gh-queue
GitHub管理Breeze仓库GitHub待办队列,自动识别并处理PR、Discussion、Issue及安全建议。支持状态持久化、贡献者画像及自动化审查流程(如草稿跳过、外部PR威胁筛查)。
触发场景
安装
npx skills add LanternOps/breeze --skill gh-queue -g -y
SKILL.md
Frontmatter
{
"name": "gh-queue",
"description": "Use when reviewing, triaging, or managing the incoming GitHub backlog on the Breeze repo — PRs, Discussions, Issues, AND Security Advisories. Stateful manager with a shared queue file, contributor profiles, parallel pr-review-toolkit dispatch, GraphQL discussion recipes, comment-style conventions, and session-log bookkeeping. Auto-detects object type from the request; bare invocation triages all four. Triggers on \"review PR\", \"PR queue\", \"triage PRs\/discussions\/issues\", \"security advisories\", \"GHSA\", \"vulnerability reports\", \"review discussions\", \"check open issues\", \"what's waiting on me\", \"gh queue\", \"repo queue\", \"review"
}
Breeze GH Queue Manager
Overview
Long-running manager for the incoming GitHub backlog on the Breeze repo (LanternOps/breeze) — across four object types: Pull Requests, Discussions, Issues, and Security Advisories (community + Billy + any agent-authored; advisories = private vulnerability reports and our own repository advisories). State persists across sessions in two co-located files:
queue.md— live snapshot, split into## PRs,## Discussions,## Issuessections (each with waiting-on-me / in-flight / recently-closed), plus one rolling session log at the bottom.contributors.md— per-contributor profiles and review heuristics (apply across all object types — Billy's patterns show up in PRs and discussions).
queue.md and contributors.md are gitignored (.gitignore lists them by path; SKILL.md and scripts/ ARE tracked), so the state files are a private scratchpad that travels with the repo on this machine but doesn't leak into commits. Never put advisory content or contributor verdicts in SKILL.md.
Outgoing items (mine) are out of scope except to track own-PRs for merge. This skill manages what needs my review/action.
Object-type selection (auto-detect, all-default)
Pick scope from the request — no explicit flag required:
- "review PR #5" / "PR queue" / "triage PRs" → PRs only
- "triage discussions" / "review discussions" / "discussion #968" → Discussions only
- "any new issues" / "triage issues" / "issue #710" → Issues only
- "security advisories" / "any vuln reports" / "GHSA-…" → Advisories only
- bare "what's waiting on me" / "/pr" / "triage the backlog" / "gh queue" → all four
When ambiguous, default to all four but lead with whatever the user named.
When to invoke
- "What's waiting on me?" / "What's in the queue?" / "Triage the backlog"
- "Review PR #NNN" / "Triage discussions" / "Any new issues?"
- After a PR merges / a discussion is answered / an issue is resolved — to update
queue.md - When a new contributor shows up — to start a profile in
contributors.md - At the start of a session if the user asks about repo state
Workflow
1. Re-orient
Read queue.md first to load current state. Then list live state for the in-scope type(s):
# PRs
gh pr list --repo LanternOps/breeze --state open --limit 50 \
--json number,title,author,isDraft,updatedAt,reviewDecision
# Discussions (no gh subcommand — GraphQL; one query gets last-comment author+date)
gh api graphql -f query='
{ repository(owner:"LanternOps", name:"breeze") {
discussions(first:50, orderBy:{field:UPDATED_AT, direction:DESC}) {
totalCount
nodes { number title category{name} isAnswered closed updatedAt
author{login} comments(last:1){ totalCount nodes{ author{login} createdAt } } }
} } }'
# Issues
gh issue list --repo LanternOps/breeze --state open --limit 50 \
--json number,title,author,labels,updatedAt,comments
Cross-check against queue.md. Add anything new, mark anything closed/merged, flag drift.
Never quote the bare open-issue count as a backlog number. As of 2026-08-16 the 248 open issues are ~112 enhancement (roadmap you filed for yourself) against ~136 defects; the total moves with how fast you file features, not with how fast you fix bugs. Every open issue now carries a label, so use the views rather than the total:
# The defect list — the number that should actually trend down
gh issue list --repo LanternOps/breeze --state open --limit 200 \
--search 'label:bug,tech-debt,ci-red'
# Roadmap — feature work; triage it, don't count it as debt
gh issue list --repo LanternOps/breeze --state open --limit 200 --label enhancement
# Waiting on someone else — sweep for stale ones each round
gh issue list --repo LanternOps/breeze --state open --search 'label:pending-author,needs-info'
An unlabeled issue is invisible to all three, so label on sight — that's what let 69 issues escape triage entirely before 2026-08-16.
