Slack Channel Summarizer

What this skill does

Pulls a week's worth of messages from a Slack channel, clusters them into topic threads, deduplicates near-identical messages, and produces a "week-in-review" post with permalinks back to the source threads.

Inputs

• channel_id: Slack channel id (e.g., C0123ABC).
• since and until: ISO timestamps bounding the window (typically a Monday-to-Sunday week).
• SLACK_BOT_TOKEN env var with channels:history, users:read, chat:write scopes.
• Optional min_thread_replies: only summarize threads with at least N replies (default 1).

Steps

1. Fetch top-level messages: curl -s -H "Authorization: Bearer $SLACK_BOT_TOKEN" "https://slack.com/api/conversations.history?channel=<id>&oldest=<ts>&latest=<ts>&limit=1000". Paginate via cursor.
2. For each message with reply_count > 0, fetch the thread: conversations.replies?channel=<id>&ts=<thread_ts>.
3. Resolve user IDs to display names via users.info. Cache results.
4. Compute message permalinks: chat.getPermalink?channel=<id>&message_ts=<ts>.
5. Filter messages: drop bot-only threads (every author has is_bot), drop join/leave events, drop messages shorter than 30 characters unless they have reactions.
6. Cluster remaining threads by topic using TF-IDF cosine similarity on the thread's first message + first 3 replies. Threshold 0.4 to merge.
7. Per cluster, write a 1-2 sentence summary based on the longest reply containing a noun-verb-object pattern (real outcome statement). Include 1-3 reply quotes.
8. Sort clusters by (reaction_count + reply_count) descending; take top 8.
9. Render markdown: H2 per cluster with permalink list at the bottom of each.
10. Save to digest-<channel>-<week>.md.

Output

A markdown file per week with: header (channel name, date range, total messages), top-N cluster sections, a "Notable links" appendix listing files and links shared in the channel. Stdout prints message count and cluster count.

Verification

For three top clusters, click each permalink and verify the summary matches the actual thread direction. Recount messages by re-fetching conversations.history for the same window; the count must match the digest header. If the channel had fewer than 50 messages in the window, downgrade to a "quiet week" minimal digest rather than over-cluster small data.

Edge cases

• Private channels: bot must be invited; conversations.history returns not_in_channel otherwise.
• Threads with sensitive content (compliance hold): respect a Slack DLP policy by checking ent_required_acceptance and skipping flagged threads.
• Channels with mostly file uploads: summarize via filenames + descriptions, not message text.
• Time zone differences: weekly bounds default to UTC; allow overriding to a team's local TZ.