vuln-tracker

${var} — Scope selector for the lifecycle poll:

• empty → full lifecycle poll: PR/advisory status + PVR triage + disclosure-queue aging (default).
• prs (also pr / tracker) → Arm A only — PR/advisory status audit + stars-secured dashboard.
• pvr → Arm B only — PVR triage-state poll on submitted advisories.
• queue (also disclosures / backlog) → Arm C only — pending-disclosure queue aging + escalation.
• a bare GHSA-xxxx-xxxx-xxxx → Arm B, single-advisory mode — check just that one advisory's triage state on demand.

Today is ${today}. This skill is the daily read/poll arm of the vuln pipeline: vuln-scanner opens PRs, submits PVRs, and queues disclosure drafts, then moves on. This skill polls everything it produced and surfaces what the operator must look at — it does not open PRs or submit advisories itself (those are vuln-scanner's write actions).

Voice

If soul/SOUL.md and soul/STYLE.md are populated, read them and match the operator's voice in every notification. If empty or absent, use a clear, direct, neutral tone.

Capability mode

This skill runs mode: write deliberately. It is a read/poll arm, but three of its capabilities cannot run under read-only:

• Arm B polls private, unpublished advisory triage state via gh api repos/$REPO/security-advisories/$GHSA. Draft/triage advisories are visible only to the repo maintainers and the reporter, so the read is intrinsically authenticated; read-only strips gh, and a bare $SECRET on the command line is refused by the Bash permission layer (so a hand-rolled authenticated curl isn't an option either) — the read needs gh api, which handles auth internally.
• Arm B persists state transitions — it rewrites state:/last_checked:/resolved_at: frontmatter in memory/pending-disclosures/*.md in place and moves resolved files to memory/pending-disclosures/resolved/ (Edit/git mv — both stripped in read-only).
• Arm A leans on authenticated gh api for the PVR-state endpoint (repos/$REPO/private-vulnerability-reporting) and repo/advisory reads.

No sibling writes a repo file outside memory/, so there was nothing to relocate — the only outside-memory/ writes are the ephemeral .pending-notify-temp/ notify-staging files. Keeping write preserves every absorbed capability.

Shared preamble (run for every invocation)

1. Read memory/MEMORY.md for context.
2. Read the last ~3 days of memory/logs/ and drop anything already reported — don't re-surface the same signal twice.
3. Read soul/SOUL.md + soul/STYLE.md if populated (voice).
4. Parse ${var} → scope (deterministic; trim + lowercase, except a GHSA- value which is compared case-insensitively but preserved verbatim):
   • empty → scope = full (run Arm A, then B, then C).
   • matches ^GHSA- (case-insensitive) → scope = pvr, single_advisory = <the GHSA value> (Arm B filtered to one advisory).
   • prs / pr / tracker → scope = prs (Arm A only).
   • pvr → scope = pvr (Arm B only, all advisories).
   • queue / disclosures / backlog → scope = queue (Arm C only).
   • anything else → log VULN_TRACKER_BAD_VAR: unrecognized scope '<var>', send no notification, exit.
5. mkdir -p .pending-notify-temp and start an empty combined-notification buffer at .pending-notify-temp/vuln-tracker-${today}.md. Each arm that has signal appends its section to this buffer; at the very end (step "Notify") the skill sends the buffer once if it is non-empty. This keeps notifications tight — a full poll with signal in two arms is one message, not two.

Then run the arm(s) selected by scope, and finish with the shared Notify and Log steps.

Network Note

• Arm A & Arm B (GitHub reads): all data via gh api / gh search / gh pr view. gh handles auth internally via GH_TOKEN (and Arm B's private-advisory reads need the elevated GH_GLOBAL PAT). No env-var-authenticated curl from bash — a bare $SECRET on the command line is refused by the Bash permission layer, so gh api (auth handled internally) is the reliable path; no postprocess scripts needed. Arm B keeps a documented curl fallback for the advisory endpoint — see Arm B step B2 — but gh api is preferred.
• Arm C (local only): reads only local files (memory/pending-disclosures/, memory/issues/, memory/topics/pr-status.md). No outbound network or auth required.

---

Arm A — PR & advisory lifecycle status (scope full or prs)

Audit the lifecycle status of every disclosure vuln-scanner has produced. Without a follow-up loop, three things rot silently:

• Merged-but-uncelebrated wins — landed fixes never reach self-improve / reflect (retro) without manual aggregation.
• Maintainer questions on open PRs — a maintainer comments asking for clarification; if the bot doesn't see it, the PR ages out.
• Queued drafts past their disclosure window — entries with channel: "skipped" (no-safe-channel) vanish into vuln-scanned.json with no recurring re-probe.

