🤖 tests: phone repro editCancelKeepsDraft once showed an empty composer after Cancel

Problem

The bug-bash repro tests/bugbash/repros/editCancelKeepsDraft.e2e.ts failed once on the phone target in CI. The web target passed in the same job. It is an assertion failure, not a timeout:

Run 37541583047, attempt 1, job 112535673205 (Test / E2E (linux 1/4)), on PR 🤖 refactor: expose Token Budget counts for turn-start stages (no behavior change) #5805's head 8da02cc. 🤖 refactor: expose Token Budget counts for turn-start stages (no behavior change) #5805 changes no frontend code.
ASSERTION_FAILED: expect.toHaveValue failed: the composer textbox ("Message") should hold unsent draft after Cancel hides the edit box. It held an empty value.
Attempt 2 of the same run passed.

So after Cancel on a phone-width layout, the composer sometimes shows an empty box instead of the unsent draft. If the user then types or sends, the draft is lost.

Notes

Not reproduced on main yet.
On main, entering edit mode writes the edit text into the shared draft and restores the pre-edit draft on Cancel (applyDraftFromPending, restorePreEditDraft in ChatInput/index.tsx). A restore that races the composer remount or the draft store echo could produce this.
Open PR 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 (🤖 Reloading during a message edit replaces the unsent draft with the edit text #5672, 🤖 Composer: Edit in one window also loads the message into another window's composer #5571) removes that write/restore path: the edit text stays in a memory-only buffer, so Cancel only drops the buffer. Check whether 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801 removes the failure. If it does, close this with 🤖 fix: an edit keeps its text in memory, so the unsent draft survives a reload and stays out of other windows #5801.

Next step

Run the phone repro in a loop on main and on #5801's head (for example 20 times each) and compare the failure counts.

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high

Second occurrence, first reported as #5876. Same test, same phone target, same assertion.

Run: https://github.com/coder/xum/actions/runs/37763109369, job "Test / E2E (linux 1/4)", the merge-group run for 🤖 tests: runner library for the bug-bash sandbox, with the first image digest #5875.
FAIL phone repros/editCancelKeepsDraft.e2e.ts > Cancel in edit mode leaves edit mode and keeps the unsent draft
ASSERTION_FAILED: expect.toHaveValue failed: expected unsent draft, observed "", at repros/editCancelKeepsDraft.e2e.ts:28:5.
The same tree passed this shard on the PR run (https://github.com/coder/xum/actions/runs/37761591117). 🤖 tests: runner library for the bug-bash sandbox, with the first image digest #5875 changes only tests/bugbash/sandbox/, which no repro imports.

Two occurrences in about 10 days, both on phone, both after Cancel. Still not reproduced locally (an earlier 20-run loop passed 20/20 on main). The cause is still open: a real intermittent draft loss at phone width, or a timing gap between toBeHidden and the value check. The planned fix for #5672 and #5571 (replacing #5801) must run this repro in a loop on its branch and on main and report both counts.

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high

Another sighting, this time on the web target.

Run 37985323071, Test / E2E (linux 1/4): https://github.com/coder/xum/actions/runs/37985323071/job/114005814158
Same assertion: after Cancel, the composer ("Message") held "" instead of unsent draft (editCancelKeepsDraft.e2e.ts:28).
PR 🤖 tests: wait for the failed turn's terminal policy in G4 goal-advancement tests #5959 changes only a bun unit test (goalAdvancement.test.ts), so it is unrelated.
A rerun of the failed jobs passed: https://github.com/coder/xum/actions/runs/37985323071/job/114011416724

Third sighting, again on web. This time it removed a PR from the merge queue.

Merge-group run 38082874892, Test / E2E (linux 1/4) (job 114303270432), commit 449fbe9846 (main ea19a44101 plus 🤖 tests: bug-bash sandbox runtime fixes from the readiness reviews (D1) #6082). 🤖 tests: bug-bash sandbox runtime fixes from the readiness reviews (D1) #6082 changes only tests/bugbash/sandbox/** and docs.
Same assertion: editCancelKeepsDraft.e2e.ts:28, composer "" instead of unsent draft. The failed attempt took 6.53 s. Passing runs take about 2.2 s, and the phone target in the same job passed in 1.87 s.
Local check on 449fbe9846 with the mock app AI and --retries 0: 16 of 16 passed. 10 were the repro alone on web, and 6 were the whole web repro phase in CI order.
The job uploads no e2e artifacts, and trace: on-first-retry records nothing with retries 0. So no sighting so far left a trace or screen. The next step is a trace on CI, for example trace: 'retain-on-failure' for the repro phase, so the next failure shows its state.

The coordinator desk owns this issue now.

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high

Navigation Menu

Repository navigation

🤖 tests: phone repro editCancelKeepsDraft once showed an empty composer after Cancel #5810

Problem

Notes

Next step

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

🤖 tests: phone repro editCancelKeepsDraft once showed an empty composer after Cancel #5810

Description

Problem

Notes

Next step

Activity

ThomasK33 commented on Oct 8, 2026

ThomasK33 commented on Oct 9, 2026

ThomasK33 commented on Oct 10, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions