Summary

• Problem: dreaming-generated narrative/report artifacts could loop back into short-term promotion, while prompt-text-only transcript detection could also suppress ordinary chats that merely quoted the dream prompt.
• Why it matters: synthetic dreaming output should not become durable memory, and real user transcripts should not be dropped from session recall ingestion.
• What changed: session transcript skipping now relies on internal dreaming run markers with prefix-only matching instead of prompt text; short-term promotion rejects dreaming/promotion-shaped snippets during normalization, recording, ranking, and apply; the dreaming lead detector was hardened with a linear parser that handles markdown/diff wrappers without the backtracking-prone regex shape.
• What did NOT change (scope boundary): this PR does not add end-to-end provenance plumbing or a new persisted schema for first-class origin tracking.

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 #
• Related #
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: dreaming-generated text and grounded recall snippets converged into downstream plain-text surfaces, so promotion-path filtering had to infer provenance from content patterns.
• Missing detection / guardrail: transcript ingestion also treated quoted prompt text as an authoritative dreaming marker, which could falsely classify ordinary chats as internal dreaming runs.
• Contributing context (if known): dreaming reports are emitted as markdown bullets/diff-like snippets, so narrow start-of-string heuristics were insufficient to catch all contamination shapes.

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:
  • src/memory-host-sdk/host/session-files.test.ts
  • extensions/memory-core/src/short-term-promotion.test.ts
  • extensions/memory-core/src/dreaming-phases.test.ts
• Scenario the test should lock in: real chats that quote the dream prompt still ingest normally, actual dreaming transcripts checkpoint as skipped, and dreaming markdown/diff candidate snippets never enter short-term promotion.
• Why this is the smallest reliable guardrail: the bug spans one transcript-ingestion seam plus the short-term promotion insertion/ranking/apply surfaces.
• Existing test that already covers this (if any): N/A
• If no new test is added, why not: N/A

User-visible / Behavior Changes

• Dreaming diary/report artifacts no longer feed durable memory promotion.
• Ordinary transcripts that quote the dream-diary prompt continue to participate in session recall ingestion.

Diagram (if applicable)

Before:
[dreaming transcript or dreaming-shaped snippet] -> [session recall / short-term store] -> [promotion candidate] -> [durable memory pollution]
[ordinary chat quoting prompt] -> [misclassified as dreaming] -> [session ingestion skipped]

After:
[dreaming transcript] -> [checkpoint as skipped] -> [no session recall ingestion]
[dreaming-shaped snippet] -> [filtered from short-term promotion paths] -> [no durable promotion]
[ordinary chat quoting prompt] -> [normal session ingestion] -> [eligible grounded recall]

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)
• If any Yes, explain risk + mitigation:

Repro + Verification

Environment

• OS: macOS
• Runtime/container: local repo workspace
• Model/provider: N/A
• Integration/channel (if any): N/A
• Relevant config (redacted): default memory-core dreaming/session ingestion paths

Steps

1. Create a dreaming transcript with internal dreaming bootstrap/run markers, or create a short-term snippet shaped like a dreaming candidate/report.
2. Run the memory-core transcript ingestion / short-term promotion path.
3. Create a normal transcript that merely quotes the dream-diary prompt and run the same ingestion path.

Expected

• Dreaming-generated transcripts and dreaming-shaped snippets are excluded from promotion inputs.
• Ordinary prompt-quoting transcripts are still ingested.

Actual

• Matches expected on the current PR head.

Evidence

Attach at least one:

• 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:
  • targeted session-files regression for prompt-quoting chats vs dreaming-marker transcripts
  • targeted short-term promotion filtering for direct, stored, markdown-bullet, and diff-prefixed dreaming snippets
  • targeted dreaming-phases checkpoint expectation for skipped dreaming transcripts
• Edge cases checked:
  • prefix-only runId/sessionKey matching
  • markdown bullet wrappers
  • diff-prefix wrappers
  • bracket wrappers
• What you did not verify:
  • full end-to-end provenance redesign beyond the current heuristic/path-level fix
  • full repo test suite

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)
• If yes, exact upgrade steps:

Risks and Mitigations

• Risk: content-based contamination heuristics remain a compatibility backstop rather than first-class provenance.
  • Mitigation: transcript-side false positives were narrowed to internal markers, promotion-path filters now cover the known dreaming formats, and a provenance-first follow-up can replace the remaining heuristics later.