lythoskill-project-scribe-weekly

Files (3)

• SKILL.md
• verify-companion.md · references/
• weekly-format.md · references/

SKILL.md (first lines)

# Project Scribe — Weekly

> Distill, don't replay. The core thread matters; the 17 commits that touched it don't.

## Value Boundary

Weekly is a **memory offloading mechanism**, not a menu listing.

| Direct queries (skip these) | Scribe-weekly must synthesize |
|-----------------------------|-------------------------------|
| `git log --since="7 days ago"` | The *one* decision that mattered most |
| `ls daily/2026-05-*` | Why the planned vs actual divergence happened |
| `cat cortex/INDEX.md` | Which emergent threads earned attention |
| `bunx @lythos/project-cortex list` | What pattern is hardening into a project trait |

If the next agent can find it via `ls daily/`, `git log`, or `cortex index` — don't replicate it. Link to it.

**Why this matters**: A future agent reading a weekly cold (no daily access) should get the *compressed representation* of the week's cognitive work — the pattern, not the commits. Without this, every agent must reconstruct the narrative from 6 daily files + 29 commits + 10 ADRs. Weekly offloads that reconstruction cost once, permanently.

## Weekly Prep (like `cortex probe` — gather before you write)

**Do NOT start writing from memory.** Run prep first. This is the write-side counterpart to onboarding's read-side: just as onboarding reads daily → verifies git → presents summary, weekly prep gathers all sources → surfaces anomalies → ranks importance.

### Step 1: Gather (deterministic)

```bash
# ISO week alignment check — confirm what week we're in and what span to cover
python3 -c "import datetime; d = datetime.date.today(); iso = d.isocalendar(); \
  mon = d - datetime.timedelta(days=d.weekday()); \
  sun = mon + datetime.timedelta(days=6); \
  print(f'Today: {d} ({d.strftime(\"%A\")}) — ISO W{iso.week}'); \
  print(f'Standard span: {mon}_to_{sun}')"

# Daily surface — which days had activity? (filter to dated names before sorting)
ls daily/ | grep -E '^[0-9]{4}-[0-9]{2}-[0-9]{2}\.md$' | sort | tail -7

# Git surface — what actually landed?
git log --since="7 days ago" --oneline

# Cortex surface — what state changed?
bunx @lythos/project-cortex@{{PACKAGE_VERSION}} probe
bunx @lythos/project-cortex@{{PACKAGE_VERSION}} stats

# ADR surface — what decisions were accepted this period?
ls -lt cortex/adr/02-accepted/ | head -15

# Delta from last weekly — what's new since the last snapshot?
ls weekly/ | sort | tail -1
```

### Step 2: Surface Anomalies

Scan the gathered data for:
- **Superseded decisions**: later ADR that contradicts an earlier one (e.g., combo redefinition May 6 superseded May 1 section semantics)
- **Renamed/removed commands**: CLI surface changed (e.g., `agent-run` → `single`, `deck sync` → `deck to-symlink`)
- **Reversed priorities**: a task/epic that was `priority_at_start` last week but never appeared in this week's dailies
- **Silent gaps**: days with no daily file but git shows commits (someone committed outside agent session)
- **Missing ADRs**: significant code/architecture change in git log with no corresponding ADR accepted in the same window. A rename with 30+ commits and no ADR is a red flag.

### Step 3: Importance Ranking (simulated annealing)

1. **High temperature**: dump every event you can recall. Don't filter. Include everything.
2. **Cool**: group related events into clusters (e.g., "path convention fixes" = 5 commits, 2 ADRs, 1 ZK sweep)
3. **Rank clusters** by downstream confusion cost: "if a future agent missed this cluster, how wrong would their decisions be?"
4. **Freeze**: top 1-2 clusters → `core_thread`. Clusters that involve superseded docs → `docs_now_stale`.

**Test**: can you identify the single most important decision of the week? If the combo redefinition (May 6) would be invisible in your weekly, you didn't rank correctly.

### Step 4: Prep Report (show to user before writing)

Present a 5-line summary:
```
📊 Weekly Prep — 2026-W22
📄 Dailies: 2026-05-21, 2026-05-28 (gap May 22-27)
🔀 ADRs: 3 proposed, 0 accepted this period
⚠️  Anomalies: combo redefinition (May 6) not captured in any weekly
💡 Top cluster: site documentation debt cleanup (P0 → done)
📋 docs_now_stale candidates: ADR-20260501160000000, external-skill-governance-bridge.md
```

User confirms → proceed to write. User flags missing items → add to cluster list.

## When to Run

- User explicitly asks for weekly synthesis
- End of week (typically Sunday) for retrospective
- **Friday afternoon**: if the week has already produced substantial work and user wants to capture it before weekend
- **Sunday night**: if weekend work continued the week's thread — append to existing weekly, don't create new
- **Monday**: if weekend was quiet and weekly wasn't written yet
- Before epic planning to understand what unlocked/paused
- **Do NOT auto-schedule** — the player decides timing

## Sunday Check (agent-initiated suggestion)

If today is Sunday and you observe any of these signals during session close:
- Epic/task cleanup feeling ("this phase feels done")
- Goal reached at a natural milestone (milestone completed, tests pass, push to remote)
- User says "LGTM", "就这样" ("that's it"), "先到这里" ("let's stop here") on a Sunday

**Suggest to the user, in their own language**: this week's work seems to have
reached a natural close — do they want to consolidate a weekly and record the
week's pattern?

User says yes → run weekly prep. User says no or ignores → write daily scribe as usual, don't push.

## Weekly Writing Modes

| Mode | When | Action |
|------|------|--------|
| **Fresh** | No weekly exists for current week | Write new weekly with standard period |
| **Append** | Weekly already exists, weekend had more work | Append to existing weekly, update core_thread if needed |
| **Catch-up** | Gap week(s) detected | Write "no active work" weekly for gap, then current week |

**Daily scribe continues regardless.** Weekly does not replace daily; they are different abstraction layers.

## Inputs (collect before writing)

```bash
# 1. Daily files from the past 7 days

From lythos-labs/lythoskill, MIT.

More from lythos-labs/lythoskill