os-zhuang · Oct 27, 2026 · Oct 27, 2026
diff --git a/scripts/check-skills-token-ratchet.mjs b/scripts/check-skills-token-ratchet.mjs
@@ -391,7 +391,23 @@ export const CEILINGS = new Map([
   // 6004: +233 net, 9 absorbed by headroom, ceiling +224. Maintainer ruling
   // 2026-10-27 on the skills-lane decision card 「skills 修复卡的教学和 eval 都撞
   // token 天花板,怎么付?」, verbatim and untranslated: 「A:先削重复,余额按实测提高」.
-  ['skills/objectstack-automation/SKILL.md', 6004],
+  // 6004 -> 6164 (card #26347, the skills half of the `record_link` screen
+  // field): the `screen` node row teaches the read-only `type: 'record_link'`
+  // field — it shows the `reference` record whose id `defaultValue` evaluates
+  // to, both keys required, the five input keys refused, the resume refuses a
+  // value under its name — with the one result-screen example over a
+  // `create_record` `outputVariable`, and the two workarounds an author
+  // otherwise reaches for (a completed `end` `message`, a markdown link in a
+  // screen `description`). Paid down first: the Filter operands bullet
+  // restated Valid-but-silently-wrong item 1's refusal detail
+  // (`expression-invalid`, "with its envelope") and was cut to the bare
+  // refusal (-41 bytes). Arithmetic in priced bytes (stamp line excluded):
+  // 24014 -> 24654 (+681 added, -41 cut, +640 net) -> ceil(24654 / 4) = 6164;
+  // headroom was 2 bytes, 0 tokens, so the ceiling moves +160, pinned AT the
+  // measurement. Maintainer ruling 2026-10-27 on the seat's decision card
+  // 「skills 修复卡的教学和 eval 都撞 token 天花板,怎么付?」, tap A verbatim and
+  // untranslated: 「A:先削重复,余额按实测提高」.
+  ['skills/objectstack-automation/SKILL.md', 6164],
   // 13783 -> 13892 (2026-08-31 app-repo-principles raise, see the block above).
   // The invariant-vs-transition-gate choice, placed in the field-conditional-rule
   // bullet list where the tool is actually picked. Both wrong picks are silent:
@@ -737,7 +753,16 @@ export const CEILINGS = new Map([
   // +568 tokens, nothing in the file to delete, zero headroom, ceiling +568.
   // Same maintainer ruling as the SKILL.md row above (2026-10-27, verbatim and
   // untranslated): 「A:先削重复,余额按实测提高」.
-  ['skills/objectstack-automation/evals/flows-triggers-approvals.json', 1823],
+  // 1823 -> 2141 (card #26347): one eval prompt for the `record_link` result
+  // screen — a `convert_lead` screen flow that creates an account and a contact
+  // and ends on a screen linking both, the ideal answer written so it never
+  // spells the negated `type: 'record'`. Paid down first: the file's eight
+  // fixtures share no sentence, nothing to cut (0 bytes). 7289 -> 8561 bytes
+  // (+1272) -> ceil(8561 / 4) = 2141; headroom was 3 bytes, 0 tokens, so the
+  // ceiling moves +318, pinned AT the measurement. Same maintainer ruling as
+  // the SKILL.md row above (2026-10-27, verbatim and untranslated):
+  // 「A:先削重复,余额按实测提高」.
+  ['skills/objectstack-automation/evals/flows-triggers-approvals.json', 2141],
   // 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

@@ -88,7 +88,7 @@ Flows are built from **20 built-in node types** (plugins register more via
 | `notify` | Send a notification through the messaging service (inbox channel by default) |
 | `connector_action` | Invoke a pre-built integration connector |
 | `script` | Call a **registered** function named by `config.function` (see *Valid-but-silently-wrong* #3) |
-| `screen` | Display a UI form to the user (screen flows only). `config.fields[]`: `{ name, label, type, required, options, defaultValue }`, `type` an open widget hint; `'file'` renders the upload widget and resumes with the `sys_file` id bound to `name` — a later `create_record` / `update_record` writes it into a `file` field, which claims the file: collect first, write after the last screen |
+| `screen` | Display a UI form to the user (screen flows only). `config.fields[]`: `{ name, label, type, required, options, defaultValue }`, `type` an open widget hint; `'file'` renders the upload widget and resumes with the `sys_file` id bound to `name` — a later `create_record` / `update_record` writes it into a `file` field, which claims the file: collect first, write after the last screen. `'record_link'` (never `'record'`, an object field type) takes no input — it shows the `reference` record whose id `defaultValue` evaluates to, by display name, as a link to its record page: the result screen of a flow that created records, `{ name: 'account_link', type: 'record_link', reference: 'crm_account', defaultValue: { dialect: 'cel', source: 'account.id' } }` over a `create_record` `outputVariable: 'account'`. Both keys required; `required` / `options` / `placeholder` / `min` / `max` refused; a resume value under its name refused (`400 INVALID_SCREEN_INPUT`). A completed `end` `message` is refused and a markdown link in a screen `description` renders literally |
 
 #### Human Decision
 
@@ -206,8 +206,8 @@ variables: [
   evaluated before the query (`filter: { id: { dialect: 'cel', source:
   'record.id' } }`). A whole-string **query placeholder** (`{current_user_id}`,
   `{today}`, `{current_year_start}`) passes through verbatim for the query
-  engine to expand; every other `{…}` token is refused (`expression-invalid`)
-  with its envelope, and a flow variable never shadows a placeholder. Only
+  engine to expand; every other `{…}` token is refused, and a flow variable
+  never shadows a placeholder. Only
   `filter` gets this hand-off — in a text slot `{{ current_year_start }}` is a
   nonsense reference.
 - **An operand that evaluates to `null` refuses the node** (not

@@ -80,6 +80,16 @@
         "must_contain": ["type: 'screen'", "type: 'file'", "create_record", "signed_copy"],
         "must_not_contain": ["type: 'upload'", "type: 'attachment'"]
       }
+    },
+    {
+      "id": 9,
+      "prompt": "Build a screen flow `convert_lead` that asks for a company name and a contact name, creates a `crm_account` and a `crm_contact` linked to it, and ends on a screen that shows links to both new records.",
+      "expected_output": "A `screen` flow (`type: 'screen'`): a first `screen` node collecting `company_name` and `contact_name`; two `create_record` nodes with `outputVariable` (`account`, `contact`), the contact's `fields` linking it with `account: { dialect: 'cel', source: 'account.id' }`; then a last `screen` node whose fields are two `type: 'record_link'` fields, each with `reference` naming its object (`crm_account`, `crm_contact`) and `defaultValue: { dialect: 'cel', source: 'account.id' }` / `{ dialect: 'cel', source: 'contact.id' }`, carrying no `required`, `placeholder` or `options`. The `end` node carries no `message` (refused on a completed end) and the screen `description` holds no markdown link — the links are the `record_link` fields, spelled `record_link`, never the object field type.",
+      "files": [],
+      "assertions": {
+        "must_contain": ["record_link", "reference", "defaultValue", "outputVariable", "dialect: 'cel'"],
+        "must_not_contain": ["type: 'record'", "]("]
+      }
     }
   ]
 }