fix: harden partial 429 success, fallbacks, zh errors

Retry HTTP 429 with backoff; keep models beyond first 3 as
fallbacks; degrade agreement when some models fail; force-high
risk only if zero usable responses. Clearer zh-CN errors/docs for
dogfood uses: ./ rebase + workflow_dispatch re-run.

.github/workflows/gatekeep-pr.yml

@@ -3,6 +3,12 @@
 #
 # Secret: OPENROUTER_API_KEY (already configured on this repo)
 # Action: in-repo composite at repository root (uses: ./)
+#
+# Important (uses: ./):
+# - Checkout is the PR head, so Action fixes on main do NOT apply until the
+#   PR branch rebases/merges main. Stale heads can look "broken" after a fix.
+# - Re-run: push an empty commit, or Actions → this workflow → Run workflow
+#   (workflow_dispatch; uses default branch / manual inputs as configured).
 
 name: Gatekeep
 

README.md

@@ -22,6 +22,8 @@ AI-assisted PR review for **your own repos**: parallel multi-model review across
 
 `layue13/gatekeep` reviews **its own** PRs via [`.github/workflows/gatekeep-pr.yml`](.github/workflows/gatekeep-pr.yml) with `uses: ./` and `secrets.OPENROUTER_API_KEY`. Open a PR against `main` to see a review (Gitea Actions must have a runner — same pool as other `layue13` repos).
 
+Because dogfood uses `uses: ./`, **rebase/merge `main` into the PR branch** after Action fixes land, or the PR still runs the old Action. Re-run via new push or **Actions → Gatekeep → Run workflow**. Free-model 429s are retried; extras in `models:` after the first 3 are fallbacks; risk is force-high only if **zero** models succeed.
+
 Allowlists / Gitea-on-github.com limitations: [`docs/remote-uses.md`](docs/remote-uses.md).
 
 ## Providers YAML

README.zh-CN.md

@@ -28,6 +28,8 @@ AI 辅助的 PR 审查 Action：多模型并行投票、按严重级别输出可
 
 `layue13/gatekeep` 用 [`.github/workflows/gatekeep-pr.yml`](.github/workflows/gatekeep-pr.yml) 对本仓库 PR 跑 Gatekeep：`uses: ./` + `secrets.OPENROUTER_API_KEY`。向 `main` 开 PR 即可触发（需有可用的 Gitea Actions runner）。
 
+**注意：** `uses: ./` 用的是 **PR 分支上的 Action**。main 上的修复要进 PR，需要 rebase/merge main，否则看起来像“修了还挂”。重跑：再推一次提交，或 Actions → Gatekeep → **Run workflow**。免费模型 429 会退避重试；`models` 列表第 4 个及以后作备用；**只有全部模型失败**才强制抬高风险档为 high。
+
 github.com runner 若不允许非 github.com Action 源，见 `docs/remote-uses.md`。
 
 ## 多供应商（providers.yaml）

action.yml

@@ -58,10 +58,18 @@ runs:
     # Runner image / consumer workflow must provide Node >= 18.
     - name: Require Node.js
       shell: bash
+      env:
+        GATEKEEP_LANGUAGE: ${{ inputs.language }}
       run: |
         if ! command -v node >/dev/null 2>&1; then
-          echo "::error::Node.js >= 18 required for Gatekeep (merge-findings.js / publish.js)."
+          lang="${GATEKEEP_LANGUAGE:-}"
-          echo "Preinstall Node on the runner, or add a mirrored setup-node step before uses: ./."
+          if [[ "${lang,,}" == zh* ]]; then
+            echo "::error::Gatekeep 需要 Node.js ≥ 18（merge-findings.js / publish.js）。"
+            echo "请在 runner 预装 Node，或在 uses: ./ 之前加本 forge 已镜像的 setup-node；不要依赖未镜像的 actions/setup-node。"
+          else
+            echo "::error::Node.js >= 18 required for Gatekeep (merge-findings.js / publish.js)."
+            echo "Preinstall Node on the runner, or add a mirrored setup-node step before uses: ./."
+          fi
           exit 1
         fi
         node --version

docs/dogfood.md

@@ -11,12 +11,14 @@ The Action repo itself runs Gatekeep on its own pull requests.
 | Proven | [PR #2](https://git.potatolab.uk/layue13/gatekeep/pulls/2) got `<!-- gatekeep-review -->`; run 15833 **success** |
 | Report language | `defaults/providers.yaml` → `language: zh-CN` (override with Action `language:`) |
 