This arm cross-references memory/vuln-scanned.json against live GitHub state and surfaces anything the operator should look at.

A1. Load the canonical scan history

memory/vuln-scanned.json may be a flat top-level array of scan rows, not a {scans: [...]} object - vuln-scanner writes the flat shape. Read it in a way that accepts both:

# Accepts both the flat `[ {repo,scanned_at,findings,channel}, ... ]` shape
# and the legacy `{scans: [...]}` object.
jq -c 'if type=="array" then .[] else (.scans[]? // empty) end' memory/vuln-scanned.json 2>/dev/null

A bare jq -c '.scans[]' errors on a flat-array file (Cannot index array with string "scans") and, with 2>/dev/null swallowing the error, silently yields zero rows - so the arm reads an empty scan history and reports only what other passes happen to surface. The if type=="array" guard is the fix; it keeps the legacy object shape working too.

If memory/vuln-scanned.json doesn't exist or the parsed row count is 0, log VULN_TRACKER_SKIP: no scan history for this arm and skip Arm A (no notification section - first runs of vuln-scanner haven't happened yet). In a full poll, continue to Arm B/C.

Each scan entry has at minimum: repo, scanned_at, findings, channel, severity. Public-PR entries also have pr (URL). Pending-disclosure entries have draft_at and patch_branch. Skipped entries have reason.

Retro-active coverage: if the JSON was written after vuln-scanner started running, some PRs won't be in the JSON. Pull all bot-authored security PRs from GitHub directly (next step) to fill the gap.

A2. Pull all bot-authored security PRs from GitHub

vuln-scanner opens security-fix PRs on a security/ branch. The title varies: dependency-bump PRs use fix(deps): (the A5a template is fix(deps): bump <pkg> to patch <CVE>), and older or hand-written ones use fix(security): or a bare security:. Match all three title prefixes, and treat the security/ branch prefix as the real invariant - the branch name is mandated by A5a, the title wording is not.

Do not filter on fix(security): alone. vuln-scanner's dominant output is fix(deps):, so a fix(security):-only filter silently drops most in-flight security PRs - that is how alibaba/open-code-review#541 (conflicted, lint-failing, one unaddressed review) was missed by a poll while other PRs were tracked. Also exclude docs(security): - those are security-sync website PRs, not disclosures.

gh search prs --json does not expose headRefName - that field is only available via GraphQL. So run both passes below every time and union them on repository + number; neither alone is complete.

The bot author is whoever the workflow uses (typically github-actions[bot] or a dedicated account configured in the workflow). Determine the author from aeon.yml or the workflow file; default to whatever account opened the most recent security-fix PR (any of the prefixes above) you can find.

Pass 1 (title prefixes, works everywhere):

BOT_AUTHOR="<resolved bot author>"
gh search prs --author "$BOT_AUTHOR" --json number,title,url,state,createdAt,closedAt,repository --limit 200 \
  | jq '[.[] | select(
      (.title | startswith("fix(deps):")) or
      (.title | startswith("fix(security):")) or
      (.title | startswith("security:"))
    )]'

Pass 2 (GraphQL, required - picks up every security/ branch whatever the title says):

gh api graphql -f query='
{
  search(query: "author:'"$BOT_AUTHOR"' is:pr sort:created-desc", type: ISSUE, first: 100) {
    nodes { ... on PullRequest {
      number title url state createdAt closedAt mergedAt
      repository { nameWithOwner }
      headRefName
    }}
  }
}' | jq '[.data.search.nodes[]
    | select((.headRefName // "") | startswith("security/"))]'

Union the two result sets, dedup by repository + number (equivalently, by URL).

Cross-reference with vuln-scanned.json:

• PR present in JSON → use JSON's severity / cwe / note for the row.
• PR not in JSON → mark as pre-history; fill severity from the PR title if obvious.
• JSON entry with channel != "public-pr" → no PR to fetch; goes in the "queued" / "skipped" sections.

A3. Fetch live state for each open PR

For each open PR (state from step A2):

gh pr view "$REPO/$NUM" --json state,merged,closedAt,createdAt,reviews,comments,reviewDecision,author

Per-PR signals:

• Maintainer-needs-answer: any comment whose author.login != $BOT_AUTHOR posted after the most recent comment by $BOT_AUTHOR (or after PR creation if the bot hasn't commented). Also reviewDecision == "CHANGES_REQUESTED" always counts.
• Stale-no-review: state == "OPEN" AND no review AND no maintainer comment AND createdAt > 7d ago.
• Aging-with-engagement: state == "OPEN" AND any maintainer activity AND open > 14d.

