From 9a1ea31bc94629cde109ebc9f75963a375a06681 Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 07:15:53 +0300
Subject: [PATCH 01/10] feat(ai-review): show model used and re-run hint on
 review comments

Refs #51
---
 ai-review/lib/publish.js      | 25 ++++++++++++++++++++++---
 ai-review/lib/publish.test.js | 23 +++++++++++++++++++++++
 2 files changed, 45 insertions(+), 3 deletions(-)

diff --git a/ai-review/lib/publish.js b/ai-review/lib/publish.js
index df279a6..a9f1f89 100644
--- a/ai-review/lib/publish.js
+++ b/ai-review/lib/publish.js
@@ -45,9 +45,21 @@ function stripLeadingBannerArtifacts(markdown) {
 /**
  * @param {{verdict: string, confidence: number, mergeRisk: string,
  *   counts: {p0:number,p1:number,p2:number,p3:number}, intentDeviated: boolean,
- *   modelVerdict: string|undefined, blockers: string[], commentBody: string}} args
+ *   modelVerdict: string|undefined, blockers: string[], commentBody: string,
+ *   modelUsed?: string|null}} args
  *   `commentBody` must already be run through stripLeadingBannerArtifacts.
  */
+function modelLine(modelUsed) {
+  if (!modelUsed || typeof modelUsed !== "string" || !modelUsed.trim()) return [];
+  return ["", `Model: \`${modelUsed.trim()}\``];
+}
+
+function modelFooter(modelUsed) {
+  const line = modelLine(modelUsed);
+  if (!line.length) return [];
+  return [...line, "_Re-run this job if you need another review pass._"];
+}
+
 function buildReviewBody({
   verdict,
   confidence,
@@ -57,6 +69,7 @@ function buildReviewBody({
   modelVerdict,
   blockers,
   commentBody,
+  modelUsed,
 }) {
   const verdictLine = verdict === "pass" ? "**✅ PASS**" : "**❌ FAIL**";
   const rejectedBanner = intentDeviated ? "❌ **Rejected — wrong solution**\n\n" : "";
@@ -91,13 +104,18 @@ function buildReviewBody({
     "",
     `Confidence: ${confidence} · Merge risk: ${mergeRisk}`,
     `P0: ${counts.p0} · P1: ${counts.p1} · P2: ${counts.p2} · P3: ${counts.p3}`,
+    ...modelFooter(modelUsed),
     "",
     commentBody || "_No review content returned._",
   ].join("\n");
 }
 
-/** @param {string} salvaged possibly-empty text recovered from a missed structured output. */
-function buildInconclusiveBody(salvaged) {
+/**
+ * @param {string} salvaged possibly-empty text recovered from a missed structured output.
+ * @param {{modelUsed?: string|null}} [opts]
+ */
+function buildInconclusiveBody(salvaged, opts = {}) {
+  const modelUsed = opts && opts.modelUsed;
   return [
     "<!-- ai-review -->",
     "### ⚠️ AI Review — inconclusive (re-run required)",
@@ -109,6 +127,7 @@ function buildInconclusiveBody(salvaged) {
     "fails closed.",
     "",
     "**Re-run the `ai-review` job** to get a verdict.",
+    ...modelLine(modelUsed),
     ...(salvaged
       ? [
           "",
diff --git a/ai-review/lib/publish.test.js b/ai-review/lib/publish.test.js
index 4b23307..eea4f8f 100644
--- a/ai-review/lib/publish.test.js
+++ b/ai-review/lib/publish.test.js
@@ -121,6 +121,20 @@ test("always leads with the <!-- ai-review --> marker", () => {
   assert.match(body, /^<!-- ai-review -->\n/);
 });
 
+test("buildReviewBody includes model used and re-review hint", () => {
+  const body = buildReviewBody({
+    ...BASE_ARGS,
+    modelUsed: "claude/claude-sonnet-5",
+  });
+  assert.match(body, /Model: `claude\/claude-sonnet-5`/);
+  assert.match(body, /Re-run this job if you need another review pass/);
+});
+
+test("buildReviewBody omits Model line when modelUsed is empty", () => {
+  const body = buildReviewBody({ ...BASE_ARGS, modelUsed: "" });
+  assert.doesNotMatch(body, /^Model:/m);
+});
+
 // --- buildInconclusiveBody ---------------------------------------------------
 
 test("without salvaged text there is no details block", () => {
@@ -136,6 +150,15 @@ test("with salvaged text the details block contains it", () => {
   assert.match(body, /<\/details>/);
 });
 
+test("buildInconclusiveBody includes Model line but not a second italic re-run hint", () => {
+  const body = buildInconclusiveBody("salvaged text", {
+    modelUsed: "claude/cursor/composer-2.5",
+  });
+  assert.match(body, /Model: `claude\/cursor\/composer-2.5`/);
+  assert.match(body, /\*\*Re-run the `ai-review` job\*\*/);
+  assert.doesNotMatch(body, /_Re-run this job if you need another review pass\._/);
+});
+
 // --- tickVerifiedBoxes --------------------------------------------------------
 
 test("ticks a single matching unchecked box", () => {

From 68a2d98a37b226b905a1cbb72e99afd8a252cd82 Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 07:17:50 +0300
Subject: [PATCH 02/10] feat(ai-review): attribute PR review to the model that
 actually ran

Refs #51
---
 ai-review/action.yml          | 25 ++++++++++++++++++++++++-
 ai-review/lib/metrics.js      | 25 ++++++++++++++++++++++++-
 ai-review/lib/metrics.test.js | 20 +++++++++++++++++++-
 3 files changed, 67 insertions(+), 3 deletions(-)

diff --git a/ai-review/action.yml b/ai-review/action.yml
index 17af413..ddea172 100644
--- a/ai-review/action.yml
+++ b/ai-review/action.yml
@@ -1515,6 +1515,11 @@ runs:
           || steps.review.outputs.structured_output) }}
         RECOMPUTE_PATH: ${{ github.action_path }}/lib/recompute.js
         PUBLISH_LIB_PATH: ${{ github.action_path }}/lib/publish.js
+        METRICS_PATH: ${{ github.action_path }}/lib/metrics.js
+        ROUTED_MODEL: ${{ steps.route.outputs.model }}
+        REVIEW_LOG: ${{ runner.temp }}/ai-review-exec-review-snapshot.json
+        REPAIR_LOG: ${{ runner.temp }}/ai-review-exec-repair-snapshot.json
+        RETRY_LOG: ${{ steps.review_retry.outputs.execution_file }}
       with:
         github-token: ${{ steps.identity-refresh.outputs.author-token }}
         script: |
@@ -1528,6 +1533,23 @@ runs:
             buildStatusBlock,
             upsertStatusBlock,
           } = require(process.env.PUBLISH_LIB_PATH);
+          const { resolveModelUsed } = require(process.env.METRICS_PATH);
+
+          const readLog = (p) => {
+            try {
+              return JSON.parse(fs.readFileSync(p, "utf8"));
+            } catch {
+              return null;
+            }
+          };
+          const modelUsed = resolveModelUsed({
+            logs: [
+              readLog(process.env.RETRY_LOG),
+              readLog(process.env.REPAIR_LOG),
+              readLog(process.env.REVIEW_LOG),
+            ],
+            fallback: process.env.ROUTED_MODEL || "",
+          });
 
           const prNumber = Number(process.env.PR_NUMBER);
           const passLabel = process.env.PASS_LABEL;
@@ -1552,7 +1574,7 @@ runs:
             } catch (e) {
               // No salvage file (step skipped, or nothing recoverable).
             }
-            const inconclusiveBody = buildInconclusiveBody(salvaged);
+            const inconclusiveBody = buildInconclusiveBody(salvaged, { modelUsed });
             try {
               await github.rest.pulls.createReview({
                 owner: context.repo.owner,
@@ -1624,6 +1646,7 @@ runs:
             modelVerdict: review.verdict,
             blockers,
             commentBody,
+            modelUsed,
           });
 
           await github.rest.pulls.createReview({
diff --git a/ai-review/lib/metrics.js b/ai-review/lib/metrics.js
index de44c59..99be815 100644
--- a/ai-review/lib/metrics.js
+++ b/ai-review/lib/metrics.js
@@ -207,4 +207,27 @@ function renderSummary(metrics) {
   );
 }
 
-module.exports = { parseExecutionLog, collectMetrics, renderSummary, formatDuration, isStalled, STALL_MIN_MS };
+/**
+ * Prefer the first execution log that names a model (retry → repair → review
+ * order is the caller's responsibility). Fall back to the routed primary when
+ * logs are missing or silent.
+ * @param {{logs: unknown[], fallback: string}} args
+ * @returns {string}
+ */
+function resolveModelUsed({ logs, fallback }) {
+  for (const log of logs || []) {
+    const m = parseExecutionLog(log).model;
+    if (typeof m === "string" && m.trim()) return m.trim();
+  }
+  return typeof fallback === "string" ? fallback : "";
+}
+
+module.exports = {
+  parseExecutionLog,
+  collectMetrics,
+  renderSummary,
+  formatDuration,
+  isStalled,
+  STALL_MIN_MS,
+  resolveModelUsed,
+};
diff --git a/ai-review/lib/metrics.test.js b/ai-review/lib/metrics.test.js
index 9210341..5d339c4 100644
--- a/ai-review/lib/metrics.test.js
+++ b/ai-review/lib/metrics.test.js
@@ -3,7 +3,7 @@
 const test = require("node:test");
 const assert = require("node:assert/strict");
 
-const { parseExecutionLog, collectMetrics, renderSummary } = require("./metrics.js");
+const { parseExecutionLog, collectMetrics, renderSummary, resolveModelUsed } = require("./metrics.js");
 
 // A minimal execution log in claude-code-action's shape: a top-level array of
 // stream entries, terminated by a `result` entry. The shape is pinned by the
@@ -236,3 +236,21 @@ test("renderSummary surfaces a stall rather than burying it in the table", () =>
   assert.match(md, /stall/i);
   assert.match(md, /review/);
 });
+
+// --- resolveModelUsed ------------------------------------------------------
+
+test("resolveModelUsed prefers first log that names a model", () => {
+  const retry = logFor({ turns: 1, cost: 0, ms: 10, model: "oc/mimo-v2.5-free" });
+  const review = logFor({ turns: 1, cost: 0, ms: 10, model: "claude/claude-opus-5" });
+  assert.equal(
+    resolveModelUsed({ logs: [retry, review], fallback: "claude/claude-opus-5" }),
+    "oc/mimo-v2.5-free",
+  );
+});
+
+test("resolveModelUsed falls back when logs empty", () => {
+  assert.equal(
+    resolveModelUsed({ logs: [null, []], fallback: "claude/claude-sonnet-5" }),
+    "claude/claude-sonnet-5",
+  );
+});

From ebaceeb0669e5df311afbb0c00b2e85e9cdfb957 Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 07:19:31 +0300
Subject: [PATCH 03/10] feat(ai-review): lock Claude primaries with Cursor then
 free fallbacks

Refs #51
---
 ai-review/README.md             | 12 +++---
 ai-review/action.yml            | 69 +++++++++++++--------------------
 ai-review/lib/write-manifest.js |  6 +--
 3 files changed, 35 insertions(+), 52 deletions(-)

diff --git a/ai-review/README.md b/ai-review/README.md
index 2475487..aea28d4 100644
--- a/ai-review/README.md
+++ b/ai-review/README.md
@@ -59,8 +59,9 @@ injection-safety rule.
    derives the diff base from a false premise reviews the wrong range and
    reports confidently on it.
 
-   The same step routes ordinary diffs to `sonnet-model`, escalating to
-   `opus-model` once a diff exceeds **either** `sonnet-files-threshold`
+   The same step routes ordinary diffs to the locked Sonnet primary
+   (`claude/claude-sonnet-5`), escalating to Opus (`claude/claude-opus-5`)
+   once a diff exceeds **either** `sonnet-files-threshold`
    (25) or `sonnet-churn-threshold` (800). These were briefly 3/60, which
    sent nearly every real PR to Opus and moved the review stage from
    ~10-13 min to a 35-min median. Widened 15/400 → 25/800 after measuring
@@ -204,11 +205,8 @@ injection-safety rule.
 | `qa-pass-label` | Post-merge `ai-qa` pass label; cleared (not applied) by this action on every new commit. | No | `✓ /ai-qa` |
 | `qa-fail-label` | Post-merge `ai-qa` fail label; cleared (not applied) by this action on every new commit. | No | `✗ /ai-qa` |
 | `confidence-threshold` | Minimum **blocking-finding** confidence (0-100) required for a pass. The Publish step recomputes confidence from the review stage's P0/P1 counts and test-quality signals and compares it against this threshold. P2/P3 findings lower the *reported* confidence but are advisory and never block. | No | `90` |
-| `sonnet-files-threshold` | Max changed-file count for a diff to still route to `sonnet-model` (must hold together with `sonnet-churn-threshold`); larger diffs route to `opus-model`. | No | `25` |
-| `sonnet-churn-threshold` | Max changed-line count (adds + deletes) for a diff to still route to `sonnet-model`. | No | `800` |
-| `sonnet-model` | Model the routing step selects for diffs within **both** thresholds. Override when a gateway aliases model names. | No | `claude-sonnet-5` |
-| `opus-model` | Model the routing step selects for every larger diff. Override when a gateway aliases model names. | No | `claude-opus-5` |
-| `haiku-model` | Model used by the context stage, and stamped on the roster's `history`/`scorer` roles by the prep step. Note: Haiku 4.5 does not accept the `effort` parameter, so no stage or role running it passes `--effort` — the roster resolves that against the model id, not the tier, so overriding another tier to a literal Haiku id (e.g. `sonnet-model: claude-haiku-4-5`) is also covered. A gateway alias that *routes* to Haiku under an unrelated string is not detected; the check is a substring match, not alias resolution. | No | `claude-haiku-4-5` |
+| `sonnet-files-threshold` | Max changed-file count for a diff to still route to the locked Sonnet primary (must hold together with `sonnet-churn-threshold`); larger diffs route to Opus. | No | `25` |
+| `sonnet-churn-threshold` | Max changed-line count (adds + deletes) for a diff to still route to Sonnet. | No | `800` |
 | `enable-context-stage` | When `false`, skips the Haiku context stage (and its `context.md` verification) entirely. The stage is best-effort and its output optional, so disabling it removes a wall-clock risk without changing the gate contract. | No | `true` |
 | `api-timeout-ms` | Per-request timeout (ms) for every Claude stage, passed as `API_TIMEOUT_MS` (CLI default `600000`). **Does not bound the ~27.5-min stall** — a run with this set to `180000` still stalled 27m36s. It is a genuine per-request bound and fails a wedged request faster than the default, nothing more. | No | `180000` |
 | `test-command` | **DEPRECATED — accepted but ignored.** The Review stage no longer runs tests; see [Why the review no longer runs tests](#why-the-review-no-longer-runs-tests). | No | — |
diff --git a/ai-review/action.yml b/ai-review/action.yml
index ddea172..7afdcf0 100644
--- a/ai-review/action.yml
+++ b/ai-review/action.yml
@@ -94,30 +94,12 @@ inputs:
       docs/adr/0004-non-blocking-findings-and-structured-output-repair.md.
     required: false
     default: "90"
-  sonnet-model:
-    description: >
-      Model ID used by the diff-size routing step ONLY for tiny diffs
-      (≤`sonnet-files-threshold` files AND ≤`sonnet-churn-threshold`
-      changed lines). Every larger diff routes to `opus-model`. Consumed
-      by the routing step, whose output feeds the Review stage as the
-      model to run. Override when a gateway aliases model names.
-    required: false
-    default: "claude-sonnet-5"
-  opus-model:
-    description: >
-      Model ID used by the diff-size routing step for all non-tiny diffs
-      (the default — the review stage prefers Opus, dropping to
-      `sonnet-model` only for trivially small changes). Consumed by the
-      routing step, whose output feeds the Review stage as the model to
-      run. Override when a gateway aliases model names.
-    required: false
-    default: "claude-opus-5"
   sonnet-files-threshold:
     description: >
       Max changed-file count for a diff to still be considered "tiny" and
-      routed to `sonnet-model`. A diff qualifies for Sonnet only when it
+      routed to the locked Sonnet primary. A diff qualifies for Sonnet only when it
       is at or under BOTH this and `sonnet-churn-threshold`; anything
-      larger routes to `opus-model`. Widened to 25 (from 15) — measured over
+      larger routes to the locked Opus primary. Widened to 25 (from 15) — measured over
       682 review jobs, 89% of all traffic routed to Opus, and Opus runs cost
       3x the wall-clock (28.1 min median vs 9.2) and 4x the spend ($13.22 vs
       $3.27) of Sonnet ones. This is NOT a depth concession: the parallel
@@ -129,23 +111,11 @@ inputs:
   sonnet-churn-threshold:
     description: >
       Max changed-line count (additions + deletions) for a diff to still
-      be considered "tiny" and routed to `sonnet-model`. Applied together
+      be considered "tiny" and routed to the locked Sonnet primary. Applied together
       with `sonnet-files-threshold` (both must hold for Sonnet). Widened to
       800 alongside `sonnet-files-threshold` — see the note there.
     required: false
     default: "800"
-  haiku-model:
-    description: >
-      Model ID used by the context stage (this phase) to read the diff and changed
-      files and write context.md for the review stage to consume. Also stamped by
-      the deterministic-prep step onto the roster's `history` and `scorer` roles,
-      so an override here reaches those too. Override when a gateway aliases model
-      names. NOTE: Haiku 4.5 does not accept the `effort` parameter, so no stage
-      running this model may pass `--effort` — the roster resolves that against
-      the model id rather than the tier, so pointing another tier at a Haiku id is
-      covered as well.
-    required: false
-    default: "claude-haiku-4-5"
   enable-context-stage:
     description: >
       When 'true' (default) the Haiku context stage runs and writes
@@ -438,12 +408,8 @@ runs:
       shell: bash
       env:
         BASE_REF: ${{ steps.pr-state.outputs.base-ref }}
-        SONNET: ${{ inputs.sonnet-model }}
-        OPUS: ${{ inputs.opus-model }}
-        # Read by lib/write-manifest.js only (lib/roster.js is pure and never
-        # touches the environment), to stamp the OSH tiering onto the roster
-        # with the consumer's overrides rather than hardcoded ids.
-        HAIKU: ${{ inputs.haiku-model }}
+        # Locked at action level — keep in sync with ai-qa cascade comments.
+        # Claude primary → Cursor if Claude blocked → free if Cursor blocked.
         FILES_MAX: ${{ inputs.sonnet-files-threshold }}
         CHURN_MAX: ${{ inputs.sonnet-churn-threshold }}
         MANIFEST_CLI: ${{ github.action_path }}/lib/write-manifest.js
@@ -454,6 +420,16 @@ runs:
         set -euo pipefail
         mkdir -p .ai-review
 
+        # Locked model IDs (not inputs). Keep in sync with ai-qa.
+        SONNET="claude/claude-sonnet-5"
+        OPUS="claude/claude-opus-5"
+        HAIKU="claude/claude-haiku-4-5-20251001"
+        FREE="oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free"
+        SONNET_FALLBACK="claude/cursor/claude-4.6-sonnet-medium-thinking,${FREE}"
+        OPUS_FALLBACK="claude/cursor/claude-opus-4-8-medium-fast,${FREE}"
+        HAIKU_FALLBACK="claude/cursor/composer-2.5,${FREE}"
+        export SONNET OPUS HAIKU
+
         # Resolve the review range once, deterministically. `git diff A...B`
         # already means "merge base of A and B", but resolving the SHA
         # explicitly puts it in the manifest and the job log, so a wrong-range
@@ -499,14 +475,19 @@ runs:
         FILES_MAX="${FILES_MAX:-25}"
         CHURN_MAX="${CHURN_MAX:-800}"
         if [ "${FILES}" -le "${FILES_MAX}" ] && [ "${CHURN}" -le "${CHURN_MAX}" ]; then
-          MODEL="${SONNET:-claude-sonnet-5}"
+          MODEL="${SONNET}"
+          FALLBACK="${SONNET_FALLBACK}"
         else
-          MODEL="${OPUS:-claude-opus-5}"
+          MODEL="${OPUS}"
+          FALLBACK="${OPUS_FALLBACK}"
         fi
         {
           echo "files=${FILES}"
           echo "churn=${CHURN}"
           echo "model=${MODEL}"
+          echo "fallback-model=${FALLBACK}"
+          echo "haiku-model=${HAIKU}"
+          echo "haiku-fallback-model=${HAIKU_FALLBACK}"
         } >> "${GITHUB_OUTPUT}"
 
     - name: Resolve linked issues
@@ -586,7 +567,8 @@ runs:
         # Review stage's relative-path read of `context.md` at the repository root
         # resolves to the same file.
         claude_args: |
-          --model ${{ inputs.haiku-model }}
+          --model ${{ steps.route.outputs.haiku-model }}
+          --fallback-model ${{ steps.route.outputs.haiku-fallback-model }}
           --allowedTools "Read,Grep,Glob,Write,Bash(git diff:*),Bash(git symbolic-ref:*),Bash(git remote show:*),Bash(git merge-base:*),Bash(git log:*),Bash(git show:*)"
         prompt: |
           Read `.ai-review/manifest.json` first: a prior deterministic step
@@ -748,6 +730,7 @@ runs:
         plugin_marketplaces: https://github.com/obra/superpowers-marketplace.git
         claude_args: |
           --model ${{ steps.route.outputs.model }}
+          --fallback-model ${{ steps.route.outputs.fallback-model }}
           --max-turns 200
           --allowedTools "Read,Grep,Glob,Bash(gh pr view:*),Bash(gh issue view:*),Bash(git diff:*),Bash(git symbolic-ref:*),Bash(git merge-base:*),Bash(git remote show:*),Bash(git log:*),Bash(git show:*),Bash(cat:*),Bash(head:*),Bash(tail:*),Bash(ls:*),Bash(find:*),Bash(wc:*)"
           --json-schema '{"type":"object","additionalProperties":false,"required":["verdict","confidence","merge_risk","intent","counts","review_event","comment_markdown"],"properties":{"verdict":{"type":"string","enum":["pass","fail"]},"confidence":{"type":"integer","minimum":0,"maximum":100},"merge_risk":{"type":"string","enum":["low","med","high"]},"intent":{"type":"string","enum":["aligned","partial","deviated","skipped"]},"counts":{"type":"object","additionalProperties":false,"required":["p0","p1","p2","p3"],"properties":{"p0":{"type":"integer","minimum":0},"p1":{"type":"integer","minimum":0},"p2":{"type":"integer","minimum":0},"p3":{"type":"integer","minimum":0}}},"review_event":{"type":"string","enum":["APPROVE","REQUEST_CHANGES"]},"comment_markdown":{"type":"string"},"findings":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["id","file","line","severity","summary","failure_scenario","reason","evidence","confidence"],"properties":{"id":{"type":"string"},"file":{"type":"string"},"line":{"type":"integer","minimum":0},"severity":{"type":"string","enum":["P0","P1","P2","P3"]},"summary":{"type":"string"},"failure_scenario":{"type":"string"},"reason":{"type":"string"},"evidence":{"type":"string"},"confidence":{"type":"integer","enum":[0,25,50,75,100]},"severity_confirmed":{"type":"string","enum":["P0","P1","P2","P3"]}}}},"files_reviewed":{"type":"array","items":{"type":"string"}},"tests_failing":{"type":"boolean"},"coverage_below_threshold_on_critical_paths":{"type":"boolean"},"no_tests_for_changed_logic":{"type":"boolean"},"test_execution":{"type":"string","enum":["passed","failed","skipped","not_run"]},"verification_evidence":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["claim","command","result"],"properties":{"claim":{"type":"string"},"command":{"type":"string"},"result":{"type":"string"}}}},"checklist":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["text","status"],"properties":{"text":{"type":"string"},"status":{"type":"string","enum":["verified","failed","unverifiable"]},"evidence":{"type":"string"}}}}}}'
@@ -957,6 +940,7 @@ runs:
         claude_args: |
           --resume ${{ steps.review.outputs.session_id }}
           --model ${{ steps.route.outputs.model }}
+          --fallback-model ${{ steps.route.outputs.fallback-model }}
           --allowedTools "Read"
           --json-schema '{"type":"object","additionalProperties":false,"required":["verdict","confidence","merge_risk","intent","counts","review_event","comment_markdown"],"properties":{"verdict":{"type":"string","enum":["pass","fail"]},"confidence":{"type":"integer","minimum":0,"maximum":100},"merge_risk":{"type":"string","enum":["low","med","high"]},"intent":{"type":"string","enum":["aligned","partial","deviated","skipped"]},"counts":{"type":"object","additionalProperties":false,"required":["p0","p1","p2","p3"],"properties":{"p0":{"type":"integer","minimum":0},"p1":{"type":"integer","minimum":0},"p2":{"type":"integer","minimum":0},"p3":{"type":"integer","minimum":0}}},"review_event":{"type":"string","enum":["APPROVE","REQUEST_CHANGES"]},"comment_markdown":{"type":"string"},"findings":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["id","file","line","severity","summary","failure_scenario","reason","evidence","confidence"],"properties":{"id":{"type":"string"},"file":{"type":"string"},"line":{"type":"integer","minimum":0},"severity":{"type":"string","enum":["P0","P1","P2","P3"]},"summary":{"type":"string"},"failure_scenario":{"type":"string"},"reason":{"type":"string"},"evidence":{"type":"string"},"confidence":{"type":"integer","enum":[0,25,50,75,100]},"severity_confirmed":{"type":"string","enum":["P0","P1","P2","P3"]}}}},"files_reviewed":{"type":"array","items":{"type":"string"}},"tests_failing":{"type":"boolean"},"coverage_below_threshold_on_critical_paths":{"type":"boolean"},"no_tests_for_changed_logic":{"type":"boolean"},"test_execution":{"type":"string","enum":["passed","failed","skipped","not_run"]},"verification_evidence":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["claim","command","result"],"properties":{"claim":{"type":"string"},"command":{"type":"string"},"result":{"type":"string"}}}},"checklist":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["text","status"],"properties":{"text":{"type":"string"},"status":{"type":"string","enum":["verified","failed","unverifiable"]},"evidence":{"type":"string"}}}}}}'
         prompt: |
@@ -1045,6 +1029,7 @@ runs:
         plugin_marketplaces: https://github.com/obra/superpowers-marketplace.git
         claude_args: |
           --model ${{ steps.route.outputs.model }}
+          --fallback-model ${{ steps.route.outputs.fallback-model }}
           --max-turns 200
           --allowedTools "Read,Grep,Glob,Bash(gh pr view:*),Bash(gh issue view:*),Bash(git diff:*),Bash(git symbolic-ref:*),Bash(git merge-base:*),Bash(git remote show:*),Bash(git log:*),Bash(git show:*),Bash(cat:*),Bash(head:*),Bash(tail:*),Bash(ls:*),Bash(find:*),Bash(wc:*)"
           --json-schema '{"type":"object","additionalProperties":false,"required":["verdict","confidence","merge_risk","intent","counts","review_event","comment_markdown"],"properties":{"verdict":{"type":"string","enum":["pass","fail"]},"confidence":{"type":"integer","minimum":0,"maximum":100},"merge_risk":{"type":"string","enum":["low","med","high"]},"intent":{"type":"string","enum":["aligned","partial","deviated","skipped"]},"counts":{"type":"object","additionalProperties":false,"required":["p0","p1","p2","p3"],"properties":{"p0":{"type":"integer","minimum":0},"p1":{"type":"integer","minimum":0},"p2":{"type":"integer","minimum":0},"p3":{"type":"integer","minimum":0}}},"review_event":{"type":"string","enum":["APPROVE","REQUEST_CHANGES"]},"comment_markdown":{"type":"string"},"findings":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["id","file","line","severity","summary","failure_scenario","reason","evidence","confidence"],"properties":{"id":{"type":"string"},"file":{"type":"string"},"line":{"type":"integer","minimum":0},"severity":{"type":"string","enum":["P0","P1","P2","P3"]},"summary":{"type":"string"},"failure_scenario":{"type":"string"},"reason":{"type":"string"},"evidence":{"type":"string"},"confidence":{"type":"integer","enum":[0,25,50,75,100]},"severity_confirmed":{"type":"string","enum":["P0","P1","P2","P3"]}}}},"files_reviewed":{"type":"array","items":{"type":"string"}},"tests_failing":{"type":"boolean"},"coverage_below_threshold_on_critical_paths":{"type":"boolean"},"no_tests_for_changed_logic":{"type":"boolean"},"test_execution":{"type":"string","enum":["passed","failed","skipped","not_run"]},"verification_evidence":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["claim","command","result"],"properties":{"claim":{"type":"string"},"command":{"type":"string"},"result":{"type":"string"}}}},"checklist":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["text","status"],"properties":{"text":{"type":"string"},"status":{"type":"string","enum":["verified","failed","unverifiable"]},"evidence":{"type":"string"}}}}}}'
diff --git a/ai-review/lib/write-manifest.js b/ai-review/lib/write-manifest.js
index c704471..0fdcc22 100644
--- a/ai-review/lib/write-manifest.js
+++ b/ai-review/lib/write-manifest.js
@@ -182,9 +182,9 @@ function writeRoster(manifest, sizes, io) {
     roster = buildRoster({
       files: manifest.changed_files.map((p) => ({ path: p, bytes: sizes[p] ?? 0 })),
       models: {
-        opus: process.env.OPUS || "claude-opus-5",
-        sonnet: process.env.SONNET || "claude-sonnet-5",
-        haiku: process.env.HAIKU || "claude-haiku-4-5",
+        opus: process.env.OPUS || "claude/claude-opus-5",
+        sonnet: process.env.SONNET || "claude/claude-sonnet-5",
+        haiku: process.env.HAIKU || "claude/claude-haiku-4-5-20251001",
       },
       importEdges: resolveImportEdges(specifiers, manifest.changed_files),
       symbolManifest: manifest.symbol_manifest,