Repeated --label is AND, not OR. --label bug --label tech-debt returns only issues carrying both, which for these labels is zero — a silently empty result that reads exactly like "no defects." Use --search 'label:a,b,c' (comma = OR) whenever you want any-of semantics, and sanity-check the count against --state open before believing a small number.
The universal staleness trap — "I responded but the queue memory is stale" applies to ALL THREE types. For every open item, compare the last-activity author + date against my last engagement. If someone else spoke after me with a date past my last action, I'm the blocker — read the full thread before deciding what's open. This is the single most common failure mode and it is type-agnostic:
- PRs:
reviewDecisionresets fromCHANGES_REQUESTED→REVIEW_REQUIREDwhenever new commits land, so a PR you bounced that's since been fixed looks identical to one the author hasn't touched. For every "REQUEST CHANGES posted" PR, verifycommits.last.committedDatevs the bounce date:
Iffor pr in <bounced PR numbers>; do gh pr view $pr --repo LanternOps/breeze --json title,updatedAt,reviewDecision,commits \ | jq -r '"#\($pr): updated \(.updatedAt) | last commit \(.commits|last.committedDate) — \(.commits|last.messageHeadline)"' donelast commit > bounce date, it's waiting on me → re-review (Tier-1 manual; scope is bounded to the original feedback). - Discussions: the
comments(last:1)author in the re-orient query.isAnswered=falseon a Q&A I already answered = I still owe an "accept answer." Do NOT readisAnswered=nullon non-Q&A categories as "open" — that field is null for Ideas/Show-and-tell by design; use last-comment author instead. (Misread this once and flagged a pile of already-closed discussions as open.) - Issues: last comment author + date. A reporter who replied after my last comment is waiting on me.
- Advisories: last comment author + date on the advisory page — see §1a. A reporter who replied after me is waiting on me, and the advisory's
updated_atwill NOT tell you.
Issues: the staleness sweep is a FULL pass, never an activity window. Iterate every open issue whose author is not me, regardless of updatedAt, and compare the last comment's author+date against my last comment. A round that only looks at items updated "since the last round" misses everything that landed during a gap between rounds — that is exactly how #2987 (08-18), #3876 (08-25), #4000 (08-25) and #3386 (08-27) went unanswered for 18 days after a two-week outage in rounds, while the log claimed "every community issue last-comment = Todd." Two corollaries: (1) a community issue with ZERO comments is waiting on me — a last-comment check returns nothing and silently drops it, which is what hid the p1 #4000; (2) never age a prior round's "not waiting on me" verdict — re-run the comparison on every item every round, because a reporter can nudge after a sign-off (#3386). The one-liner:
gh issue list --repo LanternOps/breeze --state open --limit 400 --json number,author --jq '.[] | select(.author.login != "ToddHebebrand") | .number' \
| while read n; do gh issue view $n --repo LanternOps/breeze --json number,author,comments \
--jq '"#\(.number) by \(.author.login) last=\(if (.comments|length)>0 then "\(.comments[-1].author.login)@\(.comments[-1].createdAt[:10])" else "NONE (waiting on me)" end)"'; done \
| grep -v 'last=ToddHebebrand'
Anything that prints is 🔴 until the thread is opened and read.
1a. Security advisories (private — every round, full pass)
Private vulnerability reports arrive as repository security advisories, not issues, so the PR/issue/discussion queries never see them. SECURITY.md promises acknowledgment within 48 hours, and the OpenSSF Best Practices badge requires a first response within 14 days for every report in the last 6 months.
List them (API works for state, reporter, dates):
gh api repos/LanternOps/breeze/security-advisories --paginate \
--jq '.[] | select(.state!="closed") | [.ghsa_id,.state,.created_at[:10],(.submission!=null|tostring),([.credits_detailed[]?|select(.type=="reporter")|.user.login]|join(","))] | @tsv'
state=triage is a new inbound report nobody has accepted yet. submission=true (or a reporter credit) means an outside person filed it, so they are owed replies.