If a scan entry has advisory_ids (one or more GHSA IDs), check each one's published state:

gh api "repos/$ORIGIN_REPO/security-advisories/$GHSA_ID" --jq '.state // "not found"'

state == "published" → public advisory visible. 404 → osv-scanner referenced it but the upstream repo never published its own advisory.

A4. Fetch star counts for every secured repo

For every unique repo across the union from step A2 (JSON history + bot-authored security PRs):

gh api "repos/$REPO" --jq '{stars: .stargazers_count, archived: .archived}' 2>/dev/null

Refetch every run. Do NOT carry star counts forward from the previous memory/topics/vuln-followup.md — per-repo counts drift between runs and the secured-stars headline is the operator's load-bearing metric. Cache only within a single run, keyed by nameWithOwner, so a repo with multiple PRs is fetched once.

Repo-state handling:

• 200 with stars: use .stargazers_count (raw integer).
• 200 with archived: true: still use the star count, but suffix the repo cell with (archived) so the operator knows the maintainer isn't responsive.
• 404 / 403: repo was deleted, renamed, or made private. Record null and render as repo-deleted. Exclude from total_stars_* aggregates entirely so dead repos don't quietly zero out the totals.
• Other non-2xx: record null and render ★?. Flag in the run log for operator follow-up.

These per-repo counts power both the Stars Secured aggregate (step A6) and the Stars column on every per-repo table in step A6.

A5. Re-probe channel: "skipped" and channel: "pending-disclosure" repos

For each historical entry where the disclosure couldn't ship, re-check whether the situation changed:

• channel: "skipped" with reason containing "no PVR":

  PVR_NOW=$(gh api "repos/$REPO/private-vulnerability-reporting" --jq .enabled 2>/dev/null || echo "false")
  

  If PVR_NOW=true and the original was false, surface as newly-actionable.
• channel: "skipped" with reason containing "no SECURITY.md" — re-check gh api repos/$REPO/contents/SECURITY.md and .github/SECURITY.md. If now present, surface as newly-actionable.
• channel: "pending-disclosure" — cross-reference with memory/pending-disclosures/ to see if the draft is still on disk. If the file is gone but the JSON entry says pending-disclosure, mark as lost-draft so it stops being escalated forever.

A6. Categorize every entry, then rewrite the dashboard

Status	Meaning
merged	PR merged. One-time celebration — drop from notifications after 30d.
open-clean	PR open, no maintainer activity yet, < 7d old. Wait.
needs-answer	Maintainer commented or requested changes. Operator action.
stale-no-review	Open > 7d, zero maintainer activity. Consider polite ping or close.
aging-engaged	Open > 14d with engagement. Operator should triage.
closed-no-merge	PR closed without merging. Capture the reason for review.
queued	pending-disclosure draft on disk, not yet shipped.
skipped-rechecked	Channel was "skipped" originally; re-probe still shows no channel.
newly-actionable	Skipped originally; PVR or SECURITY.md now present. Operator action.
lost-draft	JSON says pending-disclosure but draft file is gone. Display once, then suppress.
pre-history	PR found via search but predates vuln-scanned.json. Fill what we can.

Then rewrite memory/topics/vuln-followup.md (rewrite — don't append; this file is a living dashboard, not a log).

The Stars Secured block goes at the top so the operator sees aggregate impact before drilling into rows. total_stars_secured = sum of stargazers across every unique repo where vuln-scanner has landed at least one merged PR. total_stars_in_flight = sum across repos with an open PR. total_stars_tracked = sum across the full union. Track all three because celebration uses secured, prioritization uses in_flight, and historical review uses tracked.

Round star counts to abbreviated form for the headline (12.4k, 1.8k, 940). Keep raw integers in the per-repo tables so sort/diff stays exact.

# Vuln Tracker Status

*Last updated: ${today}*

## Stars Secured

- **Merged-PR repos (secured):** ★ <total_stars_secured> across <secured_repo_count> repos
- **Open-PR repos (in flight):** ★ <total_stars_in_flight> across <in_flight_repo_count> repos
- **All tracked repos:** ★ <total_stars_tracked> across <total_repo_count> repos

### Secured leaderboard — every merged PR ranked by repo stars
| Rank | Repo | Stars | PR | Merged | Severity | Title |
|------|------|-------|----|--------|----------|-------|

### Per-repo breakdown — secured (sorted by stars desc)
| Repo | Stars | Merged PRs | First merge | Latest merge | Severities landed |
|------|-------|------------|-------------|--------------|-------------------|

