Skip to content

fix(streaming): drop silent delta-discard in _merge_stream_text - #4085

Merged
WillemJiang merged 1 commit into
bytedance:mainfrom
nankingjing:fix/delta-stream-accumulation-cjk
Jul 12, 2026
Merged

fix(streaming): drop silent delta-discard in _merge_stream_text#4085
WillemJiang merged 1 commit into
bytedance:mainfrom
nankingjing:fix/delta-stream-accumulation-cjk

Conversation

@nankingjing

@nankingjing nankingjing commented Jul 11, 2026

Copy link
Copy Markdown
Contributor

Problem

The dual-mode _merge_stream_text used short-circuits chunk == existing and existing.endswith(chunk) that silently dropped legitimate delta chunks in messages-tuple mode:

  • CJK reduplication: '谢谢' tokenized as ['谢','谢'] → buffer stays '谢'
  • Repeated tokens: 'gogo' as ['go','go'] → buffer stays 'go'
  • Suffix-matching tails: 'hello' as ['hel','l','o'] → drops 'l'

Delta payloads must always append; only strictly-longer cumulative snapshots should replace.

Fix

Both copies of _merge_stream_text are fixed:

  • app/channels/manager.py (IM streaming to Feishu/Telegram/etc.): Gates cumulative-replace with len(chunk) > len(existing), keeps exact-resend guard for values-snapshot no-ops.
  • deerflow/tui/view_state.py (TUI client rendering): Simplified to always-append delta logic; exact-re-send detection moved to caller reduce() with len(action.text) > 1 guard so single-char CJK deltas aren't mistaken for re-sends.

Tests

  • test_tui_view_state.py: +8 regression tests (CJK reduplication, repeated tokens, suffix-matching, cumulative, empty, newline split, genuine append)
  • test_channels.py: +8 equivalent regression tests

All existing tests continue to pass (27/27 TUI, 258/258 channels).

🤖 Generated with Claude Code

@github-actions github-actions Bot added area:backend Gateway / runtime / core backend under backend/ risk:medium Medium risk: regular code changes size/M PR changes 100-300 lines labels Jul 11, 2026
willem-bd
willem-bd reviewed Jul 11, 2026
View reviewed changes

@willem-bd willem-bd left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for this fix. The view_state.py change is correct, but I found one issue in the manager.py path that breaks the PR's own tests.

Comment thread backend/app/channels/manager.py Outdated
if len(chunk) > len(existing) and chunk.startswith(existing):
return chunk
# Exact re-send (values snapshot re-emitting history): no-op.
if len(chunk) == len(existing) and chunk == existing:

@willem-bd willem-bd Jul 11, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P1] Exact-match no-op check drops legitimate equal-chunk deltas — the PR's own tests fail

  • Location: backend/app/channels/manager.py:379 (_merge_stream_text)
  • Problem: The added exact-match check if len(chunk) == len(existing) and chunk == existing: return existing fires before the delta-append at line 384. For an incoming delta that happens to equal the accumulated buffer (CJK reduplication '谢' + '谢', repeated tokens 'go' + 'go', split newlines '\n' + '\n'), it returns existing unchanged, discarding the delta. The PR's own added comment at lines 381-383 states this case should append ('谢' + '谢' = '谢谢'), and the new tests in test_channels.py assert it — but they fail against this implementation.
  • Evidence: _get_merge_stream_text() imports _merge_stream_text from app.channels.manager (test_channels.py:7320), and test_merge_stream_text_cjk_reduplication asserts _merge("谢", "谢") == "谢谢" (test_channels.py:7328). Tracing the head code: for ("谢","谢"), len(chunk) == len(existing) is 1 == 1 → True and chunk == existing → True, so line 380 returns "谢" instead of "谢谢". The same breaks test_merge_stream_text_repeated_token_append ("go"+"go" → returns "go") and test_merge_stream_text_newline_split ("\n"+"\n" → returns "\n"). The view_state.py version was fixed correctly (no exact-match check); manager.py was not.
  • Suggested fix: Remove the exact-match check (lines 378-380) from manager.py's _merge_stream_text to match view_state.py. The channels path feeds only delta (messages-tuple) events to this function — values snapshots are handled in the separate elif event == "values" branch at line 1732 — so the no-op guard serves no purpose here and only drops legitimate equal-chunk deltas. The strictly-longer cumulative check at line 376 already handles true cumulative re-delivery.
  • Test: The three tests above (test_merge_stream_text_cjk_reduplication, test_merge_stream_text_repeated_token_append, test_merge_stream_text_newline_split) already cover this — they need to pass. cd backend && make test.