From 05eeb190f4010454f91b5f0fa15bf0e64a45463a Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 07:20:11 +0300
Subject: [PATCH 04/10] feat(ai-qa): lock Claude primary with Cursor then free
 fallbacks

Refs #51
---
 ai-qa/README.md                 |  3 ++-
 ai-qa/action.yml                | 47 ++++++++++++++++++++-------------
 ai-qa/lib/report-footer.js      | 28 ++++++++++++++++++++
 ai-qa/lib/report-footer.test.js | 27 +++++++++++++++++++
 4 files changed, 86 insertions(+), 19 deletions(-)
 create mode 100644 ai-qa/lib/report-footer.js
 create mode 100644 ai-qa/lib/report-footer.test.js

diff --git a/ai-qa/README.md b/ai-qa/README.md
index 1855bad..6c7034d 100644
--- a/ai-qa/README.md
+++ b/ai-qa/README.md
@@ -77,7 +77,8 @@ prompt.)
 7. **Stage QA rubric** — copies the action's own `rubric.md` into the
    workspace so the review can read it with a stable path.
 8. **Post-merge QA review (agentic)** — only runs when an Anthropic
-   credential is configured. Claude (`qa-model`, Sonnet by default) reads the
+   credential is configured. Claude (locked Sonnet primary with
+   Cursor → free fallbacks) reads the
    rubric, inspects the merged diff via `git`, **smoke-tests the deployed app
    over HTTP** (the health URL plus any routes the diff touches), **evaluates
    the PR's Test Plan** if one is present (running each item it can against
diff --git a/ai-qa/action.yml b/ai-qa/action.yml
index 7f9290e..a8a0a36 100644
--- a/ai-qa/action.yml
+++ b/ai-qa/action.yml
@@ -52,12 +52,6 @@ inputs:
       before this action runs.
     required: false
     default: ""
-  qa-model:
-    description: >-
-      Model used for the agentic post-merge QA review. Defaults to Sonnet;
-      override with a repo var if a custom gateway aliases model names.
-    required: false
-    default: claude-sonnet-5
   allowed-tools:
     description: >-
       Tool allowlist passed to the QA review's `--allowedTools`. The default
@@ -308,19 +302,14 @@ runs:
       with:
         anthropic_api_key: ${{ inputs.anthropic-api-key != '' && inputs.anthropic-api-key || inputs.anthropic-auth-token }}
         github_token: ${{ inputs.github-token }}
-        # --fallback-model: escape hatch for HTTP 529 (Overloaded). Without it
-        # the CLI retries the primary internally with no exit. Measured on the
-        # sibling ai-review action, that backoff ran ~28 min per stage and
-        # produced 85-min jobs that did zero work (turns:3, $0, 0 tool calls).
-        # This action has a single model stage, so the same overload costs it
-        # one long stall rather than three — but it had no exit either.
-        # Neither entry may equal the primary or the fallback is a no-op:
-        # `qa-model` defaults to Sonnet 5, so the list is same-tier first
-        # (Sonnet 4.6), then Opus. The list is ordered and the primary is
-        # re-tried at the start of each turn.
+        # --fallback-model: Claude primary → Cursor if Claude blocked → free
+        # if Cursor blocked. Keep the free tail in sync with ai-review.
+        # Neither entry may equal the primary or the fallback is a no-op.
+        # The list is ordered and the primary is re-tried at the start of
+        # each turn. Does not bound silent gateway stalls (ADR 0005).
         claude_args: |
-          --model ${{ inputs.qa-model }}
-          --fallback-model claude-sonnet-4-6,claude-opus-4-8
+          --model claude/claude-sonnet-5
+          --fallback-model claude/cursor/claude-4.6-sonnet-medium-thinking,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free
           --allowedTools "${{ inputs.allowed-tools }}"
           --json-schema '{"type":"object","additionalProperties":false,"required":["verdict","confidence","merge_risk","deploy_status","counts","summary","report_markdown"],"properties":{"verdict":{"type":"string","enum":["pass","fail"]},"confidence":{"type":"integer","minimum":0,"maximum":100},"merge_risk":{"type":"string","enum":["low","medium","high"]},"deploy_status":{"type":"string","enum":["healthy","unhealthy","unknown"]},"counts":{"type":"object","additionalProperties":false,"required":["p0","p1","p2","p3"],"properties":{"p0":{"type":"integer","minimum":0},"p1":{"type":"integer","minimum":0},"p2":{"type":"integer","minimum":0},"p3":{"type":"integer","minimum":0}}},"summary":{"type":"string"},"report_markdown":{"type":"string"},"test_plan":{"type":"object","additionalProperties":false,"required":["present"],"properties":{"present":{"type":"boolean"},"items":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["text","status"],"properties":{"text":{"type":"string"},"status":{"type":"string","enum":["passed","failed","unverifiable"]},"evidence":{"type":"string"}}}}}}}}'
         prompt: |
@@ -395,9 +384,25 @@ runs:
         UPDATE_PR_BODY: ${{ inputs.update-pr-body }}
         UPDATE_LINKED_ISSUES: ${{ inputs.update-linked-issues }}
         RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+        REVIEW_EXEC_FILE: ${{ steps.review.outputs.execution_file }}
+        REPORT_FOOTER_PATH: ${{ github.action_path }}/lib/report-footer.js
       with:
         github-token: ${{ steps.identity.outputs.author-token }}
         script: |
+          const fs = require('fs');
+          const { formatModelFooter, modelFromExecutionLog } = require(process.env.REPORT_FOOTER_PATH);
+
+          let modelUsed = '';
+          try {
+            const log = JSON.parse(fs.readFileSync(process.env.REVIEW_EXEC_FILE || '', 'utf8'));
+            modelUsed = modelFromExecutionLog(log);
+          } catch {
+            // No execution log (review skipped / missing).
+          }
+          if (!modelUsed && process.env.QA_OUTPUT) {
+            modelUsed = 'claude/claude-sonnet-5';
+          }
+
           const prNumber = Number(process.env.PR_NUMBER);
           const healthOk = process.env.HEALTH_STATUS === 'healthy';
 
@@ -515,6 +520,12 @@ runs:
             body.push('> ⚠️ The QA review step reported a non-zero outcome — the report above may be partial.');
           }
 
+          const footer = formatModelFooter(modelUsed);
+          if (footer) {
+            body.push('');
+            body.push(footer);
+          }
+
           body.push('');
           body.push(`_Posted by \`${process.env.AUTHOR_LOGIN}\` · [workflow run](${process.env.RUN_URL})_`);
 
diff --git a/ai-qa/lib/report-footer.js b/ai-qa/lib/report-footer.js
new file mode 100644
index 0000000..67c9583
--- /dev/null
+++ b/ai-qa/lib/report-footer.js
@@ -0,0 +1,28 @@
+"use strict";
+
+/**
+ * @param {string|null|undefined} modelUsed
+ * @returns {string} empty when modelUsed is blank
+ */
+function formatModelFooter(modelUsed) {
+  if (!modelUsed || typeof modelUsed !== "string" || !modelUsed.trim()) return "";
+  return [
+    `Model: \`${modelUsed.trim()}\``,
+    "_Re-run this job if you need another review pass._",
+  ].join("\n");
+}
+
+/**
+ * First named model in a claude-code-action execution log.
+ * @param {unknown} entries
+ * @returns {string}
+ */
+function modelFromExecutionLog(entries) {
+  if (!Array.isArray(entries)) return "";
+  for (const e of entries) {
+    if (e && typeof e.model === "string" && e.model.trim()) return e.model.trim();
+  }
+  return "";
+}
+
+module.exports = { formatModelFooter, modelFromExecutionLog };
diff --git a/ai-qa/lib/report-footer.test.js b/ai-qa/lib/report-footer.test.js
new file mode 100644
index 0000000..01fee8a
--- /dev/null
+++ b/ai-qa/lib/report-footer.test.js
@@ -0,0 +1,27 @@
+"use strict";
+
+const test = require("node:test");
+const assert = require("node:assert/strict");
+const { formatModelFooter, modelFromExecutionLog } = require("./report-footer.js");
+
+test("formatModelFooter renders model and re-run hint", () => {
+  const s = formatModelFooter("claude/claude-sonnet-5");
+  assert.match(s, /Model: `claude\/claude-sonnet-5`/);
+  assert.match(s, /Re-run this job if you need another review pass/);
+});
+
+test("formatModelFooter returns empty for blank", () => {
+  assert.equal(formatModelFooter(""), "");
+  assert.equal(formatModelFooter(null), "");
+});
+
+test("modelFromExecutionLog reads the first named model", () => {
+  assert.equal(
+    modelFromExecutionLog([
+      { type: "system", subtype: "init", model: "oc/mimo-v2.5-free" },
+      { type: "result", subtype: "success" },
+    ]),
+    "oc/mimo-v2.5-free",
+  );
+  assert.equal(modelFromExecutionLog(null), "");
+});

