@@ -0,0 +1,12 @@
+---
+'@objectstack/skills': patch
+---
+
+fix(skills): objectstack-automation eval fixtures 5, 7 and 8 agree with their own assertions
+
+Clause-②: no
+Metadata-protocol: none
+
+- **Before.** Three ideal answers in `evals/flows-triggers-approvals.json` failed the assertions that grade them: eval 5 spelled the negated error-label string inside its answer, eval 7 never spelled the `type: 'position'` approver it requires and spelled the negated positions check, eval 8 never spelled the `type: 'screen'` flow kind it requires. Graded against those assertions, a correct answer scored wrong.
+- **Now.** The three ideal answers say each negation in words, eval 7 names `approvers: [{ type: 'position', value: 'finance_manager' }]`, and eval 8 opens with the flow kind. No fixture added or removed; the assertions and the skill's teaching are unchanged.
+- No schema, key or behaviour changes.
@@ -829,7 +829,16 @@ export const CEILINGS = new Map([
   // = 2724; headroom was 3 bytes, 0 tokens, so the ceiling moves +297, pinned
   // AT the measurement. Same maintainer ruling as the SKILL.md row above
   // (2026-10-27, tap A verbatim and untranslated): 「A:先削重复,余额按实测提高」.
-  ['skills/objectstack-automation/evals/flows-triggers-approvals.json', 2724],
+  // 10893 -> 10934 (#26515): no new fixture — three ideal answers rewritten
+  // to agree with their own assertions (eval 5 spelled the negated error-label
+  // string, eval 7 lacked the `type: 'position'` approver and spelled the
+  // negated positions check, eval 8 lacked `type: 'screen'`). Paid down first:
+  // evals 5 and 7 shortened by 25 and 7 bytes; the `approvers` line eval 7
+  // owed costs more than that. +41 bytes -> ceil(10934 / 4) = 2734; headroom
+  // was 3 bytes, 0 tokens, so the ceiling moves +10, pinned AT the
+  // measurement. Same maintainer ruling as the SKILL.md row above (2026-10-27,
+  // tap A verbatim and untranslated): 「A:先削重复,余额按实测提高」.
+  ['skills/objectstack-automation/evals/flows-triggers-approvals.json', 2734],
   // 894 -> 1095 (#26179): one regression case for the field permission `when`
   // (the contract prompt: lock the legal terms for finance from approval on,
   // legal keeps the edit) — expected: `when` on the finance set's entries with
 
@@ -44,7 +44,7 @@
     {
       "id": 5,
       "prompt": "Add a step that calls our `pricing.recalc` function to compute the discount and saves it on the quote, then posts the quote to our tax service over HTTP; if the tax call fails, route to a handler node instead of failing the run.",
-      "expected_output": "Emits a `script` node with `config.function: 'pricing.recalc'` (registered via `defineStack({ functions })`), `inputs`, and `outputVariable: 'recalc'`; the function is a pure compute step, so a following `update_record` persists it with `fields: { discount: { dialect: 'cel', source: 'recalc.discount' } }`. The `http` node sets `timeoutMs`. Failure routing is a `type: 'fault'` edge — a `label: 'error'` alone routes nothing.",
+      "expected_output": "Emits a `script` node with `config.function: 'pricing.recalc'` (registered via `defineStack({ functions })`), `inputs`, and `outputVariable: 'recalc'`; a pure compute step, so a following `update_record` persists it with `fields: { discount: { dialect: 'cel', source: 'recalc.discount' } }`. The `http` node sets `timeoutMs`. Failure routing is a `type: 'fault'` edge; a bare error label routes nothing.",
       "files": [],
       "assertions": {
         "must_contain": ["function:", "outputVariable", "type: 'fault'", "timeoutMs", "update_record", "dialect: 'cel'"],
@@ -64,7 +64,7 @@
     {
       "id": 7,
       "prompt": "An expense approval only holders of the `expense.approve_request` capability may start; route to the `finance_manager` position, and the submitter must never approve their own expense.",
-      "expected_output": "Flow-level `requiredPermissions: ['expense.approve_request']` (checked at every door that starts the flow, not a `current_user.positions` check inside it); the `approval` node sets `excludeSubmitter: true`, and its `onEmptyApprovers` is anything but `auto_approve`, which is refused beside it.",
+      "expected_output": "Flow-level `requiredPermissions: ['expense.approve_request']`, checked at every door that starts the flow (never a run-user positions check inside it); the `approval` node names `approvers: [{ type: 'position', value: 'finance_manager' }]`, sets `excludeSubmitter: true`, and its `onEmptyApprovers` is anything but `auto_approve`, refused beside it.",
       "files": [],
       "assertions": {
         "must_contain": ["requiredPermissions", "excludeSubmitter: true", "type: 'position'"],
@@ -74,7 +74,7 @@
     {
       "id": 8,
       "prompt": "A screen flow: the user uploads the signed contract PDF on the last step, then the flow creates the `contract` record with the file in its `signed_copy` field.",
-      "expected_output": "A `screen` field `{ name: 'signed_copy', type: 'file', required: true }`; the screen resumes with the uploaded `sys_file` id bound to `signed_copy`, and a `create_record` after the last screen writes `fields: { signed_copy: { dialect: 'cel', source: 'signed_copy' } }`, which claims the file for the record.",
+      "expected_output": "A `type: 'screen'` flow whose last `screen` carries `{ name: 'signed_copy', type: 'file', required: true }` and resumes with the uploaded `sys_file` id bound to `signed_copy`; a `create_record` after it writes `fields: { signed_copy: { dialect: 'cel', source: 'signed_copy' } }`, which claims the file for the record.",
       "files": [],
       "assertions": {
         "must_contain": ["type: 'screen'", "type: 'file'", "create_record", "signed_copy"],