---
type: routine
schedule: Sunday 17:00
updated: 2026-09-21
---

# Weekly review

**Sunday 17:00** (ADR-15). Two hours before Ruth calls the landline at 19:00.

The daily brief is the surface. **This is where the system's health is
maintained.** Delegations age, projects stall, ventures go quiet — none of it
fits in a 4,000-character morning message and none of it should.

**Skipped weeklies are the clearest signal the system is breaking.** Two in a row
triggers a shrink of the review, not a nag.

---

## Step 0 — Bea's Sunday email. Before anything else.

**Open it. Scroll to the bottom. Read the "needs you" section first.**

It has three parts: a payables list with dates, balances across four accounts,
and a short "needs you" section at the very bottom that Rowan has, to his
knowledge, **never once scrolled to.** Reading it bottom-first is the whole point.

> "If Bea's email gets read as part of the review then that's one of my failure
> modes closed on day one without anyone doing anything heroic."

Until an email curator is chartered (ADR-20), Rowan opens it and reads the bottom
section aloud; Hobbs files what is in it.

---

## Step 1 — Get current (automated, no input from Rowan)

1. **Drain `Efforts/inbox/`.** Every capture becomes a task, a note, or a
   question. Processed items move to `Efforts/inbox/_archive/<date>/`.
2. **Age the waiting list.** Any `waiting` task past its follow-up date gets
   bumped and noted. Anything over fourteen days with no follow-up date is
   flagged.
3. **Flag stalled projects** — active, no movement in seven days, no `waiting`
   reason.
4. **Flag stale ventures** — active, hub note untouched in seven days.
   **Activity keys off file edits, never task counts.** A venture with task churn
   and no note edits is still stale.
5. **Age someday items** past ninety days — delete-prompt each.
6. **Run `./scripts/guides-watch.sh`** (ADR-16).
   **If the feed is unreachable, say so. Never report "nothing new" from a failed
   fetch** — that is a silent lie of completeness, and the script exits non-zero
   rather than let it happen.
7. **Sweep every task handle in the vault against the task graph.** Added
   2026-09-13 after the fourth and fifth invented handles (`Efforts/feedback.md`).
   SL-01 governs the moment of writing and cannot see a handle already on disk;
   `hobbs-3v9` sat in a venture file for seven days because of that.

   ```bash
   grep -rhoE "hobbs-[a-z0-9]{3}" --include="*.md" . | sort -u > /tmp/refs.txt
   { bd list --status=open --limit 0; bd list --status=closed --limit 0; } \
     | grep -oE "hobbs-[a-z0-9]{3}" | sort -u > /tmp/all.txt
   comm -23 /tmp/refs.txt /tmp/all.txt
   ```

   **Corrected 2026-09-20. The version written on 13 September never worked.** It
   anchored the pattern with `^hobbs-`, and `bd list` prints a status glyph before
   the handle, so nothing matched, `/tmp/all.txt` came back empty, and **every
   handle in the vault reported as dangling** — 79 lines. A check that returns the
   whole corpus is not a check; it is noise that teaches the reader to skim it,
   which is worse than not having run it. Run correctly on 2026-09-20 it returns
   **eight**, all legitimate: the invented handles quoted inside `Efforts/feedback.md`,
   the daily notes and the 13 September report, which is what those files are for.
   **Sanity gate: if the output length is within a few of `wc -l /tmp/refs.txt`, the
   command is broken, not the vault.**

   **Non-empty output is a defect, not a judgment call.** It reports rather than
   auto-fixes: `Efforts/feedback.md` and the daily notes legitimately quote
   invented handles when describing the mistakes they record, and those must not
   be rewritten. Every hit is checked by hand against `bd list` matched on title
   text, and a correction is committed naming itself.

8. **Disposition every task handle cited in anything written this session.**
   Added 2026-09-20 after `hobbs-z4f` — a P0 capital call Rowan had already paid —
   stayed open overnight while ADR-28, written that same evening, cited it **by
   handle** as the worked example of its own clause 7.