From 87dc3dd8c8062bc7ec260519c1d883b73761313f Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 07:20:39 +0300
Subject: [PATCH 05/10] =?UTF-8?q?docs:=20note=20locked=20Claude=E2=86=92Cu?=
 =?UTF-8?q?rsor=E2=86=92free=20model=20cascade?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Refs #51
---
 ai-qa/README.md     | 1 -
 ai-review/README.md | 6 +++++-
 docs/plan.md        | 4 ++--
 3 files changed, 7 insertions(+), 4 deletions(-)

diff --git a/ai-qa/README.md b/ai-qa/README.md
index 6c7034d..9031b5a 100644
--- a/ai-qa/README.md
+++ b/ai-qa/README.md
@@ -115,7 +115,6 @@ prompt.)
 | `health-url` | URL polled with `curl --fail` until healthy or `deploy-timeout` elapses; also smoke-tested directly by the review. No sensible generic default exists. | **Yes** | — |
 | `deploy-timeout` | Seconds to keep polling `health-url` before giving up. | No | `180` |
 | `test-hint` | Optional free-text describing how to build/test this repo. Handed to the review as context — Claude MAY run it at its discretion to confirm a suspected regression, never mechanically. Consumer must provision the toolchain first. | No | `""` |
-| `qa-model` | Model used for the agentic QA review. | No | `claude-sonnet-5` |
 | `allowed-tools` | Tool allowlist passed to the review's `--allowedTools` (read/grep the code, `curl` the deploy, `git` the diff, optionally run a JS/TS build/test). Override to widen or narrow. | No | *(read/grep/glob + curl/git + node/npm/npx/yarn/pnpm/corepack)* |
 | `pass-label` | Label applied when the overall QA signal (health + review) passes. Also applied to linked issues when `update-linked-issues` is on. | No | `✓ /ai-qa` |
 | `fail-label` | Label applied when the overall QA signal fails. Also applied to linked issues (and the merge-auto-closed issue is reopened) when `update-linked-issues` is on. | No | `✗ /ai-qa` |
diff --git a/ai-review/README.md b/ai-review/README.md
index aea28d4..07b3e42 100644
--- a/ai-review/README.md
+++ b/ai-review/README.md
@@ -10,7 +10,11 @@ workflow can gate its own heavier build/test/deploy jobs on
 The verdict comes from a real two-stage AI review: a Haiku context stage
 summarizes the diff, then a diff-size-routed Sonnet/Opus review stage
 performs the full rubric scan (see `ai-review/rubric.md`) and returns a
-schema-validated structured result. The `Publish review` step never trusts
+schema-validated structured result. Model IDs are **locked in the action**
+(Claude primary → Cursor if Claude is blocked → pinned OpenCode free models,
+then `auto/best-free`). The posted PR review footer names the model that
+actually ran and hints to re-run the job for another pass. The `Publish review`
+step never trusts
 the model's self-reported verdict directly — it deterministically
 recomputes confidence, verdict, and merge risk from the model's reported
 P0-P3 finding counts and test-quality signals, then posts that as a native
diff --git a/docs/plan.md b/docs/plan.md
index d2c162c..e215c27 100644
--- a/docs/plan.md
+++ b/docs/plan.md
@@ -29,8 +29,8 @@ Today `ai-review`/`ai-qa` are autopilot **skills** a developer runs in a local C
 - **Public & generalized:** the repo is **public** and org-agnostic. No action ships an eduly-specific default — eduly values (App ID/key, runner label, health URL, test command, staging URL) are **caller-supplied**; the action's own defaults are neutral (`runner-label: ubuntu-latest`, no health-url/test-command default). READMEs, ADR 0001, and `docs/consumer-integration.md` are written generically with **eduly as one example consumer**, not the subject. A `LICENSE` is a **deferred, blocking decision**: a public repo with no license is "all rights reserved," so external reuse cannot begin until a license is chosen (see D12) — internal EdulyCom consumption is unaffected.
 - **Process:** all changes via PR (no direct push to protected branches); **merge commits, not squash**. TypeScript/no-`any` for any committed scripts; prefer inline `github-script`. Docs land in the same PR as their change.
 - **Secrets & variables (org-level, human-supplied):** every secret/var the actions consume is set **at the EdulyCom org level** (`gh secret set --org EdulyCom …` / `gh variable set --org EdulyCom …`, repo-visibility scoped to the consuming repos incl. `github-actions` itself so its `selftest.yml` can run). The **implementer never invents secret values**: at each point one is needed (Phase 1 selftest, Phase 4 eduly adoption) the implementer **pauses and asks the human for the value**, then sets it at org level. Org-level secrets: `ANTHROPIC_AUTH_TOKEN` (eduly gateway bearer), `MTM_BOT_APP_PRIVATE_KEY`. Org-level vars: `ANTHROPIC_BASE_URL` (eduly gateway URL), `MTM_BOT_APP_ID`, `RUNNER_LABEL`, `AI_REVIEW_LABEL` (+ model-ID overrides if a gateway aliases them). (`GITHUB_TOKEN` is auto-provided per repo; not set here.)
-- **Model IDs** (repo-var overridable): context `claude-haiku-4-5-20251001`; review defaults to `claude-opus-4-8` (account HAS Opus access), dropping to `claude-sonnet-5` only for **tiny** diffs. Tiny = ≤`sonnet-files-threshold` (default 3) files AND ≤`sonnet-churn-threshold` (default 60) changed lines. (Opus-by-default per ADR 0003; earlier the routing was Sonnet-default for ≤15 files AND ≤400 lines.)
-- **Anthropic endpoint — caller-supplied, defaults to Anthropic:** `anthropic-base-url` is an **optional caller input**; when omitted the action leaves `ANTHROPIC_BASE_URL` unset so `claude-code-action` uses the **standard Anthropic endpoint** (no gateway baked in — keeps the action generic per D12). When the caller passes it (eduly → its **custom gateway**), the value is plumbed to **every** `claude-code-action` invocation — Haiku context, Sonnet/Opus review, ai-qa triage — together with the caller's auth: either `anthropic-api-key` (→ `ANTHROPIC_API_KEY`/`x-api-key`) or `anthropic-auth-token` (→ `ANTHROPIC_AUTH_TOKEN` bearer). Because a gateway may expose **aliased** model names, model IDs stay repo-var overridable (a gateway lacking Opus falls back to the Sonnet ID via the override).
+- **Model IDs** (locked in the action, not repo-var overridable): context `claude/claude-haiku-4-5-20251001`; review routes tiny diffs to `claude/claude-sonnet-5` and larger diffs to `claude/claude-opus-5`. Cascade on overload: Cursor (`claude/cursor/...`) then pinned OpenCode free models then `auto/best-free`. Tiny = ≤`sonnet-files-threshold` (default 25) files AND ≤`sonnet-churn-threshold` (default 800) changed lines. ai-qa uses locked Sonnet with the same Cursor→free fallback list. See `docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md` / #51.
+- **Anthropic endpoint — caller-supplied, defaults to Anthropic:** `anthropic-base-url` is an **optional caller input**; when omitted the action leaves `ANTHROPIC_BASE_URL` unset so `claude-code-action` uses the **standard Anthropic endpoint** (no gateway baked in — keeps the action generic per D12). When the caller passes it (eduly → its **custom gateway**), the value is plumbed to **every** `claude-code-action` invocation — Haiku context, Sonnet/Opus review, ai-qa triage — together with the caller's auth: either `anthropic-api-key` (→ `ANTHROPIC_API_KEY`/`x-api-key`) or `anthropic-auth-token` (→ `ANTHROPIC_AUTH_TOKEN` bearer). Model IDs are locked in the action with a Claude→Cursor→free `--fallback-model` cascade rather than caller overrides.
 
 ## Locked decisions
 

From 4d3d83bc438c9e93398b0ca435d95f80848f9fcf Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 07:44:51 +0300
Subject: [PATCH 06/10] fix(ai-qa): derive model footer fallback from locked
 primary output

Refs #51
---
 ai-qa/action.yml | 19 +++++++++++++++++--
 1 file changed, 17 insertions(+), 2 deletions(-)

diff --git a/ai-qa/action.yml b/ai-qa/action.yml
index a8a0a36..285df08 100644
--- a/ai-qa/action.yml
+++ b/ai-qa/action.yml
@@ -286,6 +286,21 @@ runs:
       run: | # zizmor: ignore[github-env] BASE_URL is this action's own anthropic-base-url input (caller-supplied static config), never attacker-influenceable event data — see ADR-0010's env-binding rule
         echo "ANTHROPIC_BASE_URL=${BASE_URL}" >> "${GITHUB_ENV}"
 
+    # Locked model IDs (not inputs). Keep in sync with ai-review cascade.
+    # Claude primary → Cursor if Claude blocked → free if Cursor blocked.
+    - name: Resolve locked models
+      id: models
+      if: >-
+        steps.merge.outputs.skip != 'true' &&
+        (inputs.anthropic-api-key != '' || inputs.anthropic-auth-token != '')
+      shell: bash
+      run: |
+        set -euo pipefail
+        {
+          echo "primary=claude/claude-sonnet-5"
+          echo "fallback=claude/cursor/claude-4.6-sonnet-medium-thinking,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free"
+        } >> "${GITHUB_OUTPUT}"
+
     # Workaround for anthropics/claude-code-action#1294: the action requires
     # *some* value in `anthropic_api_key` even when authenticating to a custom
     # gateway via a bearer token, so fall back to the auth token itself and
@@ -308,8 +323,8 @@ runs:
         # The list is ordered and the primary is re-tried at the start of
         # each turn. Does not bound silent gateway stalls (ADR 0005).
         claude_args: |
-          --model claude/claude-sonnet-5
-          --fallback-model claude/cursor/claude-4.6-sonnet-medium-thinking,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free
+          --model ${{ steps.models.outputs.primary }}
+          --fallback-model ${{ steps.models.outputs.fallback }}
           --allowedTools "${{ inputs.allowed-tools }}"
           --json-schema '{"type":"object","additionalProperties":false,"required":["verdict","confidence","merge_risk","deploy_status","counts","summary","report_markdown"],"properties":{"verdict":{"type":"string","enum":["pass","fail"]},"confidence":{"type":"integer","minimum":0,"maximum":100},"merge_risk":{"type":"string","enum":["low","medium","high"]},"deploy_status":{"type":"string","enum":["healthy","unhealthy","unknown"]},"counts":{"type":"object","additionalProperties":false,"required":["p0","p1","p2","p3"],"properties":{"p0":{"type":"integer","minimum":0},"p1":{"type":"integer","minimum":0},"p2":{"type":"integer","minimum":0},"p3":{"type":"integer","minimum":0}}},"summary":{"type":"string"},"report_markdown":{"type":"string"},"test_plan":{"type":"object","additionalProperties":false,"required":["present"],"properties":{"present":{"type":"boolean"},"items":{"type":"array","items":{"type":"object","additionalProperties":false,"required":["text","status"],"properties":{"text":{"type":"string"},"status":{"type":"string","enum":["passed","failed","unverifiable"]},"evidence":{"type":"string"}}}}}}}}'
         prompt: |

From 91f2d8b439825a542c9e0711d7d3a912439bb4aa Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 07:46:24 +0300
Subject: [PATCH 07/10] fix(ai-qa): fall back publish Model line to locked
 primary

Prefer the execution-log model when present; otherwise use
steps.models.outputs.primary instead of a second Sonnet literal.

