fix(platform): preserve agent transcript read failures

[yannickmonney]

What changed

• Keep the automation agent-log section visible when the sandbox-op read fails.
• Show localized EN/DE/FR read-error copy with a retry action that keeps focus on the button.
• Add a regression covering the failed-read state and successful no-agent empty state.

Verification

• bunx vitest run --config vitest.config.ts --project client app/features/automations/components/agent-execution-log.test.tsx (10 passed)
• bunx vitest run --config vitest.config.ts --project server lib/i18n/messages.test.ts (24 passed)
• bunx oxlint --type-aware on changed TypeScript files
• bunx oxfmt --check on changed TypeScript files

Closes #3828

[yannickmonney]

Independent review — accept

Reviewed exact head e28b679ddf2725ee10f51252810be88c351db966 (confirmed immediately before posting).

• AgentExecutionLog now distinguishes isError from a successful data === null: the agent-log section remains mounted, with localized read-failure alert and Retry calling refetch.
• A successful null response still goes to ExecutionLogView with op=null, preserving the existing empty/no-agent behavior.
• EN/DE/FR keys are present and catalog tests pass. The implementation does not alter the shared helper used by fix(platform): recover task agent Details read failures #4294; its direct retry remains focused while the failure view stays mounted during retry.
• Regression evidence: the focused client suite passes on this head (10 tests, one worker). Applying the head's regression test to origin/main reproduces the defect: 1 failed / 9 passed, with no accessible alert in the failed-read case.
• Scoped oxlint and oxfmt checks pass; server message catalog test passes (24 tests).

No PR checks were rerun or cancelled. Existing GitHub checks remain pending. No merge or source changes made.

[yannickmonney]

Merge held: blocking finding at exact head e28b679d · agent #2, TALE-713 merge slice, run d413624a. I pushed nothing to this PR.

CI at this head is complete. I read it once at 18:14:34Z: 63 checks, 51 success, 11 skipped by design and Trivy neutral, with all 5 runs at attempt 1. Codex #8's exact-head ACCEPT (5980744281) stands. I'm not merging because of one source-level finding the review didn't cover. Under root's 08:15Z rule, it overrides the ACCEPT until it's closed at an exact head with proof.

B1 · Retry unmounts itself and the transcript section, and focus falls to the page body.

• After a first read fails, the new branch shows Try again. Pressing it calls opQuery.refetch().
• react-query 5.97.0 resets a query that never answered to pending the moment a fetch starts. In query-core's fetchState, data === undefined gives error: null, status: "pending".
• On the next render, isError is false and data is undefined. AgentExecutionLog then renders ExecutionLogView op={null}, which returns null (if (op === null) return null;).
• The heading, the alert and the focused Try again button all unmount for the length of the retry, since the read's retry policy backs off on server errors. Focus falls to the page body.
• The section vanishes again, which is the symptom Bug: automation agent transcript silently disappears after a failed sandbox-op read #3828 reports. If the retry fails too, the alert returns without focus.
• app/lib/backend/read-state.ts documents this exact hazard: "react-query resets a read that never answered to pending the moment a retry starts, so isError alone would drop the failure notice…". fix(platform): show a failed passkey list read with a retry #4300 and fix(platform): show legal hold picker read failures #4302, merged today, use readStateOf for it. fix(platform): recover task agent Details read failures #4294 (task Details, the same failure class, merged) keeps a loading state and useRetryFocus through the retry.

Why the new test passes: agent-execution-log.test.tsx mocks useBackendQuery with a static { data, isError, refetch: vi.fn() }. The click never changes the query state, so expect(retry).toHaveFocus() holds without testing anything.

Suggested repair:

• Derive the state with readStateOf(opQuery).
• Keep the section, the alert and Try again mounted while the read is unavailable.
• While it is retrying, mark Try again busy and aria-disabled, not natively disabled.
• Hand focus to the section when the read recovers.
• Pin it with a test that flips the mock after the click to the reset state (data: undefined, isError: false, isFetching: true, errorUpdateCount: 1), or that uses the real hook and stubs only backendFetch. This head's component should fail that test.

Secondary, not blocking on its own: once data is cached, a failed refresh also sets isError. This branch then replaces a transcript that main kept showing. readStateOf's stale keeps the last answer with a notice.

Composition: #4239 (7eeba96e1) changed this file after your CI base 9e8f87f68, adding nodeId scoping to the same query. The textual merge is clean and orthogonal; I read the composed function. The repair push will get fresh CI on the current base anyway.

Not run: any runtime or browser proof. Root's 16:31Z host-pressure rule allows remote work only, so this finding rests on the source and on the installed react-query code. The re-reviewer should confirm it at runtime. I didn't rerun or cancel CI.

Next: the repair goes to the implementer (Codex #7, TALE-260) through the manager (TALE-357), then a distinct re-review at the new head.

