Menú
Read GitHub for live PR state, classify each open authored PR, and deliver one message per PR to the matching downstream actor's inbox (qa, investigate, drip, retro). OTP-style routing — pr-state is the dispatcher.
Instalar con Codex o Claude Copia este prompt, pégalo en Codex, Claude u otro asistente, y deja que revise la página de la skill y la instale por ti.
Basado en la clasificación ocupacional SOC
Filter recent Hacker News for live threads on June's topics and hand him the list. Separately, flag the rare subset where work he has already done settles a specific quoted claim, and draft that comment as a plain argument with no link to his own writing. Use when June asks to check HN or find threads worth reading or commenting on.
Scan blog posts for AI writing tics and missed opportunities for human voice. Reports subtractions (AI patterns to remove) and additions (wordplay, arc, and claims to strengthen).
Poll June's Hacker News feeds, report what is new since the last run, and idempotently arm a background watch for replies to his comments. State persists in ~/Documents/sweep/ so it survives across sessions. Use when June asks what is new on HN, whether anyone replied, or to start watching a thread.
Argument-level compression by a strengthened absence test (does the argument stand with the same force, clarity, evidence, scope, pacing, and misreading-resistance?). Removes passages that fail it (self-referential cruft, cross-section restatement, self-recap, re-derivation, scaffolding, expired reader-orientation, stakes inflation, inert hedges, redundant examples); grafts good rhetoric onto an adjacent essential point rather than deleting. Idempotent. The structural cousin of /tighten (word-level) and /not-but (negation).
Audit a technical candidate's resume/CV by verifying its claims against the public contribution graph instead of taking the text at face value. Sorts every claim onto the self/peer/world attestation lattice, applies a cost-and-stake filter so cheap proxies (stars, downloads, backlinks) don't count, runs LIVE checks (GitHub, arXiv, package registries, CVEs), and reports two separate axes — merit (staked acceptance of the candidate's specific work) and attention (timestamped output) — plus a verifiability score in [0,1]. For technical recruiters and hiring managers. Invoke when given a candidate — a GitHub handle, a resume/CV, an arXiv author, a personal site — and asked to verify, audit, vet, or check their claims. Runs at two depths — a quick triage scan or a deep per-claim audit — and elicits which. Checks the provenance of a personal site or self-hosted resume (Internet Archive + git history) as an un-backdatable anti-tailoring anchor. Needs tool access (Bash/curl or WebFetch).
Find and defuse "not X but Y" / "isn't X; it's Y" / "X — not Y" constructions. Deterministic grep for detection (no LLM blind spot), the not-but subagent for triage. Split-or-cut tail negations ("X, not Y") aggressively — promote the point to its own sentence or delete it; recast in place only mid-clause "not X but Y" that genuinely can't be split; keep just the rare iconic thesis line literal.
Read every open PR you've authored from GitHub, classify each into the next-action bucket, and write one message per PR to the matching downstream actor's inbox.
Each skill in the pipeline is an actor with a mailbox. The mailbox is an append-only JSONL file at ~/.sweep/inbox/<actor>.jsonl. The actor reads its mailbox on each takt and processes one message (WIP=1).
pr-state is the dispatcher between GitHub (the world) and the pipeline (the actors). It has no inbox of its own — its input is gh. Its output is N inboxes.
GitHub (gh)
|
v
[ pr-state ] --classify--> [ qa.inbox ] --> /qa takt
\-> [ investigate.inbox ] --> /investigate takt
\-> [ drip.inbox ] --> /drip takt
\-> [ retro.inbox ] --> /retro batch (no-op messages go here for the audit trail)
The actor never reaches into another actor's state. Skills communicate only by appending to mailboxes.
| Input | Output | Valid alone? |
|---|---|---|
gh open-PR set (live) + optional drip-queue context | ~/.sweep/pr-state.jsonl snapshot per PR + ~/.sweep/PR_STATE.md rendered punch list | Yes — read-only classifier, no state mutated upstream |
Identity: if no open PRs match the filter, write a snapshot saying so and exit. No-op on empty input.
Composition: pr-state(pr-state(x)) == pr-state(x) for the same live gh state. The classifier is a pure function of (gh response, drip-queue state, clock). Repeat invocations within the same minute produce identical classifications.
Preconditions: gh auth status passes. ~/.sweep/drip-queue/ may or may not exist — pr-state degrades gracefully without it.
