vibing-with-ntm

<!-- TOC: One Rule | Cold Start | Tick Loop | Intervention Score Matrix | Operator Proof Card | Swarm Pathology Triggers | Pattern Tiers | Metrics | Checklist | Decision Tree | Quick Start | Attention-Feed Loop | Marching Orders | Swarm Loop | Operator Loop | Autonomous Unstick | Observability | Steady-State Cadence | Quality Loops | Coordination | Anti-Patterns | Troubleshooting | References | Related Skills -->

> **If you are tending a swarm right now:** jump to the [Orchestrator Decision Tree](#orchestrator-decision-tree) below. Drop into [Autonomous Unstick](#autonomous-unstick--dont-wait-for-the-human) for recovery recipes. Everything else is context.

# Vibing With NTM

> **The One Rule:** Observe real state before every nudge. A swarm is not stuck, done, or blocked until pane truth, robot state, work graph, and artifact/git evidence agree.

## Outcome — When a Tending Session Has Delivered

An operator tending session is complete (for now) when **all** of the following hold:

- Every pane is in one of three explicit states: **making progress** (recent useful output / git activity / br/beads movement), **explicitly blocked** with a logged blocker and a handoff path, or **standing down** by policy (queue-dry, convergence, rate-limit cool-down). "Idle and you don't know why" is not a terminal state — it is unfinished tending work.
- Every intervention you applied during the tick was **observed to land** via tail/event/git/br/mail evidence within one observation window. Sends that produced no downstream signal are presumed lost, not silently successful.
- Any policy lever you adjusted (review-only, rate-limit posture, ensemble member set) has been **logged** so the next operator can see what the swarm's contract currently is. Undocumented policy changes are how swarms drift.
- For convergence: the **triple-check** has fired — ready queue empty AND no in-flight work AND no expected upstream signals — before you declare done. Two of three is not convergence; it is observation lag.
- For queue-dry / stand-down: you have **stopped manufacturing work** rather than papering over the dry queue with synthetic tasks. Tending discipline includes the discipline of doing nothing.

If you find yourself sending the same nudge twice without movement, the failure mode has escalated past nudging — drop into [Autonomous Unstick](#autonomous-unstick--dont-wait-for-the-human), don't keep nudging.

## Grounding — Where Operator Truth Lives

Operator decisions must be grounded in observable artifacts, never in agent self-report:

- **Pane truth:** `ntm --robot-tail=<session> --lines=N`. What the agent says it's doing is a hypothesis; what the pane shows is the data.
- **Work graph:** `br ready --json`, `br show <id> --json`. "Ready" status is canonical; if the agent thinks something is ready that br disagrees with, the agent is wrong.
- **Mail / coordination:** the Agent Mail inbox and threads — not Slack screenshots, not summary cards. The thread is the source.
- **Git state:** `git log --since=`, `git status` on the project worktrees. Commits and uncommitted changes are ground truth for "is real work being produced."
- **Snapshot deltas:** compare two `--robot-snapshot` outputs N seconds apart. The diff is what actually changed; everything else is narrative.

Never base an intervention on a swarm participant's self-report when one of these grounding sources can confirm or refute it. If they disagree, the artifact wins.

## Cold Start For Operators

When you open this skill cold, do not read every reference first. Run one bounded operator tick:

1. `ntm --robot-snapshot` and read `sources` / `degraded_sources`, pane states, and cursor.
2. If the state is unclear, use `ntm --robot-tail=<session> --lines=40` for the suspect panes.
3. Match exactly one card or branch in the [Orchestrator Decision Tree](#orchestrator-decision-tree).
4. Act on one pane or one policy lever, then verify tail/event/git/br/mail movement.
5. If no branch fires, wait on the attention feed instead of inventing work.

Use `/ntm` only when command syntax or schema details are unclear; return here for the intervention decision.

## The Tick Loop (Mandatory)

```
1. BASELINE    -> ntm --robot-snapshot; capture cursor + sources/degraded_sources
2. ATTEND      -> ntm --robot-attention / --robot-wait; read only actionable deltas
3. CLASSIFY    -> match one OC card or AP red flag; no card means observe more
4. SCORE       -> choose the smallest reversible intervention with highest action score
5. ACT         -> one targeted send/assign/lock/restart/review instruction; never blanket-nudge
6. VERIFY      -> tail/event/git/br/mail/pipeline state changed; otherwise escalate one rung
7. STOP CHECK  -> convergence triple-check or queue-dry; stop instead of manufacturing work
8. LOG         -> record blocker, degraded source, or handoff when the operator loop changes policy
```

This is the operator counterpart to `/ntm`. `/ntm` tells you what command exists; this skill tells you when to use it, how to avoid thrashing panes, and when to stop.

## Intervention Score Matrix

When several actions are possible, score them before acting:

```
Score = (Evidence x Impact x Reversibility) / BlastRadius

Evidence       1-5: independent sources agree this state is real
Impact         1-5: likely to move code/beads/artifacts, not just produce prose
Reversibility  1-5: can undo/stop/retry without losing work
BlastRadius    1-5: one pane/message is low; all panes/session kill is high
```

Only take Score >= 2.0. If every action scores below 2.0, wait on the attention feed or gather better evidence.

| Candidate action | Evidence | Impact | Reversible | Blast | Decision |
|---|---:|---:|---:|---:|---|
| Terse prompt to one idle pane with no output and no commits | 4 | 3 | 5 | 1 | yes |
| Smart-restart one pane with identical tail across 3 ticks | 5 | 4 | 4 | 2 | yes |
| Broadcast "status?" to all panes | 1 | 1 | 3 | 4 | no |
| Kill a pane because it says "Cogit