Comments are NOT in the API, and updated_at does not move when someone comments. A triage advisory answered two days after filing still shows updated_at == created_at (2026-10-12: a report that had been answered on time was misread this way as never answered). The only source of truth is the advisory page, read logged in as me. Use the ego-browser skill:
gh api repos/LanternOps/breeze/security-advisories --paginate \
--jq '.[] | select(.state!="closed") | .ghsa_id' > /tmp/ghsa.txt
ego-browser nodejs <<'EOF'
const fs = await import("node:fs/promises");
const ids = (await fs.readFile("/tmp/ghsa.txt", "utf8")).trim().split("\n");
const task = await taskSpace("gh-queue advisories");
const page = task.page("p1");
for (const id of ids) {
await page.goto(`https://github.com/LanternOps/breeze/security/advisories/${id}`);
const c = await page.evaluate(() => [...document.querySelectorAll(".timeline-comment")]
.map(x => ({ who: x.querySelector(".author")?.innerText.trim(),
at: x.querySelector("relative-time")?.getAttribute("datetime")?.slice(0, 10) }))
.filter(x => x.who));
const last = c.at(-1);
console.log(`${id} comments=${c.length} last=${last ? last.who + "@" + last.at : "NONE"}`);
}
EOF
About 1 s per advisory. Do the whole list every round; never only the triage ones, since a reporter can comment on a draft or published advisory too.
Bucketing:
- 🔴 Waiting on me: an externally reported advisory with no comment from ToddHebebrand, OR any advisory whose last comment is from someone else. Note the days since that comment against the 48 h / 14 day clocks.
- 🟡 Their court: my comment is last and asks the reporter for something (hashes, repro, version).
- 🟢 Nothing owed: advisories we filed ourselves with no outside comments (most published ones).
Privacy — this file is public. Advisory content never goes in a PR, issue, discussion, commit, or this skill. In queue.md (gitignored) record only the GHSA id, state, and last-comment author + date. Triage notes, draft replies, and verdicts live in the private intake record ~/breeze-security/remediation/<YYYY-MM-DD>-<GHSA-id>-triage.md; read it before replying. Advisory comments can only be posted in the GitHub UI (no API), so draft the reply there and hand it to me to paste.
Drafts — not reviewed (Todd, 2026-10-29). A PR with isDraft=true gets no review of any kind: no code review, no design-level COMMENT review, no step-0 screen, no subagents, even when its body asks for early feedback. It enters the queue when the author marks it ready for review. In the report, list open drafts on one line ("Drafts, not reviewed: #A, #B") and do not bucket them. One exception to watch for: if a draft's last comment hands the merge or review decision to me (as on #859, which sat 18 days as a draft after the author had addressed every blocker), flag that draft on the same line so I can decide whether to ask the author to mark it ready. Do not review it.
Dependabot — "will auto-rebase" is a recurring queue.md lie. Dependabot only auto-rebases trivial conflicts, NOT lockfile conflicts. For any "auto-rebase pending" PR, check commits.last.committedDate vs the note date; if stale, post @dependabot rebase explicitly, then merge sequentially (10-15s between admin-merges) so GH's mergeability recompute keeps up.
2. Triage round
Bucket every in-scope item into:
- 🔴 Waiting on me — needs my review/answer/decision
- 🟡 Ball in their court — author/reporter to act; note what we're waiting for
- 🟢 Closeable — done, just needs the close action (per close rules in §5)
Verify against code before responding (load-bearing rule). When a thread claims a gap, bug, regression, or "X is broken" — read the actual code and confirm before you answer. Plausible-sounding claims are wrong often enough that posting on them unverified produces public mistakes. Two real misses came from repeating a claim instead of checking: a "security/patches ignore orgId" gap that was already fixed by #973, and a "filters are a scale regression" claim that was actually the existing page architecture. Grep/read first, then respond. A false-negative grep is also a trap — confirm with the real symbol, not a guessed string (searched query('orgId') and got zero hits when the code used query.orgId).