@nankingjing @claude
fix(streaming): drop silent delta-discard in _merge_stream_text
8d007e4 
The dual-mode _merge_stream_text used short-circuits chunk==existing
and existing.endswith(chunk) that silently dropped legitimate delta
chunks in messages-tuple mode:

- CJK reduplication: '谢谢' tokenized as ['谢','谢'] -> buffer stays '谢'
- Repeated tokens: 'gogo' as ['go','go'] -> buffer stays 'go'
- Suffix-matching tails: 'hello' as ['hel','l','o'] -> drops 'l'

Delta payloads must always append; only strictly-longer cumulative
snapshots should replace. The fix gates the cumulative-replace branch
with len(chunk) > len(existing) and removes the dropped-shape guards.

Both copies are fixed:
- app/channels/manager.py (IM streaming to Feishu/Telegram/etc.)
- deerflow/tui/view_state.py (TUI client rendering)

The TUI reduce() caller now handles exact re-sends (values snapshot
re-emitting history) before the merge, and the len>1 guard prevents
single-char CJK deltas from being mistaken for re-sends.

Co-Authored-By: Claude <noreply@anthropic.com>
@nankingjing
nankingjing force-pushed the fix/delta-stream-accumulation-cjk branch from b778bd2 to 8d007e4 Compare July 12, 2026 01:46
@nankingjing

nankingjing commented Jul 12, 2026

Copy link
Copy Markdown
Contributor Author

Thanks for the careful review! Fixed in commit 8d007e4 — removed the exact-match no-op at manager.py:378-380 so channels/manager.py now matches view_state.py. All 8 tests in test_channels.py (test_merge_stream_text_*) pass locally with this change.

@WillemJiang WillemJiang added this to the 2.1.0 milestone Jul 12, 2026
@WillemJiang
WillemJiang merged commit b650456 into bytedance:main Jul 12, 2026
11 checks passed
@ivan-skvorc ivan-skvorc mentioned this pull request Jul 12, 2026
Merged
6 of 9 tasks
lllyfff added a commit to qin-chenghan/deer-flow that referenced this pull request Jul 13, 2026
@lllyfff @claude
Merge upstream/main (bytedance/deer-flow) into MemoryManager
65703f4 
Brings in 4fd521e..2bd0f56: sandbox fixes (bytedance#4051/bytedance#4053/bytedance#4058/bytedance#4078/bytedance#4079/bytedance#4110),
CVE-2026-33128/bytedance#4089/bytedance#4093/bytedance#4106, subagents (bytedance#4056/bytedance#4059/bytedance#4088), skillscan (bytedance#4087),
channels (bytedance#4055/bytedance#4069/bytedance#4100), stream_bridge (bytedance#4071), frontend (bytedance#4058/bytedance#4068/bytedance#4094/
bytedance#4114/bytedance#4126), event-store serialization (bytedance#4077), streaming (bytedance#4082/bytedance#4085), mcp (bytedance#4084),
guardrails (bytedance#4067), and the memory fixes bytedance#4073/bytedance#4074/bytedance#4075/bytedance#4076/bytedance#4119.

