fix(platform): preserve automation edits made during saves

[yannickmonney]

What changed

• Only clear an automation draft when it is still the immutable snapshot submitted by that save. Node edits made while Save anyway is pending remain visible and dirty after the append succeeds.
• Use the append response version as the retained draft’s next base version, including when the detail query has not refreshed yet.
• Add deferred-response UI regressions for both query timings and a stale-dialog cancellation control; retain the existing ordinary-save, Save anyway and reload controls. Document the automated/manual proof boundary in the automation coverage register.

Verification

• Rechecked origin/main at 9e8f87f6832cf352eb70a0ad666bbec6fd9d1a57; both new pending-append cases fail there by replacing Later unsaved draft with the stored/submitted prompt (regression-main.log). No open PR covered Bug: Save anyway discards newer node edits made while the automation version is saving #3607.
• One-worker Node/Vitest jsdom: editor 47/47 passed; inspector and detail-shell 38/38 passed. Existing axe audits included. An initial fixture run exposed its non-reactive pending mock; the final fixture explicitly rerenders the hook result after completion.
• Scoped oxlint --type-aware --type-check --threads=1: exit 0. Changed-file oxfmt --check, git diff --check, conflict-marker check and manual-layer lint: passed.
• Platform catalog parity/usage checks: 8 passed, 16 non-selected checks skipped (EN/DE/FR full catalogs and sparse locale handling).
• Scoped pinned Opengrep: 0 findings (test fixtures excluded by repository policy).
• Boundary sweep: no new rendered controls or strings; existing localized labels/accessibility stay unchanged. No backend, authorization, dependency, infrastructure or schema changes; no migration needed. Product guides’ save/version semantics remain unchanged.

The local worktree reuses installed dependencies; its delivery-only Vitest config extends the repository UI config, permits that dependency path and uses a 30s timeout on this shared host. No repository config or dependency change.

Remaining proof and coordination

• Real-browser two-session/network timing, layout/focus/contrast, browser component tests, E2E and backend-stack proof were not run, as required by the light-only dispatch. jsdom verifies the corrected local control behavior with the real editor/inspector, mocked data hooks and canvas layout seam; it does not claim network or geometry measurements.
• TALE-165 / Bug: Automation General tab silently replaces unsaved trigger and project edits on refresh #3620 is changing General-tab editors in parallel; no production-file overlap. My only shared documentation candidate is services/platform/tests/manual/reference/automation.md, where this PR adds one separate row. Coordination posted on TALE-165.
• No merge or self-approval; independent review and remaining browser acceptance belong to the review lane.

Closes #3607

[yannickmonney]

Verification at exact head aab8eb732b624ca238d3951b6a7f7e1ecf7509b4: the final regressions were also rerun against the exact main editor source and both fail at the later-Prompt assertion. Restoring the PR source yields 3/3 focused passes (both query timings and stale Cancel); the tracked tree is clean. Full editor suite remains 47/47, adjacent inspector/detail-shell 38/38, locale parity/usage 8 passes, scoped lint/type diagnostics/format/manual/security and commitlint pass.

CI handoff: after approximately 13 minutes of watch, all five source-resolution jobs remain queued; the inspected ubuntu-latest job has no runner assigned and no failure diagnostic. Only the local watcher was stopped; no CI cancellation/rerun. CI is not green and acceptance is not complete. Browser/two-session/network/layout/focus/contrast, browser component tests, E2E and backend stack remain unrun under the light-only dispatch; independent review/remaining acceptance belongs to the review lane.

TALE-165 / #3620 production files are disjoint. This PR adds one independent row to services/platform/tests/manual/reference/automation.md; its author has been notified of that possible shared documentation file. No peer PR for #3620 was available at this handoff. No merge or self-approval.

[yannickmonney]

TALE-677: REQUEST CHANGES for PR #4313

Exact reviewed head: aab8eb732b624ca238d3951b6a7f7e1ecf7509b4. Its merge base is 9e8f87f68. It merges cleanly with origin/main 3af9e704d, whose automation-editor.tsx is byte-identical to the base's.

Reviewer: agent #6 53fe2a33, run 417d3836 (TALE-677). That is independent of the author, Codex #7 ecfaae72 (run f99d5af6). I read TALE-152 and #3607 as the contract.

The fix works for the sequence the issue reports:

• An edit made while Save anyway is pending survives as a dirty draft.
• The submitted version carries only the submitted draft.
• Save anyway sends baseVersion from the refusal's latestVersion.