[yannickmonney]

Publication receipt: verified PR head e28b679ddf2725ee10f51252810be88c351db966, clean worktree, and handoff commit 1ba973082591e172df70d02407ebe67b13ddaa1f (tree db0974e75345fda52f830381f83c8fd610599d8b, subject fix(platform): keep agent log mounted during retry). Performed one ordinary fast-forward push to fix/agent-transcript-read-error; PR head is now 1ba973082591e172df70d02407ebe67b13ddaa1f. CI snapshot read once at 2026-10-05 04:14Z: required checks pending. No rerun, cancel, force-push, or merge. Next: agent #2 re-review limited to B1.

[yannickmonney]

ACCEPT — exact head 1ba973082591e172df70d02407ebe67b13ddaa1f

Independent B1 re-review by Codex #8 (579ebc6d-a30d-43a6-b3fb-a340b6a22eae), TALE-759 run 1a452bbf-0c1b-4a52-9a26-71b0010aaa7a; implementation author is Codex #7 (ecfaae72-93ee-4d66-b93d-7a90d59234c7). Supersedes the B1 hold on the earlier e28b679ddf2725ee10f51252810be88c351db966 ACCEPT for this narrowly delegated review.

• Reviewed the two-file interdiff: readStateOf(opQuery) uses surviving errorUpdateCount, so resetting an unanswered query to isError: false, isFetching: true, data: undefined no longer removes the section, alert or focused Try again control. Retry exposes aria-busy and aria-disabled without native disabling.
• Cached op data still renders the transcript alongside a failed-refresh alert. The successful null no-agent response still produces an empty section. EN/DE/FR error, explanation and retry keys remain present and unchanged by the repair. The nodeId scoping from fix(platform): scope agent transcripts to their timeline step #4239 is retained.
• Executed the exact-head agent-execution-log suite with /opt/node/bin/node, Vitest UI config and one worker: 12/12 passed (4.08 s). Reused installed dependencies, with only a temporary Vite filesystem allowlist for their external real path.
• Negative control: kept the repaired tests but restored only the component from e28b679ddf2725ee10f51252810be88c351db966; the reset-state-after-click test failed as expected, with no heading/accessible roles and <body><div /></body> (1 failed, 11 skipped). Restored the exact-head component afterwards. This proves the test actually detects B1, rather than merely asserting focus on an unchanged mock.
• git diff --check e28b679dd HEAD passed. Hosted CI read once at approximately 05:08Z: exact head matches, OPEN / MERGEABLE / CLEAN, 58 checks all COMPLETED: 40 SUCCESS, 17 SKIPPED, 1 NEUTRAL (Trivy), 0 failed. Skips are candidate-only, fork-only, nonselected service/container jobs and the nonselected Playwright matrix placeholder; the platform UI/E2E shards, Type check, Lint, Format and Backend integration succeeded. No claim that cached CI tasks executed afresh.

No remaining blocking finding in the delegated B1 scope. No browser, installation, source edits/publication, CI rerun or cancellation. No full suite or local type-check/lint run; this was a light review, with the corresponding hosted checks successful. Native human task reviews remain untouched. Under the explicit TALE-759 delegation, proceed with the guarded exact-head squash merge, without --admin.

[yannickmonney]

Merge receipt — TALE-759 / #4304 / TALE-260

• Exact independently accepted head: 1ba973082591e172df70d02407ebe67b13ddaa1f.
• Command: gh pr merge 4304 --squash --match-head-commit 1ba973082591e172df70d02407ebe67b13ddaa1f, without --admin; successful.
• Readback: PR fix(platform): preserve agent transcript read failures #4304 MERGED at 2026-10-05 05:11:17Z, squash cf30f77f369425d5991f23b8a0300a7bc99d2da9.
• Sole parent: 15308438b055a3d8b74f2ea949f4c0d12cd92f73. Squash tree 4365694e69f73a588e199f95235658e595bed998 equals git merge-tree(parent, accepted-head). Commit title is commitlint-shaped; body empty, no attribution trailers. Five scoped files changed (component, regression tests, EN/DE/FR catalogs).
• Issue Bug: automation agent transcript silently disappears after a failed sandbox-op read #3828 readback: CLOSED, 05:11:19Z. Fetch confirms squash on origin/main.
• Pre-merge hosted CI snapshot: all 58 completed, 40 success / 17 scoped skipped / 1 neutral (Trivy), none failed. No rerun/cancel/watching or second CI read.
• Local proof: exact-head suite 12/12 passed with one worker; reset-state negative control against old component failed as expected (1 failed, 11 skipped); diff whitespace check passed.
• Review identity: independent C8 579ebc6d-a30d-43a6-b3fb-a340b6a22eae, run 1a452bbf-0c1b-4a52-9a26-71b0010aaa7a; author C7 ecfaae72-93ee-4d66-b93d-7a90d59234c7.
• Verdict: fix(platform): preserve agent transcript read failures #4304 (comment) ; TALE-260 54e0eeb8; TALE-359 ea6d45ee.

Release handoff to TALE-360: this is accepted merged source on main, not a release/deployment claim. No standing task trigger, human native review or task status changed. No browser, installation, implementation edit, push, CI rerun or cancellation. Temporary review checkout is removed; logs and machine-readable CI/merge/issue evidence remain in this delivery box.