Each PR lands in exactly one bucket. Order matters: the first matching rule wins.
| Bucket | Trigger | Next action |
|---|---|---|
| close | Competing PR merged (superseded), or maintainer review state is CHANGES_REQUESTED with explicit close-this-PR language in the latest review body, or our drip-queue entry has status: "superseded" | Human reviews, decides to close. Never auto-closes (see rule below). |
| investigate | reviewDecision == "CHANGES_REQUESTED", or maintainer left a comment with a question (? in body, authorAssociation is MEMBER/OWNER/COLLABORATOR) and we haven't replied since | Spawn /investigate or address manually. |
| rebase | mergeable == "CONFLICTING", or PR base branch advanced > 50 commits since PR head's merge-base | git pull --rebase, force-push, re-trigger CI. |
| qa | Any CI check has conclusion == "FAILURE", or our gate attestation is missing/older than 24 h while PR is still open | Re-run /qa on the entry. |
| ship | reviewDecision == "APPROVED" + mergeable == "MERGEABLE" + all CI checks SUCCESS + no open conversations | Maintainer can merge. Optional ping after 48 h of inactivity. |
| wait | None of the above — recently opened (< 24 h), no maintainer activity yet, CI green or pending | No action. The world hasn't responded yet. |
Tie-breakers between buckets (in priority order, stop at first match):
If the bucket is ambiguous (e.g., approved but CI failing), the higher-priority bucket wins. CI failing on an approved PR → qa, not ship.
--repo owner/repo — classify only this repo's open PRs.--from-queue — classify only PRs referenced in ~/.sweep/drip-queue/*.jsonl with status: "open". Use when you only care about the pipeline's PRs, not manual ones.--limit N — cap the number of PRs fetched (default 50, max 200).--bucket NAME — after classification, print only PRs in this bucket. --bucket close is the common one.Each bucket has a single destination. pr-state appends one message per classified PR to the destination mailbox. Bucket and inbox are not the same axis — close and ship both route to drip because drip is the actor that handles "advance an open PR's status."
| Bucket | Destination inbox | Message intent |
|---|---|---|
| close | ~/.sweep/inbox/drip.jsonl | "this PR should be closed — competing PR merged or maintainer asked" |
| investigate | ~/.sweep/inbox/investigate.jsonl | "a reviewer is asking a question — read it, decide a response" |
| rebase | ~/.sweep/inbox/drip.jsonl | "this PR needs a rebase against base before anything else can happen" |
| qa | ~/.sweep/inbox/qa.jsonl | "gates are stale or CI failed — re-attest" |
| ship | ~/.sweep/inbox/drip.jsonl | "approved + green — maintainer can merge; optional ping after 48 h" |
| wait | ~/.sweep/inbox/retro.jsonl | "no action; audit trail only" |
The intent field in each message names what the receiver should do:
# ~/.sweep/inbox/qa.jsonl (appended by pr-state)
{"ts":"2026-05-14T17:45:00Z","from":"pr-state","intent":"reattest","repo":"sharkdp/bat","pr":3734,"branch":"zsh-color-never","signals":{"ci":"failing","check":"Test stable on windows-latest"},"reason":"CI failure on windows","msg_id":"prstate-2026-05-14T17:45:00Z-sharkdp-bat-3734"}
# ~/.sweep/inbox/drip.jsonl
{"ts":"...","from":"pr-state","intent":"close","repo":"hyperium/hyper","pr":4002,"reason":"superseded by #4067","msg_id":"..."}
{"ts":"...","from":"pr-state","intent":"ship","repo":"hyperium/hyper","pr":4068,"reason":"approved + green CI","msg_id":"..."}
# ~/.sweep/inbox/investigate.jsonl
{"ts":"...","from":"pr-state","intent":"respond","repo":"...","pr":...,"reviewer":"...","comment_url":"...","reason":"maintainer asked a question","msg_id":"..."}
Mailbox contract:
msg_id; senders never edit.msg_id is <sender>-<ts>-<repo-slug>-<pr> — deterministic, deduplicates re-sends.from names the sender actor. The receiver uses it to route audit/retro records.intent is the action verb. Each receiver advertises which intents it consumes (e.g., qa consumes reattest).Idempotence: if pr-state runs twice in the same minute, the second run produces messages with the same msg_id (the ts is rounded to minute, not second, for snapshot stability). Receivers MUST dedupe on msg_id — first delivery wins, subsequent ones are no-ops. This is the OTP "at-least-once delivery, idempotent receiver" pattern.
Receiver ack (optional, recommended): when an actor processes a message, it appends {"ts":"...","from":"qa","intent":"ack","msg_id":"...","outcome":"..."} to ~/.sweep/inbox/_acks.jsonl. pr-state and retro read acks for backpressure and post-mortem. No ack = receiver is down or the message is unprocessable; surface to human.
In addition to delivering to inboxes, pr-state writes one snapshot for humans:
~/.sweep/PR_STATE.md — rendered from the same classifications. Rebuilt every run.
# PR State — 2026-05-14 17:45 UTC
## close (1) → drip
- hyperium/hyper#4002 — superseded by #4067 (merged 2026-05-13)
## investigate (3) → investigate
- sharkdp/bat#3734 — windows test failing on edge case "color=never with explicit decorations"
- ...