One blocking defect remains.

B1 (blocking, data integrity): a draft restarted during the append gets the append's version as its base

submitSave now sets draftBaseRef.current = saved.version (automation-editor.tsx:557) for whatever draft is on screen when the append lands. That base is only right for a draft that grew out of the submitted one.

While Save anyway is pending, the version picker (:758) and its Discard and switch confirm (:1006) still work. So the draft on screen can be a new one, started on another version with its own pinned base (:360).

Repro (jsdom probe with the real editor, inspector and version picker):

1. Edit Prompt to Submitted draft → Save → Save version → stale refusal (v4 landed) → Save anyway. The append stays pending.
2. While it's pending: Version → v2 → Discard and switch → edit Prompt to Edited on v2. This pins the base to 2.
3. The append lands as v5. Then Save → Save version.

Observed at the head:

P1 save-anyway call: baseVersion=4 prompt="Submitted draft"
P1 during append: picker=v2 prompt="Edited on v2"
P1 after append: picker=v5 prompt="Edited on v2" dirty=true
P1 next save: baseVersion=5 prompt="Edited on v2"

Why it matters:

• The store accepts baseVersion: 5, because the latest is 5 (store.ts:227 refuses only when latest !== baseVersion).
• So the next Save appends v2-plus-edit as v6, with no warning. It silently replaces what v5 holds (the draft Save anyway just saved) and every change since v2, while the picker reads v5.
• Without a pending save, the same v2-based draft is refused and gets the stale-version dialog.
• This breaks AUTO-B7's judgment: "Nothing is ever reverted silently".

On main, the same sequence drops Edited on v2 (the #3607 loss). So neither main nor the head handles it.

Fix, either way:

• Settle only the submitted draft's own line. Rebase or clear only when the draft on screen descends from the submitted one. A local sketch (p1-repair-sketch.diff, three hunks) does this: it bumps an epoch counter where the base is pinned and checks it after the await. With it, the probe sends baseVersion: 2, and the PR suite (47/47) and the probe (2/2) are green. No new copy.
• Or lock editing during the append. That is the issue's first option, with fix(platform): block edits during project-agent saves #4256 (project agents) as precedent. Keep the stale dialog open with isLoading, as the save dialog already is, or make the inspector and the version switch inert while save.isPending.

Either fix needs a regression test for this sequence.

Criteria

Check	Result
Later edit during a pending Save anyway survives as a dirty draft	✅ Both new cases (query refreshed or not) fail on main (2 failed / 45 passed; received One sentence, please. / Submitted draft) and pass at the head
Submitted version carries only the submitted draft; baseVersion from latestVersion	✅ mock.calls[1] keeps Submitted draft and sends baseVersion: 4
Controls: stale Cancel keeps the draft and Save; an ordinary save clears dirty	✅ New Cancel test (also green on main, as a control should be); existing appends the version … and clears the draft
Retained draft's next base	❌ B1
Mutant: drop draftBaseRef.current = saved.version	Both new cases fail on the next-save baseVersion, so the tests pin the rebase
Composition with #4321 (TALE-165)	No shared files
Composition with #4278 (TALE-252)	Shares automation-editor.tsx and automation-editor.test.tsx. Main + #4313 + #4278 (new head db96fff2, updated 14:59Z) + #4321 merges cleanly, and the 6 automations suites pass 134/134. Probe case P1 reproduces there too (baseVersion=5). #4278's new readStateOf(...).unavailable gate fires only when nothing has loaded, so it doesn't hide a retained draft. An earlier run against #4278 4a2ae222 also passed, 130/130
EN/DE/FR	No new copy, so no catalog change is needed. The EN/DE/FR editor guides stay accurate

Minor, non-blocking:

• The identity check keeps an edit-then-undo draft dirty with no diff (probe P2). It's harmless.
• The new register row could cite AUTO-B7.
• No manual box covers an edit during a pending Save anyway, so Bug: Save anyway discards newer node edits made while the automation version is saving #3607's "verify in the local UI" step is still manual.

Evidence

All runs are local: installed Node v24.21.0, Vitest 4.1.11, jsdom, one worker.

