[Juliusolsson05]

Fixes #1285.

Plan: docs/plans/2026-09-27-ai-workspace-write-status-warning.md, the first commit on the branch.

Problem

• The sequence. The registry owes a preservation copy of rows it could not read (fix(ai-workspace): one malformed row no longer fails every AI Workspace operation (#1246) #1260), and that copy is blocked, for example by a directory on ai-workspaces.json.invalid-<digest>.json. writeFile replaces the user's file on disk, then refreshes the entry status, which saves state, which must make the owed copy first. That step fails.
• The wrong answer. The call returned { ok: false, error: "is a directory" } although the file was written. The editor (or an agent) told "failed" retries, or reports a failure that did not happen.
• Cosmetic. Refused saves relayed the raw errno text, with no hint of how to unblock them.
• Evidence. The sequence was reproduced by fix(ai-workspace): one malformed row no longer fails every AI Workspace operation (#1246) #1260 review B (round 2) and recorded as residual AI Workspace writeFile reports failure after the file landed, when the owed-copy save fails #1285 (steering q24).

Change

• AiWorkspaceRegistry.writeFile. A failed status refresh after a successful write no longer fails the write. It returns ok: true with an optional warning: new on AiWorkspaceWriteFileResult's success branch, documented as "the file IS written; do not retry". It logs the error, and still emits file-written to every workspace holding the file. The owed-copy invariant is unchanged: state is still never saved while a copy is owed.

• preserveOwedCopy. A copy failure now says what is wrong and how to unblock it, keeping only the errno code: "AI Workspace storage needs attention: N unreadable row(s) must be copied aside before saving, and the copy could not be written (EISDIR). "

• Review a (589d42a9):

  • The editor keeps a retained storageWarning from Save and Overwrite, shown through the file list's existing alert and cleared by the next clean write.
  • The warning is main's fixed sentence (AI_WORKSPACE_STATUS_NOT_SAVED), never a raw filesystem message.
  • changed emits are guarded, so a throwing listener cannot fail a landed write or starve other workspaces.
  • The refusal advice now depends on the errno: occupied path, missing folder, permissions or read-only, or full disk.

• Review b (d8f1eae6):

  • The storage notice is now durable in main. get returns a runtime-only storageWarning while an owed copy cannot be written, and the editor reads it on every load, so a workspace switch or remount shows it again.
  • ENOTDIR gets the generic advice, and EROFS gets read-only-volume advice.

• Verification b (5ea05e1d): get also reports AI_WORKSPACE_STATUS_NOT_SAVED while the last state save failed, even when the owed copy itself succeeded, until a later save succeeds.

Tests (AiWorkspaceRegistry.test.ts)

The input is the real recorded registry state, testing/fixtures/ai-workspace/real-workspaces-2026-09-25.json (two workspaces from the owner's file). Its paths are redacted, so one real entry is pointed at a temp file. One real workspace is made malformed so a copy is owed, as the #1260 tests do. The copy path is blocked with a directory.

• The write. writeFile puts v2 on disk, returns ok: true with a warning that the file was saved but the status was not, and leaves the state file unchanged.
• Red without the fix. The test is committed on its own, before the fix: { ok: false, error: 'is a directory' }.
• The refused save. A following create is refused with the actionable message.
• Mutations. Dropping the warning, and misreporting the write as failed, each fail the test.
• Review a cases. The file reads back after the warning; an ordinary write has no warning; a throwing listener still gives ok: true; a missing folder gets ENOENT advice. The always-warn, unguarded-emit and wrong-advice mutations each fail.
• Review b cases. get reports the blocked storage, clears it once the copy is written, and never persists it. The refusal carries no raw error text. A file shared by two workspaces notifies both, even when a listener throws.
• Verification b case. The copy succeeds, then the state write fails: get keeps reporting the unsaved status, and a successful save clears it.
• Not covered at component level: the editor's display of the notice. There is no AiWorkspaceEditor renderer harness, and the change there is three lines.

Verification

• npx tsc -b clean.
• src/main/aiWorkspace: 23 tests, run with --testTimeout 60000 under a machine load of about 60.

🤖 Generated with Claude Code

[Juliusolsson05]

Process gate q107 (W1). History rebuilt, force-pushed over my own c4fb5e74: plan first (da06c8ce), then the fail-first test (ba94fef3, red), then the fix (462a2e22). The test now uses the real recorded registry state, not a hand-built one. It pins that the file is written and that ok:true carries the warning: dropping the warning or misreporting failure each fail it. The body is updated. Reviews start at this head. The one reviewer started earlier was closed before it reported.