-Gitea notes: no `actions/setup-node` mirror — Action requires runner Node ≥18. PyYAML installs via workspace venv (PEP 668). Free OpenRouter models may 429; summary softens that messaging and prefers cohere/nvidia in the first-3 set.
+Gitea notes: no `actions/setup-node` mirror — Action requires runner Node ≥18. PyYAML installs via workspace venv (PEP 668). Free OpenRouter models may 429: Action retries with backoff, tries `fallback_models` (list entries after the first 3), and only force-elevates risk to high when **zero** models succeed. Summary is Chinese when `language: zh-CN`.
 
 ## How to see a review
 
-1. Branch off `main`, make a tiny change, open a PR on https://git.potatolab.uk/layue13/gatekeep
+1. Branch off **current** `main` (or merge/rebase main into the PR branch — `uses: ./` loads Action code from the PR head)
-2. Wait for the **Gatekeep** workflow (same Actions runner pool as other `layue13` repos)
+2. Open/push a PR on https://git.potatolab.uk/layue13/gatekeep
-3. Look for the summary comment `<!-- gatekeep-review -->` on the PR
+3. Wait for the **Gatekeep** workflow (same Actions runner pool as other `layue13` repos)
+4. Look for the summary comment `<!-- gatekeep-review -->` (updated in place on re-run)
+5. Manual re-run: Actions → Gatekeep → **Run workflow** (`workflow_dispatch`), or push another commit
 
 Consumer sample workflows remain under top-level `workflows/` for copy-paste into *other* repos’ `.github/workflows/`.

scripts/call-model.sh

@@ -2,6 +2,7 @@
 # Call one model via provider protocol.
 # Protocols: openai | openai-completions | openai-responses | anthropic
 # (openai / openai-compatible → chat completions)
+# Retries HTTP 429 with exponential backoff (GATEKEEP_HTTP_RETRIES, default 3).
 # Args: <provider> <model> <protocol> <base_url> <prompt.json> <out.json> <secrets.json>
 set -euo pipefail
 
@@ -21,6 +22,7 @@ if [[ -z "$KEY" ]]; then
 fi
 
 MAX_TOKENS="${GATEKEEP_MAX_TOKENS:-4096}"
+RETRIES="${GATEKEEP_HTTP_RETRIES:-3}"
 SYSTEM=$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1]))["system"])' "$PROMPT_FILE")
 USER=$(python3 -c 'import json,sys; print(json.load(open(sys.argv[1]))["user"])' "$PROMPT_FILE")
 
@@ -54,13 +56,9 @@ print(json.dumps({
 }))
 PY
 )
-  HTTP_CODE=$(curl -sS -o "$OUT_FILE.raw" -w "%{http_code}" \
+  URL="${BASE_URL%/}/chat/completions"
-    "${BASE_URL%/}/chat/completions" \
+  EXTRA_HDR=(-H "HTTP-Referer: https://github.com/${GITHUB_REPOSITORY:-gatekeep}" -H "X-Title: Gatekeep")
-    -H "Authorization: Bearer $KEY" \
+  AUTH=(-H "Authorization: Bearer $KEY")
-    -H "Content-Type: application/json" \
-    -H "HTTP-Referer: https://github.com/${GITHUB_REPOSITORY:-gatekeep}" \
-    -H "X-Title: Gatekeep" \
-    -d "$BODY")
   EXTRACT=openai
 elif [[ "$WIRE" == "openai-responses" ]]; then
   BODY=$(python3 - "$MODEL" "$SYSTEM" "$USER" "$MAX_TOKENS" <<'PY'
@@ -77,11 +75,9 @@ print(json.dumps({
 }))
 PY
 )
-  HTTP_CODE=$(curl -sS -o "$OUT_FILE.raw" -w "%{http_code}" \
+  URL="${BASE_URL%/}/responses"
-    "${BASE_URL%/}/responses" \
+  EXTRA_HDR=()
-    -H "Authorization: Bearer $KEY" \
+  AUTH=(-H "Authorization: Bearer $KEY")
-    -H "Content-Type: application/json" \
-    -d "$BODY")
   EXTRACT=openai-responses
 elif [[ "$WIRE" == "anthropic" ]]; then
   BODY=$(python3 - "$MODEL" "$SYSTEM" "$USER" "$MAX_TOKENS" <<'PY'
@@ -96,21 +92,36 @@ print(json.dumps({
 }))
 PY
 )
-  HTTP_CODE=$(curl -sS -o "$OUT_FILE.raw" -w "%{http_code}" \
+  URL="${BASE_URL%/}/v1/messages"
-    "${BASE_URL%/}/v1/messages" \
+  EXTRA_HDR=(-H "anthropic-version: 2023-06-01")
-    -H "x-api-key: $KEY" \
+  AUTH=(-H "x-api-key: $KEY")
-    -H "anthropic-version: 2023-06-01" \
-    -H "Content-Type: application/json" \
-    -d "$BODY")
   EXTRACT=anthropic
 fi
 
