Vibing With Ntm

If you are tending a swarm right now: jump to the Orchestrator Decision Tree below. For recovery recipes see RECOVERY.md. 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.

This skill is the operator layer above the tools — the decisions, ticks, nudges, and recoveries an orchestrator performs. It deliberately does NOT re-document the ntm command surface: the binary self-describes via ntm --help and ntm robot-docs (and the $ntm skill catalogs it). Always re-query those for syntax, flags, and schemas; come back here for when to act and when to stop.

Outcome — When a Tending Session Has Delivered

A 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 / bead movement), explicitly blocked with a logged blocker and handoff path, or standing down by policy (queue-dry, convergence, rate-limit cool-down). "Idle and you don't know why" is unfinished tending work.
• Every intervention was observed to land via tail/event/git/br/mail evidence within one observation window. Sends with no downstream signal are presumed lost, not silently successful.
• Any policy lever you adjusted (review-only, rate-limit posture, ensemble set) is logged — 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. Two of three is observation lag, not convergence.
• For queue-dry: you have stopped manufacturing work. Tending discipline includes the discipline of doing nothing.

If you send the same nudge twice without movement, the failure has escalated past nudging — go to RECOVERY.md, don't keep nudging.

Grounding — Where Operator Truth Lives

Ground every decision in observable artifacts, never in agent self-report:

• Pane truth: ntm --robot-tail. What the agent says is a hypothesis; what the pane shows is data.
• Work graph: br ready --json, br show <id> --json. The tracker is canonical; a disagreeing agent is wrong.
• Mail / coordination: the Agent Mail inbox and threads — the thread is the source.
• Git state: git log --since=, git status on project worktrees — ground truth for "is real work being produced."
• Snapshot deltas: two --robot-snapshot outputs N seconds apart. The diff is what changed; everything else is narrative.

If a grounding source and a self-report disagree, the artifact wins.

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 loop changes policy

If you cannot name the phase you are in and the evidence behind it, do not nudge — re-observe.

Cold start: run one bounded tick (--robot-snapshot → tail suspect panes → match exactly one decision-tree branch → act on one pane or lever → verify). Do not read every reference first; wait on the attention feed instead of inventing work.

Gate-Request Interrupts

A peer gate request is an interrupt, not background watch traffic. Treat these phrases as a first-class ACT trigger even if another pane looks busy: ACTION NEEDED, Hey! Listen!, merge gate, unblock-condition, dry-run verdict, PASS / REVISE verdict, ACK-with-id, or an explicit request to verify a SHA/PR/bead before merge or close.

When this trigger fires:

1. Stop broad watching and answer the gate first.
2. Run the smallest authoritative verifier named by the request (PR check, br show, dry-run command, test, or pane capture).
3. Surface the result where the requesting peer can actually read it. Agent Mail is enough only if their read path is known-good; otherwise write a bead note, PR comment, or pane relay with tmux send-keys ... C-m and capture evidence.
4. Do not claim the peer was answered until the visible artifact or capture proves the verdict landed.

This is the fix for ag-navi-gate-interrupt-ggque: a navigator that leaves an unblock request sitting in mail or pane scrollback has not tended the swarm, even if the underlying watch loop is otherwise healthy.

Liveness Truth Stack (what to believe, in order)

Before acting on any "pane is stuck / rate-limited / done / idle" judgment, verify in this order — each layer catches lies from the one above:

1. tmux list-panes … -F '#{pane_current_command} #{pane_pid}' — is the agent CLI even running? Silent exits back to zsh are invisible to --robot-tail. (OC-026)
2. tmux capture-pane -p -S -20 — ground truth for transient state; --robot-tail can sample stale buffer content for several ticks. (AP-41) Caveat: a full-screen TUI agent (codex especially) can return almost entirely ANSI/alt-screen control sequences — stripping leaves empty. Capture-unreadable is NOT evidence of idle; fall through to layer 3.
3. git log --since='15 minutes ago' + pgrep -af cargo|rustc|go|bun — are commits landing, are builds running? Timer labels like "Cogitated for 35m" are display artifacts, not activity. (AP-42)
4. ps -p $(tmux display -p -t <s>:<w>.<pane> '#{pane_pid}') -o pcpu= — the signal-of-last-resort. 0.0% on the pane process tree + no growing artifact = genuinely idle, even when the meter AND capture both lie. The one signal that survives when everything above it is unreadable (e.g. codex-TUI capture returns empty). (AP-48, AP-54)
5. ntm --robot-is-working / --robot-health-oauth — provider-side state (rate-limit, context, quota).
6. ntm --robot-snapshot | jq '{sources, degraded_sources}' — data-source freshness before acting on any derived state.

