Polaris boilerplate: PARA vault, AGENTS.md rulebook, eight polaris skills

.agents/skills/polaris-close/SKILL.md

@@ -0,0 +1,104 @@
+---
+name: polaris-close
+description: "Close a finished workstream in an Polaris vault and return its deliverable to the Area that spawned it. Verifies done_when is actually met, splits the output into deliverable (→ 2. Areas/<Area>/knowledge/) and process record (→ 5. Archive/<Area>/<ID>/), updates the Area hub, releases follow-ups tied to the work, and repoints inbound links. Not reversible. Distinct from polaris-archive, which retires a whole Area. Triggers on: close workstream, close [ID], this is done, finish workstream, ship it."
+---
+
+# polaris-close — return a finished workstream to its Area
+
+**An Area spawns a workstream; when the workstream finishes, its deliverable comes home.** The Area is the memory; the workstream is the episode. If finished output archives together with its process, the Area learns nothing and the next workstream starts cold.
+
+```
+2. Areas/<Area>/                ──spawns──▶  1. Projects/<Area>/<Name>/
+  knowledge/  ◀── deliverable ──────────────────┤
+                                                 └── process ──▶ 5. Archive/<Area>/<ID>/
+```
+
+## Close or archive?
+
+| | `polaris-close` | `polaris-archive` |
+|---|---|---|
+| Trigger | One workstream **finished** | A whole Area **ended or paused** |
+| The Area | **Stays, gains knowledge** | Moves into the archive |
+| Reversible | **No** | Yes |
+
+**Is the Area staying? → close. Is it going away? → archive.** If there's any chance the work resumes, it's `paused`, not closed.
+
+---
+
+## 1. Verify `done_when:` against reality
+
+Read the hub's `done_when:` and check it against what actually exists — not against how finished the work feels.
+
+**Not met → stop and say which part is outstanding.** Never rewrite `done_when:` at close time to match what happened; a done-when edited at close never tested anything. Legitimate paths:
+
+| Situation | Action |
+|---|---|
+| Work is done, the sentence was wrong | **The owner's call.** Record the correction and why on the hub, then close |
+| Work stopped and won't resume | Not a close — `paused` with a reason |
+| Scope changed | Old workstream closes as superseded; a new one opens with a new ID |
+
+## 2. Split the output
+
+| Pile | Goes to | What |
+|---|---|---|
+| **Deliverable** | `2. Areas/<Area>/knowledge/` | What the Area now knows: conclusions, specs, finished documents, reusable findings |
+| **Process record** | `5. Archive/<Area>/<ID>/` | How it got there: session notes, drafts, dead ends, the hub itself |
+
+**One question per file:** *would someone working in this Area six months from now want this without knowing the workstream existed?* Yes → knowledge. No → archive.
+
+**When genuinely unclear, ask.** This is the one checkpoint that stops execution.
+
+## 3. Migrate the deliverable
+
+Move each deliverable file into `knowledge/`. **Keep the filename** (renaming breaks links for no gain). Add a provenance line if missing: *"Produced by `<ID>` — `<Name>`, closed YYYY-MM-DD."*
+
+## 4. Archive the process record
+
+Move everything else, **including `<ID>_hub.md`**, to `5. Archive/<Area>/<ID>/`. Set `status: closed`, add `closed: YYYY-MM-DD`. No tombstone needed — nothing to restore.
+
+## 5. Update the Area hub
+
+Mark the workstream closed in `## Workstreams`, with its ID and one line on what it produced. Link the deliverable from the hub or `knowledge/_index.md`. ⚠ **Never decrement `workstream_counter:`.** IDs are never reused.
+
+## 6. Release follow-ups
+
+List every `tier: relationship` person whose follow-up exists *for this workstream* (they link the hub, are named by its `next:`, or their page says so). **Change nothing on their pages** — whether the relationship outlives the work is the owner's call. Record them on a `Released:` line in the log entry; the next weekly sweep lists them.
+
+## 7. Repoint inbound links
+
+1. **Scan** every `.md` in the vault for `[[…]]` targets pointing into the old folder (full-path and bare-name forms). Ignore links inside code blocks.
+2. **Resolve each hit against the real file tree** — a grep hit is a hypothesis.
+3. **Change only the link target, never the prose.** Keep display aliases: `[[new/path/File|Original Text]]`. Ledgers (`4. Decisions/`, `wiki/log.md`, `wiki/folds/`) included — pointer only.
+4. **Rescan from scratch and report that number**, not a tally of your edits.
+
+## 8. Log and hand off
+
+Append to `wiki/log.md`: `[build]` with a `Shipped:` line if the deliverable is a shipped thing, otherwise `[vault]`. Include `Released:` if any. Then suggest `cascade`.
+
+---
+
+## Output
+
+```
+## Close — <ID> · <Name>
+
+done_when: "<sentence>" — ✅ met / ⚠ met with correction (why)
+Deliverable → 2. Areas/<Area>/knowledge/
+- <file> — what the Area now knows
+Process → 5. Archive/<Area>/<ID>/ — N files incl. hub
+Area hub: <what changed>
+Links repointed: N — rescan shows N dead (was N)
+Released: <Name — follow_up_by, what it was for> · or none
+
+Next: run cascade.
+```
+
+## Rules
+
+- **Verify done_when before anything moves.** Never rewrite it to fit.
+- **Split deliverable from process.** Ask only when genuinely unclear.
+- **Every migrated file carries its origin.**
+- **Never decrement the counter. Never reuse an ID.**
+- **Closing isn't reversible.** If it might resume, pause it.
+- **Only link targets change. Rescan and report the rescan.**
+- **`area: none` workstreams archive whole** — there's no knowledge base to receive a deliverable.