chat-summary

chat-summary

Produce a concise, grouped summary of what happened in a chat (channel, group, or DM) over a user-specified window. The skill pulls message history in pages, summarises each page, then merges the page-level summaries into a single grouped bullet list — it does not keyword-search across chats (that's chat-search) and it does not send anything (that's notify).

When to use

• The user asks "catch me up on the deploy channel", "what did #eng discuss this morning?", or "summary of my DMs with Alice this week".
• Another skill needs a short narrative of what a channel has been talking about before making a decision.
• The user was offline and wants a quick briefing before opening the messenger UI.

Procedure

1. Resolve the chat

If the user named the chat by a humanised handle (#eng, "the deploy channel", "DMs with Alice"), look up the chat_id:

devboy tools call get_messenger_chats '{"search": "eng", "limit": 10}'
# Pick the match whose name / topic clearly corresponds to what the user said.

If the user already handed over a chat_id, skip this step.

2. Convert the window to provider timestamps

get_chat_messages takes since / until as provider-native timestamps (for Slack: floating-point epoch seconds, as strings — e.g. "1712448000.000000"). Convert the user's natural-language window ("today", "this week", "since Monday") into those before calling. Default to the last 24 hours when the user does not name a window.

3. Pull messages in narrow slices — do not load the whole history

The tool returns formatted text and does not surface a next_cursor in the output today, so you cannot page by cursor from the tool result. Instead, split a large window into several narrow since / until calls and summarise each slice independently:

# Morning slice
devboy tools call get_chat_messages '{
  "chat_id": "C0123456789",
  "since": "1713052800.000000",
  "until": "1713091200.000000",
  "limit": 200
}'

# Afternoon slice
devboy tools call get_chat_messages '{
  "chat_id": "C0123456789",
  "since": "1713091200.000000",
  "until": "1713139200.000000",
  "limit": 200
}'

Cap limit at 200 per call to keep each response small enough to summarise cleanly. If a slice comes back close to the cap, halve the window and re-run. If a thread is referenced by thread_id and the user asked about a specific thread, pull its replies separately:

devboy tools call get_chat_messages '{"chat_id": "C0123456789", "thread_id": "1712450000.001500", "limit": 200}'

4. Summarise chunk by chunk, then merge

Rather than feeding the entire history into one summary prompt:

1. Per page, draft 3–6 bullet points: decisions, open questions, action items, notable quotes (with author handle).
2. Merge the page-level bullets into a single list, collapsing duplicates and merging threads that span pages.
3. Group the final list by topic, not by time. Suggested top-level groups: Decisions, Action items, Open questions, Context / FYI. Omit groups that have no bullets.
4. Keep each bullet to one line, prefixed with the author when it matters (alice: shipped v2.4.2). Cite threads by their chat_id + thread_id pair — the unified messenger output does not supply a permalink field.

5. Render the summary

Target shape (adjust headings based on what actually came up):

Summary of #eng — 2026-04-14 00:00 UTC → 2026-04-15 00:00 UTC (312 messages, 4 threads)

Decisions
- Rolled back v2.4.1 after staging regression (alice, bob)
- Ship v2.4.2 with the patched migration tomorrow at 09:00 UTC

Action items
- @alice: draft incident-204 post-mortem by Thursday
- @carol: add regression test covering the migration case

Open questions
- Is the feature flag cutover still on for Friday?

Context
- Customer X asked about the GitLab SAML change — redirected to #support

If the window had fewer than ~20 messages, skip the grouping and render a single flat list.

Success criteria

• The summary covers the requested window — the earliest and latest messages are reflected.
• Bullets are grouped, not a flat timestamp dump. Duplicates across threads are collapsed.
• Author handles are present where they add signal (decisions, action items).
• Pagination was actually used — the skill did not truncate silently at the first page.

Guardrails

• Messages frequently reference confidential business decisions, customer names, or unreleased work. Never forward the raw message content to external systems — the issue tracker, a PR description, an email, a public channel — without the user's explicit ask. A summary surfaced in the conversation with the requesting user is fine; cross-posting it is not.
• If the Slack token lacks channels:history, groups:history, im:history, or mpim:history, the corresponding chat types will return empty. When a summary feels suspiciously short, run devboy test slack and check the "Missing scopes" line before blaming the users.

Non-goals

• This skill does not search across many chats for a keyword — that is chat-search.
• It does not post the summary anywhere — if the user wants that, hand the summary to notify after confirming the target.