### Per-repo breakdown — in flight (sorted by stars desc)
| Repo | Stars | Open PRs | Oldest open | Severities open |
|------|-------|----------|-------------|-----------------|

### Per-repo breakdown — queued / skipped / closed (sorted by stars desc)
| Repo | Stars | Status | Severity | Note |
|------|-------|--------|----------|------|

## Operator-action queue

### Needs answer (<count>)
| Repo | Stars | PR | Title | Last activity | Latest commenter |
|------|-------|----|----|--------------|------------------|

### Newly actionable — channel opened up since the original scan (<count>)
| Repo | Stars | Original date | Original blocker | Now |
|------|-------|---------------|------------------|-----|

### Stale or aging
| Repo | Stars | PR | Age | Status | Suggested action |
|------|-------|----|----|-----|------------------|

## Recently merged (last 30d, <count>)
| Date merged | Repo | Stars | PR | Severity | Title |
|-------------|------|-------|----|----------|-------|

## Open / clean (no operator action — wait, < 7d) (<count>)
| Repo | Stars | PR | Severity | Opened | Age |
|------|-------|----|----------|--------|-----|

## Closed without merge (last 30d, <count>)
| Date | Repo | Stars | PR | Severity | Title | Likely reason |
|------|------|-------|----|----------|-------|---------------|

## Queued (no PR yet) (<count>)
| Severity | Repo | Stars | Original channel | Original blocker | Days queued |
|----------|------|-------|------------------|------------------|-------------|

## Lost-draft ghosts (suppressed from notifications)
| Date | Repo | Stars | Severity |

A7. Decide whether Arm A has signal

No signal (contribute nothing to the notification) if all are true:

• Zero needs-answer
• Zero newly-actionable
• Zero items moved to merged since the last run
• Zero items moved to closed-no-merge since the last run
• Zero items aged into stale-no-review or aging-engaged since the last run

To detect "since last run," diff today's categorization against the previous memory/topics/vuln-followup.md. If the file doesn't exist (first run), treat all entries as new and surface the full backlog.

A8. Emit Arm A's notification section (when it has signal)

Append this section to the combined buffer (.pending-notify-temp/vuln-tracker-${today}.md). Keep the whole message under 4000 chars; if tight, drop the merged/opened/stale section bodies and keep counts only.

*PR & advisory status*

★ secured: <total_stars_secured> across <secured_repo_count> repos  (in flight: <total_stars_in_flight> / <in_flight_repo_count>)

needs answer: <N>
<repo> (★<stars>) #<num> — <latest_commenter>: "<comment_excerpt_first_120_chars>"

newly actionable: <N>
<repo> (★<stars>, queued <days>d) — <what_changed>

merged this week: <N>  <list_with_stars>
opened, waiting: <N>
stale: <N>
queued: <N> (<critical>C / <high>H / <other>M+L)

leaderboard top-3 (PRs by ★): #1 <repo1> ★<s1> (PR #<pr1>) — #2 <repo2> ★<s2> (PR #<pr2>) — #3 <repo3> ★<s3> (PR #<pr3>)

dashboard: memory/topics/vuln-followup.md (full leaderboard inside)

Record Arm A status VULN_TRACKER_OK for the log.

---

Arm B — PVR triage state (scope full, pvr, or a bare GHSA-…)

vuln-scanner's re-submit arm (var=resubmit, the former pvr-watchlist) monitors repos waiting to open PVR. This arm monitors PVRs that have already been submitted and tracks their lifecycle: triage → draft (accepted) → published (public) or withdrawn (rejected). Without this, submitted advisories sit unmonitored until manually recalled from memory.

Source of truth: memory/pending-disclosures/*.md files with channel: pvr frontmatter. Each such file must have ghsa, repo, state, submitted_at fields (full schema below).

Configuration — optional tracking issue. The arm can reference an optional tracking issue in the operator's own repo — useful for cross-linking advisory state with an internal issue board. Resolve from (priority order):

1. aeon.yml top-level key pvr_triage.tracking_issue: (e.g. pvr_triage: { tracking_issue: "owner/repo#123" })
2. environment variable AEON_PVR_TRACKING_ISSUE
3. unset → skip cross-linking entirely.

If a tracking issue is configured, mention its URL in the notification and the per-advisory write-up so the operator can navigate to the canonical tracker.

B1. Discover in-flight PVRs

Scan memory/pending-disclosures/ for all .md files. Parse the YAML frontmatter. Keep only those with channel: pvr.

If single_advisory is set (a bare GHSA-… was passed), filter to just the matching ghsa value (one-off mode).