• Head: automation-editor.test.tsx 47/47. node-inspector + automation-detail-shell 38/38.
• Red check: main's automation-editor.tsx with the PR's tests gives 2 failed / 45 passed.
• Probe automation-editor.review-probe.test.tsx, which is the PR's harness plus 2 cases:
  • head: P1 fails, expected 5 to be 2;
  • main: P1 drops the v2 edit;
  • sketch: 49/49;
  • composition: P1 fails, expected 5 to be 2.
• Artifacts: the probe, p1-repair-sketch.diff and all logs are in the TALE-677 delivery box /agent/output/83271ddc-720b-4ada-9867-c85c7b8b8c82/.
• Static checks: oxfmt --check clean (2 TS files). oxlint --type-aware --type-check --threads=1 reports 0 diagnostics; I checked that the flag reports TS2322/TS2551 on a planted error. lint:manual OK (5 trees, 1448 boxes).

Unrun

• Real browser: two sessions, throttled network, focus and layout.
• *.browser.test.tsx, E2E and the backend stack.
• tsc (type errors were covered by oxlint's type check), the full platform suite, bun run check, SAST, knip and commitlint.

CI

At 15:04Z, all five exact-head workflows (Checks, Build, E2E, SAST, Commitlint) had been queued since 14:07Z. 254 runs were queued repo-wide and 3 were running. CI is not green. I didn't rerun or cancel anything.

Under root's 08:15Z rule, B1 holds until it is closed on an exact head with proof. I didn't merge, push or move a card.

[yannickmonney]

Admitted publication receipt — 2026-10-04 22:33Z

TALE-152 / PR #4313, author #7, publication run 028a4505-279e-4977-99e6-c2460e8aba3b.

Consumed manager Answer 4bf01553-3ef1-48b3-8877-418ef0775fb9, manager run 097e735f-ed36-453d-9879-5795478ad85c: item 3 of the single publication batch admitted at the 00:00 Zurich full pass (checkpoint e45460d8).

Verified exact local commit 98cd7537c63d664c14aae8d2ccef26a0aaa87e6e, clean task-owned branch checkout, parent and remote PR head aab8eb732b624ca238d3951b6a7f7e1ecf7509b4. One ordinary fast-forward push on the existing branch succeeded. Remote PR head now matches 98cd7537c63d664c14aae8d2ccef26a0aaa87e6e. No new branch, PR, commit or force push; unrelated standing checkout untouched; temporary verification checkout removed.

One short readback captured workflow IDs: Checks 37240457510, SAST 37240457512, E2E 37240457514, Commitlint 37240457524, Build 37240457536. At the snapshot SAST was in progress and the other workflows queued; substantive CI is not yet green. Exact check-run IDs/statuses are retained in publication-check-runs.json (21 checks: 16 queued, 1 in progress, 4 candidate-source checks skipped). No polling, rerun or cancellation.

B1 repair and prior local evidence are unchanged: two regressions fail at the reviewed parent; 87 focused UI and 167 composition UI tests pass with the repair; scoped local gates pass. No install, build or tests repeated in this publication run.

Next owner: manager/review coordinator to arrange distinct exact-head B1 re-review and follow CI. B1 remains open until independent closure; no self-review, approval or merge. Browser two-session/network timing, visual/focus/contrast, E2E/backend stack and full-workspace proof remain unrun locally. Separate #4278 orphan-key failure stays with its existing owner. #4278 shares both editor files; #4321 shares none; #4288 shares only the original manual register.

Outcome: published for independent review, not accepted or merged.

[yannickmonney]

TALE-722: ACCEPT for PR #4313 at 98cd7537

Exact reviewed head: 98cd7537c63d664c14aae8d2ccef26a0aaa87e6e, one commit on aab8eb73 with merge base 9e8f87f68.

• GitHub reads it as MERGEABLE.
• git merge-tree with origin/main d5e6f2ca2 is clean.
• Main hasn't touched automation-editor.tsx or its test since the base.

Reviewer: agent #6 53fe2a33 (TALE-722, manager dispatch d5c5c8c2). I raised B1 (4f426cad, record), and I'm independent of the author, Codex #7 ecfaae72.

B1 is closed

How the repair works. submitSave captures draftEpochRef before the await.

• The epoch goes up where a draft starts (its first edit, next to the base pin) and wherever one is dropped. Drops all go through discardDraft, which Discard and switch and the stale reload now use too.
• After the append, the dialog, base, draft, message and picker are settled only if the epoch hasn't changed.
• The file has only three setDraft calls, and all of them go through these paths. So a draft can't be swapped any other way while an append is in flight.

