Check Open PRs

> Usage: /check-open-prs

All mechanical data collection lives in a deterministic script — never loop gh pr view / gh api / Slack calls per PR yourself. The model's job is only the judgment layer on top: Slack thread state for ambiguous PRs, what needs the user's attention, and suggested next actions.

The script is also directly usable from a terminal for zero tokens:

bash
bunx tsx script/utils/check-open-prs.ts          # human-readable dashboard
bunx tsx script/utils/check-open-prs.ts --json   # compact JSON (what this skill consumes)
bunx tsx script/utils/check-open-prs.ts --quick  # own PRs only, skips incoming queue

GitHub scope is env-overridable (LI.FI defaults baked in), so the script is org-agnostic without code edits:

• PR_DASH_ORGS — comma list of owners to search for your own PRs (default lifinance,lifinance-tron).
• PR_DASH_INCOMING_REPOS — comma list of owner/repo whose PRs form your review queue (default lifinance/contracts,lifinance/contracts-tron).

Phase 1 — collect (ONE Bash call)

bash
bunx tsx script/utils/check-open-prs.ts --json

(Add --quick for quick mode.) The script returns, per PR: repo/number/title/url, kind (own/incoming), bucket, draft, CI rollup + failing checks tagged review/audit vs restartable vs core-dev-gate, conflicts, reviewDecision, mergeStateStatus, last commit, last non-author human comment, my last review (myReview: {state, at, sha} — sha is the commit my review was submitted against, used by Phase 2D), and slackCheck: true where Slack thread state is needed to finish classification. Archived-repo PRs and stale incoming PRs are already excluded (listed under excluded).

Refresh mode ("refresh the dashboard"): just re-run the script — it's cheap. Never re-render from data fetched earlier in the conversation; a PR may have been replied to, merged, or re-pinged since. Stamp every dashboard with the script's asOf (convert to +07).

Bucket vocabulary (script native → dashboard display)

The script assigns each PR exactly one native bucket (the bucket field). Rows flagged slackCheck: true are _provisional_ until Phase 1.5 resolves them into the display sections rendered in Phase 1.6; every other bucket maps 1:1 to its section with no enrichment. If Phase 1.5 is skipped, a slackCheck PR renders under its raw native bucket name.

| Script native bucket | Needs Phase 1.5 (Slack)? | Phase 1.6 display section | | -------------------------------------------------------- | ------------------------------ | ----------------------------------------------------------- | | WAITING-ON-TEAM | yes — 48h cooldown split | 🔔 REMIND-DUE (≥48h) / ⏳ REMIND-RECENTLY-PINGED (<48h) | | WAITING-ON-ME | yes — confirm ball-in-court | 👀 YOUR-ACTION | | CI-RED | no | 🔴 CI-RED | | CONFLICTS | no | ⚔️ CONFLICTS | | APPROVED-BLOCKED | no | 🚧 APPROVED-BLOCKED | | READY-TO-MERGE | no | ✅ READY-TO-MERGE | | OWN-DRAFT/{READY-TO-FLIP,NEEDS-WORK,DORMANT,SYNC-PR} | no | 📝 OWN-DRAFTS | | STALE | no | 🧹 STALE | | INBOX-UNREVIEWED | no | 📥 INCOMING-UNREVIEWED | | MAYBE-REREVIEW | yes — confirm author hand-back | 🔁 INCOMING-REREVIEW (else suppressed) | | WAITING-ON-AUTHOR / WAITING-ON-OTHERS / DONE-BY-ME | no | suppressed counts (Phase 1.6 §6) |

--quick filters at the source — the script omits the incoming queue entirely (INBOX-UNREVIEWED / MAYBE-REREVIEW / WAITING-ON-AUTHOR / WAITING-ON-OTHERS / DONE-BY-ME are never collected), and the skill then renders only display groups 1–2. It is not post-filtering of the full JSON.

Phase 1.5 — Slack cross-reference (ONLY slackCheck: true PRs)

Skip this entirely for PRs without the flag. For the flagged set:

1. Locate threads. One scrape of #dev-sc-review (C088UJWC8PR) history (last 6 weeks) serves all contracts-repo PRs — match by PR URL. Fall back to slack_search_public (in:#dev-sc-review pull/<NUMBER>) only for PRs not in the scraped window. Only assert "not posted for review yet" when the scan window provably covers the PR (its created is inside the 6-week scrape and the fallback search returned nothing) — otherwise the absence is unconfirmed, so emit no Slack note rather than a false negative. Never word a note as "Posted today" (ambiguous with the Slack action): for a freshly-opened PR say "Created today", reserving "posted" for a confirmed #dev-sc-review message.
2. Read each found thread with slack_read_thread (response_format="detailed"), batched 5–8 per message in parallel. Paginate until reply_count is exhausted — the missed reply is usually the newest, which decides the bucket.
3. Classify own PRs mechanically by last human (non-bot) message author — content never overrides this rule:

• Last author is the user → REMIND: <48h since that message → REMIND-RECENTLY-PINGED (cooldown), ≥48h → REMIND-DUE. Zero replies → apply the 48h rule to the parent post.
• Last author is anyone else → YOUR-ACTION.

1. Incoming MAYBE-REREVIEW PRs become INBOX-REREVIEW only if the PR author posted an explicit hand-back after the user's review (ready for re-review | PTAL | addressed | rebased | all comments addressed, case-insensitive) AND the last human reply isn't the user's. New commits alone are not a signal; a re-ping the user already answered is not a signal.