Refs #51
---
 ai-qa/action.yml | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/ai-qa/action.yml b/ai-qa/action.yml
index 285df08..fdf4c16 100644
--- a/ai-qa/action.yml
+++ b/ai-qa/action.yml
@@ -401,6 +401,10 @@ runs:
         RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
         REVIEW_EXEC_FILE: ${{ steps.review.outputs.execution_file }}
         REPORT_FOOTER_PATH: ${{ github.action_path }}/lib/report-footer.js
+        # Same locked primary the review stage was started with (steps.models).
+        # Prefer the execution-log model when present; otherwise the requested
+        # primary — never a second hardcoded literal that can drift from --model.
+        QA_PRIMARY_MODEL: ${{ steps.models.outputs.primary }}
       with:
         github-token: ${{ steps.identity.outputs.author-token }}
         script: |
@@ -414,8 +418,12 @@ runs:
           } catch {
             // No execution log (review skipped / missing).
           }
+          // Only attribute a model when we have structured QA output. Prefer the
+          // log's resolved model (captures Cursor/free fallback). If the log is
+          // silent, fall back to the locked primary that this run requested —
+          // same shape as ai-review's resolveModelUsed({ fallback: ROUTED_MODEL }).
           if (!modelUsed && process.env.QA_OUTPUT) {
-            modelUsed = 'claude/claude-sonnet-5';
+            modelUsed = process.env.QA_PRIMARY_MODEL || '';
           }
 
           const prNumber = Number(process.env.PR_NUMBER);

From 98261acfbc51a9b660fbfad727146e79c701a73c Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 08:17:26 +0300
Subject: [PATCH 08/10] fix(ai-review): use composer-2.5 for Cursor fallback
 tier

claude/cursor/claude-* hangs Claude Code CLI (aborted_streaming) even when
/v1/messages succeeds, so a Claude-off cascade never reached free. Cap the
chain at three entries to match the CLI limit.

Refs #51
---
 ai-qa/action.yml                              |  6 +++--
 ai-review/README.md                           |  4 +--
 ai-review/action.yml                          | 16 +++++++++---
 ...-08-17-claude-cursor-free-model-cascade.md | 25 +++++++++++--------
 4 files changed, 33 insertions(+), 18 deletions(-)

diff --git a/ai-qa/action.yml b/ai-qa/action.yml
index fdf4c16..de8e7fe 100644
--- a/ai-qa/action.yml
+++ b/ai-qa/action.yml
@@ -287,7 +287,9 @@ runs:
         echo "ANTHROPIC_BASE_URL=${BASE_URL}" >> "${GITHUB_ENV}"
 
     # Locked model IDs (not inputs). Keep in sync with ai-review cascade.
-    # Claude primary → Cursor if Claude blocked → free if Cursor blocked.
+    # Claude primary → Cursor (composer-2.5) if Claude blocked → free.
+    # Do not use claude/cursor/claude-* here: CLI hangs on those IDs (see
+    # ai-review Resolve model routing comment). Cap chain at 3 (CLI limit).
     - name: Resolve locked models
       id: models
       if: >-
@@ -298,7 +300,7 @@ runs:
         set -euo pipefail
         {
           echo "primary=claude/claude-sonnet-5"
-          echo "fallback=claude/cursor/claude-4.6-sonnet-medium-thinking,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free"
+          echo "fallback=claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free"
         } >> "${GITHUB_OUTPUT}"
 
     # Workaround for anthropics/claude-code-action#1294: the action requires
diff --git a/ai-review/README.md b/ai-review/README.md
index 07b3e42..983f1f1 100644
--- a/ai-review/README.md
+++ b/ai-review/README.md
@@ -11,8 +11,8 @@ The verdict comes from a real two-stage AI review: a Haiku context stage
 summarizes the diff, then a diff-size-routed Sonnet/Opus review stage
 performs the full rubric scan (see `ai-review/rubric.md`) and returns a
 schema-validated structured result. Model IDs are **locked in the action**
-(Claude primary → Cursor if Claude is blocked → pinned OpenCode free models,
-then `auto/best-free`). The posted PR review footer names the model that
+(Claude primary → Cursor `composer-2.5` if Claude is blocked → OpenCode free /
+`auto/best-free`; Claude Code caps the fallback chain at 3). The posted PR review footer names the model that
 actually ran and hints to re-run the job for another pass. The `Publish review`
 step never trusts
 the model's self-reported verdict directly — it deterministically
diff --git a/ai-review/action.yml b/ai-review/action.yml
index 7afdcf0..f064a07 100644
--- a/ai-review/action.yml
+++ b/ai-review/action.yml
@@ -421,13 +421,21 @@ runs:
         mkdir -p .ai-review
 
         # Locked model IDs (not inputs). Keep in sync with ai-qa.
+        #
+        # Cursor tier MUST be composer-2.5 (or another CLI-proven Cursor ID).
+        # claude/cursor/claude-* IDs accept raw /v1/messages but hang Claude Code
+        # CLI (unrecognized_model → aborted_streaming), so a Claude-off cascade
+        # never reaches free if those IDs are first in --fallback-model.
+        # Claude Code caps fallback chains at 3 entries after dedupe.
         SONNET="claude/claude-sonnet-5"
         OPUS="claude/claude-opus-5"
         HAIKU="claude/claude-haiku-4-5-20251001"
-        FREE="oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free"
-        SONNET_FALLBACK="claude/cursor/claude-4.6-sonnet-medium-thinking,${FREE}"
-        OPUS_FALLBACK="claude/cursor/claude-opus-4-8-medium-fast,${FREE}"
-        HAIKU_FALLBACK="claude/cursor/composer-2.5,${FREE}"
+        CURSOR="claude/cursor/composer-2.5"
+        FREE="oc/nemotron-3.5-lightning-free,auto/best-free"
+        FALLBACK_CHAIN="${CURSOR},${FREE}"
+        SONNET_FALLBACK="${FALLBACK_CHAIN}"
+        OPUS_FALLBACK="${FALLBACK_CHAIN}"
+        HAIKU_FALLBACK="${FALLBACK_CHAIN}"
         export SONNET OPUS HAIKU
 
         # Resolve the review range once, deterministically. `git diff A...B`
diff --git a/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md b/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md
index 7087680..488e062 100644
--- a/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md
+++ b/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md
@@ -17,7 +17,9 @@
 - Keep `sonnet-files-threshold` / `sonnet-churn-threshold` inputs on `ai-review`.
 - Prefer `claude/` and `claude/cursor/` prefixes (Anthropic Messages–compatible).
 - Free tail = **pinned usable `oc/*-free` IDs**, then **`auto/best-free` last** (hybrid). Do not use `auto/*` as Claude or Cursor slots. Do not use `auto/coding:free` (unsupported upstream). Prefer `oc/` over duplicate `opencode/` aliases.
-- Shared free tail (all roles): `oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free`
+- Shared fallback chain (all roles, CLI max 3): `claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free`
+  - Cursor tier is **composer-2.5 only** — not `claude/cursor/claude-*` (CLI hang)
+  - Longer free inventories are ignored past the 3-entry CLI cap
 - `--fallback-model` must not include the primary (CLI no-op otherwise).
 - Do not claim `--fallback-model` fixes silent ~27m gateway stalls (ADR 0005).
 - Injection safety: bind values via `env:`, never interpolate attacker-controlled text into `run:`/`script:` bodies with `${{ }}`.
@@ -78,16 +80,19 @@ Catalog free-ish IDs (excluding video `veo*`):
 Shared free tail constant (paste identically everywhere):
 
 ```text
-FREE=oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free
+FREE=oc/nemotron-3.5-lightning-free,auto/best-free
+# Full chain (Cursor + FREE) must stay ≤3 entries — Claude Code ignores the rest.
 ```
 
 Do **not** add `oc/big-pickle` while billing is unknown (not labeled free in the catalog). Revisit only after confirming it is zero-cost on the Eduly gateway.
 | Action / role | Primary | `--fallback-model` (ordered) |
 | --- | --- | --- |
-| ai-review context | `claude/claude-haiku-4-5-20251001` | `claude/cursor/composer-2.5,` + FREE |
-| ai-review review (tiny) | `claude/claude-sonnet-5` | `claude/cursor/claude-4.6-sonnet-medium-thinking,` + FREE |
-| ai-review review (large) | `claude/claude-opus-5` | `claude/cursor/claude-opus-4-8-medium-fast,` + FREE |
-| ai-qa review | `claude/claude-sonnet-5` | `claude/cursor/claude-4.6-sonnet-medium-thinking,` + FREE |
+| ai-review context | `claude/claude-haiku-4-5-20251001` | `claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free` |
+| ai-review review (tiny) | `claude/claude-sonnet-5` | same Cursor→free chain (CLI max 3) |
+| ai-review review (large) | `claude/claude-opus-5` | same Cursor→free chain (CLI max 3) |
+| ai-qa review | `claude/claude-sonnet-5` | same Cursor→free chain (CLI max 3) |
+
+> **Do not** put `claude/cursor/claude-*` in `--fallback-model`: raw `/v1/messages` may 200, but Claude Code CLI hangs (`unrecognized_model` → `aborted_streaming`) and never reaches free. Use `composer-2.5` as the Cursor tier.
 
