herdr-goal

name	herdr-goal
description	Operate bounded Herdr goal loops with an explicit Luna verifier, sealed epoch packets, and watcher callbacks. Use when the user names herdr-goal, wants a Ralph loop, or asks for verifier-gated continuation.

Herdr Goal

Run a main Codex worker and a separate read-only Luna verifier as a bounded, durable loop. The main seals an exact epoch packet while working; watcher wakes Luna only after that pinned main session settles, then returns one authenticated verdict to the same main terminal.

Boundary

• herdr-goal owns goal epochs, verifier launch, sealed packets, convergence, stalls, and final completion reconciliation.
• herdr-watcher owns settled wakes, epoch callbacks, terminal pinning, delivery, and unknown write boundaries.
• herdr owns pane creation and agent transport. Apply its targeting rules; never substitute the focused pane.
• goal-protocol owns authoring or changing a durable GOAL.md; this skill executes an already measurable verification loop.

Preflight

1. Require HERDR_ENV=1, nonempty Herdr workspace/tab/pane IDs, and direct invocation from the main agent pane. Reject plugin-action context.
2. Resolve herdr-goal and herdr-watcher from live plugin state or validated paths. Read their live help and watcher capabilities --json.
3. Run herdr integration status; require the Codex integration to be current and verify both panes expose native agent_session.value identity.
4. Require watcher capabilities settled_wake_v1 and epoch_callback_v1.
5. Inspect current goal state before mutation. Never infer an epoch or verdict from pane text.

Route

Reference	Read when
references/start-loop.md	Starting a loop, selecting bounds, or launching the Luna verifier
references/checkpoint-and-report.md	Sealing an epoch, executing a verifier packet, or returning a verdict
references/lifecycle.md	Inspecting, completing, cancelling, recovering, or diagnosing a stalled loop

Read only the selected reference. Read more than one only when the requested operation crosses those lifecycle surfaces.

Workflow

1. Confirm the goal and acceptance criteria are measurable. Do not give Luna a bare goal; produce an exact read-only verification packet.
2. Start with explicit maximum epochs, overall deadline, settle timeout, wake timeout, and verifier timeout. Record the returned goal and verifier IDs.
3. Before each main work turn ends, seal the current epoch with checkpoint. Re-read state and require durable watcher acknowledgements before ending.
4. Let watcher wake Luna after the recorded main session reaches idle or done. Luna executes only the sealed packet and submits one report.
5. On changes-requested, address only failing criteria and checkpoint the next epoch. On pass, let the pinned main reconcile evidence and run complete.
6. Treat timeout, identity loss, blocked/failed verification, exhausted epochs, or unconfirmed delivery as a durable stall, not another automatic attempt.

Hard rules

• Launch the verifier explicitly as Codex gpt-5.6-luna with effort max. Luna executes the exact packet; it never plans, edits source, mutates Git, delegates, or broadens scope.
• Pin main and verifier by terminal ID, agent kind, and native session. Pane IDs are addresses only; never route by focus, layout order, or a reused name.
• Never interrupt an agent in working, blocked, or unknown. A goal settle requires the recorded main to transition from working to idle or done.
• Never place a callback capability in a packet, prompt, report, or log. Pass only its protected invite-file path to the pinned Luna terminal.
• A Luna pass is necessary but insufficient. Only the pinned main may mark the goal complete after validating accepted evidence.
• Never reuse an epoch after an ambiguous prompt write, resend an unconfirmed delivery, hand-repair a journal, or weaken bounds to keep a loop alive.
• Do not cancel or close resources the plugin did not create.

Stop conditions

• Stop if either agent identity is missing, ambiguous, replaced, or no longer matches its pinned native session.
• Stop if live watcher capabilities or CLI syntax do not match the required protocol, or if a packet has ambiguous steps or forbidden side effects.
• Stop after the configured epoch/deadline/error boundary. Report the durable stall rather than inventing recovery.

Output

Report the goal ID, main and verifier terminal IDs, current epoch/state, watcher resource IDs, packet digest, verdict and evidence, delivery state, bounds remaining, resources left active, and the exact next action or stall.