phoenixclaw

PhoenixClaw: Zero-Tag Passive Journaling

PhoenixClaw automatically distills daily conversations into meaningful reflections using semantic intelligence.

Automatically identifies journal-worthy moments, patterns, and growth opportunities.

🛠️ Core Workflow

[!critical] MANDATORY: Complete Workflow Execution This 9-step workflow MUST be executed in full regardless of invocation method:

• Cron execution (10 PM nightly)
• Manual invocation ("Show me my journal", "Generate today's journal", etc.)
• Regeneration requests ("Regenerate my journal", "Update today's entry")

Never skip steps. Partial execution causes:

• Missing images (session logs not scanned)
• Missing finance data (Ledger plugin not triggered)
• Incomplete journals (plugins not executed)

PhoenixClaw follows a structured pipeline to ensure consistency and depth:

1. User Configuration: Check for ~/.phoenixclaw/config.yaml. If missing, initiate the onboarding flow defined in references/user-config.md.
2. Context Retrieval:
   • Scan memory files (NEW): Read memory/YYYY-MM-DD.md and memory/YYYY-MM-DD-*.md files for manually recorded daily reflections. These files contain personal thoughts, emotions, and context that users explicitly ask the AI to remember via commands like "记一下" (remember this). CRITICAL: Do not skip these files - they contain explicit user reflections that session logs may miss.
   • Scan session logs: Call memory_get for the current day's memory, then CRITICAL: Scan ALL raw session logs and filter by message timestamp. Session files are often split across multiple files. Do NOT classify images by session file mtime:

     # Read all session logs from ALL known OpenClaw locations, then filter by per-message timestamp
     # Use timezone-aware epoch range to avoid UTC/local-day mismatches.
     TARGET_DAY="$(date +%Y-%m-%d)"
     TARGET_TZ="${TARGET_TZ:-Asia/Shanghai}"
     read START_EPOCH END_EPOCH < <(
       python3 - <<'PY' "$TARGET_DAY" "$TARGET_TZ"
     

from datetime import datetime, timedelta from zoneinfo import ZoneInfo import sys

day, tz = sys.argv[1], sys.argv[2] start = datetime.strptime(day, "%Y-%m-%d").replace(tzinfo=ZoneInfo(tz)) end = start + timedelta(days=1) print(int(start.timestamp()), int(end.timestamp())) PY )