If any two layers disagree, resync before acting. See OBSERVABILITY.md → "Liveness Signals That Can Lie" for the full catalog.

Never-engaged vs stuck-mid-work (use the stack to tell them apart). A never-engaged pane = CLI running (layer 1 OK) + 0.0% CPU on pane_pid (layer 4) + no working/thinking indicator + no artifact + no record of an acknowledged first dispatch. This is almost always a dropped first send (the boot race — see OC-047): the pane booted but the send arrived before its input box rendered. The fix is RE-DISPATCH (it never received the order), NOT restart. Distinguish from stuck-mid-work (CPU/build active, or an indicator present — leave it, see AP-54) and from standing-down (queue-dry by policy). Restarting a never-engaged pane throws away a perfectly healthy CLI.

Intervention Discipline

Two gates before any state-changing action:

1. Score it: Score = (Evidence × Impact × Reversibility) / BlastRadius, each 1–5. Only act on Score ≥ 2.0; below that, wait on the attention feed or gather better evidence.
2. Proof-card it: evidence, card matched, target, expected state change, reversibility, verification command, next escalation rung. No proof card → no nudge, restart, force-release, or shutdown.

The full scoring matrix with worked examples, the proof-card template, the swarm-pathology trigger table, the three pattern tiers (low-risk tending → directed recovery → session policy changes), the red-flag-phrase classifier, recovery shortcuts, and the troubleshooting table all live in DECISION-AIDS.md.

Orchestrator Decision Tree

Run one tick. Pick the FIRST branch whose condition fires.

Is CURSOR expired (or missing)?
  → ntm --robot-snapshot  (resync, get new cursor); continue next tick.

Is ANY pane rate_limited?  (check via --robot-health-oauth, NOT pane buffer text)
  → OC-001/OC-002 in OPERATOR-CARDS.md: ping-probe first; rotate or switch account
    only if the limit is confirmed provider-side.

Is ANY pane stuck (identical tail ≥3 ticks, zero output growth)?
  → Climb the stuck-pane ladder (OC-003):
    wake-ping → C-u + send → smart-restart → hard-kill → restart-pane → add+kill.

Is there prose-without-commits? (pane is_working=true but git log 1h=0)
  → Dispatch OC-004 Ship-or-Surface prompt (PROMPTS.md).

Is context >85% on any pane?
  → Dispatch handoff-then-restart (OC-009).

Is there a file-reservation conflict or coordinator-reported collision?
  → Force-release too-broad patterns (OC-008); mediate via bead status-flip (OC-015).

Is ANY pane never-engaged?  (CLI running AND 0.0% pane_pid CPU AND no indicator AND no artifact AND no acknowledged first dispatch — typically a bare spawn whose first send lost the boot race)
  → RE-DISPATCH, don't restart (OC-047): confirm input-ready, then re-send the marching orders.

Does convergence triple-check hold?
  ( git log 1h=0 AND br ready=0 AND in-flight unchanged ≥2 ticks AND convergence language in every pane )
  → STOP. Do not nudge. Exit the loop; report final state.

Otherwise — one specific-terse nudge per genuinely-idle pane (OC-010). Then wait.

Every card (OC-###) and anti-pattern (AP-###) is documented with recipe, prompt module, and validator in OPERATOR-CARDS.md and ANTI-PATTERNS.md.