 Comment footer copy (**successful** reviews / QA comments only):
 
@@ -305,9 +310,9 @@ In **Deterministic prep and model routing**, replace `inputs.*-model` with:
 SONNET="claude/claude-sonnet-5"
 OPUS="claude/claude-opus-5"
 HAIKU="claude/claude-haiku-4-5-20251001"
-SONNET_FALLBACK="claude/cursor/claude-4.6-sonnet-medium-thinking,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free"
-OPUS_FALLBACK="claude/cursor/claude-opus-4-8-medium-fast,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free"
-HAIKU_FALLBACK="claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free"
+SONNET_FALLBACK="claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free"
+OPUS_FALLBACK="claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free"
+HAIKU_FALLBACK="claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free"
 ```
 
 Export `HAIKU` / `OPUS` / `SONNET` to the environment for `write-manifest.js` (same as today). After choosing `MODEL`, also set `FALLBACK` to the matching list and write:
@@ -432,7 +437,7 @@ In `ai-qa/action.yml`, delete `qa-model`. Replace review `claude_args` model lin
 # Locked at action level — keep in sync with ai-review cascade.
 # Claude primary → Cursor if Claude blocked → free if Cursor blocked.
 --model claude/claude-sonnet-5
---fallback-model claude/cursor/claude-4.6-sonnet-medium-thinking,oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free
+--fallback-model claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free
 ```
 
 Update the comment that currently describes `claude-sonnet-4-6,claude-opus-4-8`.

From 447846e0b47ff1214c704d72415e3f7fd6df7d67 Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 08:45:08 +0300
Subject: [PATCH 09/10] fix(ai-review): recover session_id for SO repair;
 schema free fallbacks
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Repair was skipped when claude-code-action failed on missing
structured_output without exporting session_id. Recover it from the
execution snapshot. Schema stages fall back to structured_output free
models only — Cursor has no SO and can retract schema mid-stream.

Refs #51
---
 ai-qa/action.yml                              |  9 ++-
 ai-review/README.md                           |  5 +-
 ai-review/action.yml                          | 66 +++++++++++++++----
 ...-08-17-claude-cursor-free-model-cascade.md | 11 ++--
 4 files changed, 66 insertions(+), 25 deletions(-)

diff --git a/ai-qa/action.yml b/ai-qa/action.yml
index de8e7fe..c7dfaab 100644
--- a/ai-qa/action.yml
+++ b/ai-qa/action.yml
@@ -286,10 +286,9 @@ runs:
       run: | # zizmor: ignore[github-env] BASE_URL is this action's own anthropic-base-url input (caller-supplied static config), never attacker-influenceable event data — see ADR-0010's env-binding rule
         echo "ANTHROPIC_BASE_URL=${BASE_URL}" >> "${GITHUB_ENV}"
 
-    # Locked model IDs (not inputs). Keep in sync with ai-review cascade.
-    # Claude primary → Cursor (composer-2.5) if Claude blocked → free.
-    # Do not use claude/cursor/claude-* here: CLI hangs on those IDs (see
-    # ai-review Resolve model routing comment). Cap chain at 3 (CLI limit).
+    # Locked model IDs (not inputs). Keep in sync with ai-review's *schema*
+    # cascade: Claude → structured_output free only (no Cursor — none advertise
+    # structured_output, and --json-schema reviews go inconclusive on them).
     - name: Resolve locked models
       id: models
       if: >-
@@ -300,7 +299,7 @@ runs:
         set -euo pipefail
         {
           echo "primary=claude/claude-sonnet-5"
-          echo "fallback=claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free"
+          echo "fallback=oc/nemotron-3.5-lightning-free,oc/deepseek-v4-flash-free,auto/best-free"
         } >> "${GITHUB_OUTPUT}"
 
     # Workaround for anthropics/claude-code-action#1294: the action requires
diff --git a/ai-review/README.md b/ai-review/README.md
index 983f1f1..574563d 100644
--- a/ai-review/README.md
+++ b/ai-review/README.md
@@ -11,8 +11,9 @@ The verdict comes from a real two-stage AI review: a Haiku context stage
 summarizes the diff, then a diff-size-routed Sonnet/Opus review stage
 performs the full rubric scan (see `ai-review/rubric.md`) and returns a
 schema-validated structured result. Model IDs are **locked in the action**
-(Claude primary → Cursor `composer-2.5` if Claude is blocked → OpenCode free /
-`auto/best-free`; Claude Code caps the fallback chain at 3). The posted PR review footer names the model that
+(Claude primary → for context, Cursor `composer-2.5` then free; for schema
+reviews, structured_output free models only — Cursor has no SO on this
+gateway). The posted PR review footer names the model that
 actually ran and hints to re-run the job for another pass. The `Publish review`
 step never trusts
 the model's self-reported verdict directly — it deterministically
diff --git a/ai-review/action.yml b/ai-review/action.yml
index f064a07..78f7c35 100644
--- a/ai-review/action.yml
+++ b/ai-review/action.yml
@@ -422,20 +422,22 @@ runs:
 
         # Locked model IDs (not inputs). Keep in sync with ai-qa.
         #
-        # Cursor tier MUST be composer-2.5 (or another CLI-proven Cursor ID).
-        # claude/cursor/claude-* IDs accept raw /v1/messages but hang Claude Code
-        # CLI (unrecognized_model → aborted_streaming), so a Claude-off cascade
-        # never reaches free if those IDs are first in --fallback-model.
-        # Claude Code caps fallback chains at 3 entries after dedupe.
+        # Two cascades (Claude Code caps --fallback-model at 3 entries):
+        #   Context (no --json-schema): Claude → composer-2.5 → free.
+        #     composer is CLI-safe; claude/cursor/claude-* hang the CLI.
+        #   Review (has --json-schema): Claude → structured_output free only.
+        #     No Cursor model advertises structured_output on this gateway, and
+        #     Agent SDK docs: a mid-stream model fallback can retract an already
+        #     completed structured_output — landing on composer then yields the
+        #     inconclusive "success but no structured_output" failure mode.
         SONNET="claude/claude-sonnet-5"
         OPUS="claude/claude-opus-5"
         HAIKU="claude/claude-haiku-4-5-20251001"
         CURSOR="claude/cursor/composer-2.5"
-        FREE="oc/nemotron-3.5-lightning-free,auto/best-free"
-        FALLBACK_CHAIN="${CURSOR},${FREE}"
-        SONNET_FALLBACK="${FALLBACK_CHAIN}"
-        OPUS_FALLBACK="${FALLBACK_CHAIN}"
-        HAIKU_FALLBACK="${FALLBACK_CHAIN}"
+        FREE_SO="oc/nemotron-3.5-lightning-free,oc/deepseek-v4-flash-free,auto/best-free"
+        HAIKU_FALLBACK="${CURSOR},oc/nemotron-3.5-lightning-free,auto/best-free"
+        SONNET_FALLBACK="${FREE_SO}"
+        OPUS_FALLBACK="${FREE_SO}"
         export SONNET OPUS HAIKU
 
         # Resolve the review range once, deterministically. `git diff A...B`
@@ -870,11 +872,18 @@ runs:
       if: always() && steps.fork-guard.outputs.is-fork != 'true' && steps.pr-state.outputs.skip != 'true'
       shell: bash
       env:
+        # Prefer the action output; fall back to the fixed path the pinned
+        # claude-code-action always writes (outputs can be missing when the
+        # step fails on a missing structured_output after logging session_id).
         EXEC_FILE: ${{ steps.review.outputs.execution_file }}
       run: |
         set -uo pipefail
-        if [ -n "${EXEC_FILE:-}" ] && [ -s "${EXEC_FILE}" ]; then
-          cp "${EXEC_FILE}" "${RUNNER_TEMP}/ai-review-exec-review-snapshot.json" || true
+        CANDIDATE="${EXEC_FILE:-}"
+        if [ -z "${CANDIDATE}" ] || [ ! -s "${CANDIDATE}" ]; then
+          CANDIDATE="${RUNNER_TEMP}/claude-execution-output.json"
+        fi
+        if [ -s "${CANDIDATE}" ]; then
+          cp "${CANDIDATE}" "${RUNNER_TEMP}/ai-review-exec-review-snapshot.json" || true
         fi
         exit 0
 
