PR #66568: fix(streaming): prevent block chunker from splitting markdown tables

• Repository: openclaw/openclaw
• Author: zongqichen
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/66568

Description (problem / solution / changelog)

Summary

• Problem: Block streaming (flushOnParagraph) treats \n\n around markdown tables as valid split points, splitting tables across multiple Telegram messages.
• Why it matters: Tables are unreadable when fragmented across messages.
• What changed: Added parseTableSpans() / isSafeTableBreak() mirroring the existing FenceSpan pattern, wired into all break-finding paths in EmbeddedBlockChunker. Moved table span detection to standalone table-spans.ts to avoid import cycle.
• What did NOT change: No changes to deliver.ts, config resolution, or non-table chunking behavior.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Closes #66614
• Related #47454
• Related #17679
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: findNextParagraphBreak() and all findSafe*BreakIndex() functions only checked isSafeFenceBreak(). Tables have no equivalent protection.
• Missing detection / guardrail: No table span detection existed.
• Contributing context: The fence-protection pattern was correct and extensible — tables just were not covered.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file: pi-embedded-block-chunker.test.ts, markdown/tables.test.ts
• Scenario the test should lock in: text+table+text under maxChars → single chunk; table stays intact when flushOnParagraph splits around it; thematic breaks not falsely detected as tables
• Why this is the smallest reliable guardrail: Unit tests on the chunker directly verify the break-finding logic without needing a full channel roundtrip.
• Existing test that already covers this: None — new tests added in this PR.
• If no new test is added, why not: N/A — new tests added.

User-visible / Behavior Changes

Markdown tables in streaming replies are no longer split into separate messages.

Diagram (if applicable)

Before:
[text] → msg1 | [table] → msg2 | [text] → msg3

After:
[text + table + text] → msg1

Security Impact (required)

• New permissions/capabilities? No
• Secrets/tokens handling changed? No
• New/changed network calls? No
• Command/tool execution surface changed? No
• Data access scope changed? No

Repro + Verification

Environment

• OS: macOS 26.4.1 (arm64)
• Runtime/container: Node 22+ / pnpm global
• Model/provider: anthropic/claude-sonnet-4
• Integration/channel: Telegram (block streaming + flushOnParagraph)
• Relevant config: default streaming config

Steps

1. Send a prompt that triggers a response containing text, a markdown table, then more text
2. Observe the Telegram delivery

Expected

• One message containing the full response with the table inline

Actual

• Three separate messages: text before table, the table itself, text after table

Evidence

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

Unit tests: 15+ new tests all passing. Live verified: applied patched build to local OpenClaw, sent table reply via Telegram, confirmed single message delivery.

Human Verification (required)

• Verified scenarios: Unit tests + live Telegram end-to-end
• Edge cases checked: Table inside code fence (ignored), thematic break --- (not falsely detected), table near maxChars
• What you did not verify: Other channels (Slack, WhatsApp) — behavior should be identical since they share the same chunker code path.

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

• Backward compatible? Yes
• Config/env changes? No
• Migration needed? No

Risks and Mitigations

• Risk: Very large tables exceeding maxChars will force-split at maxChars instead of \n\n.
  • Mitigation: Matches existing code fence behavior — correct fallback.

Changed files

• src/agents/pi-embedded-block-chunker.test.ts (modified, +85/-4)
• src/agents/pi-embedded-block-chunker.ts (modified, +30/-13)
• src/auto-reply/chunk.ts (modified, +3/-1)
• src/markdown/table-spans.ts (added, +116/-0)
• src/markdown/tables.test.ts (modified, +47/-0)