Memory conflict resolution (upstream is still monolithic agents/memory/, we are
vendored backends/deermem/deermem/core/ - git rename-detected host->vendored):
- core/prompt.py, core/updater.py: keep ours (vendored with all ports); upstream's
  host-side bytedance#4119/bytedance#4074/bytedance#4075/bytedance#4076 changes are the same fixes already ported.
- core/queue.py: auto-merged cleanly (both sides had identical bytedance#4073 change).
- test_memory_queue.py: replaced the stale pre-bytedance#4073 test with upstream's three
  bytedance#4073 tests, adapted from host API (MemoryUpdateQueue()) to vendored (_queue()).
- test_memory_search.py: kept our vendored tests + added a vendored
  test_null_confidence_does_not_crash_sort (bytedance#4076 coverage).

Post-merge: 171 memory tests pass (+2 module-skipped), 7164 collected, no
import/collection errors. All 8 upstream memory fixes since the 8be7411
divergence are adapted (bytedance#4044/bytedance#4028/bytedance#4097/bytedance#4119/bytedance#4073/bytedance#4074/bytedance#4075/bytedance#4076).

Co-Authored-By: Claude <noreply@anthropic.com>
@DaoyuanLi2816 DaoyuanLi2816 mentioned this pull request Jul 15, 2026
Merged
8 tasks
marvin9551 pushed a commit to marvin9551/deer-flow that referenced this pull request Aug 21, 2026
@nankingjing @claude
fix(streaming): drop silent delta-discard in _merge_stream_text (byte…
51881bf 
…dance#4085)

The dual-mode _merge_stream_text used short-circuits chunk==existing
and existing.endswith(chunk) that silently dropped legitimate delta
chunks in messages-tuple mode:

- CJK reduplication: '谢谢' tokenized as ['谢','谢'] -> buffer stays '谢'
- Repeated tokens: 'gogo' as ['go','go'] -> buffer stays 'go'
- Suffix-matching tails: 'hello' as ['hel','l','o'] -> drops 'l'

Delta payloads must always append; only strictly-longer cumulative
snapshots should replace. The fix gates the cumulative-replace branch
with len(chunk) > len(existing) and removes the dropped-shape guards.

Both copies are fixed:
- app/channels/manager.py (IM streaming to Feishu/Telegram/etc.)
- deerflow/tui/view_state.py (TUI client rendering)

The TUI reduce() caller now handles exact re-sends (values snapshot
re-emitting history) before the merge, and the len>1 guard prevents
single-char CJK deltas from being mistaken for re-sends.

Co-authored-by: Claude <noreply@anthropic.com>
jihtsan pushed a commit to jihtsan/dnx-deer-flow that referenced this pull request Aug 29, 2026
@nankingjing @claude
fix(streaming): drop silent delta-discard in _merge_stream_text (byte…
1e3ba42 
…dance#4085)

The dual-mode _merge_stream_text used short-circuits chunk==existing
and existing.endswith(chunk) that silently dropped legitimate delta
chunks in messages-tuple mode:

- CJK reduplication: '谢谢' tokenized as ['谢','谢'] -> buffer stays '谢'
- Repeated tokens: 'gogo' as ['go','go'] -> buffer stays 'go'
- Suffix-matching tails: 'hello' as ['hel','l','o'] -> drops 'l'

Delta payloads must always append; only strictly-longer cumulative
snapshots should replace. The fix gates the cumulative-replace branch
with len(chunk) > len(existing) and removes the dropped-shape guards.

Both copies are fixed:
- app/channels/manager.py (IM streaming to Feishu/Telegram/etc.)
- deerflow/tui/view_state.py (TUI client rendering)

The TUI reduce() caller now handles exact re-sends (values snapshot
re-emitting history) before the merge, and the len>1 guard prevents
single-char CJK deltas from being mistaken for re-sends.

Co-authored-by: Claude <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@WillemJiang WillemJiang WillemJiang approved these changes
+1 more reviewer
@willem-bd willem-bd willem-bd left review comments
Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

area:backend Gateway / runtime / core backend under backend/ risk:medium Medium risk: regular code changes size/M PR changes 100-300 lines

Projects

None yet

Milestone

2.1.0

Development

Successfully merging this pull request may close these issues.

3 participants

@nankingjing @WillemJiang @willem-bd