Approve Exemption

Approve one or more Pending STO security exemptions. This skill covers the approval workflow after someone has requested an exemption (see /exempt-vuln for creation).

Pending scope is only PROJECT, PIPELINE, or TARGET. Exemptions are requested at project, pipeline, or target scope only (All Issues → project; Vulnerabilities tab → project, pipeline, or target). STO does not create pending requests at org or account scope — those wider scopes exist only after elevation on approve (body.scope ORG or ACCOUNT). You will not see scope: ORG or scope: ACCOUNT on rows from status: Pending in normal flows. If a list row shows ORG/ACCOUNT, it is already approved/widened (re-approve or a different workflow), not a fresh pending request this skill targets.

Harness exposes two different "scope" ideas — do not mix them up:

Users almost always say approve, even when they want org- or account-wide coverage. Treat "approve for org", "org-wide", and "at account level" as approve with elevation — the MCP server routes those to the promote endpoint internally. You never need a separate action='promote' call; always use action='approve' with the right body.scope.

Keep the Harness UI out of the loop. Do not ask the user to copy issue titles, CVEs, or exemption IDs from the platform. Your default opening move is to list pending exemptions in chat, show a numbered table, report total, and let them pick by row number (or narrow with search if they volunteer a keyword). Pagination surfaces more rows on request — you never dump the full queue at once.

Instructions

Step 1 — Establish project context

You need org and project unless they are already in session defaults or a pasted Harness URL provides them.

Exemptions list URL shape (auto-extracts org/project):

…/ng/account/{accountId}/all/orgs/{org}/projects/{project}/sto/exemptions

Step 2 — Surface pending exemptions (default opening move)

Unless the user already picked row numbers from a table you showed in this same session, call harness_list before asking them to choose anything. Do not send them to the Harness UI.

Show a paginated preview, not the whole queue

Chat is a narrow surface — dumping 50 pending rows is unreadable. Default behavior:

1. Call harness_list with status: Pending, size: 5, page: 0 (unless org/project are still unknown — resolve those first).
2. Read total from the response.
3. Render the table from items[] following _display_hint. Number rows 1–N to match _action_id_by_row (1-based).
4. Tell the user how many pending exemptions exist and how to proceed:
   • If total ≤ 5: show every row; ask which to approve.
   • If total > 5: show the first page and say something like:

     {total} pending exemption(s) in this project. Showing rows 1–5 below. Pick by number (e.g. 1, 1 and 3, 2-4), say next for the next page, or give a search term to narrow (CVE, requester, issue title).

5. For next page, follow _nextPageHint exactly — keep size, status, and any search identical; only increment page.

harness_list
  resource_type: "security_exemption"
  org_id: <org>
  project_id: <project>
  filters:
    status: "Pending"
    size: 5
    page: 0
    search: "<optional — only when user narrows>"

Rules:

• status must be a single value (Pending), not comma-separated.
• Default size: 5 inside filters on every call in a pagination session. Do not bump size mid-session unless the user explicitly asks for a larger page (e.g. "show 10").
• Keep size and all other filters identical when paginating; follow _nextPageHint verbatim.
• Never pass resource_scope, and never set org_id / project_id to words like org or account — those are approval-scope phrases, not list overrides.

The list response is optimized for chat:

• items[] — display fields only (issue_title, severity, type, requested_by, target, scope, reason, …). No exemption IDs in the table.
• _action_id_by_row — maps row number (1-based on the current page) → exemption_id for execute calls.
• _display_hint — column layout; follow it.
• Each row's scope is the exemption's requested scope — expect only TARGET, PIPELINE, or PROJECT for pending rows.

The selection prompt

After showing the table, always include:

1. total pending count (project-wide, not just this page).
2. How to pick — by row number from the table you just rendered.
3. Sample inputs — copy-pasteable examples covering scope and comments, so the user does not have to guess the syntax (this is the part users get stuck on).
4. How to see more — "next" / "next page" (pagination) or a search term to narrow.

Render the prompt with a sample-inputs block every time. Do not paraphrase it down to "pick a number" — users need to see that they can mix scope and comments in one message. Template:

12 pending exemptions in this project. Showing rows 1–5 below.