If no PVR files found:

PVRT_SKIP: no submitted PVRs on disk

Record that Arm B status and skip Arm B (no notification section). In a full poll, continue to Arm C.

B2. Probe each advisory's triage state

For each entry, determine repo and ghsa from frontmatter.

REPO="owner/repo"
GHSA="GHSA-xxxx-xxxx-xxxx"

gh api "repos/${REPO}/security-advisories/${GHSA}" \
  --jq '{state: .state, cve_id: .cve_id, published_at: .published_at}' 2>&1

Expected outcomes:

Response	Meaning
{state: "triage", ...}	Maintainer hasn't reviewed yet
{state: "draft", ...}	Accepted — maintainer is working on it
{state: "published", ...}	Published — fully resolved
{state: "withdrawn", ...}	Rejected or withdrawn by reporter
HTTP 403	Private advisory, we don't have read access — state unknown, treat as still triage
HTTP 404	Advisory deleted / repo private / GHSA invalid — flag as not-found

Fallback: gh api uses GH_TOKEN internally (workflow wires GH_GLOBAL for the elevated advisory read). If gh is somehow unavailable, fall back to ./secretcurl with the {GH_TOKEN} placeholder (same token; GH_GLOBAL itself is not a substitutable name):

./secretcurl -s -H "Authorization: Bearer {GH_TOKEN}" \
  "https://api.github.com/repos/${REPO}/security-advisories/${GHSA}" \
  | grep -o '"state":"[a-z]*"'

(Never write a raw curl with $GH_GLOBAL / $GH_TOKEN: a bare $SECRET on the command line is refused by the Bash permission layer. gh api stays the primary path.)

B3. Detect state changes

Compare the probed state to the state in the frontmatter.

• No change: note it, continue.
• Changed: this is the primary event. Log old → new state.

Also flag:

• Aged triage: state=triage AND (today − submitted_at) > 30 days → escalate. Most maintainers respond within 30 days; silence past that is actionable.
• Accepted (draft): surface the patch branch from the patch_branch frontmatter field — maintainer may want a PR instead of a private advisory.
• Published: advisory is live. The finding is closed. Update state and mark for removal.
• Withdrawn: rejected. Note the reason if visible. Mark for cleanup.

B4. Update frontmatter in-place

For each file with a state change, rewrite just the state field in the YAML frontmatter. Also update a last_checked field (add it if absent).

Do NOT modify the body of the advisory file — only update frontmatter.

Example frontmatter update:

state: draft          # was: triage
last_checked: 2026-05-21

For published or withdrawn entries, add:

resolved_at: 2026-05-21

B5. Decide whether Arm B has signal

• All entries still triage, no changes, none aged: no notification section. Record status and continue.
• Any state change, aged entry, or action item: emit the section below.

B6. Emit Arm B's notification section (when it has signal)

Append to the combined buffer:

*PVR triage*

pvr triage: {total} advisories in flight. {changed_count} changed.

CHANGED:
- {repo} {ghsa} — {old_state} → {new_state}
  {action_item}

AGED (>30d no response):
- {repo} {ghsa} — {days}d in triage. {severity}. escalate or close.
  patch: {patch_branch}

STILL TRIAGE:
{n} advisories waiting. oldest: {repo} ({days}d).

{if tracking_issue configured}
tracker: {tracking_issue_url}
{end}

Action items by transition:

• triage → draft → "maintainer accepted — offer to PR the patch branch: {patch_branch}"
• triage → published → "published as {cve_id}. remove from tracking."
• triage → withdrawn → "rejected. remove from tracking and note in vuln-scanned.json."
• aged triage (>30d) → "30d+ no response. consider pinging maintainer or withdrawing."

B7. Clean up resolved entries

For entries where state=published or state=withdrawn AND resolved_at is set: move the file from memory/pending-disclosures/ to memory/pending-disclosures/resolved/ (create the directory if needed; use git mv).

Do NOT delete — keep as a historical record.

Record Arm B status PVRT_OK for the log.

---

Arm C — Disclosure-queue aging (scope full or queue)

Monitor the pending vulnerability disclosure backlog. vuln-scanner queues draft advisories to memory/pending-disclosures/ when PVR auto-submission fails or the disclosure path is email-only. Without daily visibility, CRITICAL/HIGH advisories silently age past responsible-disclosure windows. This arm surfaces the queue state and escalates when findings sit too long. (Arm B tracks the lifecycle of the channel: pvr subset; Arm C ages the whole queue by severity and disclosure status.)