-if [[ "$HTTP_CODE" != "200" ]]; then
+attempt=0
+HTTP_CODE="000"
+while true; do
+  HTTP_CODE=$(curl -sS -o "$OUT_FILE.raw" -w "%{http_code}" \
+    "$URL" \
+    "${AUTH[@]}" \
+    -H "Content-Type: application/json" \
+    "${EXTRA_HDR[@]}" \
+    -d "$BODY" || echo "000")
+  if [[ "$HTTP_CODE" == "200" ]]; then
+    break
+  fi
+  if [[ "$HTTP_CODE" == "429" && "$attempt" -lt "$RETRIES" ]]; then
+    sleep_s=$((2 ** attempt))
+    echo "$LABEL HTTP 429 — retry $((attempt + 1))/${RETRIES} after ${sleep_s}s" >&2
+    sleep "$sleep_s"
+    attempt=$((attempt + 1))
+    continue
+  fi
   echo "$LABEL failed HTTP $HTTP_CODE" >&2
   head -c 800 "$OUT_FILE.raw" >&2 || true
   echo "{\"error\":\"http\",\"model\":\"$MODEL\",\"provider\":\"$PROVIDER\",\"http\":$HTTP_CODE}" > "$OUT_FILE"
   exit 0
-fi
+done
 
 python3 - "$OUT_FILE.raw" "$OUT_FILE" "$MODEL" "$PROVIDER" "$EXTRACT" <<'PY'
 import json,sys
@@ -123,7 +134,6 @@ if kind == "anthropic":
         if c.get("type") == "text"
     )
 elif kind == "openai-responses":
-    # Best-effort extract from Responses API
     text = raw.get("output_text") or ""
     if not text:
         chunks = []

scripts/merge-findings.js

@@ -167,10 +167,22 @@ for (const m of models) {
   }
 }
 
+// Prefer successful model tiers; only if every call failed, fail-closed high.
+const okModels = models.filter((m) => !m.error);
+const modelErrors = models.filter((m) => m.error).length;
+const degraded = modelErrors > 0 || okModels.length < agreementMin;
+
+// When some models 429/fail, lower the bar to surface usable findings from survivors
+// (still require at least 1 successful model). Full fail-closed high only if zero ok.
+let effectiveAgreement = agreementMin;
+if (okModels.length > 0 && okModels.length < agreementMin) {
+  effectiveAgreement = Math.max(1, okModels.length);
+}
+
 const agreed = [];
 const disputed = [];
 for (const f of bucket.values()) {
-  if (f.models.length >= agreementMin) agreed.push(f);
+  if (f.models.length >= effectiveAgreement) agreed.push(f);
   else disputed.push(f);
 }
 
@@ -178,8 +190,6 @@ const bySeverity = (a, b) => (sevRank[a.severity] ?? 9) - (sevRank[b.severity] ?
 agreed.sort(bySeverity);
 disputed.sort(bySeverity);
 
-// Prefer successful model tiers; only if every call failed, fail-closed high.
-const okModels = models.filter((m) => !m.error);
 const tierSources = okModels.length ? okModels : models;
 let risk_tier = floor;
 for (const m of tierSources) risk_tier = maxTier(risk_tier, m.risk_tier);
@@ -193,13 +203,13 @@ if (scanHardFail) risk_tier = maxTier(risk_tier, "high");
 const openSecurityBlocking =
   scanHardFail || agreed.some((f) => (f.kind === "security" || f.severity === "Critical") && f.blocking);
 
