Summary

• Problem: rehydratePromotionCandidate re-reads the raw daily-note source file and feeds the unredacted lines into relocateCandidateRange, whose fuzzy window search will latch onto a window that straddles a human bullet and the adjacent managed <!-- openclaw:dreaming:light:* --> / <!-- openclaw:dreaming:rem:* --> block, leaking ## Light Sleep headings and - Candidate: … / confidence: … / status: staged scaffolding into MEMORY.md.
• Why it matters: the chunk builder already strips managed blocks before snippet capture (stripManagedDailyDreamingLines in dreaming-phases.ts), so the post-rehydration isContaminatedDreamingSnippet check never sees a bare Candidate: lead and lets straddle snippets through; users see staged dreaming scaffolding promoted into long-term memory.
• What changed: added redactManagedDreamingLines(lines) in extensions/memory-core/src/short-term-promotion.ts and call it on the raw source split inside rehydratePromotionCandidate before relocation. The helper mirrors dreaming-phases.ts's contract — zeroes start..end markers AND walks upward to also zero the ## Light Sleep / ## REM Sleep heading directly above the start marker — and unterminated managed blocks redact through end of file.
• What did NOT change (scope boundary): no changes to the chunk builder, the contamination predicate, the relocate algorithm, the apply path, the MEMORY.md writer, or the CJK dedupe path (the second sub-bug from [Bug]: dreaming pipeline leaks raw candidate content into MEMORY.md and CJK dedup is ineffective in tokenizeSnippet #80613 is being addressed separately by fix(memory-core): use CJK-aware tokenizer for dreaming dedupe (#80613) #80620).

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 [Bug]: dreaming pipeline leaks raw candidate content into MEMORY.md and CJK dedup is ineffective in tokenizeSnippet #80613 (promotion-leak half)
• Related fix(memory-core): use CJK-aware tokenizer for dreaming dedupe (#80613) #80620 (CJK dedupe half — separate PR, separate diff)
• This PR fixes a bug or regression

Real behavior proof (required for external PRs)

• Behavior or issue addressed: managed <!-- openclaw:dreaming:light:* --> / <!-- openclaw:dreaming:rem:* --> blocks no longer leak into MEMORY.md when promotion rehydration straddles a human bullet next to a managed block.
• Real environment tested: local OpenClaw source checkout on Linux 6.17 with Node 22 / pnpm 11; daily-note fixtures written via the same recordShortTermRecalls → rankShortTermPromotionCandidates → applyShortTermPromotions pipeline used by memory_search recall capture, with workspace state on disk and a real MEMORY.md round-tripped through fs.readFile.
• Exact steps or command run after this patch: pnpm exec vitest run extensions/memory-core/src/short-term-promotion.test.ts extensions/memory-core/src/dreaming-phases.test.ts
• Evidence after fix (screenshot, recording, terminal capture, console output, redacted runtime log, linked artifact, or copied live output): terminal output reported Test Files 2 passed (2) | Tests 88 passed (88) | Duration 19.76s (49 short-term-promotion + 39 dreaming-phases). The new fuzzy-window-straddle integration regression (commit c97b3633) was verified to FAIL against current main without the sanitizer — the produced MEMORY.md was captured verbatim as <!-- openclaw-memory-promotion:memory:memory/2026-04-13.md:20:30 --> - - Plan switches use exRule, not abConfig ## Light Sleep <!-- openclaw:dreaming:light:start --> - Candidate: Plan toggle field summary - confidence: 0.00 - evidence: memory/.dreams/session-corpus/2026-04-13.txt:1-3 - recalls: 0 - status: staged <!-- openclaw:dreaming:light:end --> [score=0.611 recalls=1 avg=0.940 source=memory/2026-04-13.md:2-12] (matches the exact leak shape on issue [Bug]: dreaming pipeline leaks raw candidate content into MEMORY.md and CJK dedup is ineffective in tokenizeSnippet #80613). With the sanitizer in place, MEMORY.md keeps only - Plan switches use exRule, not abConfig.
• Observed result after fix: promoted MEMORY.md entries derived from daily notes that interleave human content with managed dreaming blocks contain only the human content; no ## Light Sleep heading, no <!-- openclaw:dreaming:* --> markers, no - Candidate: / confidence: / evidence: / recalls: / status: staged scaffolding ever lands.
• What was not tested: a live multi-day dreaming-promotion run against a real Anthropic/OpenAI provider — the regression is exercised end-to-end through the in-process pipeline against on-disk daily notes and real MEMORY.md writes, not against a streaming model session. Operator confirmation against a real long-running workspace is welcome.
• Before evidence (optional but encouraged): the leak was already documented on [Bug]: dreaming pipeline leaks raw candidate content into MEMORY.md and CJK dedup is ineffective in tokenizeSnippet #80613 with the reporter's MEMORY.md showing - ## 教训：Plan 实验开关字段 - Plan 配置中实验开关字段是 \... ## Light Sleep <!-- openclaw:dreaming:light:start --> - Candidate: 教训… (human bullet + heading + start marker + dreaming Candidate: line all promoted); ClawSweeper's earlier review on the issue confirmed the source-level path through promotion rehydration.