PR step 0 — hostile-intent screen (EVERY external PR, before any tier decision)
Correctness review does not ask "what could this diff do that the body doesn't say". Until 2026-10-20 nothing in this process did; Todd asked and the honest answer was "no". Now, for every PR whose author is not ToddHebebrand or dependabot, run the mechanical screen first and read its hits before bucketing:
.claude/skills/gh-queue/scripts/pr-threat-screen.sh <N> # from the repo root
It prints only sections with hits: sensitive paths (workflows, Dockerfiles, scripts/, manifests, lockfiles, migrations, agent/helper/viewer, auth/signing/secrets), dependency manifest changes, lockfile entries with the integrity hash checked against the npm registry and any non-registry source, install hooks / shell-outs in package.json scripts, outbound hosts not on the known list, dynamic execution / decoding / raw DOM, long opaque strings, non-ASCII in code lines (bidi/homoglyph), env/secret/privilege/RLS-bypass touches, listeners/file writes/timers, and tests removed or lint suppressed. An empty report means no mechanical signal, not safe. Paste the hits that matter (or "screen: clean") into the review comment so the check is visible, and treat these as hard stops regardless of tier:
- Any
.github/workflows/*, Dockerfile, compose,scripts/,.npmrc,turbo.json, orpackage.jsonscriptschange from an external author → read the whole change yourself, not a subagent, and ask what it runs and with which secrets (merge_group runs with full secrets; forkpull_requestdoes not). - Any new dependency, override,
npm:alias, git/tarball source, or integrity mismatch → verify the package on the registry (publisher, age, downloads, repo) before anything else. A typosquat looks like a routine bump. - Any new outbound host, decoded blob,
eval/Function/exec, or non-ASCII in a code line → explain it or bounce it. - Anything under
agent/,apps/helper,apps/viewer, the installer/updater, signing, SSO/auth, orSECURITY DEFINER/ system-scope DB context → Tier 2 with the hostile-reading section, no exceptions for trusted contributors; this code runs on customer machines or bypasses tenant isolation by design.
Dependency bumps and "trivial" config from external authors are never green-on-sight. They are exactly the diffs an adversary would choose. The old Tier 1 list below is for own/dependabot PRs and for external diffs that the screen has already shown to be clean.
Contributor trust tiers live in contributors.md (§ Trust tiers). Reputation is built in weeks and is not provenance; a tier changes how much you read yourself, never whether the screen runs.
PR-specific tiering — don't waste subagent budget on green-on-sight
Tier 1 — fast manual scan (~30s, no subagents). gh pr diff <N> | head -150 or gh pr view <N> --json files,additions,deletions. Green-on-sight (own/dependabot PRs, or external PRs with a clean screen): test-only diffs, dependency bumps, explicit verifiable cleanups, trivial config mirroring an existing pattern, dated-TODO removal (verify date + grep for live consumers).
Tier 2 — parallel pr-review-toolkit subagents (only the 🟡 pile). Heuristics: touches auth/RBAC/RLS/tenant isolation; new schema/migration; diff >~200 lines or >5 files; novel subsystem; frontend PR from an AI-assisted contributor (check Dialog.tsx/runAction/helper reinvention); title makes a hard-to-skim claim; any step-0 hard stop. Send one message with multiple Agent tool uses, subagent_type: pr-review-toolkit:code-reviewer:
Review PR #<NUM> on LanternOps/breeze. Fetch the diff with:
gh pr view <NUM> --repo LanternOps/breeze --json title,body,additions,deletions,files
gh pr diff <NUM> --repo LanternOps/breeze
Context: <2-3 sentences on what it claims + prior discussion>
Focus on: correctness bugs at file:line; tenant isolation / RLS (every new tenant table needs
policies — see CLAUDE.md); whole-system coherence (does the diff match the body? does it reinvent
something?); for Billy's PRs check globally-disconnected patterns (re-implemented modals, body-sniff
vs helper — see contributors.md); comment style (bold leads, file:line, no tables).
HOSTILE READING (separate section, required): assume the author is adversarial and the body is
cover. Name every thing this diff could do that the body does not say: data it can read or send
outside the tenant or the host, code it can execute (install hooks, dynamic imports, exec,
templates reaching the shell or SQL), privilege it gains (system DB scope, SECURITY DEFINER,
platform-admin paths, RLS predicates removed or widened, auth/signing/verification weakened,
TLS checks relaxed), persistence it adds (timers, workers, cron, startup paths, agent/helper/
updater/installer changes), and anything obfuscated (encoded strings, homoglyphs, dead-looking
code that is reachable). Check that tests still cover what they covered before (deleted/skipped
tests, loosened assertions). Say "hostile reading: nothing found" explicitly if so. The screen
output is: <paste step-0 hits or "clean">.