C1. Scan the backlog

ls memory/pending-disclosures/ 2>/dev/null

If the directory doesn't exist or is empty: log DISCLOSURE_TRACKER_SKIP: no pending advisories for this arm and skip Arm C (no notification section).

C2. Parse each advisory file

For each .md file in memory/pending-disclosures/ (ignore the resolved/ subdir):

From the filename (pattern: {repo-slug}-{YYYY-MM-DD}.md or {repo-slug}-{YYYY-MM-DD}-{ampm}.md):

• Extract target repo slug (everything before the last date segment)
• Extract filed date

From the YAML frontmatter (if present) parse:

• repo: — overrides the filename slug when present (canonical target)
• severity: — CRITICAL / HIGH / MEDIUM / LOW
• status: — see step C2.5 for the controlled vocabulary

From the file content (fallback for files without frontmatter), look for these near the top of the file:

• Severity: or **Severity:** — one of CRITICAL / HIGH / MEDIUM / LOW
• CVE/CWE: or similar identifier
• Short title (first non-blank heading line)

If severity is not parseable, treat as MEDIUM. Compute age: today - filed date in days.

C2.5. Classify each advisory's disclosure state

Before counting any draft as a past-threshold escalation, decide whether the draft is genuinely pending or already covered. A draft can be in one of these states:

• escalate — pending, no canonical PR found, past the severity-tier threshold
• pending — pending, no canonical PR found, within the threshold window
• operator-todo — needs operator-only action (email send, PVR enable nudge); not an agent failure
• covered-by-pr — a canonical disclosure PR has already been filed against the target repo and is OPEN or recently merged
• superseded-upstream — the bypass / vuln is fixed in upstream already, draft is dead-weight
• submitted — already submitted via PVR / GHSA; awaiting maintainer response

Resolution rules:

1. Check frontmatter status: first — map the literal value to a state:

   • superseded-upstream → superseded-upstream
   • submitted, submitted-via-pvr, disclosed-via-pr-{N}, email-sent → submitted or covered-by-pr
   • pending-operator-send with auto_send: true → pending (vuln-scanner's disclose arm, var=disclose, sends it autonomously - not a human task)
   • pending-operator-send (without auto_send: true), queued for operator manual send, email-failed, any string mentioning "operator" → operator-todo
   • pending, blank, or missing → fall through to rule 2

2. Cross-reference memory/topics/pr-status.md (if present) — grep for the {repo} slug (frontmatter repo: or filename) in the Open section and Recent Merges section. If a row exists with a fix(security) or chore(security) title against that repo, opened on or after the draft's detected_at / reconstructed_at / filed-date, classify as covered-by-pr and capture the PR number / title for the summary. If memory/topics/pr-status.md doesn't exist, skip this lookup and fall to rule 3.

3. Fall through — if no status hint and no canonical PR found, classify as pending. Then check age vs the severity-tier threshold (CRITICAL 3d / HIGH 7d / MED-LOW 14d) — if past, promote to escalate.

This is the load-bearing step. Without cross-referencing already-merged fix PRs the tracker generates false-positive escalations for drafts that have already been resolved.

C3. Build the summary

Group advisories by state first, then by severity within each state. The three buckets are:

• Escalate — escalate state only (truly stuck, past threshold, no canonical PR)
• Operator-todo — operator-todo state (email-only sends, PVR-enable nudges, anything awaiting human action)
• Cleanup candidates — covered-by-pr, submitted, superseded-upstream (draft files that can be removed from memory/pending-disclosures/)

Severity tiers and thresholds (only apply to escalate and pending states):

• CRITICAL (age threshold: 3 days — escalate immediately)
• HIGH (age threshold: 7 days — escalate at 7d)
• MEDIUM / LOW (threshold: 14 days)

For each advisory, produce one line:

- {repo-slug} | {severity} | {age}d | {short title}{state-suffix}

where {state-suffix} is empty for escalate / pending, [operator-todo: {reason}] for operator-todo, and [covered: PR #{N}] / [superseded-upstream] / [submitted] for cleanup candidates.

Count totals. Identify advisories in the escalate state — those are the only ones that drive the urgent notification path.

C4. Check for upstream PVR / token issues

Look in memory/issues/INDEX.md for any open issues tagged with pvr, repository_advisories, or missing-secret that explain why advisories are stuck. If such an issue exists, note:

• Number of consecutive PVR failures (if logged)
• Fix estimate from the issue notes
• How many of the backlogged advisories are blocked by it

If no such issue exists, treat the queue as routine and skip the "blocked by" line in the notification.

