows paused by quota resume when quota resets. The significance is in how these features combine: code can be reached from more places, and work can continue after the developer has stopped watching. The product notes describe capabilities, not measured outcomes or guarantees that an old approval is re-evaluated in every case. ZCode release history
Our GLM-Legal analysis studied a specialized model and how to assess domain claims. ZCode’s release is a harness and desktop product update, not a new GLM checkpoint. Its evaluation should check what actually ran, where, and under whose authority.
Use the worktree for source separation
Suppose two agents update a table filter and its accessibility labels. Worktrees can give each lane its own checkout. Each lane’s normal file reads then see its own checkout and uncommitted changes. Both branches may still change the same component, rely on a running local server, or write to a shared fixture store. One lane can pass its tests while the combined changes fail. A separate checkout does not create a separate process, credential, or network boundary.
For a pilot, give each lane a base revision, assigned worktree, allowed directories, and a definition of what it can change. Record temporary resources and choose test fixtures that can be safely isolated by lane. Rebase or merge only through the project’s normal review and tests. The build that matters is the one produced from the candidate merge, not the last green result in each lane. If the task crosses both lanes’ ownership, stop and choose an owner rather than asking two agents to update the same contract independently.
Keep these as evaluation requirements. The November release notes establish the addition of worktree support; they do not specify the full host, plugin, or credential separation a particular deployment provides. Verify the actual selected workspace and execution identity before relying on either.
Distinguish a controlling device from the execution host
The remote development guide distinguishes the desktop, which retains settings such as account and provider configuration, from the environment where the agent, file operations, and commands run. Its manual configuration sync can copy user-level skills, plugins, and MCP configuration, and it cautions that MCP secrets can transfer with that configuration. SSH, WSL, and containers are different target environments with different dependencies and permissions. Remote Development
The Remote Control guide separately describes the phone as a control surface, not a runtime. The application uses the already selected desktop workspace’s environment. It also treats the connection link as sensitive and distinguishes closing the pairing dialog from explicitly stopping the session. Those semantics make the review of execution identity more important than the physical device in the user’s hand.
One practical record is a small run card: the task and lane, checkout and base revision, host and account, installed tools, authority scope, and expiry. Keep it outside the model’s editable narrative. When the UI reconnects, compare the observed runtime with that card before accepting a new intervention. A user pressing stop on the remote-control surface should not be told that all work on the host has stopped until the runtime has actually confirmed the relevant jobs are no longer active.
Treat resumption as a decision point
The permission documentation distinguishes task-scoped permission requests from ordinary questions that can continue after a countdown; it says permission and plan approvals wait. The release notes also describe an option to remember Full Access. A remembered UI setting cannot prove that an action remains appropriate after a role change, a repository change, or a tool update. Where that distinction matters, the application or surrounding controls must enforce it.
Imagine a documentation agent paused by quota while collecting review material. The workflow can continue successfully after a reset and still be wrong about the thing it should update: a new reviewer may own the page, the base commit may have changed, or the source may have been withdrawn. In the pilot, classify the resumed work first. A read of an approved frozen fixture can often continue. A write against current project state requires checking the target and the actor’s current authorization. Preserve old observations as history without treating them as new permission.
If a previous operation’s outcome is uncertain, reconcile it before repeating it. A rebuilt interface needs to distinguish “no acknowledgement received” from “nothing happened.” This discipline is also useful for tasks that only prepare changes. Repeating a long build or report can still burn hours of compute and leave multiple outputs competing for attention.
A proposed parallel-work pilot
Use a disposable repository with synthetic inputs, test credentials limited to the pilot, and a harmless action that records an acceptance marker. Compare a single-lane baseline to two parallel tasks; grade the actual combined result. Treat the following comparisons as test cases, not claims about documented ZCode behavior.
| Situation | Acceptance evidence | Metric |
|---|---|---|
| Parallel changes against one base | Combined diff passes; no unassigned files changed | Lead time and review minutes per accepted change |
| Disconnect and reconnect mid-build | Same intended host, lane, and identifiable attempt | Recovery time, duplicate attempts, unfinished jobs |
| Resume after quota reset and changed base | Stale work held or correctly revalidated | Stale approvals or outdated changes accepted |
| Stop control or expire a share | Expected access ends; runtime status is reported separately | Time to revocation and unresolved background work |
The economics should cover the whole change. If two tasks each take 20 minutes and can run in parallel, that removes at most 20 minutes from a 40-minute serial build before review. An added 25 minutes of merge repair would erase the gain. Measure the review and rework, and segment independent changes from changes that touch one interface. Use parallel worktrees for independent lanes with a clear merge gate. Permit persistent work only where the team can identify its actual host, observe its outcome, and recheck its authority when conditions change.