Report under 300 words: must-fix vs nice-to-have, file:line + one-line rationale each, hostile
reading last. Do NOT post comments — return findings as text for me to consolidate.
Consolidate findings before posting. Group must-fix vs nice. When in doubt, lean Tier 1 for own/dependabot PRs; for external PRs that hit a step-0 stop, when in doubt lean Tier 2.
Discussions & Issues — mostly manual
No subagent tiering. Read the thread, verify claims against code (§2 rule), bucket, respond. Use a subagent only for genuinely large investigations (e.g. "is this proposed feature already half-built?" — like checking whether a filter engine already exists before greenlighting a filter PR). Issues follow the github-issues skill for etiquette/lifecycle; this manager owns the queue/triage.
3. Single-item deep review
For one PR or an architectural item, skip dispatch and review directly: gh pr view/gh pr diff first, check linked issues/discussions for context, verify tenancy changes against the RLS shape table in CLAUDE.md, check Go slog calls use err.Error() (raw err serializes as {}), verify migration naming YYYY-MM-DD-<slug>.sql + same-day -a-/-b- ordering.
4. Posting
Style (from feedback_issue_comment_style memory): bold section leads (**Root cause:**, **Fix:**), file:line refs (apps/api/src/routes/devices.ts:142), commit SHAs for prior work, prose not tables, trailing status line.
Escaping (from feedback_gh_comment_escaping): \n does NOT expand in bash double-quotes — use a heredoc or a temp file. Nested backticks in inline code spans break the comment; rephrase. For long/structured comments, write the body to a temp file and pass it via -F (most robust — avoids all shell escaping):
# PR / Issue comment
gh pr review <N> --repo LanternOps/breeze --request-changes --body "$(cat <<'EOF'
**Root cause:** ...
EOF
)"
gh issue comment <N> --repo LanternOps/breeze --body "$(cat <<'EOF' ... EOF)"
# Discussion comment — GraphQL mutation, body from a temp file
DID=$(gh api graphql -f query='{repository(owner:"LanternOps",name:"breeze"){discussion(number:<N>){id}}}' --jq '.data.repository.discussion.id')
gh api graphql -f query='mutation($id:ID!,$body:String!){addDiscussionComment(input:{discussionId:$id,body:$body}){comment{url}}}' \
-f id="$DID" -F body=@/tmp/comment.md --jq '.data.addDiscussionComment.comment.url'
Handles (from feedback_never_invent_github_handles): NEVER @-mention a handle you didn't see verbatim in the thread. Verify with gh issue view --json author,comments / the discussion query first.
Cite docs by PUBLISHED URL, code by repo path. file:line is the right currency for code — a reader can find apps/api/src/config/validate.ts:1198 on GitHub. It is the wrong currency for documentation: apps/docs/src/content/docs/deploy/upgrades.mdx:326 is unusable to the self-hoster who asked. That tree publishes to https://docs.breezermm.com, mapping apps/docs/src/content/docs/<path>.mdx → https://docs.breezermm.com/<path>/ (Starlight, site set in apps/docs/astro.config.*, no base). Link the page, not the source file, and curl each URL for a 200 before posting — a dead link in a public answer is worse than a repo path. Repo docs outside that tree (docs/operations/*.md, docs/registers/*.md) are not published; name them as repo files and say so. Internal issues we file for ourselves are the exception — a dev fixing backup.mdx wants the file path.
Always verify the post landed (count comments / re-read last comment). The discussion mutation can succeed while a follow-up verify query errors on a missing-first/last pagination boundary — that's a query bug, not a post failure; check before re-posting or you'll double-comment (happened on #982; deleted the dup via deleteDiscussionComment(input:{id:"DC_..."})).
5. Closing — rules differ by type
- PRs:
mainis behind the merge queue (since 2026-09-07) — a baregh pr merge <N> --repo LanternOps/breezeenqueues; the queue squashes and lands it serially after the fullCI Successgate. Never--admin(bypasses the queue and rebuilds every entry behind it), per CLAUDE.md "PR Merge Process". - Discussions: I can close my own. Comment first, then close with a reason:
gh api graphql -f query='mutation($id:ID!){closeDiscussion(input:{discussionId:$id,reason:RESOLVED}){discussion{number closed}}}' -f id="$DID"reason: RESOLVEDfor answered/shipped,OUTDATEDfor superseded. Leave open anything with genuine outstanding work even if I wrote "closing as resolved" earlier — verify the work actually shipped (e.g. a discussion whose prerequisite PR merged but whose feature PR never did → stays open). - Issues: defer to the
github-issuesskill, which splits the rule by who filed it. NEVER close community issues yourself — reporter verifies and closes, or owner closes after reporter confirms, or owner closes stale ones with a note. A commit/deploy is not a verification. Internally-filed issues are the opposite: close them once their fix is merged, since no third party is waiting to verify.
