---
name: thermograph-kylene-summaries
description: The periodic job that mirrors kylene's Drive-hosted call summaries into the docs repo so they are searchable. Read this when running the loop, or setting it up.
---

# kylene call-summary mirror

**kylene** records `war-room`, and when the room empties it summarises the
whole call, posts to `#voice-summaries`, and files a Google Doc in the team's
Drive folder. That output is good — and invisible. `docs_search` and
`notes_search` index the docs repo **only**; neither can see Google Drive. So a
decision made on a kylene call cannot be found by anyone, human or agent, who
does not already know to go and open Drive.

This job closes that gap: it mirrors each finished call into `summaries/` in the
docs repo, enriched with estate context, so the existing search tools reach it.

## There is only one voice pipeline

kylene is it. Earlier versions of these docs described a second pipeline running
beside it — raw transcript streamed into `#voice-transcripts` and digested every
30 minutes into `summaries/` — and **no such job ever existed**. The only thing
that ever posted to `#voice-transcripts` was a third-party bot trialled briefly
and since removed from the server; nothing in this estate wrote there.

It matters here because that second pipeline was the reason this skill warned
about duplicate and empty digests, and the reason the `(kylene)` marker read as a
collision guard. There is nothing to collide with. Keep the marker — it is how a
re-run recognises what it already mirrored — but do not go looking for the other
side of it.

| | kylene (this skill) |
|---|---|
| Voice channel | `war-room` |
| Posts to | `#voice-summaries` |
| What it posts | a **finished summary**, once per call |
| Durable copy | a Google Doc in Drive |
| This job's work | **mirror and enrich** an existing summary |

Because kylene has already summarised the call, **do not re-summarise from
scratch and do not pad.** The value added here is durability, searchability and
estate cross-referencing — not a second opinion on what was said.

## What the loop does, each run

1. **Find what is already mirrored.** `summaries_list` + `docs_read` on recent
   `summaries/` files. Sections mirrored by this job are headed
   `war-room <HH:MM>–<HH:MM> UTC (kylene)`, so an existing section for a
   call means that call is done. This is the idempotency anchor — a re-run or a
   missed run must not produce a second copy of the same call.

2. **Find the calls.** `voice_calls` with no `file` argument lists the Drive
   documents newest first, one per call. Any call with no matching section is unmirrored. If
   every call is already mirrored, **stop — do not write an empty summary.**

   If the Drive tools report no credential configured, fall back to reading
   `#voice-summaries` with `discord_get_messages` (its id is in
   `community_channels` under `voice_summaries`, or `CENTRALIS_VOICE_SUMMARY_CHANNEL`).
   The Discord post carries the summary but not the full transcript, so say so
   in the mirrored section rather than implying you had the whole call.

3. **Read the call.** `voice_calls` with `file` set to its id. It holds kylene's
   summary, the decisions and action items, and the transcript those were drawn
   from.

4. **Mirror it.** Carry across, tightened but not reinterpreted:
   - What was discussed and, above all, **what was decided**.
   - Action items, with owners where kylene identified them.
   - Open questions left unresolved.
   - Keep the grade-language rule: never call a percentile "hot" or "cold".
   - Link the Drive document, so the verbatim record is one click away. The
     mirror is a searchable index of the call, not a replacement for it.

5. **Enrich with estate context.** Where the call *genuinely* connects to
   something the estate already knows, look it up and fold in a brief, **cited**
   note under a short **Related context** section, clearly separated from what
   was said. Never invent a connection. Draw on:
   - **Repos, PRs** — `onedev_read`, `onedev_projects`, `onedev_prs`. OneDev
     has no issue tracker of its own; a bug or feature request lives in Taiga
     instead (`taiga_story_list`, `taiga_story_get`).
   - **Docs & notes** — `docs_search`, `docs_read`, `notes_search`, `notes_read`
     (is there a prior record? does a note contradict this?).
   - **The other pipeline** — `summaries_list` / `docs_search`. A topic raised on
     a kylene call may already have a war-room summary; cross-reference rather
     than duplicate.
   - **App stats** — `sql_query` (read-only, never `write:true`).
   - **Climate stats** — `lake_query` / `lake_tables`, `grade_location`,
     `geocode`. Percentiles versus local history, never absolutes.
   - **What broke** — `logs_query`, `fleet_status` when an incident is mentioned.
   - **Outside world** — `WebSearch`, `WebFetch`.

6. **Write it.** `summaries_write` with a `heading` of
   `war-room <HH:MM>–<HH:MM> UTC (kylene)`, taking the times from the call
   document. It appends to `summaries/YYYY-MM-DD.md` for the **call's** date —
   pass `date` explicitly rather than defaulting to today, or a call summarised
   after midnight UTC lands on the wrong day. Commits straight to main; do not
   open a PR.

   The `(kylene)` marker is what this job anchors on to tell which sections it
   has already mirrored, so a re-run is harmless and a missed run self-heals.

7. **Surface it.** Post a short version into **#voice-summaries** with the
   Discord tools, linking the committed file and the Drive document. kylene
   posts there too, so it is the single place to
   see "what was decided on a call", whichever room it happened in. Name the
   room in the post so the two are distinguishable.

## Cost and gating, worth a glance

Every kylene summary reports minutes of speech, minutes billed, and the request
count. If billed minutes run far above spoken minutes, capture gating or
batching has regressed — mention it in the mirrored section and in the
`#voice-summaries` post, because nothing else on the estate watches that number.

## Idempotency and safety

- **Anchor on mirrored sections, never wall-clock.** That is what makes a re-run
  harmless and a missed run self-healing.
- **One section per call.** A call is the natural unit — never merge two calls
  into one section, and never split one.
- **Never invent content.** If the call document is thin or the transcript
  garbled, mirror what is legible and say it was thin.
- **Only `#voice-summaries` and the Drive folder, for call content.** Do not pull
  from other Discord channels — those two are where consent was given.
  `war-room` is announced-on-join and gated to Owner/Staff/dev; the
  community `Lounge` is never recorded by anything.
- **Enrich read-only.** No `sql_query(write:true)`, no `docs_write` /
  `onedev_write`. The only writes are `summaries_write` and the
  `#voice-summaries` post.
- If neither the Drive folder nor `#voice-summaries` is configured, there is
  nothing to do — report that the source is not configured and stop.