Hard invariants

1. Never estimate a Slack timestamp — every cooldown figure comes from an actual Message TS of the actual last reply.
2. Search-result context snippets are NOT thread content — they preview only the FIRST replies and silently omit the newest. Discovery only; every classification needs a slack_read_thread from THIS invocation.
3. One audit line per Slack-classified PR — emitted by _this skill layer_ (the script only sets slackCheck: true; it never produces audit lines) as a visible line in the conversation, printed during Phase 1.5 immediately above the Phase 1.6 dashboard, one per PR that carried slackCheck. Mandatory format and fields: Audit: #<pr> last_reply_ts=<ISO Message TS> (by=<me|login>, <Δh> ago) replies=<read>/<reply_count> → <DISPLAY-BUCKET>. Each field must trace to a slack_read_thread result from THIS invocation: last_reply_ts is a real Message TS (never estimated — invariant 1), replies=<read>/<reply_count> proves pagination was exhausted (read must equal the thread's reply_count), and → <DISPLAY-BUCKET> is the resolved Phase 1.6 section. A Slack-classified PR rendered without its audit line — or with read < reply_count — is a bug.

Phase 1.6 — render the dashboard

Header: As of <YYYY-MM-DD HH:MM +07> — full | quick | refresh. Sections in order, tables with PR / title / key timestamps / notes:

1. 🔔 REMIND-DUE · ⏳ REMIND-RECENTLY-PINGED · 👀 YOUR-ACTION (from Phase 1.5)
2. 🔴 CI-RED (flag review/audit-gated vs restartable failures, from script) · ⚔️ CONFLICTS · 🚧 APPROVED-BLOCKED (approved + green but mergeStateStatus != CLEAN; render the script's note verbatim) · ✅ READY-TO-MERGE (only mergeStateStatus == CLEAN — actually mergeable now)
3. 📝 OWN-DRAFTS (script sub-buckets: READY-TO-FLIP / NEEDS-WORK / DORMANT / SYNC-PR)
4. 🧹 STALE
5. 📥 INCOMING-UNREVIEWED · 🔁 INCOMING-REREVIEW
6. Excluded + suppressed counts (WAITING-ON-OTHERS / WAITING-ON-AUTHOR / DONE-BY-ME: N each) for auditability.

Each section's table must be fully closed (its last row followed by a blank line) before the next section heading — otherwise the following separator gets swallowed and the prior table's header bleeds into the next section. Group 6 (Excluded + suppressed counts) is a plain bullet/inline list, never a table.

In quick mode render only group 1–2 and note: "Quick mode — drafts, stale, and incoming queue not scanned."

End with the action menu and STOP — no actions without explicit approval:

text
Proceed?
  (a) Bump REMIND-DUE threads          (e) Open INCOMING-UNREVIEWED for review
  (b) Investigate YOUR-ACTION PRs       (f) Open INCOMING-REREVIEW with diff-since-my-review
  (c) Walk through OWN-DRAFTS           (g) All of the above
  (d) Walk through STALE PRs            (h) Cancel

Phase 2A — bump REMIND-DUE threads

Immediately before posting, re-read each target thread; skip (and say why) if the last message is no longer the user's or is now <48h old. Post one reply per PR via slack_send_message with thread_ts=<parent_ts> — exact text, do not paraphrase:

text
friendly bump <!subteam^S096X6MCB0C>

(The <!subteam^…> syntax is mandatory — plain @smartcontract_core doesn't notify.) S096X6MCB0C is LI.FI's @smartcontract_core subteam ID — org-specific, like the #dev-sc-review channel this skill targets; another org swaps in its own reviewer subteam ID, looked up once via Slack usergroups.list (or the group's "Copy mention" handle). Confirm each post with its permalink.

Phase 2B/2C — walk OWN-DRAFTS / STALE

Per PR, one short prompt with the script's bucket and suggested next step; execute the chosen disposition:

• Drafts: (1) flip ready (gh pr ready <n> + invoke post-pr-for-review) (2) keep working (3) close with reason (4) investigate (Phase 3) (5) skip.
• Stale: (1) close with reason (2) post "Still active — target: <D>" (3) re-assign (4) convert to draft (gh pr ready --undo) (5) investigate (6) skip.

Phase 2D — incoming queue

• INBOX-UNREVIEWED: show link + 2-line preview; offer open-in-browser / sub-agent triage (≤200 words) / skip.
• INBOX-REREVIEW: fetch diff since my last review (gh pr diff <n> --commit-range <sha>..HEAD, where <sha> is the PR's myReview.sha from the script JSON — the commit my last review was submitted against). If myReview.sha is null (review predates the commit or it was detached), fall back to the full gh pr diff <n>. Quote the dev's hand-back message, list my unresolved inline comments; offer browser / sub-agent verify-comments-addressed / skip.

Never auto-approve, auto-comment, or auto-request-changes — the skill routes; the human reviews.

Phase 3 — sub-agent investigation (on request)

One sub-agent per PR, in parallel, self-contained prompt: PR URL + pasted Slack thread + "pull review state via gh pr view --json reviewDecision,reviews,comments,statusCheckRollup,mergeable + inline comments, identify what the team is blocking on and ONE concrete next step, report <300 words". Consolidate, then offer: implement interactively / in parallel / list only / done.

Phase 4 — implementation (per-PR approval required)

gh pr checkout <n> → apply → lint/test → push → reply in the Slack thread linking the commit.

What this skill never does

• Touches closed or merged PRs; auto-reminds on STALE; re-pings inside the 48h cooldown.
• Loops per-PR gh/Slack calls when the script (or a batch) covers it.
• Pushes code or submits any review without explicit per-PR approval.