fix(tui): preserve streamed text when final payload regresses (#15452) by TsekaLuk · Pull Request #15573 · openclaw/openclaw

TsekaLuk · 2026-02-13T16:36:41Z

Summary

Fixes a TUI regression where the finalized assistant message can overwrite a richer streamed message with a shorter payload.

Closes #15452.

Problem

When an assistant turn emits text, then tool calls, then more text, the TUI can show the correct full streamed content during delta events, but replace it with a shorter text on final.

Root Cause

TuiStreamAssembler.finalize() rebuilt display state from the final payload and accepted shorter subsets (e.g. post-tool blocks only), which regressed already-correct streamed content.

Changes

src/tui/tui-stream-assembler.ts
- Added mergeTextPreferRicher() to prefer richer cumulative text and reject subset regressions.
- In finalize(), pass the pre-final streamed display as streamedText fallback.
src/tui/tui-stream-assembler.test.ts
- Added regression test: keep full streamed text when final payload drops earlier blocks.
- Added guard test: still accept richer final payload when final extends streamed content.

Tests

pnpm vitest run src/tui/tui-stream-assembler.test.ts src/tui/tui-event-handlers.test.ts src/tui/tui.test.ts
- Passed (18/18).

Broader Validation Notes

I also ran repo-level checks in this environment:

pnpm lint and pnpm check fail on pre-existing unrelated files (src/infra/provider-usage.auth.normalizes-keys.test.ts, src/plugins/discovery.test.ts) due unused vi imports.
pnpm build passed.
pnpm test has unrelated pre-existing failures/timeouts in this environment (e.g. src/security/audit.test.ts, src/gateway/tools-invoke-http.test.ts, src/browser/pw-session.get-page-for-targetid.extension-fallback.test.ts, lobster extension timeouts).

No failures were observed in the touched TUI tests.

Sign-Off

lobster-biscuit

Greptile Overview

Greptile Summary

This PR updates TuiStreamAssembler to avoid a regression where a shorter final assistant payload can overwrite the richer text accumulated during delta streaming. It introduces mergeTextPreferRicher() to retain previously accumulated thinking/content when the next update appears to be a subset, and it changes finalize() to pass the pre-final streamed display text as the fallback into resolveFinalAssistantText().

Tests add two new cases covering the reported regression (final drops earlier text blocks) and ensuring richer finals still win when they extend streamed content.

Confidence Score: 3/5

This PR is close, but the new substring-based merge can produce incorrect final text in some realistic cases.
The targeted regression fix makes sense and tests cover the reported scenario, but mergeTextPreferRicher() uses broad includes checks that can cause the assembler to ignore a legitimate non-empty final payload whenever it is shorter and happens to be a substring of the streamed text, changing the intended “final wins unless empty” behavior.
src/tui/tui-stream-assembler.ts

_{Last reviewed commit: ffcd96e}

greptile-apps

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-02-13T16:38:27Z

src/tui/tui-stream-assembler.ts

+function mergeTextPreferRicher(currentText: string, nextText: string): string {
+  const current = currentText.trim();
+  const next = nextText.trim();
+  if (!next) {
+    return current;
+  }
+  if (!current || current === next) {
+    return next;
+  }
+  if (next.includes(current)) {
+    return next;
+  }
+  if (current.includes(next)) {
+    return current;
+  }
+  return next;
+}


Final text can be wrong

mergeTextPreferRicher() decides “richer” via substring checks (includes). That means a non-empty but shorter final payload can be ignored if it happens to be a substring of the streamed text (e.g. streamed "NOT OK", final "OK"), even when the final payload is the canonical value. This is a behavior change vs resolveFinalAssistantText() (src/tui/tui-formatters.ts:8-15), which prefers any non-empty final text and only falls back to streamed when final is empty. Consider restricting the “prefer streamed” behavior to the specific regression case (final is an exact suffix/prefix caused by dropped blocks) rather than general substring inclusion.

Prompt To Fix With AI

This is a comment left during a code review. Path: src/tui/tui-stream-assembler.ts Line: 14:30 Comment: **Final text can be wrong** `mergeTextPreferRicher()` decides “richer” via substring checks (`includes`). That means a *non-empty but shorter* final payload can be ignored if it happens to be a substring of the streamed text (e.g. streamed `"NOT OK"`, final `"OK"`), even when the final payload is the canonical value. This is a behavior change vs `resolveFinalAssistantText()` (`src/tui/tui-formatters.ts:8-15`), which prefers any non-empty final text and only falls back to streamed when final is empty. Consider restricting the “prefer streamed” behavior to the specific regression case (final is an exact suffix/prefix caused by dropped blocks) rather than general substring inclusion. How can I resolve this? If you propose a fix, please make it concise.

已修复：收敛了 merge 逻辑，仅在“行块前后缀丢失”回归形态下保留 streamed；不再用通用 substring 覆盖非空 final。并新增 NOT OK -> OK 回归测试。

TsekaLuk · 2026-02-13T16:56:01Z

Quick note on the cancelled checks: they were from older head SHAs and got auto-cancelled by workflow concurrency (cancel-in-progress: true) after I pushed follow-up fixes.

Current head SHA is 1694a71, and the latest check suite has been re-triggered for this commit. I will keep monitoring and address any new findings immediately.

TsekaLuk · 2026-02-13T17:18:02Z

I tracked down the CI failure and pushed a minimal follow-up fix in commit 5bd515194.

Root cause was shared lint drift (no-unused-vars) in two test files:

src/plugins/discovery.test.ts
src/infra/provider-usage.auth.normalizes-keys.test.ts

This commit only removes unused vi imports (no behavior change) and retriggers CI.

TsekaLuk · 2026-02-13T17:24:28Z

Heads-up: if the new run fails on (), that is a base-branch lint regression now tracked in #15610.\n\nThis PR already has its own branch-specific fixes applied.

TsekaLuk · 2026-02-13T17:24:42Z

Correction with exact text:

potential external blocker is CI check
failing lint: eslint(no-control-regex)
file path: src/gateway/server/ws-connection.ts

Tracked/fixed in #15610.

TsekaLuk · 2026-02-13T17:31:44Z

Applied the baseline lint fix directly to this branch.

New commit: b672cbd6f

fixes eslint(no-control-regex) in src/gateway/server/ws-connection.ts
keeps branch pnpm check green locally

This should unblock the check job from the shared base-branch failure mode.

TsekaLuk · 2026-02-13T17:35:39Z

Rebased this branch onto latest main and force-pushed to clear merge conflicts.

Current head: ef873aafd
Status now shows mergeable=MERGEABLE; CI has restarted.

…TsekaLuk

…TsekaLuk

…ekaLuk

…ekaLuk

…TsekaLuk

…TsekaLuk

…ekaLuk

…ekaLuk

@TsekaLuk

… thanks @TsekaLuk

…TsekaLuk

…TsekaLuk

…ekaLuk

…ekaLuk

steipete · 2026-02-13T18:13:03Z

Merged via squash.

Prepared head SHA: e4a5e3c
Merge commit: 5cd9e21

Thanks @TsekaLuk!

@steipete