Root Cause (if applicable)

• Root cause: rehydratePromotionCandidate (extensions/memory-core/src/short-term-promotion.ts) reads the raw daily-note file and passes the unredacted lines into relocateCandidateRange, which then fuzzy-matches against any window of 1..maxSpan lines. When the candidate's stored startLine/endLine does not match exactly (stale capture, file drift, partial-line snippet), the fuzzy search prefers higher-quality windows and can pick a multi-line window that includes the managed block. The relocated snippet starts with a clean human bullet, so the post-rehydration isContaminatedDreamingSnippet check (which keys on a leading Candidate: after diff-prefix stripping) does not flag it.
• Missing detection / guardrail: no parallel sanitizer for the rehydration boundary — only the chunk builder's stripManagedDailyDreamingLines was protecting daily ingestion.
• Contributing context (if known): the leak only surfaces when a daily note interleaves a human bullet directly next to a managed block, which is the dominant shape produced by the dreaming pipeline itself, so any user with the dreaming feature on hits this.

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: extensions/memory-core/src/short-term-promotion.test.ts — three new regressions: two unit tests for redactManagedDreamingLines (terminated light + REM blocks; unterminated light block through EOF), one integration test for the full record→rank→apply chain (exact-match path), and one integration test for the fuzzy-window straddle path (commit c97b3633, addresses ClawSweeper P3).
• Scenario the test should lock in: a daily note where a human bullet sits directly above ## Light Sleep + <!-- openclaw:dreaming:light:start --> + staged dreaming scaffolding + <!-- openclaw:dreaming:light:end -->, captured into a recall whose startLine/endLine no longer point at the human bullet, so rehydration must traverse the fuzzy window search and must NOT relocate to a window that includes the managed block.
• Why this is the smallest reliable guardrail: it exercises the exact rehydration boundary touched by this fix, against real on-disk fixtures and the real apply path, and is verified to fail without the sanitizer.
• Existing test that already covers this (if any): None — dreaming-phases.test.ts only covers the chunk-builder strip; the rehydration path was untested for managed-block leaks.
• If no new test is added, why not: N/A (4 new regressions added).

User-visible / Behavior Changes

MEMORY.md will no longer accumulate ## Light Sleep / ## REM Sleep headings, <!-- openclaw:dreaming:* --> markers, or - Candidate: / confidence: / status: staged scaffolding from short-term promotion. Existing MEMORY.md files are not rewritten — only new promoted entries are sanitized.

