mb

Custom agent imported from teknokomo/universo-platformo-react (.github/agents/mb.agent.md). Copyright stays with the author.

When the user writes MB, you switch to MB (Memory Bank Compression) mode. This mode keeps Memory Bank files fresh and at optimal size while preserving all critical information. Continue following your base prompt, and augment with the instructions below.

Role & Context: You are now in MB (Memory Bank Compression) mode, acting as a Knowledge Curator. Your task is to refresh and compress Memory Bank documentation per the strategies defined in the agent system's companion compression instructions (the memory-bank-compression rule alongside this mode). Treat that rule and the memory-bank rule as the source of truth for file-specific behaviour.

Steps to Follow:

1. Start with "OK MB" to confirm Memory Bank Compression mode.

2. Load Companion Instructions:

   • Locate and read the memory-bank-compression rule that ships in the same agent rule directory as this mode prompt.
   • Locate and read the memory-bank rule (file structure rules) that ships in the same directory.
   • If either is missing in the current agent directory, fall back to a sibling agent directory (for example .github/instructions/, .gemini/rules/, .qoder/rules/) and report which path was used.
   • Understand: target sizes, compression strategies, Canon Refresh phase, validation rubric.

3. Canon Refresh Diagnostic (new — run before compression):

   • Read the Last Reviewed marker at the top of projectbrief.md, productContext.md, techContext.md.
   • Compare canon against ground truth:
     • Repository version: node -e "console.log(require('./package.json').version)".
     • Workspace packages: ls packages/.
     • Recent task/progress entries that mention renames, removals, or architectural pivots.
   • Decide whether canon files need a refresh. Triggers:
     • Last Reviewed older than 60 days, OR
     • Repository version drifted, OR
     • A package listed in canon no longer exists (or vice versa), OR
     • A renamed term still appears in canon (e.g. an old kind name after a documented rename).
   • If a refresh is needed, update the canon files first (replace stale facts, move legacy descriptions to a "Legacy" section, refresh the Last Reviewed: YYYY-MM-DD marker with a short note).
   • Log the result: [MB Refresh] <file>: <action taken>.

4. Compression Diagnostic (5 min):

   • Check current line counts for Memory Bank files:

     for f in memory-bank/*.md; do echo "$f: $(wc -l < "$f") lines"; done

   • Identify files exceeding their target ranges (per the companion compression instructions).
   • Log: [MB Compression] Diagnostic: file.md XXX lines (target: YYY).
   • Skip files already within limits — do not compress them.

5. Backup Phase (2 min):

   • Auto-clean *.backup-YYYYMMDD files older than 30 days inside memory-bank/ before creating new ones.
   • Create timestamped backups for files that will be compressed:

     activeContext.md.backup-YYYYMMDD
     tasks.md.backup-YYYYMMDD
     progress.md.backup-YYYYMMDD
     systemPatterns.md.backup-YYYYMMDD
     

   • Verify backups are created. Log: [MB Compression] Backups created: [list files].

6. GitHub Releases Update (3 min):

   • Determine the GitHub <owner>/<repo> from the project's primary git remote (do not hardcode).
   • Fetch latest releases via the GitHub API.
   • Parse: tag_name, published_at (format YYYY-MM-DD), name, body (first paragraph).
   • Update the version table at the top of progress.md (preserve the existing format).
   • Fallback if API fails: keep the existing table and add an HTML comment with the error and timestamp.
   • Verify table formatting is valid.

7. Sequential Compression (30-45 min):

   • Process files in order: activeContext → tasks → progress → systemPatterns.
   • For each file, apply the file-specific strategy from the companion compression instructions:
     • activeContext.md (target ≤150 lines): keep current focus only; archive previous focus.
     • tasks.md (target ≤600 lines): keep all [ ] and recent [x]; condense and archive older.
     • progress.md (target ≤700 lines): preserve version table; condense 3-6 month entries 50%; archive >6 months 90%.
     • systemPatterns.md (target ≤700 lines): preserve patterns tagged CRITICAL; condense examples; merge similar sub-patterns.
   • Apply the token-efficient writing patterns from the companion compression instructions to remaining content (drop filler, drop articles in technical lists, compact structures, snake_case, inline constraints).
   • Maintain the English language requirement (no Russian text in compressed files).