C5. Decide whether Arm C has signal

Compute counts from step C3:

• escalate_count — drafts in the escalate state
• pending_count — drafts in pending (in-window) state
• operator_todo_count — drafts awaiting operator action
• cleanup_count — drafts in covered-by-pr / submitted / superseded-upstream

Decision:

• Queue empty: log DISCLOSURE_TRACKER_SKIP: queue empty, no section.
• escalate_count > 0: emit the urgent section.
• escalate_count == 0 but cleanup_count > 0: emit the daily digest section (includes the cleanup-candidate list so the operator can prune memory/pending-disclosures/).
• All pending / operator-todo, nothing past threshold, no cleanup candidates: emit the daily digest section.

Coverage from covered-by-pr / submitted / superseded-upstream is never counted as an escalation — informational only. Operator-todo is never an escalation either; it's surfaced separately so the operator knows their inbox.

C6. Emit Arm C's notification section

Append to the combined buffer.

Urgent format (escalate_count > 0):

*Disclosure queue*

disclosure queue: {escalate_count} past threshold (of {total} drafts).

ESCALATE:
- {repo} — {severity}, {age}d old (threshold: {N}d)
[... others in `escalate` state ...]

operator-todo ({operator_todo_count}):
- {repo} — {severity}, {age}d — {operator-reason}

cleanup candidates ({cleanup_count}):
- {repo} — [covered: PR #{N} / superseded-upstream / submitted] — safe to delete from memory/pending-disclosures/

{IF blocking issue tracked in memory/issues/INDEX.md}
blocked by {ISS-ID} — {short reason}
fix: {fix estimate} unblocks {N} of {escalate_count + pending_count}
{end}

Daily digest format (no escalation):

*Disclosure queue*

disclosure queue: {total} drafts. {critical_count} CRITICAL, {high_count} HIGH, {other_count} MED/LOW.
{pending_count} in-window, {operator_todo_count} operator-todo, {cleanup_count} cleanup candidates.
oldest in-window: {repo} ({age}d).
{cleanup section if cleanup_count > 0}
{IF blocking issue tracked in memory/issues/INDEX.md}
blocked by {ISS-ID} — {short reason}.
{end}

Record Arm C status DISCLOSURE_TRACKER_OK for the log.

---

Notify (shared, runs once at the end)

If the combined buffer .pending-notify-temp/vuln-tracker-${today}.md has at least one arm section, prepend the header line *Vuln Tracker — ${today}* and send once:

./notify -f .pending-notify-temp/vuln-tracker-${today}.md

If the buffer is empty (every arm that ran was silent), send nothing — a clean poll is silent. Keep the whole message under 4000 chars; if over, trim Arm A's body first (keep counts), then Arm C's cleanup list, keeping every needs-answer / newly-actionable / ESCALATE line.

Log (shared, runs once at the end)

Append to memory/logs/${today}.md under ONE heading (the health loop parses ### <skill-name>), with a discriminator line naming the scope/arms that ran:

### vuln-tracker
- scope: <full | prs | pvr | queue | single-advisory:GHSA-…>
- Arm A (PR/advisory): <status> — <total_in_json> JSON + <pre_history_prs> pre-history = <total> tracked; states: <merged> merged / <open_clean> open-clean / <needs_answer> needs-answer / <stale> stale / <aging> aging / <closed> closed / <queued> queued / <skipped> skipped-rechecked / <newly_actionable> newly-actionable
- Stars: ★<total_stars_secured> secured (<secured_repo_count> repos) / ★<total_stars_in_flight> in flight (<in_flight_repo_count> repos) / ★<total_stars_tracked> tracked (<total_repo_count> repos)
- Arm B (PVR triage): <status> — checked <total>; changed <changed_count> (<list>); aged>30d <aged_count>; still-triage <waiting_count>; tracking issue <url|none>
- Arm C (disclosure queue): <status> — <total> drafts (<critical_count> CRITICAL / <high_count> HIGH / <other_count> MED/LOW); <escalate_count> escalate / <pending_count> in-window / <operator_todo_count> operator-todo / <cleanup_count> cleanup; oldest in-window <repo> (<age>d); blocking issue <ISS-ID|none>
- Operator queue this run: <needs_answer + newly_actionable + aged PVRs + escalate_count>
- Notification: <sent | skipped (no movement)>
- VULN_TRACKER_OK

Only include the Arm B / Arm C lines for arms that actually ran under the current scope. Use each arm's own status token (VULN_TRACKER_OK / VULN_TRACKER_SKIP, PVRT_OK / PVRT_SKIP, DISCLOSURE_TRACKER_OK / DISCLOSURE_TRACKER_SKIP) inline, and close with VULN_TRACKER_OK for the hub run (or VULN_TRACKER_BAD_VAR if the var was unrecognized).