[Juliusolsson05]

Disposition for review a (head 589d42a9)

• P2: the editor discards the warning. Fixed. A retained storageWarning, set from the result on both Save and Overwrite and cleared by the next write without a warning, is shown via the file list's existing alert. It is separate from error, which loadWorkspace clears after every save. The text is main's fixed AI_WORKSPACE_STATUS_NOT_SAVED sentence, which says "could not save its status" as you suggested. No raw filesystem message reaches the UI, and the cause is logged. There is no component-level test, because no AiWorkspaceEditor harness exists; this is stated in the body.
• P3: a throwing changed listener turns a landed write into ok:false. Fixed. Each emit is guarded and logged. Test: a throwing listener gives ok:true, the file is v2, and the listener was called.
• P3: wrong advice for non-collision errors. Fixed. The advice now depends on the code: EISDIR, EEXIST or ENOTDIR (occupied path); ENOENT (missing folder); EACCES, EPERM or EROFS (not writable); ENOSPC (full disk); anything else (check the folder is writable). Test: ENOENT gets the missing-folder advice.
• Mutation survivors.
  • "Warning on every write" is now killed (an ordinary write has no warning).
  • "Suppress emits" is killed by the listener test.
  • "Drop the errno" is killed, since the EISDIR and ENOENT codes are asserted.
  • Reads after the warning are asserted.
  • "Synthetic refresh throw" and "clear knownFilePaths" are covered in practice by the read-back assertion, which needs knownFilePaths intact.

[Juliusolsson05]

Disposition for review b (head d8f1eae6)

• P2: the warning disappears on a workspace switch, and is never shown if the switch happens mid-write. Fixed at the source.
  • Main now owns the fact. get returns a runtime-only storageWarning (AI_WORKSPACE_STORAGE_BLOCKED) while a copy is owed and its last attempt failed. The flag is set on the copy failure and cleared on success, and it is never persisted: get returns a copy.
  • The editor's loadWorkspace sets the notice from it on every load. A remount therefore shows it, and so does an editor that missed the write result because the switch happened mid-write.
  • Tests: get reports it; once the path is unblocked and a save succeeds, it clears; the state file never contains it. Dropping get's notice fails the test.
• P3: ENOTDIR and EROFS advice. Fixed. ENOTDIR (a component of the folder path is a file) now gets the generic advice. EROFS gets "on a read-only volume".
• Survivors. Fixed. Appending the raw error message to the refusal now fails the test, since raw text is asserted absent. Stopping after the first workspace's emit fails the new two-workspace test. The editor wiring is still untested at component level; that is stated in the body.

[Juliusolsson05]

Disposition for review c (MERGE-READY at 589d42a9; head is now 2bb9d19c)

• 1–2: the emit on the warning path and the fan-out were unpinned. Fixed in d8f1eae6 (review b's round): a file shared by two workspaces notifies both even when the listener throws. The "stop after the first" mutation now fails.
• 3: four advice branches are untested (EACCES/EPERM, EROFS, ENOSPC, generic). No change. Producing those errnos needs filesystem mocking in this real-filesystem suite. EISDIR and ENOENT are tested end to end.
• 4: the editor display is untested. Acknowledged, as stated in the body. d8f1eae6 moved the durable signal into main's get, which is tested.
• Nits. Fixed (2bb9d19c): the quoted refusal sentence in the plan and body now matches the code, and the test title reads "gives no warning on an ordinary write…".

[Juliusolsson05]

Disposition for verification b (head 5ea05e1d)

• The status-save notice cleared when the copy succeeded but the state write failed. Fixed.
  • The registry now also tracks lastSaveFailed, set by any failed save step and cleared by a successful save.
  • get reports AI_WORKSPACE_STORAGE_BLOCKED while the copy is blocked, otherwise AI_WORKSPACE_STATUS_NOT_SAVED while the last save failed. The editor's next load, or a workspace switch, keeps the notice until a save really succeeds.
• Test: your exact sequence on the real recorded state: blocked copy, unblock, state path made a directory, write gives ok:true with the warning and the copy exists, get still reports it, then a successful save clears it. Red on 2bb9d19c.

[Juliusolsson05]

READY (W1). Head 5ea05e1d. Exact-head CI is green. Every reviewer is MERGE-READY: a (verify-report-a), b (verify2-report-b), c (report-c). History is plan, then red test, then fix (manager gate q107). Dispositions for every round are posted above. merge-gate.sh --member: GATE PASS. Mergeable and clean.