ctxr-cli 0.29.0 → 0.31.0

package/templates/skills/ctxr-circleback.md

@@ -0,0 +1,70 @@
+Capture the Circleback meetings this store owes, on a review branch, without deciding the list by hand. One
+Circleback account can feed several stores, and a tag on each meeting says which store it belongs to. This store's
+routing is in `settings.json` beside this file, rendered from the `circleback` block in `contexture.yaml`.
+
+## When
+
+On a request such as "sweep Circleback" or "capture my meetings", and from a scheduled job. A scheduled job runs the
+runner before you start and puts its report in your prompt, so begin at step 2.
+
+## 1. Run the runner
+
+From the store root:
+
+```
+node __SKILLS_PATH__/ctxr-circleback/sweep.mjs
+```
+
+It needs Circleback's command-line tool (`cb`) signed in, plus `git` and `gh`. It reads every meeting from the last
+35 days with its tags, compares them with this store's default branch and its open pull requests, and prints a report.
+When the store owes any meeting, it also prepares a review-branch worktree and writes each owed meeting's capture into
+`__INBOX_PATH__` there, under CAPTURES RENDERED.
+
+The report is authoritative. Do not work the list out again yourself, and do not rebuild the files it wrote. If it
+prints `COLLECTION FAILED`, report that line and stop.
+
+What the runner writes is the mechanical part of a capture: the source type and id, the meeting's metadata, its
+attendees, Circleback's action items and summary (marked as the service's derivation, not the record), and the
+transcript verbatim. It never writes the recording URL, which is a signed, expiring link and so a credential.
+
+## 2. Add the judgment
+
+For each file under CAPTURES RENDERED, read it, and after `## Attendees` add a `### Speaker caveat` when the speaker
+labels cannot be trusted: generic "Participant N" labels, or a summary that gives a position or a role to the wrong
+person. Say what is unreliable. Name who a label most likely is only where the transcript itself makes it clear.
+
+Never edit the transcript or Circleback's summary, and never add `source_hash` or `ingested`; ingest assigns both.
+
+Then, from the worktree, run `ctxr source check <file> --source-id circleback/<id>`. If the verdict is not `new`,
+leave that file unstaged and report it.
+
+## 3. Publish
+
+Stage exactly the capture files and commit on the review branch. The store's commit hook is the validation; add no
+checks of your own on top of it. Push the branch, and open a pull request against `__DEFAULT_BRANCH__` if none is
+open; a later run adds to the open one. Its body lists each meeting (date, title, id) and says the captures still need
+an interactive ingest through `ctxr-ingest-orchestration`.
+
+Never merge the pull request and never ingest here.
+
+## 4. Report
+
+Give the counts (captured, already present), the pull request link, and the runner's other sections copied as they
+are:
+
+- **LOCKED** — tagged for this store but no longer readable in full, so it cannot be captured.
+- **UNTAGGED**, **TAGGED FOR MORE THAN ONE STORE** and **DECLINED IN THE LAST 8 DAYS**, when the runner prints them.
+
+If nothing was owed and every section is empty, say "Nothing missing".
+
+## How a meeting is routed
+
+- A meeting is this store's when it carries this store's tag, does not carry the decline tag, and carries no other
+  store's tag.
+- An untagged meeting is undecided, not skipped. It gets a tag in Circleback; ask the operator when the home is
+  unclear.
+- A meeting tagged for two stores is captured nowhere until the operator resolves it. A store that already holds it
+  keeps it.
+- Closing a capture pull request does not reject a meeting. Retagging it with the decline tag does.
+- Circleback's free plan hides a meeting's full notes after about 30 days. The runner looks back 35 days, so a weekly
+  run sees each meeting several times before then.