  # Recursively scan all session directories (multi-agent architecture support)
  for dir in "$HOME/.openclaw/sessions" \
             "$HOME/.openclaw/agents" \
             "$HOME/.openclaw/cron/runs" \
             "$HOME/.agent/sessions"; do
    [ -d "$dir" ] || continue
    find "$dir" -type f -name "*.jsonl" -print0
  done |
    xargs -0 jq -cr --argjson start "$START_EPOCH" --argjson end "$END_EPOCH" '
      (.timestamp // .created_at // empty) as $ts
      | ($ts | split(".")[0] + "Z" | fromdateiso8601?) as $epoch
      | select($epoch != null and $epoch >= $start and $epoch < $end)
    '
  ```
  Read **all matching files** regardless of their numeric naming (e.g., file_22, file_23 may be earlier in name but still contain today's messages).
- **EXTRACT IMAGES FROM SESSION LOGS**: Session logs contain `type: "image"` entries with file paths. You MUST:
  1. Find all image entries (e.g., `"type":"image"`)
  2. Keep only entries where message `timestamp` is in the target date range
  3. Extract the `file_path` or `url` fields
  4. Copy files into `assets/YYYY-MM-DD/`
  5. Rename with descriptive names when possible
- **Why session logs are mandatory**: `memory_get` returns **text only**. Image metadata, photo references, and media attachments are **only available in session logs**. Skipping session logs = missing all photos.
- **Activity signal quality**: Do not treat heartbeat/cron system noise as user activity. Extract user/assistant conversational content and media events first, then classify moments.
- **FILTER HEARTBEAT MESSAGES (CRITICAL)**: Session logs contain system heartbeat messages that MUST be excluded from journaling. When scanning messages, SKIP any message matching these criteria:
  1. **User heartbeat prompts**: Messages containing "Read HEARTBEAT.md" AND "reply HEARTBEAT_OK"
  2. **Assistant heartbeat responses**: Messages containing ONLY "HEARTBEAT_OK" (with optional leading/trailing whitespace)
  3. **Cron system messages**: Messages with role "system" or "cron" containing job execution summaries (e.g., "Cron job completed", "A cron job")
  
  Example jq filter to exclude heartbeats:
  ```jq
  # Exclude heartbeat messages
  | select(
      (.message.content? | type == "array" and 
        (.message.content | map(.text?) | join("") | 
          test("Read HEARTBEAT\.md"; "i") | not))
      and
      (.message.content? | type == "array" and 
        (.message.content | map(.text?) | join("") | 
          test("^\\s*HEARTBEAT_OK\\s*$"; "i") | not))
    )
  ```
- **Edge case - Midnight boundary**: For late-night activity that spans midnight, expand the **timestamp** range to include spillover windows (for example, previous day 23:00-24:00) and still filter per-message by `timestamp`.

• Merge sources: Combine content from both memory files and session logs. Memory files capture explicit user reflections; session logs capture conversational flow and media. Use both to build complete context.
• Fallback: If memory is sparse, reconstruct context from session logs, then update memory so future runs use the enriched memory. Incorporate historical context via memory_search (skip if embeddings unavailable)

3. Moment Identification: Identify "journal-worthy" content: critical decisions, emotional shifts, milestones, or shared media. See references/media-handling.md for photo processing. This step generates the moments data structure that plugins depend on. Image Processing (CRITICAL):

   • For each extracted image, generate descriptive alt-text via Vision Analysis
   • Categorize images (food, selfie, screenshot, document, etc.)

   Filter Finance Screenshots (NEW): Payment screenshots (WeChat Pay, Alipay, etc.) should NOT be included in the journal narrative. These are tool images, not life moments.

   Detection criteria (check any):

   1. OCR keywords: "支付成功", "支付完成", "微信支付", "支付宝", "订单号", "交易单号", "¥" + amount
   2. Context clues: Image sent with nearby text containing "记账", "支付", "付款", "转账"
   3. Visual patterns: Standard payment app UI layouts (green WeChat, blue Alipay)

   Handling rules:

   • Mark as finance_screenshot type

   • Route to Ledger plugin (if enabled) for transaction recording

   • EXCLUDE from journal main narrative unless explicitly described as part of a life moment (e.g., "今天请朋友吃饭" with payment screenshot)

   • Never include raw payment screenshots in daily journal images section

   • Match images to moments (e.g., breakfast photo → breakfast moment)

   • Store image metadata with moments for journal embedding

4. Pattern Recognition: Detect recurring themes, mood fluctuations, and energy levels. Map these to growth opportunities using references/skill-recommendations.md.

5. Plugin Execution: Execute all registered plugins at their declared hook points. See references/plugin-protocol.md for the complete plugin lifecycle:

   • pre-analysis → before conversation analysis
   • post-moment-analysis → Ledger and other primary plugins execute here
   • post-pattern-analysis → after patterns detected
   • journal-generation → plugins inject custom sections
   • post-journal → after journal complete

6. Journal Generation: Synthesize the day's events into a beautiful Markdown file using assets/daily-template.md. Follow the visual guidelines in references/visual-design.md. Include all plugin-generated sections at their declared section_order positions.

   • Embed curated images only, not every image. Prioritize highlights and moments.
   • Route finance screenshots to Ledger sections (receipts, invoices, transaction proofs).
   • Use Obsidian format from references/media-handling.md with descriptive captions.
   • Generate image links from filesystem truth: compute the image path relative to the current journal file directory. Never output absolute paths.
   • Do not hardcode path depth (../ or ../../): calculate dynamically from daily_file_path and image_path.
   • Use copied filename as source of truth: if asset file is image_124917_2.jpg, the link must reference that exact filename.

7. Timeline Integration: If significant events occurred, append them to the master index in timeline.md using the format from assets/timeline-template.md and references/obsidian-format.md.

8. Growth Mapping: Update growth-map.md (based on assets/growth-map-template.md) if new behavioral patterns or skill interests are detected.

9. Profile Evolution: Update the long-term user profile (profile.md) to reflect the latest observations on values, goals, and personality traits. See references/profile-evolution.md and assets/profile-template.md.

⏰ Cron & Passive Operation

PhoenixClaw is designed to run without user intervention. It utilizes OpenClaw's built-in cron system to trigger its analysis daily at 10:00 PM local time (0 22 * * *).

• Setup details can be found in references/cron-setup.md.
• Mode: Primarily Passive. The AI proactively summarizes the day's activities without being asked.

Rolling Journal Window (NEW)

To solve the 22:00-24:00 content loss issue, PhoenixClaw now supports a rolling journal window mechanism:

Problem: Fixed 24-hour window (00:00-22:00) misses content between 22:00-24:00 when journal is generated at 22:00.

Solution: scripts/rolling-journal.js scans from last journal time → now instead of fixed daily b

---

Content truncated.