diff --git a/scripts/summary-comment.mjs b/scripts/summary-comment.mjs
index 5c0e4ba..f8faff6 100644
--- a/scripts/summary-comment.mjs
+++ b/scripts/summary-comment.mjs
@@ -3,7 +3,7 @@
 //
 //   node summary-comment.mjs <result.json> --tier cheap|strong --push <n>
 //     --gate pass|blocked [--prev <file with the previous comment body>]
-//     [--passes <n>] [--quiet] [--block-on P0,P1] [--held] [--fix-first P0,P1]
+//     [--passes <n>] [--quiet] [--block-on P0,P1]
 //
 // Prints the summary MARKDOWN to stdout. The driver (action.yml) writes it into
 // a marker-delimited region of the PR DESCRIPTION body (scripts/inject-summary.mjs),
@@ -41,9 +41,15 @@
 // a cheap pass withholding the strong review over a fix-first finding — counted
 // over the fix-first set instead, because block-on need not contain those
 // severities and "❌ 0 findings block merge" beside a "held" tier line
-// contradicts itself. There is no held run and no tier line now, so the count
+// contradicts itself. The cascade went, and the tier line with it, so the count
 // and the gate read the same set, which is the property that matters: the
 // summary cannot claim something the merge gate does not enforce.
+//
+// --held and --fix-first went with it here, having been parsed and then read by
+// nothing for as long as the exception has been gone. fix-first REMAINS an action
+// input — it stops the exhaustive loop early, see action.yml — and must not be
+// wired back into this count on the strength of that name: the two answer
+// different questions, which is what the paragraph above is about.
 
 import fs from "node:fs";
 import { SEVERITIES, countSeverities } from "./severity.mjs";
@@ -54,14 +60,13 @@ const STATE_RE = /<!-- orca-cr-state: (\{.*?\}) -->/;
 const usage = () => {
   console.error(
     "usage: node summary-comment.mjs <result.json> --tier cheap|strong --push <n> " +
-      "--gate pass|blocked [--prev <file>] [--passes <n>] [--quiet] [--block-on P0,P1] " +
-      "[--held] [--fix-first P0,P1]",
+      "--gate pass|blocked [--prev <file>] [--passes <n>] [--quiet] [--block-on P0,P1]",
   );
   process.exit(2);
 };
 
 const [file, ...rest] = process.argv.slice(2);
-const opts = { passes: "1", blockOn: "P0,P1", fixFirst: "P0,P1" };
+const opts = { passes: "1", blockOn: "P0,P1" };
 for (let i = 0; i < rest.length; i += 1) {
   if (rest[i] === "--tier") opts.tier = rest[++i];
   else if (rest[i] === "--push") opts.push = rest[++i];
@@ -70,20 +75,21 @@ for (let i = 0; i < rest.length; i += 1) {
   else if (rest[i] === "--passes") opts.passes = rest[++i];
   else if (rest[i] === "--quiet") opts.quiet = true;
   else if (rest[i] === "--block-on") opts.blockOn = rest[++i];
-  else if (rest[i] === "--held") opts.held = true;
-  else if (rest[i] === "--fix-first") opts.fixFirst = rest[++i];
+  // NO else: an unrecognised flag, and the value that follows it, are skipped
+  // one token at a time. That is what makes removing --fix-first safe to ship
+  // ahead of the caller that still passes it — the pair falls through without
+  // shifting the flags after it.
 }
 const push = Number(opts.push);
 const passes = Number(opts.passes);
-// Severity-set normalization shared by --block-on and --fix-first: trim +
-// uppercase, empty = the empty set (valid — "block on nothing" / no fix-first).
+// Severity-set normalization for --block-on: trim + uppercase, empty = the
+// empty set (valid — a deliberate "block on nothing").
 const parseSet = (v) =>
   String(v ?? "")
     .split(",")
     .map((s) => s.trim().toUpperCase())
     .filter(Boolean);
 const blockOn = parseSet(opts.blockOn);
-const fixFirst = parseSet(opts.fixFirst);
 if (
   !file ||
   !["pass", "blocked"].includes(opts.gate) ||
@@ -91,8 +97,7 @@ if (
   push < 1 ||
   !Number.isInteger(passes) ||
   passes < 1 ||
-  blockOn.some((s) => !SEVERITIES.includes(s)) ||
-  fixFirst.some((s) => !SEVERITIES.includes(s))
+  blockOn.some((s) => !SEVERITIES.includes(s))
 ) {
   usage();
 }
@@ -151,7 +156,7 @@ lines.push("");
 // Counted over block-on, which is what the gate enforces. The --held variant
 // (count over fix-first instead) went with the cascade: it existed so a withheld
 // escalation would not render "❌ 0 findings block merge" beside a held tier line,
-// and there is neither withholding nor a tier line any more.
+// and there is neither withholding nor a tier line any more. See the header.
 const blockingSet = blockOn;
 const blocking = blockingSet.reduce((n, s) => n + counts[s], 0);
 lines.push(
diff --git a/scripts/summary-comment.test.mjs b/scripts/summary-comment.test.mjs
index 7cac1d6..305940e 100644
--- a/scripts/summary-comment.test.mjs
+++ b/scripts/summary-comment.test.mjs
@@ -173,12 +173,15 @@ describe("the ❌ count follows block-on", () => {
   // the summary read "❌ 0 findings block merge" beside a held tier line. There is
   // no withholding any more, so what survives is the case that was never about
   // held runs at all.
-  test("the count follows --block-on, not --fix-first", () => {
+  test("the count follows --block-on, and a removed flag cannot steer it", () => {
     // --fix-first is deliberately a DIFFERENT set here: it used to steer this
-    // count on a held run, and now it must not steer it at all.
+    // count on a held run, and now it must not steer it at all. It is no longer
+    // parsed, and it sits in the MIDDLE of the argv on purpose — the flag AND
+    // its value have to be skipped without shifting the --block-on that follows,
+    // which is what lets this script ship ahead of a caller still passing it.
     const out = run(
       ["[P0] a", "[P2] b"],
-      ["--push", "1", "--gate", "blocked", "--block-on", "P2", "--fix-first", "P0"],
+      ["--push", "1", "--fix-first", "P0", "--gate", "blocked", "--block-on", "P2"],
     );
     assert.ok(out.includes("❌ 1 finding blocks merge"), `got:
 ${out}`);
diff --git a/skills/orca-review-action/references/inputs.md b/skills/orca-review-action/references/inputs.md
index 8463cb4..fb52b59 100644
--- a/skills/orca-review-action/references/inputs.md
+++ b/skills/orca-review-action/references/inputs.md
@@ -3,6 +3,12 @@
 Every input of `Continuum-AI-Corp/orca-code-review@v1`. Read the row before
 answering a question about behavior — do not guess a default.
 
+**These rows describe this action.** The hosted OrcaCode Review GitHub App runs a
+different review engine, so a setting of the same name can behave differently
+there — `fix-first` and `exhaustive` in particular. For the App, the console at
+**OrcaRouter → Apps → OrcaCode Review** is the authority; do not answer a question
+about the App from this file.
+
 ## Required
 
 | Input | Default | What it does |