Reply with row numbers and (optionally) the approval scope + a comment per row. Examples you can adapt:

• 1 — approve row 1 at its current scope, no comment
• 1 as-is — same, more explicit
• 2 for project — widen row 2 to project scope
• 3 for org / 3 org-wide — elevate row 3 to org scope
• 4 for account — elevate row 4 to account scope
• 5 with comment "JIRA SEC-440, fix tracked" — approve at current scope with a note
• 2 for org with comment "approved by AppSec review" — scope + comment together
• 1, 3, 5 as-is — approve multiple rows at current scope
• 1 as-is, 2 for org, 3 for account with comment "exception logged" — mixed scopes + comment

Say next for rows 6–10, or give a search term (CVE, requester, title) to narrow.

Only add search to the list call when the user gives a keyword — do not require them to know CVE or title text upfront. If the user just replies 1, 2 with no scope, default to CURRENT (approve as-is) and ask once whether they want a comment before executing.

When listing returns zero rows

Say there are no pending exemptions in this project (for the current filter). Do not ask them to check the UI — offer to list without search if they had narrowed, or confirm org/project.

Step 3 — Resolve which rows to approve

Primary path: the user picks row numbers from your table. Map through _action_id_by_row on the same page they selected from. If they say "next" and pick from a later page, re-list that page (or keep page context) before mapping numbers.

When multiple rows match one search term, show the filtered table and ask which row(s) — do not guess.

If the user gives a vague approve request with no row numbers yet ("approve pending exemptions", "sign off waivers"), go back to Step 2 — list first, then ask them to pick. Do not execute until they choose from the surfaced table (or confirm "all on this page" with explicit scope).

For a mixed batch, the user may assign a different approval scope per exemption in one message, for example:

Approve #1 as-is, #2 for org, #3 for account.

Build an internal plan: { exemption_id, approval_scope, issue_title }[] — one entry per exemption, scopes may differ.

Step 4 — Map user intent to body.scope

body.scope is required on every approve call. It is the approval scope (where the exemption takes effect), not the pending row's current scope label.

Elevation paths (pending → destination)

body.scope is always one of CURRENT, PROJECT, ORG, or ACCOUNT. It names the destination after approval. The list row's scope is always the starting request scope (TARGET, PIPELINE, or PROJECT only). Elevation uses /promote when the destination is wider than “approve as-is”; the MCP routes that from action='approve'.

There is no pending row that starts at ORG or ACCOUNT — creation cannot request those scopes (canCreate is false at org/account in sto-core). Do not plan paths like “org → org” or “org → account” on the Pending queue. ORG / ACCOUNT in the table above are only body.scope destinations when widening from target, pipeline, or project.

Skip-level elevation: A pending target (or pipeline) row can go to org or account in one call — the user does not need a separate project approval step first unless they ask for project scope explicitly.

Project → project: When the row is already PROJECT and the user wants no widening, use CURRENT. Do not pass body.scope: "PROJECT" for that — PROJECT on execute means “promote to project scope” (for target/pipeline pending rows). Plain “approve this project exemption” → CURRENT.

Not supported (do not offer):

Org / account execute params: When body.scope is ORG, omit project_id on the execute call. When body.scope is ACCOUNT, omit both org_id and project_id (MCP preflight clears them). For CURRENT and PROJECT, keep normal project context.

Natural language → body.scope

Default: If the user says plain "approve" with no elevation hint, use CURRENT.

Per-row overrides: In a mixed list, honor each row's stated scope. Row 2 can be ORG while row 3 is ACCOUNT in the same session.

Optional fields (same for every row unless the user specifies otherwise):

Do not use harness_execute with action='promote'. The approve action accepts body.scope and picks /approve vs /promote automatically (CURRENT → approve endpoint; ORG / ACCOUNT / PROJECT → promote endpoint).

Step 5 — Confirm before executing

Echo a compact confirmation table before any write:

About to approve N exemption(s):
  # | Issue                          | Pending scope | Approve at   | Comment
  1 | CVE-2024-1234 in log4j-core    | TARGET        | CURRENT      | —
  2 | SQL Injection in api-gateway   | PROJECT       | ORG          | SEC-440
  3 | Spring4Shell                   | PROJECT       | ACCOUNT      | —
Proceed? (yes/no)

Wait for explicit yes (or pass confirm: true on harness_execute only after the user confirms in clients that cannot elicit interactively).

