Ntm Skill for Claude

NTM — optional pane adapter

NTM hosts explicit agent roles in persistent panes. It is transport, not an

AgentOps lifecycle controller. The caller chooses the panes, roles, commands,

write scopes, and stopping point.

Robot surfaces work because they report what the pane is doing, not what was

sent to it; a dispatch layer that only proved delivery would let every dead

worker look busy.

Judge pane liveness by the truth-stack, strongest first: new artifacts on

disk, then transcript growth, then robot state and attention flags, then bare

process existence. A successful prompt send sits below all of these and proves

nothing about work.

Named failure mode — kill-the-witness: restarting a stuck pane before

capturing its state, destroying the only evidence of why it stalled.

Anti-pattern: restarting an unresponsive pane as the default remedy.

Corrective: rescue before restart — snapshot robot state and transcript,

attempt a nudge, and restart only when the truth-stack shows no liveness at

any level.

Boundary

• Never start or probe NTM merely because it is installed.
• Discover the live command contract with ntm --help,

ntm --robot-capabilities, and ntm --robot-snapshot before unfamiliar

actions.

• Dispatch each caller-supplied command once. NTM does not select work, retry a

failed command, validate a candidate, integrate changes, or decide what runs

next.

• Pane roles are descriptive. They grant no ownership, admission, Git, release,

or delivery authority.

• Concurrent writers require caller-supplied disjoint scopes and whatever

isolation the repository requires. NTM does not infer safe concurrency.

• Safety, lock, mail, and pipeline features remain substrate facts. Their state

cannot change an RPI phase result or semantic verdict.

One-shot dispatch

• Record the explicit session, pane, role, working directory, and command.
• Inspect robot capabilities and current snapshot.
• Create or select only the named pane.
• Send the command once.
• Observe robot state, attention, and transcript until the command exits or the

caller's observation window ends.

• Return the factual result and stop.

For a software-factory layout, the caller may name producer, tester, validator,

or integrator panes. The same identity rule still applies: a validator for a

candidate must have a distinct context identity from its author. Merely placing

two roles in different panes is a declared runtime fact, not proof of semantic

independence. Mixed-model judgment panes follow

the agent-native model-dispatch recipe (probe, disclose, never

claude -p).

Output

Return:

• session and pane identifiers;
• role and exact command;
• start and observation timestamps;
• exit status when known;
• robot state and relevant transcript references;
• degraded or unavailable substrate surfaces;
• effects that were and were not observed.

Terminal outcomes are explicit, never a silent hang:

• NTM unavailable — ntm absent or the robot surface unreachable: report it

and stop; do not fall back to blind key injection or assume the pane is idle.

• Observation-window deadline reached — report the last observed robot state

and transcript reference; the window ending is a stop, not a failure verdict.

• Timeout / nonzero exit / degraded source — reported as evidence with what

was and was not observed.

• Cancellation and cleanup — the caller decides whether to invoke another

experiment; report which panes were created or left running so nothing is

orphaned silently. NTM never retries or reaps on its own.

Useful live surfaces

Prefer machine-readable robot surfaces for capability, snapshot, attention,

tail, and pipeline observations. Use interactive key injection only when the

caller explicitly requests an interactive action and no robot command provides

the needed behavior.

External NTM documentation and examples remain the authority for command syntax;

this skill owns only the AgentOps boundary above. NTM's CLI drifts across

versions, so confirm the surface at runtime (ntm --version,

ntm --robot-capabilities) rather than trusting a remembered command.