   ```bash
   # Handles newly written into the two files where naming one means "this item
   # is implicated": the ADR log and the feedback log. Added lines only.
   for id in $(git diff "$BASE" HEAD -U0 -- Atlas/decisions.md Efforts/feedback.md \
                 | grep '^+' | grep -oE "hobbs-[a-z0-9]{3}" | sort -u); do
     printf '%-12s %s\n' "$id" "$(bd show "$id" 2>/dev/null | sed -n '1p')"
   done
   ```

   **Scope, and it was narrowed the hour it was written.** The first version
   scanned the whole diff, and run against this session it returned **84 handles** —
   every task the weekly review enumerates, plus one (`hobbs-cw8`) that existed only
   on a deleted line of `.beads/issues.jsonl`. That is step 1.7's cry-wolf failure
   reproduced in the same session that fixed it, which is worth recording rather
   than quietly correcting: **a check that returns the corpus is not a check, and
   writing one twice in two hours means the pull toward it is strong.**

   A report enumerates the graph by design and naming a handle there means nothing.
   **An ADR or a feedback entry naming a handle means the item is implicated in a
   decision or a fault** — those are the two files where the invariant is real, so
   those are the two the check reads.

   **Any handle whose state is unchanged from before the document was written is a
   defect, not a judgment call.** Naming an item is discovering it. This vault must
   not be able to hold a document that identifies a problem beside a task that
   still has it. Closed, re-flagged, re-prioritised, or explicitly left alone with
   the reason written into the task — but never silently untouched.

   Every prior version of this failure in `Efforts/feedback.md` was a rule scoped
   to the wrong **artefact**. This one was scoped to the wrong **tense**: the ADR
   was correct, was written before the action per ADR-16, and writing it was taken
   as having handled the instance it cited.

---

## Step 2 — The review. Every section produces a disposition or renders empty.

| # | Section | Disposition |
|---|---|---|
| 1 | **Bea's "needs you" items** | act, delegate, or defer with a date |
| 2 | **Inbox state** — processed, dropped, still unprocessed | decide the leftovers |
| 3 | **Waiting-for**, sorted by age | follow up, escalate, or close |
| 4 | **The avoided conversations** — Priya, Dmitri, Claire, Tomas | one each: had it, or a date by which |
| 5 | **Quiet** — days since for Maya, Ruth, Claire, June, Dmitri, Adaeze | one action each |
| 6 | **Projects** — no next action / no movement / outcome achieved | unstick, park, or close |
| 7 | **Next-actions scan** by venture — orphans, mis-filed, blocked | re-label, re-prioritise |
| 8 | **Someday** — items aging out | keep, promote, or delete |
| 9 | **Ventures untouched this week** — of thirteen | say so |
| 10 | **Portfolio state** — one paragraph per venture | correct the picture |
| 11 | **The unknowns list** — which date-shaped unknowns closed this week | assign the rest |
| 12 | **New from chiefofstaff.io** — renders only when non-empty | "yes, that one" or nothing |
| 13 | **Next week's three to five priorities**, proposed | confirm or edit |
| — | *Appendix:* **FYI digest** | skim, or don't |

## Step 3 — the disposition queue, written in the same act as the review

**ADR-29.** Every section above produces a disposition or renders empty. Each of
those dispositions also gets an entry in
`Efforts/dispositions/YYYY-MM-DD-weekly.json`, written **while the section is
being written** and never afterwards from a re-reading — the report's ask and the
queue entry must not be able to disagree, which is ADR-29's third *Reverse if*.

An entry carries the ask in one sentence, the context Rowan needs to answer it,
the handles it touches, the labels for yes and no, and **the mechanical effect of
each answer**. The effect is the judgment and it is made here, with the vault
open, not by the tool at click time. The five verbs the surface will execute are
`close`, `priority`, `label`, `note`, and `hobbs` — a declared hand-off. A defer
never carries an effect: it carries his words, and words are intentions.

Then say the one line at the end of the review: **`node Tools/review-surface/server.mjs`**.

**Before the next review**, drain `Efforts/inbox/*-disposition-*.md`. Each is an
answer of his that needed judgment. Execute it, write the curation line and the
Atlas-changes line, delete the file.

## Section 12 — how it works

One line per new or updated guide: **title, track, and one sentence on whether it
applies to us.** Not a summary of the guide — a judgement about this system.
"We have no execution lane; this describes one." "We already do this; skip."

On Rowan's "yes, that one": fetch the Markdown twin, report any prerequisites not
yet met, **propose an ADR before touching anything**, then implement, commit per
edit with provenance pointing at the guide, and report it in the next brief's
Atlas-changes line.

Write the seen-file back and **commit it** — it is state, and state lives in the
vault.

## The rule that keeps it honest

A review that is *read* and produces nothing has become a newspaper — the daily
anti-pattern at a larger scale. When that starts, **shrink the review**; do not
make it more interesting.