Command Surfaces (do not re-learn from here)

• atm is ntm. atm (Bo's fork/alias, ~/.local/bin/atm) is byte-identical to ntm — same robot surface, same flags, same exit codes. Every ntm … / ntm --robot-* form in this skill and its references applies verbatim to atm; the out-of-session substrate-runner skill is named using-atm for that reason. Use whichever the operator typed; they resolve to one contract.
• Fresh Claude/Codex duel requests go through ATM panes. If the operator asks for fresh Claude and Codex peers, a cross-family duel, or fresh peer-model judgment, spawn the requested panes with atm spawn and follow the using-atm "Fresh Claude/Codex Peer Duels" pattern. Do not route this shape through claude -p / claude --print or ad hoc headless one-shots; the value is fresh interactive peer state plus observable pane engagement. Use atm codex preflight, atm send --codex-goal, and atm codex wait-goal-engaged for the Codex pane, then kill the temporary duel session after collecting output.
• Agents use robot surfaces (--robot-snapshot, --robot-attention, --robot-send, --robot-smart-restart, …); interactive TUIs (ntm dashboard, palette, view) are for humans. The authoritative catalog is ntm robot-docs / ntm --help — re-query it; see ROBOT-MODE.md for lanes, transports, and deprecations.
• Marching orders: copy-paste dispatch prompts (first dispatch, steady-state, wide-swarm domain assignment, review dispatch) live in PROMPTS.md; the fill-in template is assets/marching-orders-template.md. Keep one constraint live: first dispatch claims one scoped item and reserves files/worktree scope; steady-state asks for one commit or one explicit blocker per timebox.
• Isolation: default is Agent Mail file reservations + clear bead ownership; --worktrees when repo policy allows. Repo-local AGENTS.md always wins.
• Pending-input etiquette: a pane's input box may hold staged-but-unsent text. Submit it ONLY if it is clearly an orchestrator-staged directive awaiting send. NEVER submit text that reads as the human's live composition — when in doubt, leave it and log the observation.
• Destructive dialogs: decline by default; require explicit justification to accept.

Cadence And Stop Conditions

Tick every 4 minutes during nucleation, 10–17 minutes in steady state, 30 minutes when panes are deep in real work. Stop tending when the convergence triple-check holds across repeated observations. If the queue is dry, do not manufacture tasks — report state, blockers, degraded sources, and residual risk. If coordination surfaces degrade, continue with explicit bead ownership and backfill mail later.

Review-Only Mode

When the swarm audits rather than implements, flip panes to review-only: no Agent Mail registration, no bead claims, read recent diffs, tag findings by severity, rotate reviewers through study → fresh-eyes → cross-review → continuation passes, kill+relaunch between rounds. Full spec (phase cycle, mixed-swarm ratios, kill-relaunch rhythm, quality rubric, mode-switch prompts) is in REVIEW-MODE.md; Gemini-specific tuning is the code-review-gemini-swarm-with-ntm skill.

Metrics You Report

At closeout, summarize the swarm in concrete deltas:

Checklist Before Ending A Tending Turn

• No command sessions needed for the user's request are still running.
• Active panes are working, blocked with a named blocker, or intentionally stopped.
• br ready / ntm work queue-dry state is known.
• Reservations and Agent Mail state are not silently stale.
• Any restart/force-release/shutdown has evidence and a verification result.
• Closeout names commits/beads/tests/blockers, not "the swarm seems fine."

When To Use / When To Skip

Use when: you are the orchestrator of an NTM session with ≥2 panes — unsticking panes, rotating accounts, dispatching marching orders, switching implement↔review modes, judging convergence, or diagnosing cross-session contention.

Skip when: you just need the ntm command catalog ($ntm), single-agent one-pane work, new-machine provisioning (provision-new-machine), Beads DB repair (fixing-beads-problems), or Gemini review tuning (code-review-gemini-swarm-with-ntm). MCP Agent Mail primitives are $agent-mail; bead mechanics are $beads-br; BV triage is $beads-bv; account management is $account-rotation.

Degrees of freedom: this is a medium-freedom methodology skill — prefer the specific OC/AP card when evidence fires its trigger; fall back to the decision tree otherwise. Following steps without the evidence their triggers require is worse than skipping the card.

Reference Index

Related Skills

Tri-vendor (Opus + Codex + AGY): worker-only three-pane shape and AGY observability caveats live in $dual-pane-atm (§ Tri-vendor) — reach there before tending an AGY lane.

This skill stays focused on swarm tending decisions: observe, classify, act once, verify, and stop when the evidence says stop.