Fixed-but-open sweep. Because PRs historically used Refs #N rather than Closes #N, merged fixes do not close their issue. When the open count looks wrong, run the sweep rather than reading titles:
gh issue list --state open --limit 400 --json number --jq '.[].number' > /tmp/open.txt
git log origin/main --since=<date> --pretty=format:'%H%x09%s%x0a%b%x00' > /tmp/log.txt
# match each open number against every commit subject+body; a subject of the form
# `fix(scope): summary (#ISSUE) (#PR)` is the repo's fully-fixed signature
Then read the commit body before closing, not the subject: several deliberately say "this is half of #N" or "leaving open for the reporter". Close only the ones with no such caveat, and comment with the fixing PR + SHA.
6. After every action
Update queue.md:
- Move closed/merged items out of the waiting section of the right type (
## PRs/## Discussions/## Issues/## Advisories— ids and dates only) - Add a one-line entry to the session log at the bottom with today's absolute date (
### 2026-05-30 — ...), never "today"/"yesterday" - If you learned a contributor pattern, update
contributors.md
Adding a new contributor
When an item arrives from someone not in contributors.md: skim their first few PRs/issues (gh search prs --author <handle> --repo LanternOps/breeze), and after 2-3 interactions add a 4-6 line profile (real name + handle, AI-assisted/human/mixed, recurring patterns at file:line, response style). The point is "what to check first when their next item lands." Profiles apply across PRs, discussions, and issues.
Common gh commands
# PRs waiting on me / drafts (list only, never reviewed — see §1) / single PR
gh pr list --repo LanternOps/breeze --state open --search "review-requested:@me" --limit 20
gh pr list --repo LanternOps/breeze --state open --draft --limit 20
gh pr view <N> --repo LanternOps/breeze --comments
# Discussions (all via GraphQL — see §1 list query and §4/§5 mutations)
gh api graphql -f query='{repository(owner:"LanternOps",name:"breeze"){discussion(number:<N>){id title body closed comments(first:30){nodes{author{login} createdAt body}}}}}'
# Issues
gh issue list --repo LanternOps/breeze --state open --limit 50
gh issue view <N> --repo LanternOps/breeze --comments --json title,body,comments,author,labels
# Author history
gh search prs --author <handle> --repo LanternOps/breeze --limit 20
# Merge (enqueue — the merge queue owns the squash; never --admin)
gh pr merge <N> --repo LanternOps/breeze
Defers to (not duplicated here)
pr-review-toolkit:code-reviewer— deep PR code review (Tier 2 dispatch)github-issues— issue etiquette, lifecycle, issue-body templates, the never-close-community-issues rule
Related memories
feedback_issue_comment_style— bold leads, file:line refs, prose not tablesfeedback_gh_comment_escaping— heredoc/temp-file for newlines, no nested backticksfeedback_never_invent_github_handles— only @-mention what's verbatim in the threadfeedback_verify_own_closes_keywords—Closes #X #Yonly auto-closes #X; grep the merged diff per claimed item
Out of scope
- Outgoing PRs (mine) beyond merge-tracking — use TodoWrite per session
- Project-wide code-review conventions —
pr-review-toolkit:code-reviewerowns those
版本历史
-
0b71144
当前 2026-11-03 11:01
新增跳过Draft PR逻辑;为外部PR增加恶意意图机械筛查步骤,强化依赖变更与敏感路径的安全审查。
-
8f240dd
2026-10-19 17:56
新增对安全公告(Security Advisories)的支持,将其作为第四类对象纳入队列管理,解决私有漏洞报告无法被常规API捕获的问题。
- 1c3a08e 2026-09-22 22:17
-
0ee1815
2026-08-28 19:50
新增文档引用规范,要求使用已发布URL而非源码路径;修复因工作区丢失风险而提交的未提交文档记录。
- 10bb1af 2026-08-20 08:50


