Morning Brief

Morning brief

You are producing one email for the operator (oli@tomevault.io). The goal is a trustworthy daily snapshot, not a feed of agent findings. If today is genuinely quiet, say so in two sentences and stop. Most days will be quiet.

Hard rules

1. The pulse is the source of truth, not the agent events. Agent events (pulse.events.items) are suggestions — verify each one against pulse counts and memory before promoting it.
2. Apply memory hints before drawing conclusions. pulse.memory_hints lists topics likely relevant today. Read each named memory file under /root/.claude/projects/-home-tomevault/memory/ before writing about a topic the hint flags.
3. Never promote upstream confidence. If an event says "boilerplate count may be inflated" → use roughly_even or weaker, never almost_certain.
4. No subtraction-based "stuck" claims. If you're tempted to write "X repos are stuck/unfetched/etc.", check pulse.pipeline.repos.by_status first. gone, security_blocked, skipped_*, cancelled, *_error are terminal — they're not stuck, they're done.
5. Filter bot traffic from funnels. If pulse.traffic.bot_pct > 5, any PostHog/CTR claim must explicitly note the filter or be omitted. Do not promote unfiltered CTR collapses.
6. Cross-check "stage dead" claims against pulse.cron.stages. Only flag a stage as broken if in_cron=true AND log_age_min exceeds its expected interval.
7. Orphan locks are not crises. pulse.cron.orphan_locks are leftovers from removed stages. Mention once if you've never mentioned them, then drop them.
8. Classifier health is a state-change-only signal. pulse.classifier carries Haiku triage progress. Surface it ONLY when:
   • paused: true (kill-switch fired — surface as RED with the reason)
   • calls_today.pct_used > 80 (approaching daily ceiling — AMBER)
   • last_24h.fallback_rate_pct > 10 (Haiku failing → topics-only fallbacks — AMBER)
   • suspicion_dist_24h shape shifts unusually (e.g. score≥3 jumps from ~3% baseline to >10% — calibration drift, AMBER)
   • A tier-drain milestone crosses (Tier 1 hits 100% → flag as a green positive, prompts UI flip)
   • last_classification is older than 30 minutes (stage stopped silently — RED) On a normal day, omit it. The one-line Pipeline summary already captures steady-state.
9. Verify deploy state before flagging "built but not live" claims. Before promoting any ticket as "fix shipped but undeployed", check pulse.deploys.tomevault_web. If in_sync: true, the deploy already shipped — drop the claim. If in_sync: false AND pending_lag_minutes is small (<30), the deploy is likely in-flight. This addresses the 2026-05-11 false-positive on TOM-536 where a deploy 3.5h prior was flagged as pending.
10. One operator-facing email. No more than 400 words.

Procedure

1. Read the pulse

cat /home/tomevault/reports/$(date -u +%F)/pulse.json

If it's missing, run venv/bin/python3 -m scripts.daily_pulse first.

2. Read memory index + flagged topics

cat /root/.claude/projects/-home-tomevault/memory/MEMORY.md

For each item in pulse.memory_hints, read the corresponding *.md file in the memory dir.

3. Read recent events with skepticism

For each pulse.events.items entry:

• Verify the count. If event says "50,953 stuck", check pulse.pipeline.repos.by_status.
• Apply memory. If event is about claim CTR and bot_pct > 5, the GAE scrape memory likely explains it.
• Lower confidence on hedged source language. Auditor saying "may be inflated" → confidence ≤ roughly_even.
• Drop the event if pulse contradicts it.

4. Read yesterday's briefing for continuity

cat /home/tomevault/reports/$(date -u -d yesterday +%F)/morning-brief-*.md 2>/dev/null

• If yesterday flagged something as RED, explicitly confirm whether it's resolved (don't ask the operator to confirm).
• If yesterday flagged the same item, mark it as "still open day N" rather than re-introducing it.

5. Decide the briefing mode

• GREEN (0 verified red, 0 verified amber after pulse cross-check): one-line email. Subject TomeVault daily: green. Body: one sentence + the pulse summary line.
• AMBER (1+ verified items): full briefing.
• RED (anything pulse-confirmed irreversible or operator-blocking): RED tag in subject.

6. Write the briefing

Sections in order:

# Daily — YYYY-MM-DD [GREEN | AMBER: N | RED: N]

## TL;DR
<one sentence. include the most decision-relevant fact.>

## Operator decisions (N items)
<only items that need operator action this week. Each:>
- **<title>** — <one sentence what + one sentence why now>
  - Action: <exact command or decision>
  - Memory: <which memory file informed this, if any>

(If none: "No operator decisions today.")

## Pipeline
- Repos: <total> (<active> active, <terminal> terminal)
- Claims: <claimed>/<notified> (<rate>%)
- Distributions: <today> today, <yesterday> yesterday, <total> total
- Drift: <drift> (Meili sync gap)
- New authors: <today>/<yesterday> (baseline 600-800/day)
- Classifier: tier1 <pct>% done, <review_queue> in Sonnet queue, <calls_today>/<ceiling> Haiku calls
  (one line; omit if classifier is paused or off — that goes in the AMBER/RED section instead per Hard Rule 8)

## What changed since yesterday
<diff items with actual numbers from pulse.delta>

## Notes (verified)
<2-4 bullets max. Each must be cross-checked against pulse + memory.>

7. Save the briefing

mkdir -p /home/tomevault/reports/$(date -u +%F)
write to /home/tomevault/reports/$(date -u +%F)/morning-brief-$(date -u +%F).md

8. Send the email

Use the existing factory email skill:

import sys; sys.path.insert(0, '/home/tomevault')
from scripts.lib.email_sender import send_notification
send_notification(
    to_email="oli@tomevault.io",
    to_name="Oli",
    subject=f"TomeVault daily {date} [{tag}]",
    text_body=briefing_markdown,
)

If TOMEVAULT_ENV != prod, the call fails closed and writes to logs/staging-would-have-sent.jsonl — that's expected during testing.

9. Append one line to changelog

[morning-brief YYYY-MM-DD] {tag}: N decisions. Verified M events, dropped K.

What to drop entirely from the legacy briefing format

• ❌ "PENDING YOUR DECISION" + "DECISIONS" both as separate sections (collapse into one).
• ❌ "Confidence" labels — operator doesn't act on them. State the action, or don't surface the item.
• ❌ "FACTORY" health section — this is internals; only mention if something is broken.
• ❌ "LOOP HIGHLIGHTS" — vibes, not data.
• ❌ Operator confirmation prompts ("operator should confirm"). The brief should do the confirming.

Verification examples (use as templates)

Bad (today's briefing):

50,953 repos stuck in discovered (22.9%)

Good:

No items "stuck in discovered" — verified pulse.pipeline.repos.by_status shows 0 in discovered status. The 50k delta from total is gone (41k, deleted from GitHub) + security_blocked (8k) + skipped (~1.7k). All terminal.

Bad:

Claim funnel CTR collapsed 100x — 23,944 PVs → 14 clicks (0.058%)

Good:

Claim funnel signal is corrupted by GAE bot scrape (bot_pct=14.8% in caddy, per project_gae_claim_scrape). Real human PVs ~14k of 23.9k; CTR is approximately unchanged. Recommend: bot-IP filter on claim_page_viewed event.

Source: tomevault-io/companyos — distributed by TomeVault.