repokernel

name	repokernel
description	Operate a RepoKernel-governed repo. Six verbs (plan, status, next, run, review, doctor) map to slash commands that drive the rk CLI. Lifecycle order — plan first, doctor only on drift.
version	1.16.0

RepoKernel Operator

Six verbs. Each verb is a slash command.

Intent	Slash
"plan an epic", "new feature", "scope this work"	/rk-plan
"status", "where are we"	/rk-status
"what's next"	/rk-next
"run this", "ship it", "fix bug X"	/rk-run
"review"	/rk-review
"doctor", "what's broken"	/rk-doctor

For one-line CLI lookups: reference/cheatsheet.md.

Cold-start check

Before any verb other than /rk-status or /rk-doctor, run rk status --brief --json once. If initialized: false, halt and tell the user:

Repo not initialized. Run rk init --commit (or rk init --example --commit to scaffold a starter epic), then re-invoke.

Do not proceed past this check until init exists. /rk-status and /rk-doctor are safe to run uninitialized — both handle the case explicitly.

Operational preflight

Once per session — before the first verb that dispatches work — run rk preflight --json (or rk team status --json until preflight ships). Surface any non-empty operational.collection_errors, operational.live_claims, or operational.leaked_worktrees. The preflight is session-scoped: do not re-run it per command. If the user reports state drift mid-session, run it again. Plugin commands (/rk-next, /rk-run, /rk-review) trust this single preflight; they do not invoke rk team status themselves.

Tier routing

rk route <ID> --json returns routing_hint.tier (light / standard / heavy). Map to your harness:

light    → cheap reasoning model
standard → default coding model
heavy    → strongest reasoning model

If routing_hint.fanout is present, dispatch one agent per entry in parallel (single message, multiple tool calls). If reason: "pinned", do not override.

Stop conditions

• rk status --brief --json returns initialized: false → stop, surface init guidance.
• rk validate --fail-on P0,P1 exits non-zero → route to /rk-doctor.
• rk next returns blocked → surface the reason.
• A run reaches merge_conflict / agent_failed / path_violation → run rk run inspect <RUN_ID>, surface to user.
• rk doctor surfaces operational corruption → run rk recover --preview then rk recover --apply.

Tracker bridge

When /rk-plan runs against a JIRA / Linear / GitHub Issues ticket, pass --from-tracker <source>:<ref> to rk create epic to seed title + body from the ticket and record extras.tracker_*. When /rk-run is a one-shot from a ticket, pass --from-tracker <source>:<ref> directly to rk run. Tracker ingest fails closed on offline / 401 / 404 before writing state. Use --allow-tracker-fallback only after explicit user approval. See reference/cheatsheet.md for forms.

Custom branch naming

worktrees.branchPattern is compatibility shorthand: without {sprintId} it applies to epic branches only; with {sprintId} it applies to sprint branches only. Prefer explicit worktrees.epicBranchPattern + worktrees.sprintBranchPattern for custom naming. Tokens: {branchPrefix}, {epicId}, {sprintId}. Skill should not inspect or override pattern — config-driven, no skill-level branching.