## qa (2) → qa
## rebase (0) → drip
## ship (5) → drip
## wait (11) → retro (audit only)
The snapshot is a courtesy render. The mailboxes are the contract. If the snapshot is stale, the mailboxes are still authoritative.
gh auth status — fail fast.gh search prs --author "$(gh api user --jq .login)" --state open --limit <limit> --json repository,number,title,url,createdAt,updatedAt,author,headRefName,baseRefName--repo: gh pr list --repo <repo> --author "@me" --state open --json ...--from-queue: read ~/.sweep/drip-queue/*.jsonl, take entries with status: "open" (latest per branch), use (repo, pr_number) pairs.For each PR, call once:
gh pr view <pr> --repo <repo> --json state,mergeable,reviewDecision,reviews,statusCheckRollup,isDraft,comments,headRefOid,baseRefOid,updatedAt
Batch concurrently up to 5 at a time. gh rate-limits get triggered around 30 req/min; 5-way parallelism with ~3 PRs/min average is safe.
If a gh pr view fails (transient), skip the PR, log to ~/.sweep/pr-state-errors.log, continue. Don't abort the run.
For each PR, evaluate the bucket rules in priority order. Stop at the first match.
Signals to extract:
review: reviewDecision (APPROVED / CHANGES_REQUESTED / REVIEW_REQUIRED / null)mergeable: MERGEABLE / CONFLICTING / UNKNOWNci: derived from statusCheckRollup. failing if any conclusion == "FAILURE", pending if any not yet COMPLETED, green if all SUCCESS, unknown otherwise.activity_h: hours since updatedAt.maintainer_question: did a MEMBER/OWNER/COLLABORATOR comment after our latest commit, and does the comment body contain ?? Boolean.competing_pr: optional. Skip in v1 — needs a separate gh search call per PR. Add later if close bucket under-fires.gate_stale_h: optional. From ~/.sweep/gates/<owner>-<repo>.gate mtime. Skip if file missing.Rule application: plain if/elif chain matching the priority table above. Record the matching rule's identifier in reason so the audit log is debuggable.
For each classified PR, build a message with msg_id = pr-state-<minute-rounded-ts>-<owner>-<repo>-<pr>. Append it to the inbox named by the bucket → inbox table. --dry-run prints the messages to stdout instead.
Ensure ~/.sweep/inbox/ exists. Each inbox file (qa.jsonl, investigate.jsonl, drip.jsonl, retro.jsonl) is created on first delivery.
Rebuild ~/.sweep/PR_STATE.md from the same classifications. Group by bucket → inbox, sort within a bucket by activity_h ascending. The snapshot is for humans; downstream actors read the inboxes.
To stdout, print a one-line summary per bucket with counts and destination inbox. If --bucket NAME is set, print the full filtered list.
/qa or /investigate. It classifies; humans and other skills act.gh response + same drip-queue state → same classification. Snapshots are timestamped so the jsonl grows, but reading the latest-per-PR gives a stable view.(repo, pr).gh auth status before any other work.--monitor tick can fire /pr-state as its read pass. The classifier output becomes the work queue for the next pipeline tick (qa bucket → /qa, investigate bucket → /investigate, etc.).gh (the world)
|
v
[ pr-state ] --intent=reattest--> [ qa.inbox ] --> /qa takt (WIP=1)
--intent=respond--> [ investigate.inbox ] --> /investigate
--intent=close----> [ drip.inbox ] --> /drip
--intent=rebase--> [ drip.inbox ] --> /drip
--intent=ship----> [ drip.inbox ] --> /drip
--intent=audit--> [ retro.inbox ] --> /retro batch
Each downstream actor consumes intents from its inbox:
/qa consumes intent: "reattest" — re-run gates on a PR whose CI failed or whose attestation is stale./investigate consumes intent: "respond" — read the reviewer's comment, decide a response./drip consumes intent: "close" | "rebase" | "ship" — advance the PR's status (close, force-push rebase, or no-op while waiting for merge)./retro consumes intent: "audit" — append the wait-bucket entry to the rolling audit trail. No action taken; only history kept.Acks flow back to ~/.sweep/inbox/_acks.jsonl. pr-state's next tick reads acks to detect backpressure (e.g., qa.inbox unread for >30 min → surface to human).
gh search returns nothing: write an empty snapshot, render an empty PR_STATE.md, exit 0.gh pr view rate-limited: lower concurrency, retry once with 30 s backoff. If still failing, log and skip.ci = "unknown". Treat as green for bucket purposes — absence of failure isn't failure.gh data alone. The drip queue is a hint, not the source of truth.statusCheckRollup is null on a brand-new PR: ci = "pending". Bucket: wait.mergeable == "UNKNOWN": GitHub hasn't computed it yet. Treat as MERGEABLE for classification (don't promote to rebase on uncertainty).Any actor that wants work routed to it can declare its inbox path and the intents it consumes. pr-state delivers; the actor processes on its own takt. The mailbox JSONL schema is stable:
{"ts":"...","from":"<sender>","intent":"<verb>","msg_id":"<deterministic>","repo":"...","pr":...,"...":"..."}
A one-shot /pr-state --bucket close --dry-run prints the messages it would deliver — useful for sanity-checking the close bucket without firing actual notifications.