Step 6 — Execute approvals (one call per exemption)

There is no bulk approve API. Mixed scopes require sequential harness_execute calls — one per exemption, each with its own body.scope:

harness_execute
  resource_type: "security_exemption"
  action: "approve"
  resource_id: <exemption_id from _action_id_by_row or user>
  org_id: <org>          # omit org_id/project_id when body.scope is ACCOUNT
  project_id: <project>  # omit project_id when body.scope is ORG or ACCOUNT
  body:
    scope: <CURRENT | ORG | ACCOUNT | PROJECT>
    comment: "<optional>"

After each call, record success or failure. If call 2 of 3 fails, report which succeeded and which failed — do not silently retry with a different scope.

Step 7 — Report results

For each approved exemption, return:

• Issue title (from the plan)
• Previous pending scope and the body.scope used
• New status from the API response (expect Approved when successful)
• Deep link: /ng/account/{accountId}/all/orgs/{org}/projects/{project}/sto/exemptions

If the user's RBAC does not allow the requested elevation, surface the API error verbatim and suggest a lower scope (for example CURRENT or PROJECT) or a different approver.

Examples

Example 1 — Default flow: list in chat, user picks by row (no UI copy)

User: "Approve pending exemptions."

1. Confirm org/project from session or URL.
2. harness_list with status: Pending, size: 5, page: 0.
3. Show numbered table + total (e.g. "8 pending, showing 1–5").
4. User: "#2 and #4, as-is." → map _action_id_by_row[2] and [4], both CURRENT.
5. Confirm table, then execute.

Do not ask them to open Harness or paste issue names first.

Example 2 — Pagination

User: "Next page" after seeing rows 1–5 of 12.

1. Follow _nextPageHint (page: 1, same size: 5).
2. Show rows 6–10 (numbered 1–5 on this page — clarify page-relative numbers or renumber as 6–10 in the display so picks are unambiguous).
3. User picks; map via _action_id_by_row for that page's response.

Example 3 — Approve one row at current scope (after list)

User: "Approve pending exemptions" → sees table → "#1, as-is."

1. Row 1 pending scope: TARGET → CURRENT.
2. Confirm, then harness_execute with resource_id from _action_id_by_row[1].

Example 4 — Target → project elevation (user says "approve for project")

User picks row from table: "#1 for the whole project."

1. List already showed scope: TARGET on row 1.
2. "For project" → body.scope: "PROJECT".
3. Confirm Pending scope: TARGET → Approve at: PROJECT, then execute.

Example 5 — Org-wide approval (user says "approve", not "promote")

User: "#2 for org" (after selecting from surfaced table).

1. Map row 2 → body.scope: "ORG".
2. Confirm showing Approve at: ORG, execute with action: "approve".

Example 6 — Target → org in one step (skip project)

User: "Approve row 3 for org" from the pending table.

1. Row 3 shows scope: TARGET.
2. body.scope: "ORG" — omit project_id on execute.

Example 7 — Mixed scopes in one request

User: "Approve #1 as-is, #2 for org, and #3 for account." (from the same page table).

1. Plan three rows with CURRENT, ORG, ACCOUNT.
2. One confirmation table; three sequential executes.

Example 7b — Inline scope + comment in one message

User: "Approve #2 for org with comment "AppSec reviewed, JIRA SEC-440" and #4 as-is."

1. Row 2 → body.scope: "ORG", body.comment: "AppSec reviewed, JIRA SEC-440".
2. Row 4 → body.scope: "CURRENT", no comment.
3. Confirm both lines (showing the comment column), then execute sequentially.

Do not strip or rewrite the quoted comment text.

Example 8 — Optional search narrow (user volunteers a keyword)

User: "Show pending log4j exemptions."

1. List with search: "log4j", page: 0, size: 5.
2. Show filtered table; user picks by row number.

Example 9 — Reject is out of scope unless asked

User: "Reject the XSS exemption."

This skill is for approval. Use harness_execute with action: "reject" and the same resource_id lookup pattern, or handle via a separate reject workflow if the user insists.

Performance Notes

• List-first workflow — one small harness_list preview beats asking the user to hunt in the UI. Default size: 5; paginate with _nextPageHint when they say "next".
• Listing is POST-based and paginated; keep size constant across pages in a session.
• Approvals are high_write with do_not_retry — surface errors to the user instead of auto-retrying.
• Sequential executes are correct for mixed scopes; do not parallelize writes.
• approver_id is derived once per call from the PAT — no extra user lookup needed.
• Add search only when the user narrows — not as a substitute for showing the pending queue.

Troubleshooting