-const modelErrors = models.filter((m) => m.error).length;
-
 const result = {
   risk_tier,
   policy_floor: floor,
   agreement_min: agreementMin,
+  effective_agreement_min: effectiveAgreement,
   language,
+  degraded,
   model_errors: modelErrors,
   models_ok: okModels.length,
   models: models.map((m) => ({

scripts/publish.js

@@ -81,8 +81,13 @@ const T = zh
         "看起来风险低 — Gatekeep + 安全绿灯后可不需人工（仅当 kill switch 打开才可自动合并）。",
       approveMed: "需要 **1 名人工**过目。Bot 点赞不算人工。",
       approveHigh: "需要 **2 名人工**批准。Bot 不能顶替。",
-      errNote: (n, ok) =>
+      errNote: (n, ok, degAgree) =>
-        `> **说明：** ${n} 个模型调用失败（常见于免费档限流 429）。风险档按**失败关闭**规则取高，**不等于**代码一定有问题。成功响应：${ok} 个。`,
+        `> **降级说明：** ${n} 个模型调用失败（常见免费档限流 429；已重试/尝试备用模型）。` +
+        (ok === 0
+          ? `全部失败 → 风险档**失败关闭为 high**，**不等于**代码一定有问题。`
+          : `已用 ${ok} 个成功响应继续审查` +
+            (degAgree ? `（一致门槛临时降为 ${degAgree}）` : "") +
+            `；失败关闭的 high **仅在零成功响应时**强制抬升。`),
       statusBlock: "存在未关闭的安全 / Critical 结论 — 合并阻断",
       statusOk: (tier) => `PR 风险 ${tier} — 无未关闭安全阻断`,
       modelsAgree: "一致门槛 ≥",
@@ -135,8 +140,13 @@ const T = zh
         "Looks low-risk — no human needed once Gatekeep + security are green (auto-merge only if the kill switch is on).",
       approveMed: "Needs **1 human** look. Bot thumbs-up don't count as that look.",
       approveHigh: "Needs **2 human** approvals. Bots never fill those seats.",
-      errNote: (n, ok) =>
+      errNote: (n, ok, degAgree) =>
-        `> **Note:** ${n} model call(s) failed (often free-tier 429). Risk tier is **fail-closed** high when calls fail — that does **not** mean the code is necessarily bad. Successful responses: ${ok}.`,
+        `> **Degraded:** ${n} model call(s) failed (often free-tier 429; retried / tried fallbacks). ` +
+        (ok === 0
+          ? `Zero usable responses → risk tier **fail-closed high** (not proof the code is bad).`
+          : `Reviewing from ${ok} successful response(s)` +
+            (degAgree ? ` (agreement bar temporarily ${degAgree})` : "") +
+            `; fail-closed high is forced only when **zero** models succeed.`),
       statusBlock: "Open security / Critical findings — merge blocked",
       statusOk: (tier) => `PR risk ${tier} — no open security blockers`,
       modelsAgree: "≥",
@@ -236,11 +246,16 @@ const okCount = Number(
     ? merged.models_ok
     : (merged.models || []).filter((m) => !m.error).length,
 );
-const errBanner = errCount > 0 ? `\n${T.errNote(errCount, okCount)}\n` : "";
+const effAgree = merged.effective_agreement_min;
+const degAgree =
+  merged.degraded && effAgree != null && effAgree < merged.agreement_min ? effAgree : null;
+const errBanner =
+  errCount > 0 || merged.degraded ? `\n${T.errNote(errCount, okCount, degAgree)}\n` : "";
 
 const agreementBar = zh
-  ? `${T.modelsAgree}${merged.agreement_min} 个模型`
+  ? `${T.modelsAgree}${merged.agreement_min} 个模型` +
-  : `≥${merged.agreement_min} models`;
+    (degAgree ? `（本轮降为 ${degAgree}）` : "")
+  : `≥${merged.agreement_min} models` + (degAgree ? ` (this run: ${degAgree})` : "");
 
 const body = `${marker}
 ${T.title}

scripts/resolve-providers.py

@@ -46,8 +46,14 @@ KNOWN_DEFAULTS = {
 }
 
 
-def die(msg: str, code: int = 1) -> None:
+def _zh() -> bool:
-    print(f"gatekeep: {msg}", file=sys.stderr)
+    lang = (os.environ.get("GATEKEEP_LANGUAGE") or "").strip().lower()
+    return lang.startswith("zh")
+
+
+def die(msg: str, code: int = 1, msg_zh: str | None = None) -> None:
+    text = msg_zh if (_zh() and msg_zh) else msg
+    print(f"gatekeep: {text}", file=sys.stderr)
     sys.exit(code)
 
