Scheduled Summary

Scheduled Summary

Overview

Messaging platforms are excellent for one active conversation, but they hide the activity that happened elsewhere: other sessions, scheduled jobs, saved memory, and unresolved items from earlier work. This skill creates a periodic digest that surfaces that cross-session activity in a compact format suitable for chat messaging platforms or any notification channel that accepts Markdown or plain text.

The digest is designed for a cron scheduler. It reads optional local data sources (session store, cron output directory, memory JSON files, and log file), filters activity to a configurable time window, and emits Markdown, JSON, or plain text. When no data sources are configured, it emits a template digest that an agent can fill by querying its own platform-specific session history, scheduler status, and memory tools.

Quick start

From this skill directory:

python scripts/summarize.py generate --since 24h

With explicit data sources:

python scripts/summarize.py generate \
  --since 24h \
  --sessions-db /path/to/sessions.sqlite \
  --cron-dir /path/to/cron-output \
  --memory-dir /path/to/memory-json \
  --log-file /path/to/agent.log

Show current configuration and skipped sources:

python scripts/summarize.py config

Create a config template to fill in:

python scripts/summarize.py init --output .summary-config.json

What the digest includes

1. Session activity — sessions started or completed inside the time window, plus topic titles where available.
2. Cron job status — jobs that ran successfully, jobs that failed, and jobs marked overdue by their output files.
3. Memory changes — new saved memories and updated facts from JSON records.
4. Tool usage — most-used tools, total calls, and error counts from a text or JSONL log file.
5. Outstanding items — due decision reviews, incomplete tasks, and TODO-like items found in session transcripts.
6. Time window — every section is scoped to --since (24h, 7d, 2w, etc.) unless the source explicitly marks an item overdue.

Workflow

1. Pick the delivery cadence. Daily is a good default. Weekly works for quieter systems. Avoid high-frequency summaries unless the downstream channel has strong threading or batching.
2. Choose data sources. Configure only the stores you trust the digest to read. Start with cron output and logs, then add session and memory stores once you understand their schema.
3. Generate once manually. Run generate --since 24h and inspect the result for length, sensitive fields, and duplicate items.
4. Tune sections. Use --sections sessions,cron,outstanding to keep the first scheduled digest short. Add memory and tools if they are useful.
5. Schedule the command. Run it from a cron scheduler and redirect output to the channel integration or a file consumed by your notifier.
6. Review after a week. Check whether the digest is surfacing actionable items. Remove noisy sections and add missing source paths.

Completion criterion: the scheduled command produces a digest with a generated timestamp, the requested sections, and no unexpected secrets or full transcripts.

Cron scheduler setup

Generic daily cron scheduler entry:

0 8 * * * cd /path/to/scheduled-summary && python scripts/summarize.py generate --since 24h --output /tmp/activity-digest.md

Example with environment variables instead of flags:

0 8 * * * SESSIONS_DB=/data/sessions.sqlite CRON_OUTPUT_DIR=/data/cron MEMORY_DIR=/data/memory LOG_FILE=/data/agent.log python /path/to/scheduled-summary/scripts/summarize.py generate --since 24h

To deliver to a messaging platform, pipe or post the generated file using your own notifier:

python scripts/summarize.py generate --since 24h --output /tmp/digest.md
python /path/to/send_notification.py /tmp/digest.md

Keep the notifier separate from the digest generator so the script remains portable across platforms.

Output format

Markdown output is optimized for chat platforms:

• Compact ## / ### headings.
• Bullets rather than paragraphs.
• Emoji status markers for quick scanning.
• Short quoted topics instead of transcript excerpts.
• Code blocks only for dense stats when useful.
• A final generated timestamp in UTC.

Supported formats:

python scripts/summarize.py generate --format markdown
python scripts/summarize.py generate --format json
python scripts/summarize.py generate --format text

Write to a file:

python scripts/summarize.py generate --since 7d --output weekly-digest.md

Customization

Include only specific sections:

python scripts/summarize.py generate --sections sessions,cron,outstanding

Common section sets:

• Daily operational digest: sessions,cron,outstanding
• Weekly review: sessions,cron,memory,tools,outstanding
• Low-noise health check: cron,tools
• Agent-filled template: run with no source paths configured, then fill the placeholders from platform-specific tools.

Configure paths with flags, environment variables, or a generated config file:

See references/data-sources.md for expected formats and schema notes.

Integrations

decision-log

If a decision record directory is available, pass it with --decisions-dir or DECISIONS_DIR. The digest scans Markdown decision records for accepted or proposed decisions whose Next review: YYYY-MM-DD is due, then reports the count and titles in the Outstanding section.

news-monitoring

News monitoring jobs usually write a scheduled output file. Point --cron-dir at the directory containing those outputs. The digest will report whether monitoring ran, succeeded, failed, or marked itself overdue. Keep the full news summary in its own file; the scheduled summary should only mention that monitoring ran and whether action is required.

Common pitfalls

1. Too verbose for chat. Do not paste full transcripts or full cron logs into the digest. Surface counts, titles, and one-line action items.
2. Sensitive data leakage. Session titles, memory facts, and log lines may contain secrets or personal data. Review the first outputs manually and redact upstream sources where needed.
3. Running too frequently. Hourly summaries often create notification fatigue and duplicate the active chat. Daily or weekly is usually better.
4. No deduplication across summaries. If the same overdue item appears every day, either resolve it, suppress it upstream, or add a stable decision/task owner outside this digest.
5. Treating template mode as complete data. Placeholder output means no data sources were readable. The agent or operator must fill it from platform tools.
6. Assuming one universal session schema. Session stores vary by platform. The script uses best-effort SQLite introspection, but exact counts depend on the available columns.
7. Mixing generation and delivery. Keep the digest generator independent from platform-specific webhook or bot code; this makes testing and migration easier.

What this skill is not

• Not a session search tool. Use your platform's session search or transcript browser when you need a specific conversation.
• Not a full transcript exporter. It intentionally summarizes titles, counts, and TODO-like lines only.
• Not a replacement for session browsing. It helps decide what to inspect next.
• Not a complete scheduler dashboard. It reads output files and status markers; it does not own or run the scheduler.
• Not a secrets scanner. It reduces verbosity, but it cannot guarantee that the source data is safe to send to a messaging platform.

Reference files

• references/digest-format.md — section contract, output variants, daily and weekly examples.
• references/data-sources.md — source configuration, expected formats, and schema examples.
• templates/digest-template.md — blank Markdown template for agent-filled or manually curated digests.
• scripts/summarize.py — stdlib CLI for generating, configuring, and bootstrapping summary files.

Verification checklist

• python scripts/summarize.py --help shows generate, config, and init.
• python scripts/summarize.py generate --since 24h works without sources and emits template mode output.
• A mock run with sources includes only events inside the time window, except explicitly overdue cron jobs.
• --sections excludes omitted sections from Markdown, JSON, and text.
• --format json returns parseable JSON.
• The first scheduled output has been reviewed for secrets and excessive length before connecting it to a messaging platform.