@@ -890,6 +899,11 @@ runs:
     # wins, and its duration_ms is the stage duration. Best-effort — any
     # parse failure leaves retry-ok=true, so this can only ever suppress a
     # retry that is provably too late, never block one on missing data.
+    #
+    # Also recovers session_id for the cheap repair step: when the review
+    # action fails on missing structured_output it may log "Set session_id"
+    # without exporting the output, which previously skipped repair entirely
+    # (observed on run 31997386554) and forced a full re-review.
     - name: Gate the review retry on elapsed review time
       id: retry_budget
       if: always() && steps.fork-guard.outputs.is-fork != 'true' && steps.pr-state.outputs.skip != 'true'
@@ -898,14 +912,34 @@ runs:
         # 15 min. A retry starting later than this cannot complete within the
         # 25-min timeout docs/consumer-integration.md recommends.
         RETRY_BUDGET_MS: "900000"
+        REVIEW_SESSION_ID: ${{ steps.review.outputs.session_id }}
       run: |
         set -uo pipefail
         SNAPSHOT="${RUNNER_TEMP}/ai-review-exec-review-snapshot.json"
         RETRY_OK=true
         DURATION_MS=0
+        SESSION_ID="${REVIEW_SESSION_ID:-}"
         if [ -s "${SNAPSHOT}" ]; then
           DURATION_MS="$(jq -r '[.[] | select(.type == "result")] | last | .duration_ms // 0' \
             "${SNAPSHOT}" 2>/dev/null || echo 0)"
+          if [ -z "${SESSION_ID}" ]; then
+            SESSION_ID="$(jq -r '
+              [
+                .[]
+                | select(.type == "system" and .subtype == "init")
+                | .session_id // empty
+              ]
+              | map(select(length > 0))
+              | last // empty
+            ' "${SNAPSHOT}" 2>/dev/null || true)"
+          fi
+          if [ -z "${SESSION_ID}" ]; then
+            SESSION_ID="$(jq -r '
+              [.[] | select(.type == "result") | .session_id // empty]
+              | map(select(length > 0))
+              | last // empty
+            ' "${SNAPSHOT}" 2>/dev/null || true)"
+          fi
         fi
         case "${DURATION_MS}" in
           ''|*[!0-9]*) DURATION_MS=0 ;;
@@ -914,9 +948,13 @@ runs:
           RETRY_OK=false
           echo "::warning::Review stage ran ${DURATION_MS}ms (>= ${RETRY_BUDGET_MS}ms budget); skipping the full-review retry, which could not finish inside the caller's timeout."
         fi
+        if [ -z "${SESSION_ID}" ]; then
+          echo "::warning::No review session_id in step outputs or execution snapshot; structured-output repair will be skipped."
+        fi
         {
           echo "retry-ok=${RETRY_OK}"
           echo "review-duration-ms=${DURATION_MS}"
+          echo "session_id=${SESSION_ID}"
         } >> "${GITHUB_OUTPUT}"
         exit 0
 
@@ -935,7 +973,7 @@ runs:
         steps.fork-guard.outputs.is-fork != 'true' &&
         steps.pr-state.outputs.skip != 'true' &&
         steps.review.outputs.structured_output == '' &&
-        steps.review.outputs.session_id != ''
+        steps.retry_budget.outputs.session_id != ''
       uses: anthropics/claude-code-action@6b082c41935b4c8a3b8b0ef85ba4ba4d9eeb8975 # v1.0.189
       env:
         ANTHROPIC_BASE_URL: ${{ inputs.anthropic-base-url }}
@@ -946,7 +984,7 @@ runs:
         github_token: ${{ inputs.github-token }}
         allowed_bots: ${{ inputs.allowed-bots }}
         claude_args: |
-          --resume ${{ steps.review.outputs.session_id }}
+          --resume ${{ steps.retry_budget.outputs.session_id }}
           --model ${{ steps.route.outputs.model }}
           --fallback-model ${{ steps.route.outputs.fallback-model }}
           --allowedTools "Read"
diff --git a/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md b/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md
index 488e062..9edaa87 100644
--- a/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md
+++ b/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md
@@ -88,11 +88,14 @@ Do **not** add `oc/big-pickle` while billing is unknown (not labeled free in the
 | Action / role | Primary | `--fallback-model` (ordered) |
 | --- | --- | --- |
 | ai-review context | `claude/claude-haiku-4-5-20251001` | `claude/cursor/composer-2.5,oc/nemotron-3.5-lightning-free,auto/best-free` |
-| ai-review review (tiny) | `claude/claude-sonnet-5` | same Cursor→free chain (CLI max 3) |
-| ai-review review (large) | `claude/claude-opus-5` | same Cursor→free chain (CLI max 3) |
-| ai-qa review | `claude/claude-sonnet-5` | same Cursor→free chain (CLI max 3) |
+| ai-review review (tiny) | `claude/claude-sonnet-5` | `oc/nemotron-3.5-lightning-free,oc/deepseek-v4-flash-free,auto/best-free` (SO free only) |
+| ai-review review (large) | `claude/claude-opus-5` | same SO free chain |
+| ai-qa review | `claude/claude-sonnet-5` | same SO free chain |
+
+> **Do not** put `claude/cursor/claude-*` in `--fallback-model`: raw `/v1/messages` may 200, but Claude Code CLI hangs (`unrecognized_model` → `aborted_streaming`).
+> **Do not** put Cursor (including `composer-2.5`) on `--json-schema` review/QA stages: no Cursor ID advertises `structured_output` on this gateway, and Agent SDK docs note a mid-stream fallback can retract structured output — that is the inconclusive path.
+> Context (no schema) may still use `composer-2.5` as the Cursor tier.
 
-> **Do not** put `claude/cursor/claude-*` in `--fallback-model`: raw `/v1/messages` may 200, but Claude Code CLI hangs (`unrecognized_model` → `aborted_streaming`) and never reaches free. Use `composer-2.5` as the Cursor tier.
 
 Comment footer copy (**successful** reviews / QA comments only):
 

From dd8f5ab2da0a753b444dc98b875d7544d5e4c021 Mon Sep 17 00:00:00 2001
From: Hussam Aldarwish <h.aldarwish@eduly.com>
Date: Mon, 17 Aug 2026 12:08:38 +0300
Subject: [PATCH 10/10] fix(ai-review): unescape literal \\n in
 comment_markdown before publish

Some structured outputs deliver comment_markdown with two-character
"\n" sequences (seen on PR #52 review 4949356509), which GitHub
renders as one smashed block under the banner.

Refs #51
---
 ai-review/lib/publish.js      | 17 +++++++++++++++++
 ai-review/lib/publish.test.js | 31 +++++++++++++++++++++++++++++++
 2 files changed, 48 insertions(+)

diff --git a/ai-review/lib/publish.js b/ai-review/lib/publish.js
index a9f1f89..b04a5d6 100644
--- a/ai-review/lib/publish.js
+++ b/ai-review/lib/publish.js
@@ -13,6 +13,20 @@
 const STATUS_BLOCK_START = "<!-- ai-review-status -->";
 const STATUS_BLOCK_END = "<!-- /ai-review-status -->";
 
+// Some structured-output paths deliver comment_markdown with literal
+// two-character "\n" sequences instead of real newlines (observed on
+// EdulyCom/github-actions#52 review 4949356509). GitHub then renders the
+// findings as one smashed line under the banner. Only rewrite when the
+// literal escapes dominate real newlines so a normal body that mentions
+// `\n` in prose/code is left alone.
+function unescapeLiteralNewlines(markdown) {
+  if (typeof markdown !== "string" || !markdown.includes("\\n")) return markdown;
+  const literal = (markdown.match(/\\n/g) || []).length;
+  const real = (markdown.match(/\n/g) || []).length;
+  if (literal === 0 || literal < real) return markdown;
+  return markdown.replace(/\\r\\n/g, "\n").replace(/\\n/g, "\n").replace(/\\t/g, "\t");
+}
+
 // The review-stage prompt instructs the model not to prepend its own
 // verdict token / confidence-merge-risk line / HTML marker to
 // comment_markdown (the caller owns that banner), but model
@@ -21,6 +35,7 @@ const STATUS_BLOCK_END = "<!-- /ai-review-status -->";
 // duplicate the banner above it.
 function stripLeadingBannerArtifacts(markdown) {
   if (!markdown) return markdown;
+  markdown = unescapeLiteralNewlines(markdown);
   const verdictTokenRe = /^\*\*(?:✅ PASS|❌ FAIL)\*\*\s*$/;
   const confidenceLineRe = /^Confidence:\s*\d+\s*·\s*Merge risk:\s*\S+\s*$/i;
   const htmlCommentRe = /^<!--.*-->\s*$/;
@@ -116,6 +131,7 @@ function buildReviewBody({
  */
 function buildInconclusiveBody(salvaged, opts = {}) {
   const modelUsed = opts && opts.modelUsed;
+  salvaged = unescapeLiteralNewlines(salvaged || "");
   return [
     "<!-- ai-review -->",
     "### ⚠️ AI Review — inconclusive (re-run required)",
@@ -224,6 +240,7 @@ function upsertStatusBlock(body, block) {
 }
 
 module.exports = {
+  unescapeLiteralNewlines,
   stripLeadingBannerArtifacts,
   buildReviewBody,
   buildInconclusiveBody,
diff --git a/ai-review/lib/publish.test.js b/ai-review/lib/publish.test.js
index eea4f8f..3fcd932 100644
--- a/ai-review/lib/publish.test.js
+++ b/ai-review/lib/publish.test.js
@@ -14,6 +14,37 @@ const {
 
 // --- stripLeadingBannerArtifacts --------------------------------------------
 
+test("unescapes literal \\\\n when comment_markdown was double-escaped", () => {
+  // Observed on PR #52 review 4949356509: model/schema path left `\\n` as
+  // two characters, so GitHub rendered the body as one smashed line.
+  const raw =
+    "### P0 — Blockers\\n\\n_None._\\n\\n### P1 — Should Fix\\n\\n_None._\\n\\n### P2 — Nice-to-Have\\n\\n- drift risk";
+  const out = stripLeadingBannerArtifacts(raw);
+  assert.equal(
+    out,
+    [
+      "### P0 — Blockers",
+      "",
+      "_None._",
+      "",
+      "### P1 — Should Fix",
+      "",
+      "_None._",
+      "",
+      "### P2 — Nice-to-Have",
+      "",
+      "- drift risk",
+    ].join("\n")
+  );
+  assert.equal(out.includes("\\n"), false);
+});
+
+test("leaves normal markdown with real newlines alone", () => {
+  const raw = "### P0 — Blockers\n\n_None._\n\nUse `\\\\n` in a code span occasionally.";
+  const out = stripLeadingBannerArtifacts(raw);
+  assert.equal(out, raw);
+});
+
 test("strips a leading verdict token line", () => {
   const out = stripLeadingBannerArtifacts("**✅ PASS**\n\nReal content here.");
   assert.equal(out, "Real content here.");