Skip to content

🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.

chore(deps): add skill to track dependabot resolutions (#18167)

Nimar3b02ae42e4f3
diff --git a/.agents/skills/add-to-dependabot-csv/SKILL.md b/.agents/skills/add-to-dependabot-csv/SKILL.md
new file mode 100644
index 0000000000..83ccd923a0
--- /dev/null
+++ b/.agents/skills/add-to-dependabot-csv/SKILL.md
@@ -0,0 +1,102 @@
+---
+name: add-to-dependabot-csv
+description: Append GitHub Dependabot or Snyk/code-scanning alerts to an existing vulnerability CSV after verifying their API metadata. Use for "add to Dependabot CSV" or "dismissed it, add it to the list".
+---
+
+# Add to Dependabot CSV
+
+Update the local tracker only. Read GitHub; do not dismiss alerts, change dependencies,
+or commit files as part of this workflow.
+
+## 1. Locate and inspect
+
+Use the requested CSV. Otherwise look for `Dependabot List - CLEAN CURRENT.csv`
+in the working checkout. If absent or ambiguous, ask for the path; do not create
+or replace a tracker silently.
+
+Read its header and a few recent rows, including a dismissed row. Preserve column
+order, CSV quoting, encoding, line endings, and timestamp conventions. Snapshot
+original bytes before writing. Match fields by header, not column position.
+
+Deduplicate against both existing rows and the requested batch using the canonical
+GitHub `html_url` (ignore input query strings, fragments, and trailing slashes).
+Alert numbers and CVEs alone are not unique: Dependabot and code scanning have
+separate number spaces, and one CVE can have multiple image alerts. Skip an
+existing URL and report it; updating old rows requires an explicit request.
+
+## 2. Verify GitHub metadata
+
+Take owner, repository, alert type, and number from each supplied URL. For a bare
+number, resolve the alert type from context or ask; never guess the endpoint.
+
+```sh
+gh api 'repos/OWNER/REPO/dependabot/alerts/NUMBER'
+gh api 'repos/OWNER/REPO/code-scanning/alerts/NUMBER'
+gh api 'advisories?cve_id=CVE-YYYY-NNNNN'
+gh api 'advisories/GHSA-XXXX-XXXX-XXXX'
+```
+
+Fetch each distinct requested alert. Reuse embedded Dependabot advisory metadata;
+for code scanning, query a referenced GHSA or CVE once and match the package and
+ecosystem, rather than taking the first result. If there is no advisory ID, retain
+the scanner's available metadata and leave unsupported advisory fields blank.
+If the user says they dismissed it, verify `state`, `dismissed_at`,
+`dismissed_reason`, and `dismissed_comment`. Preserve GitHub's exact reason and
+comment, including punctuation. Report discrepancies with the user's account;
+do not silently record an unverified dismissal. If API access fails, stop before
+appending those rows and state what could not be verified.
+
+## 3. Map to the existing columns
+
+- Copy repository, alert number, URL, state, lifecycle timestamps, and dismissal
+ fields from the alert. Keep actual code-scanning URLs even in a Dependabot CSV.
+- For Dependabot, use `dependency`, `security_advisory`, and
+ `security_vulnerability` for package, scope, advisory, and affected/fixed versions.
+- For code scanning, use `rule.full_description`, `rule.description`,
+ `rule.security_severity_level`, `rule.tags`, and `rule.help`. Resolve CVSS,
+ advisory dates, references, and package-specific affected/fixed versions from
+ the matching GitHub advisory. Use scanner guidance when GitHub has no package
+ range; include that source in `advisory_references`.
+- Keep similarly named packages separate (`deepmerge` versus `deepmerge-ts`).
+ Do not apply another package's version range or confuse detection dates with
+ advisory publication dates. Leave unavailable fields blank; do not invent
+ classification, dependency scope, scores, or history.
+- Copy the dismissal comment into `dismissed_comment_or_delay_reasoning` when
+ that matches the existing convention. Use the user's supplied delay reasoning
+ for an open alert only when explicitly provided.
+- Keep API timestamps in their existing UTC format. This tracker uses
+ `Europe/Berlin` with an explicit offset for `dismissed_at_web_ui` and
+ `state_list_timestamps`; use timezone-aware conversion, including DST.
+- Keep `dismissed_at` as the current API value. This tracker's historical columns
+ use the earliest dismissal for `dismissed_at_web_ui` and `dismissed_reason`,
+ and the first non-empty dismissal comment for both comment columns. For an
+ alert just dismissed with no evidence of earlier transitions, use the current
+ API dismissal; do not require a timeline visit to prove no earlier event exists.
+- Pair known state transitions with their event timestamps: creation and current
+ dismissal yield `open|dismissed`; creation and fix yield `open|fixed`. Inspect
+ the authenticated timeline when there is evidence of earlier dismissals or
+ reopenings, or a later fix cleared known dismissal fields. If that history is
+ unavailable, preserve known events, leave only unverified historical fields
+ blank, and report the gap. Never claim an incomplete history is exhaustive.
+- Match existing separators: `; ` for CWEs, ` | ` for advisory references, and
+ `|` for state history and corresponding timestamps.
+
+## 4. Append and verify
+
+Serialize new rows with a CSV-aware writer and append only. Preserve timestamp
+strings exactly; spreadsheet tooling may coerce ISO strings into dates and lose
+the original offset. Check that the file still matches the snapshot immediately
+before writing. Preserve every original byte, adding a separator newline if needed.
+
+Reparse the saved CSV and verify:
+
+- Original bytes remain an exact prefix; original rows and header are unchanged.
+- Record count increases by the number of new, unique alert URLs.
+- Every row has the header's column count; new URLs occur exactly once.
+- New fields match their sources: current API state, verified historical
+ dismissal events when applicable, and any explicit user delay reasoning.
+ State timestamps align with their states.
+
+Finish with a short, valid Markdown file link, added/skipped alert numbers, final
+record/column counts, and confirmation that existing content is unchanged. Mention
+any unavailable metadata. No application tests are needed for this CSV-only edit.
diff --git a/.agents/skills/add-to-dependabot-csv/agents/openai.yaml b/.agents/skills/add-to-dependabot-csv/agents/openai.yaml
new file mode 100644
index 0000000000..910bd0b377
--- /dev/null
+++ b/.agents/skills/add-to-dependabot-csv/agents/openai.yaml
@@ -0,0 +1,4 @@
+interface:
+ display_name: "Add to Dependabot CSV"
+ short_description: "Verify GitHub alerts and append them to the CSV"
+ default_prompt: "Use $add-to-dependabot-csv to verify these GitHub alerts and append them to the current Dependabot CSV."