Settings

# FlowVolt Operator OS — MASTER

The single operating doc for Lars's senior agent squad. Sage keeps the footer
current nightly; the rest is human-owned.

## 1. What this is

An 11-agent team that runs FlowVolt's internal operations: two directors
(Eli, Diana) and nine specialists. Brain (reasoning) on Opus, hands (execution)
on Sonnet, volume on Haiku. Output surfaces: Slack (#flowvolt-team / -drafts /
-builds) and the Mission Control dashboard. Everything Lars-impacting routes
through the approval queue. Zero automatic sends.

## 2. The squad

Directors:
- Eli ♚ Chief of Staff — weekly memo, delegates. Mon 10:00.
- Diana 🎯 GTM Director — drives deals, on-demand.

Eli's team:
- Hugo ☀️ Pipeline Reporter — daily 07:00 funnel brief.
- Felix ✨ Voice Drift QA — weekly engine-draft audit. Mon 09:00.
- Iris 🪶 Content — LinkedIn ghostwriter; research daily 06:00.
- Cleo 🛠️ Builder — on-demand builds + PRs; AI digest daily 06:30.
- Sage 🌙 Memory — real-time learner (event-driven) + nightly batch 02:00. See §6.

Diana's team (on-demand → #flowvolt-drafts):
- Maya 🔍 Prospect Brief · Nora ✏️ Sales Follow-up · Otto 🤝 Deal Closer ·
  Atlas 🧭 Customer Success (dormant until first paying customer).

## 3. Voice

`lars-voice.md` is auto-extracted from Lars's real replies. Eight agents read it.
Re-extract with `npm run extract-voice`.

## 4. Approvals (Hands)

Nora/Otto/Atlas/Cleo queue actions to `pending_approvals`. Lars approves in the
dashboard; email sends go from his Gmail on approve and are logged to
`operator_sends`. Cleo opens PRs on `cleo/<task>` branches (never merges).

## 5. Owned numbers (accountability)

Each agent owns one weekly metric. Sage/Eli track these in `agent_outcomes`.

| Agent | Metric | Target |
|---|---|---|
| Hugo | brief_on_time | 5/5 |
| Felix | team_voice_avg | 8.0 |
| Diana | new_warm_prospects | 1 |
| Maya | briefs_delivered_pct | 100 |
| Nora | reply_rate | 0.25 |
| Otto | warm_to_meeting_pct | 0.30 |
| Atlas | first_to_paying_pct | 0.50 (when active) |
| Iris | followers_gained_weekly | 250 |
| Cleo | prs_shipped | 3 |
| Sage | prompt_tunings | 5 |
| Eli | outcome_hit_rate | 0.80 |

Eli flags any metric missed two weeks running in the Monday memo.

## 6. Sage — the learning loop

Sage learns in two clocks. *Real-time* (event-driven, in-process) reacts the
moment a decision lands, so a 14:00 edit makes every agent smarter by 14:00:01.
*Nightly* (02:00) still runs for the aggregation that needs a full day: hook-score
averages, per-agent ROI, voice re-extraction.

Real-time triggers (`lib/sage-realtime.ts`, dispatched via `dispatchToSage`):

1. *Approve verbatim* — the draft is embedded into `golden_examples`.
2. *Edit then approve* — strongest signal. The edited version goes to golden +
   `lars_writing` + `lars-corpus/`; the original is logged as a near-miss in
   `failure_patterns`.
3. *Reject* — only `failure_patterns`. Never touches Lars's voice.
4. *Positive reply* — a send that drew a reply is promoted to golden. Fires
   automatically via the Gmail poller (`scripts/poll-gmail-replies.ts`, every
   30 min in business hours; sentiment on Haiku).
5. *Deal closed* — `revenue_attribution.deal_closed_at` set (dashboard
   `/api/deal`) credits every agent in the `agents_touched` chain and promotes
   their actual outputs to golden as revenue-proven examples.
6. *LinkedIn metrics* — metrics entered on `/content` (`/api/linkedin/[id]/metrics`):
   a post beating 1.5x the 7-day median is promoted to golden (viral); under
   0.5x it is kept as regen material.

Every tuning is logged to `sage_tunings` (the audit trail + the dashboard
`/team` activity feed) and each change is git-committed so Lars can revert.

Guards: tunings capped at **20/agent/day** (stops thrashing on batch approvals);
killswitch **`SAGE_REALTIME_DISABLE_COMMIT=true`** stops the per-decision git
commits (embeds + tunings still run).

---

<!-- SAGE-FOOTER -->
Last updated by Sage: (not yet run)

Changes this week: (none yet)
<!-- /SAGE-FOOTER -->