Diagram (if applicable)

Before:
daily note (human bullet | managed block) -> rehydratePromotionCandidate(raw lines)
  -> relocateCandidateRange fuzzy match -> straddle window -> MEMORY.md leak

After:
daily note (human bullet | managed block) -> rehydratePromotionCandidate(redactManagedDreamingLines(raw lines))
  -> relocateCandidateRange fuzzy match over sanitized lines -> human-only window -> clean MEMORY.md

Security Impact (required)

• New permissions/capabilities? (Yes/No) No
• Secrets/tokens handling changed? (Yes/No) No
• New/changed network calls? (Yes/No) No
• Command/tool execution surface changed? (Yes/No) No
• Data access scope changed? (Yes/No) No
• If any Yes, explain risk + mitigation: N/A

Repro + Verification

Environment

• OS: Linux 6.17 (host)
• Runtime/container: local OpenClaw source checkout, Node 22, pnpm 11
• Model/provider: N/A (in-process pipeline, no model calls)
• Integration/channel (if any): N/A (memory-core only)
• Relevant config (redacted): default short-term promotion thresholds (minScore: 0, minRecallCount: 0, minUniqueQueries: 0 for the regression fixture)

Steps

1. Write a daily memory note with the shape ## Plan toggle field / human bullet / blank / ## Light Sleep / <!-- openclaw:dreaming:light:start --> / staged scaffolding / <!-- openclaw:dreaming:light:end -->.
2. Record a short-term recall whose stored snippet is the multi-line straddle (human bullet + heading + dreaming bullets) at a stale startLine/endLine (e.g. 20/30) so relocate must use the fuzzy window search.
3. Run rankShortTermPromotionCandidates then applyShortTermPromotions, then read back MEMORY.md.

Expected

• MEMORY.md contains only the human bullet (- Plan switches use exRule, not abConfig); no ## Light Sleep, no <!-- openclaw:dreaming:* -->, no - Candidate: / confidence: / status: staged.

Actual

• Matches expected on this branch (verified by pnpm exec vitest run extensions/memory-core/src/short-term-promotion.test.ts — all 49 tests pass, including the 4 new regressions).

Evidence

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

Human Verification (required)

What you personally verified (not just CI), and how:

• Verified scenarios: ran pnpm exec vitest run extensions/memory-core/src/short-term-promotion.test.ts extensions/memory-core/src/dreaming-phases.test.ts locally (Test Files 2 passed (2) | Tests 88 passed (88) | Duration 19.76s); reverted the redactManagedDreamingLines call inside rehydratePromotionCandidate and re-ran the new fuzzy-window straddle regression — observed the leak in MEMORY.md matching [Bug]: dreaming pipeline leaks raw candidate content into MEMORY.md and CJK dedup is ineffective in tokenizeSnippet #80613's reporter shape; restored the sanitizer and re-ran — the leak is gone.
• Edge cases checked: terminated light block; terminated REM block; unterminated light block (redacts through EOF); heading-walk redacts the ## Light Sleep / ## REM Sleep heading directly above the start marker; diary marker (no heading) only redacts start..end markers; exact-match relocation path (existing test fix: add @lid format support and allowFrom wildcard handling #1 at short-term-promotion.test.ts:1249); fuzzy-match straddle path (new test at short-term-promotion.test.ts:1304).
• What you did not verify: a live multi-day dreaming-promotion run against a real provider session; a live long-running workspace where dreaming has been running for weeks. The change is additive and structural (an extra sanitization pass before relocation), with no behavior change for daily notes that contain no managed blocks.

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.

ClawSweeper's earlier P3 (exercise the straddling rehydration path) is addressed by commit c97b3633 and acknowledged in a follow-up comment with the verbatim leak captured under no-fix conditions.

Compatibility / Migration

• Backward compatible? (Yes/No) Yes
• Config/env changes? (Yes/No) No
• Migration needed? (Yes/No) No
• If yes, exact upgrade steps: N/A

Risks and Mitigations

• Risk: the heading-walk could redact a ## Light Sleep / ## REM Sleep heading that a user authored manually for their own notes (i.e. one that does NOT precede a managed start marker).
  • Mitigation: the heading walk only fires when an exact ## Light Sleep or ## REM Sleep heading sits directly above the matching managed start marker (mirrors findManagedDailyDreamingHeadingIndex in dreaming-phases.ts); a stand-alone user heading without the immediately-following start marker is left untouched.