github/lever

Turns a task you keep doing by hand into a paused GitHub Copilot Automation with an event trigger, a least-privilege tool list, a PARAMETERS + CONTRACT prompt, a quiet exit and a propose-only improver sibling. It creates automations, so it runs only on /lever automate or an explicit ask ("automate this", "run this on every PR/issue", "do this nightly", "make this run in the cloud"). When the same task has been done 3 or more times by hand, or a local self-wake has fired 5 or more times for the same job, suggest it in one line instead of running it.

Runs long-lived delivery as a coordinator with parallel lanes. Sets authority and pacing dials up front, keeps lanes alive and fed, lands work through a single-writer release train, and reports status only when something changed. Runs only on /lever deliver, "/autonomous-delivery" or an explicit request to run multi-hour, overnight or multi-lane delivery ("run this overnight", "keep N streams going", "don't stop until it ships"), or to re-arm an existing run after a restart.

Resolves how far an agent may take a change, from L0 propose to L5 release. It reads the repo's real GitHub rules (rulesets, branch protection, CODEOWNERS, merge queue) and .github/lever.yml, applies per-path and per-class limits, records every grant verbatim with its source and expiry, and exposes a may_merge check that delivery skills call before landing anything. Use when a delivery skill or automation starts, when creating an automation, before any merge or enqueue, or when the user says "merge", "ship it", "land it", "then merge", "hand off", "do not merge", "freeze" or "you can merge X".

Turns a short intent, often around 20 words, into a brief that can run unattended. The brief has the user's words verbatim, an outcome owner, an observable done-when, a feedback loop, constraints and neighbors, a budget, an autonomy level, an absence window, stop conditions and a report contract. Runs only on /lever brief, when another lever skill delegates, or on an explicit request to write a brief or hand work off ("kick off", "spin up", "hand this to", "while I'm asleep", "I'll be away"). It files issues, so it never runs just because work is happening in parallel.

Reviews a PR, diff, plan or skill with 2–3 models from different providers working blind and independently, and surfaces only the findings they agree on or that survive a contradiction-first proof; a precision mode runs one automated reviewer that posts at most two proven findings and stays silent otherwise. Use when asked to "review this with opus, codex and gemini", "rubber duck this plan", "second opinion", "only show me issues the models agree on", when a plan is about to drive a day or more of work, when work has been stuck for hours, or when setting up review-on-ready automation.

Runs many single-hypothesis "arms" as do-not-merge PRs, measures each with paired, interleaved runs against a cost-matched control, keeps only the arms that clear a keep bar set in advance, records every negative, and composes the winners into a stack that is re-measured on real traffic after merge. It opens PRs and burns CI, so it runs only on /lever arms or an explicit ask to measure ideas ("which of these ideas actually helps?", "is this change worth it?", "try that tip I saw", "run arms on this") for a CI, perf, eval, benchmark or cost hill-climb where the system can be run on demand.

Before iterating on a speedup or quality problem, build the feedback loop first and write it down as a Loop Card that names the outcome, the instrument the agent runs itself, the noise floor, the keep bar and the stop rule. Use on /lever loop, when asked to speed up, optimize, hill-climb or "make X better" against something measurable, and when a loop smell shows up mid-task: the user has pasted two or more screenshots or numbers, "still broken" has come up twice, the agent said "please try it" without running it, someone claims a win from one run, or the human is acting as the instrument. Not for one-shot fixes that existing tests already prove.

Backlog steward that keeps a pool of ready, independently mergeable issues above a target by mining real sources (failing tests, TODOs, gaps, audits, user reports), deduplicating, sizing and labeling them ready. It files issues, so it runs only on /lever backlog, inside a delivery run whose idle policy is refill, on a schedule the user set up, or on an explicit ask ("file more issues", "keep them busy", "fill the backlog", "we have nothing to do").

Turns a set of open PRs into a merge plan. It maps their dependencies, puts the PRs into merge-queue batches with strict order only where they overlap, and writes a hand-off with evidence for each PR. At L2 it gives the plan to the human who will merge; at L4 it enqueues the PRs itself. It can enqueue, so it runs only on /lever land, a hand-off from a lever coordinator, or an explicit ask ("what order should these merge in", "land the stack", "why is landing taking all day").

The one lever command. /lever <verb> [args] runs a lever skill directly (loop, arms, review, brief, deliver, refactor, backlog, shepherd, land, status, automate, health, mirror, steer, autonomy, spend, setup, help). /lever <plain words> picks the verb for you, says which in one line, then runs it. Also reads and writes the user's personal dials in ~/.copilot/lever.yml. Use when the user types /lever, asks "which lever skill should I use?", or wants to see or change their autonomy or spend dial.

Drives an open pull request to landed through an event-driven loop. On each wake it does one batched read, fixes root causes of red required checks, dispositions every review thread, and lands at the level set by the autonomy dial. It pushes and can merge, so it runs only on /lever shepherd, a hand-off from a lever coordinator, a PR labeled for shepherding, or an explicit ask ("monitor CI until green", "push fixes until green", "resolve the review threads", "merge when green", "babysit this PR", "shepherd

Pushes progress on long-running or delegated work so the human never has to poll. It publishes milestones with ETAs and the evidence each needs, sends reports only when something meaningful changed, labels every claim with an evidence level, logs ETA accuracy, watches time since last merge, and audits "done" against evidence before declaring it. Use for any delegated, multi-hour or overnight run, when a brief sets a report contract, or when the user asks "status?", "ETA?", "where are we?", "what happened overnight?", "hour by hour" or "are you done?".

Mine your own local Copilot session history and return a scorecard of how you direct agents, plus three suggestions illustrated with your own before/after messages. It reads your private history, so it runs only on /lever mirror or an explicit ask ("how am I prompting?", "coach me", "mine my sessions", "what do I keep repeating?", a weekly self-retro).

Runs a huge, mechanical refactor (crate or package splits, module extraction, API migrations, dead-code removal) as three independent lanes, producer, packager and shepherd, that coordinate only through git and a PR stack. It pushes branches and opens PRs, so it runs only on /lever refactor or an explicit ask ("split this crate", "break up this monolith", "this module is huge, cut it down", "migrate every caller").

Stall watchdog and automation health review. Pulls recent runs, merges, no-op rate, failures, rate-limit hits and stuck PRs; classifies each stall; proposes exactly one change or "no change"; and takes over orphaned work. Use when unattended delivery stalled ("why did it stall overnight", "no merges in N hours", "it was fast the first few days, now slow"), when an automation fails or its output drops, when rate limits or quotas are hit, or on a schedule while agents run unattended.

Turn a correction you keep repeating into a durable fix: mine your corrections, classify them, choose the cheapest layer that would have prevented them (instruction, skill, lint/CI check, test, automation, or product bug report), open a PR with an evidence block, then re-measure after K days and keep, escalate or revert. Opens PRs, so it runs only on /lever steer or an explicit ask ("mine my corrections", "why do I keep telling you X", "didn't I give you a skill for this"), or when prompt-mirror reports the same correction class in 3+ sessions and you say go.