Best possible solution:

Repair resolveCronPayloadOutcome so trusted recovered final assistant text can win for self-corrected all-error payload summaries while unresolved fatal payloads and structured delivery failures still report errors, with focused regression coverage.

Do we have a high-confidence way to reproduce the issue?

Yes, by source inspection. Calling resolveCronPayloadOutcome with only error payloads, finalAssistantVisibleText, and preferFinalAssistantVisibleText: true still resolves to the last error, and the current helper test asserts that shape.

Is this the best way to solve the issue?

No. The PR adds final assistant text only after payload fallbacks have already selected error text and before an unchanged fatal delivery override replaces the result with the error again.

Full review comments:

• [P2] Rework the fatal-error fallback — src/cron/isolated-agent/helpers.ts:263-270
  This still does not let the recovered all-error case use finalAssistantVisibleText: the payload fallback helpers return the last error first, and hasFatalStructuredErrorPayload still forces fatalDeliveryText to replace the output. Move the recovery/fatal classification before the fatal override and add a regression for only error payloads plus final assistant text.
  Confidence: 0.94

Overall correctness: patch is incorrect
Overall confidence: 0.92

Acceptance criteria:

• node scripts/run-vitest.mjs src/cron/isolated-agent.helpers.test.ts
• Run a real cron agentTurn scenario or equivalent packaged/local cron run showing the final assistant reply is announced instead of a recovered tool error, with private data redacted.

What I checked:

• PR diff scope: The PR only inserts normalized final assistant text into fallback summary/output selection; it does not change fatal error classification or the fatal delivery override. (src/cron/isolated-agent/helpers.ts:258, ae085b9e6bcd)
• Current resolver still treats all-error payloads as fatal: Current main computes hasFatalStructuredErrorPayload when an error payload exists with no later successful payload, blocks final assistant text, and later replaces summary/output with fatalDeliveryText. (src/cron/isolated-agent/helpers.ts:295, a6aa48350b7e)
• Regression shape is pinned by tests: The helper test passes two error payloads plus finalAssistantVisibleText: "Recovered final answer" and still expects outputText and delivery to be last error. (src/cron/isolated-agent.helpers.test.ts:284, a6aa48350b7e)
• Cron run status follows resolver output: The isolated cron runner passes payloads and finalAssistantVisibleText into resolveCronPayloadOutcome, then records run status from hasFatalErrorPayload. (src/cron/isolated-agent/run.ts:962, a6aa48350b7e)
• Linked issue and PR remain open: Live GitHub metadata shows [Bug] Cron task summary ignores agent self-correction - picks tool error over final assistant reply #74807 is open and this PR is its open closing implementation candidate.
• No contributor real behavior proof: The PR body and only current comment contain no after-fix terminal output, logs, recording, screenshot, or linked artifact demonstrating the fixed cron behavior in a real setup. (ae085b9e6bcd)

Likely related people:

• welfo-beo: Authored merged work that added finalAssistantVisibleText handling across the isolated cron resolver, runner, and tests. (role: feature introducer; confidence: high; commits: 81c7304a18b8, 01bfe4f25e3f; files: src/cron/isolated-agent/helpers.ts, src/cron/isolated-agent.helpers.test.ts, src/cron/isolated-agent/run.ts)
• jalehman: Reviewed/co-authored the final announce delivery work and appears in the follow-up final-reply scoping history for this cron surface. (role: reviewer and adjacent contributor; confidence: medium; commits: 81c7304a18b8, a97e0abb2bf0, f3928f79ebc5; files: src/cron/isolated-agent/helpers.ts, src/cron/isolated-agent.helpers.test.ts, src/cron/isolated-agent/run.ts)
• hydro13: Authored merged resolver work that preserved multi-payload announce output while keeping error-only runs falling back to the last error payload. (role: adjacent cron output contributor; confidence: medium; commits: bdd9bc93f12c, 779d2ce830ff; files: src/cron/isolated-agent/helpers.ts, src/cron/isolated-agent.helpers.test.ts)
• steipete: Recently changed the same resolver and tests around presentation warnings, denial classification, and run-level failure behavior. (role: recent classifier contributor; confidence: medium; commits: 9f62c738930a, 382e03a2d838, b0c70786fd63; files: src/cron/isolated-agent/helpers.ts, src/cron/isolated-agent.helpers.test.ts, src/cron/isolated-agent/run.ts)

Remaining risk / open question:

• The contributor has not supplied after-fix real cron evidence, so merge would rely only on source review and tests.
• A repair must distinguish recovered final assistant text from genuinely fatal tool, run-level, denial, and structured delivery failures.

Codex review notes: model gpt-5.5, reasoning high; reviewed against a6aa48350b7e.