Required Env Vars

• GH_TOKEN / GITHUB_TOKEN — GitHub read token for Arm A's PR/comment/repo/advisory reads and the PVR-state endpoint. repo scope is enough. Present by default in GitHub Actions.
• GH_GLOBAL — GitHub PAT with public_repo + repository_advisories:write scope, used by Arm B to read the triage state of private/unpublished advisories you submitted. Same token vuln-scanner uses (all arms). Without it, Arm B advisory reads may 403 (treated as still triage).
• AEON_PVR_TRACKING_ISSUE (optional) — internal tracking-issue reference for Arm B cross-linking (else resolved from aeon.yml pvr_triage.tracking_issue:, else skipped).
• Arm C requires no env vars — all its data comes from local files written by vuln-scanner.

Pending-disclosure file schema (Arm B & Arm C)

memory/pending-disclosures/*.md files. Newer drafts use YAML frontmatter:

---
repo: owner/repo
ghsa: GHSA-xxxx-xxxx-xxxx
ghsa_url: https://github.com/owner/repo/security/advisories/GHSA-xxxx-xxxx-xxxx
channel: pvr                       # Arm B tracks only channel: pvr entries
state: triage                      # triage | draft | published | withdrawn
status: pending-operator-send      # optional; Arm C vocabulary below
submitted_at: 2026-05-12T19:54:42Z
last_checked: 2026-05-15           # added/updated by Arm B
severity: HIGH
cwe: [CWE-xxx]                      # may be a string or an array — handle both
patch_branch: https://github.com/<fork-owner>/repo/tree/security/branch-name
patch_commit: abc1234
submit_url: https://github.com/owner/repo/security/advisories/new
---

# {Repo}: {Title}
...

• Arm B required fields: repo, ghsa, channel: pvr, state, submitted_at. Optional: patch_branch, patch_commit, cwe, ghsa_url.
• Older drafts use inline **Severity:** lines and no frontmatter — parse defensively: grep for severity: and Severity: case-insensitively; if unparseable, default to MEDIUM.

status: controlled vocabulary (Arm C, step C2.5)

Set by vuln-scanner / operator / cleanup chores:

• (blank or missing) — pending; falls through to the PR-status cross-ref
• pending — same as blank
• pending-operator-send with auto_send: true - queued for autonomous send by vuln-scanner's disclose arm (var=disclose); classify as pending (NOT operator-todo). It flips to email-sent on the next disclose run, which only happens when vuln-scanner is dispatched with var=disclose (the default schedule runs the scan arm). If it lingers armed-but-unsent across several days, either no disclose run was dispatched or RESEND_API_KEY isn't configured - surface it once so the operator schedules or runs the disclose arm and wires up Resend.
• pending-operator-send with auto_send: false/absent / queued for operator manual send — operator-todo (human must send; e.g. non-email contact or AI-report ban)
• email-sent - sent by vuln-scanner's disclose arm via Resend; awaiting maintainer reply. Treat like submitted (informational cleanup candidate, never an escalation).
• email-failed - the disclose arm gave up after repeated send failures (bad address / Resend error). operator-todo — the contact needs fixing or a manual send.
• submitted / submitted-via-pvr — submitted, awaiting maintainer
• disclosed-via-pr-<N> — covered-by-pr, draft can be archived
• superseded-upstream — bypass is already fixed in upstream; draft is dead
• Any string containing operator — operator-todo

When a canonical PR lands but the draft's status: was never set, Arm C falls through to step C2.5 rule 2 (cross-ref pr-status.md) and classifies as covered-by-pr automatically. The status: shortcut just bypasses the grep.

When to delete a draft: cleanup candidates in the notification can be removed by the operator with rm memory/pending-disclosures/<file>.md. Safe once the canonical PR is open — the patch branch on the fork remains the authoritative artifact.

Notes & related

• vuln-scanned.json schema is loose — cwe may be a string or an array; advisory_ids may be present or absent. Handle both.
• Sibling skill: vuln-scanner produces the records this skill audits (its write actions — open PR, submit PVR, queue draft — are deliberately NOT here). vuln-scanner's re-submit arm probes repos waiting to open PVR; Arm B picks up once a PVR is submitted. inbox-triage catches inbound maintainer replies via the GitHub notification layer (complementary, not a duplicate of Arm A's branch-name lifecycle audit).
• Arm A, Arm B, and Arm C all read memory/pending-disclosures/ from different angles; coordinate via the shared memory/topics/vuln-followup.md dashboard to avoid duplicate escalation.