Case (real editor, inspector and picker in jsdom)	aab8eb73	98cd7537
P1, my B1 repro: during the append, Discard and switch to v2, then edit	Next save sends baseVersion=5, so v5 is replaced silently	Sends baseVersion=2 and the picker stays on v2, so the store refuses it (v5 is the latest)
P3: switch to v2, then back to v3 with no draft, then edit on v3	baseVersion=5	baseVersion=3
P4, a control: Cancel on Discard and switch, then keep editing	Rebased on v5	Unchanged: rebased on v5, picker on latest, message cleared
The PR's two real-picker regressions, settles only its own draft… (edited while pending: true / false)	Red	Green

While the append is pending, the header's Save and Discard are disabled. So the picker's Discard and switch is the only reachable way to drop the submitted draft, and P1 covers it.

Controls and assertions

• Controls pass at the head:

  • ordinary success: appends the version with the typed message and clears the draft;
  • stale Cancel: keeps the draft when the stale-version decision is cancelled;
  • both cases of keeps later edits dirty after Save anyway succeeds.

• Assertions: the test diff from aab8eb73 to 98cd7537 only adds lines (+101 / −0). Against the base, the only lines removed are:

  • the import line;
  • two mockResolvedValue(undefined) calls, which now return { name, version } because submitSave reads saved.version.

  No assertion was weakened.

• Mutants, each run on the editor suite plus my probe (52 tests):

  • Dropping the epoch check fails 4: both new regressions, P1 and P3.
  • Dropping the bump in discardDraft fails 1: the new regression with pending false.
  • Dropping the bump at the first edit fails 0. That bump is redundant today, because every drop during an append already goes through discardDraft. It's harmless as a backstop.

Non-blocking

• N1: when the epoch has changed, setSaveMessage('') is skipped too. The landed version's message then stays prefilled in Version message for the replacement draft's save (P1: "Fix prompt"; P4 and aab8eb73: ""). That message belongs to the version that landed, so it could be cleared before the epoch return. You can see it and edit it, so nothing is silent.
• N2: the manual register row added at aab8eb73 still describes only the rebase. It could add that a draft dropped or switched during the append keeps its own base and picker.

Run locally

All runs used installed Node v24.21.0, Vitest 4.1.11 and jsdom, with --maxWorkers=1 and no install. No z.number collection error appeared, so /opt/node/bin/node wasn't needed.

• Suites at the head: automation-editor, node-inspector and automation-detail-shell pass 87/87.
• Probe at the head: automation-editor.rereview-probe.test.tsx (the PR's harness plus P1, P3 and P4) passes 3/3.
• Red checks: the editor suite plus the probe, with only automation-editor.tsx swapped:
  • with aab8eb73's editor: 4 failed, 48 passed (the 2 new regressions, P1 and P3);
  • with the base's editor: 7 failed, 45 passed.
• Artifacts: the probe, logs and mutant results are in TALE-722's delivery box.

Hosted CI at 98cd7537 (read once, 23:09Z)

• Passed: Build, E2E (Playwright 4/4), SAST and Commitlint.
• Checks failed, on the Unit job only. The failing test is in @tale/cli: scripts/ci-e2e-optimization.test.ts:331, "browser version lookup propagates failure instead of restoring an empty key". This PR doesn't touch that file.
• The rest of Checks passed: UI (4/4), Browser, Type check, Lint, Format, Knip and Performance.
• Likely cause: main's a22944679 (22:38Z) rewrote exactly the failing assertion, expect(result.exitCode).toBe(succeeds ? 0 : 19). That landed after this run's merge ref was built at 22:33Z, and main's own Unit job at d5e6f2ca2 passed.

So CI is not green at the exact head. A rerun or a main update is the manager's call. I didn't rerun or cancel anything.

Not run

• Real browser: two sessions, throttled network, focus and layout. TALE-152's "verify the corrected behavior in the local UI" step therefore stays open.
• *.browser.test.tsx locally, E2E and the backend stack.
• The full platform suite, bun run check and local tsc. Hosted Type check, Lint and Format passed at this head.
• A fresh composition with fix(platform): recover automation editor from detail read errors #4278, which shares both editor files. Its orphan-key issue is separate and not this PR's.

Under root's 08:15Z rule, B1 is closed on this exact head, with the proof above. I didn't merge, push, take any CI action or move a card.