Skip to content

loadViewState commits nothing when the selection moves during the reapplied lookup and back #2495

Open

Description

@easonLiangWorldedtech

Filed as a design question with a named outcome. The current behaviour is defensible and this issue is not a claim that it is wrong.

Where

src/core/webview/ClineProvider.ts, loadViewState, on head 57b766b:

  • the selection the re-resolve is keyed to is snapshotted at lines 1162 to 1165 (selectionBeforeReapply: currentApiConfigName and apiConfiguration only),
  • the reapplied profile name is derived at lines 1166 to 1172, which requires the name to have moved inside the load window and the settings object reference not to have moved,
  • the reapplied lookup and its supersession re-read live at lines 1174 to 1212: the comparison is at 1196 to 1201, the discard log at 1202 to 1204, and the return at 1205 leaves mergedState unassigned, so the load commits nothing.

The sequence the question is about

  1. A view has a persisted record naming profile P with the settings resolved for P.
  2. While the lookup of P is awaited the user switches the view to profile N, so the merge carries N.
  3. The name moved and the settings reference did not, so the load re-resolves N and snapshots the selection it resolved for.
  4. While that lookup is awaited the user switches the view back to P.
  5. The re-read at 1196 sees a selection that differs from the snapshot, logs the discard at 1202 and returns at 1205.

So the profile the view ends on is the one whose settings were already resolved successfully at the first lookup, and the load nevertheless commits nothing. The buffer keeps the user's own newest writes and the persisted record is untouched, so nothing is corrupted: in particular the view is never left naming one profile while holding another profile's settings, which is the pairing the whole guard exists to prevent.

The alternative reading

Because the view ended on P, a load could fall back to the record it started from and commit P with the settings it resolved for P, rather than abandoning the commit. That is a different discard policy: a superseded re-resolve means fall back to the persisted record, not commit nothing at all. Both policies keep the name and the settings paired. They differ in whether a load that was interrupted mid-flight still applies the state it loaded before the interruption.

Why this is a policy choice and not a defect

The current behaviour never mispairs a name with settings resolved for a different profile, never writes a value the user did not choose, and never loses a persisted record. The alternative never leaves a view on its previous in-memory state when a valid persisted record for the profile it ended on was resolved during the same load. Choosing between them is a call about what a discard should mean for this load path, and it belongs to a maintainer rather than to a fix. It is deliberately not being changed on the head above: the discard semantics for this path are the subject of published contract answers on PR #1560, and changing them here would reopen that discussion.

Provenance

Found while sweeping for the class behind review thread 4609689999 on PR #1560 - a comparison that decides ownership or precedence against the wrong snapshot - which was fixed in 57b766b by comparing the mode against the reading its carried value was derived from. The sweep found that the two neighbouring reapplies for currentApiConfigName and apiConfiguration are not instances of that class, because they compare against the same post-load reading their writes are derived from and the only asynchronous window after them is guarded by the re-read quoted above. What the sweep did find is this asymmetry in that guard, which is why it is filed rather than fixed. Related to issue 2488 step 2 (a record-level write revision): both questions are about which write a comparison is entitled to judge as superseded.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions