Configure AI implementer lanes for delegated coding work

Runs a discover-propose-approve-write flow that configures which implementer CLI (Claude, Codex, Cline, Grok, OpenCode, and 12 others) handles which kind of delegated work, for the whole family of companion *-delegate skills. It never dispatches coding work itself - only discovers CLIs and writes the lane map. Discovery (node discover.mjs) reports which implementers are installed, their auth status (true/false/null for unknown - notably agy and pi expose no status command, so null there means no probe is wired, not a failed login), and whether their model list was actually reported, curated as aliases, unsupported, or failed to load; an optional --usage pass adds session counts and last-used dates read from local session-file metadata only, never conversation content. Before proposing anything it asks one grounding question with three options - quick defaults (the orchestrator's own labeled opinion), an interview (four allocation-policy questions: what work gets delegated most, which subscriptions to burn versus spare, which CLI is already trusted or has burned the user, and fast-cheap versus slow-thorough), or a usage scan letting session numbers place the lanes - and every dial in interview or usage-scan mode is gated: set only from the user's actual answer or a schema requirement, never invented, since omitting a dial is always safe (the CLI's or relay's own configured default applies).

Every proposal shows a human-readable lane table with a mandatory Basis column (your answer / usage data / repo / my opinion / schema requirement) and the complete delegate-fleet.v1 JSON before any write, and re-shows both after every tweak; a write happens only on an explicit "yes"/"approve"/"write it", to whichever scope the user chose - global ($XDG_CONFIG_HOME/delegate-skills/config.json or ~/.config/delegate-skills/config.json) or per-project (<git-root>/.delegate/config.json), built from that scope's own raw file rather than the merged effective view, so a project write can never silently shadow a global-only lane or vice versa. A project lane only takes effect once its exact content matches an approval hash recorded at write time; any later edit or a cloned config fails closed until it is reviewed and re-approved through this same skill. Delegation economics guide the proposal itself: bounded, objectively-gated work (tests, mechanical refactors) is steered toward capable, authenticated, low-usage CLIs to keep spend off the subscriptions the user is protecting, while work whose output IS the product (debate, architecture, research) or that touches security/concurrency/migrations is bound to stronger implementers instead, since review limits damage but cannot manufacture a good first attempt.

Use it to configure which implementer CLI handles which kind of work (fleet lanes), or to discover installed implementers and write lane config after user approval. Do NOT use it when the user wants a task actually implemented - use the matching *-delegate skill instead - or for a one-off model change on a single dispatch, which just needs --model/--effort/--variant passed directly to that relay rather than a persisted lane.

Input is the user's answers to the grounding menu (or "quick defaults") plus what discover.mjs finds installed. Output is a delegate-fleet.v1 JSON document, {"version": "delegate-fleet.v1", "lanes": {"<name>": {"implementer": "<key>", ...dials}}}, written to disk only after explicit approval via:

node "<skill-dir>/scripts/config.mjs" validate "<lanes-json>"
node "<skill-dir>/scripts/config.mjs" write --scope global "<lanes-json>"

Each lane requires an implementer key from a fixed registry (claude, cline, codex, commandcode, opencode, agy, grok, kimi, qoder, vibe, cursor, pi, omp, aider, copilot, warp, zcode), and only the dials that implementer's own *-delegate relay supports - OpenCode uses variant instead of effort and requires a provider/model string; Oh My Pi's effort dial maps to its own --thinking levels; ZCode's permissionMode accepts only plan/yolo headless, since build/edit block every write tool with no permission client to answer them. Later, a *-delegate relay resolves --lane <name> through node lane.mjs resolve, which fails loudly on a missing lane, an untrusted project config, or an implementer mismatch.

Requires only Node 18+ - no implementer CLI is required to run the skill itself, since it discovers what is actually available. Produces the configuration every companion *-delegate skill (claude-delegate, codex-delegate, cline-delegate, commandcode-delegate, grok-delegate, kimi-delegate, omp-delegate, pi-delegate, aider-delegate, vibe-delegate, warp-delegate, and others) reads via --lane, and never edits a repo's own AGENTS.md, CLAUDE.md, or other agent-instruction files.

Orchestrating agents and the humans directing them who want a deliberate, evidence-labeled map of which CLI implementer handles which class of delegated coding work, instead of guessing model choice per dispatch or letting an orchestrator invent dials on the user's quota. Adapted from the community amElnagdy/delegate-skills repository under the MIT license; this catalog entry is a docs-only import, since the executable scripts/ helpers are not bundled here.