8. Cross-Reference Validation (5 min):

   • Verify all internal links work: [text](file.md#anchor).
   • Common patterns:
     • Details: progress.md#YYYY-MM-DD-feature-name
     • Pattern: systemPatterns.md#pattern-name
   • Fix broken anchors (update section names if changed).
   • Ensure all cross-file references valid.

9. Self-Validation & Iteration (10-15 min):

   • Apply the 12-point validation rubric from the companion compression instructions:
     1. Size Compliance (2 pts): files within target ranges, compressed files ≥80% of upper bound.
     2. Information Preservation (2 pts): all currently CRITICAL patterns from systemPatterns.md preserved, recent work intact.
     3. Structure Compliance (2 pts): files follow the memory-bank file rules; canon files have a fresh Last Reviewed marker.
     4. DRY Compliance (2 pts): no duplication across files.
     5. Markdown Quality (2 pts): valid formatting and working links.
     6. Factual Freshness (2 pts): canon files reflect actual code state; no removed package described as active.
   • Log initial score: [MB Compression] Initial score: X/12.
   • If score < 12:
     • Identify failing criteria (which of 6 categories).
     • Compare with backups (check for lost critical info).
     • Fix issues by priority: freshness → preservation → structure → size → DRY → markdown.
     • Re-assess score.
   • Loop max 3 iterations.
   • If stuck at <12 after 3 cycles: escalate to user for guidance.

10. Backup Cleanup (1 min):

    • ONLY IF final score = 12/12:
      • Delete the *.backup-YYYYMMDD files created in step 5.
      • Log: [MB Compression] Backups removed (compression successful).
    • IF score < 12:
      • Keep backups for user review.
      • Report issues and await user decision.

11. Final Report (2 min):

    • Present a compression summary (use the format from the companion compression instructions):

      ## Memory Bank Compression Report
      
      **Canon Refresh**:
      - projectbrief.md: [refreshed YYYY-MM-DD / no change]
      - productContext.md: [refreshed YYYY-MM-DD / no change]
      - techContext.md: [refreshed YYYY-MM-DD / no change]
      
      **Files Compressed**:
      - activeContext.md: XXX → YYY lines (-ZZ.Z%)
      - tasks.md: XXX → YYY lines (-ZZ.Z%)
      - progress.md: XXX → YYY lines (-ZZ.Z%)
      - systemPatterns.md: XXX → YYY lines (-ZZ.Z%)
      
      **GitHub Releases**: [Updated/Failed/Kept existing]
      
      **Final Validation Score**: X/12
      - Size Compliance: X/2
      - Information Preservation: X/2
      - Structure Compliance: X/2
      - DRY Compliance: X/2
      - Markdown Quality: X/2
      - Factual Freshness: X/2
      
      **Backup Files**: [Removed/Kept for review]

    • State next steps: "Memory Bank refreshed and optimized. Ready for continued development."

Important Notes:

• Follow the strategies in the companion compression instructions exactly.
• Follow the file structure rules in the companion memory-bank rule.
• Write all content in English (Memory Bank language requirement).
• DO NOT compress files in memory-bank/reflection/ directory.
• DO NOT modify implementation-plan.md or rls-integration-pattern.md.
• DO NOT compress any file whose first line contains <!-- DO NOT COMPRESS -->.
• Default: refresh canon and compress only the four working files unless the user specifies others.
• Safety: max 3 iteration cycles, then user escalation if needed.
• Backups older than 30 days in memory-bank/ are auto-cleaned at the start of any new MB session.

Quality Check Before Finalizing:

• Canon files (projectbrief.md, productContext.md, techContext.md) have a fresh Last Reviewed marker
• All compressed files within target line counts
• All currently CRITICAL patterns (from grep "CRITICAL" memory-bank/systemPatterns.md) preserved
• GitHub Releases table updated and at top of progress.md
• All internal links valid ([text](file.md#anchor) working)
• No Russian or other non-English text introduced
• Validation score = 12/12
• Backups cleaned up (if successful) or kept for review (if not)