 
@@ -57,7 +63,10 @@ def load_raw(path: Path) -> dict:
         try:
             import yaml  # type: ignore
         except ImportError:
-            die("PyYAML required to load providers.yaml (pip install pyyaml)")
+            die(
+                "PyYAML required to load providers.yaml (pip install pyyaml)",
+                msg_zh="加载 providers.yaml 需要 PyYAML（请 pip install pyyaml 或使用 Action 的 venv 步骤）",
+            )
         raw = yaml.safe_load(text) or {}
     else:
         raw = json.loads(text)
@@ -240,24 +249,40 @@ def main() -> None:
     requested = uniq
 
     if len(requested) < 2:
-        die(f"need 2–3 models for parallel review; got {len(requested)}")
+        die(
+            f"need 2–3 models for parallel review; got {len(requested)}",
+            msg_zh=f"并行审查至少需要 2–3 个模型，当前只有 {len(requested)} 个",
+        )
+    fallback_requested: list[dict] = []
     if len(requested) > 3:
-        print(f"gatekeep: truncating to first 3 of {len(requested)} models", file=sys.stderr)
+        print(
+            f"gatekeep: primary parallel set = first 3 of {len(requested)}; "
+            f"rest kept as fallbacks on 429/failure",
+            file=sys.stderr,
+        )
+        fallback_requested = requested[3:]
         requested = requested[:3]
 
     provider_meta: dict[str, dict] = {}
-    for pid in {r["provider"] for r in requested}:
+    for pid in {r["provider"] for r in requested + fallback_requested}:
         pconf = providers[pid]
         key_name = str(pconf["key"])
         api_key = secret_value(key_name)
         if not api_key:
             die(
                 f"missing secret/env for provider {pid!r}: expected env var named {key_name!r} "
-                f"(from providers.*.key — free-form; set it on the workflow step env from secrets)."
+                f"(from providers.*.key — free-form; set it on the workflow step env from secrets).",
+                msg_zh=(
+                    f"缺少供应商 {pid!r} 的密钥：需要环境变量 {key_name!r}"
+                    f"（来自 providers.*.key；请在 workflow step 的 env 里从 secrets 注入）。"
+                ),
             )
         available = discover_model_ids(pconf["base_url"], pconf["protocol"], api_key)
         if not available:
-            die(f"provider {pid}: model list empty — check base_url/protocol/key")
+            die(
+                f"provider {pid}: model list empty — check base_url/protocol/key",
+                msg_zh=f"供应商 {pid} 模型列表为空 — 请检查 base_url / protocol / key",
+            )
         provider_meta[pid] = {
             "base_url": pconf["base_url"].rstrip("/"),
             "protocol": pconf["protocol"],
@@ -268,26 +293,30 @@ def main() -> None:
             "_available": available,
         }
 
-    plan_models = []
+    def to_plan_entry(r: dict) -> dict:
-    for r in requested:
         pid, model = r["provider"], r["model"]
         avail = provider_meta[pid]["_available"]
         if model not in avail:
             sample = sorted(avail)[:8]
             die(
                 f"model {model!r} not available on provider {pid} "
-                f"({provider_meta[pid]['available_count']} listed). Examples: {sample}"
+                f"({provider_meta[pid]['available_count']} listed). Examples: {sample}",
+                msg_zh=(
+                    f"模型 {model!r} 在供应商 {pid} 上不可用"
+                    f"（列表共 {provider_meta[pid]['available_count']} 个）。示例：{sample}"
+                ),
             )
-        plan_models.append(
+        return {
-            {
+            "provider": pid,
-                "provider": pid,
+            "model": model,
-                "model": model,
+            "base_url": provider_meta[pid]["base_url"],
-                "base_url": provider_meta[pid]["base_url"],
+            "protocol": provider_meta[pid]["protocol"],
-                "protocol": provider_meta[pid]["protocol"],
+            "wire_protocol": provider_meta[pid]["wire_protocol"],
-                "wire_protocol": provider_meta[pid]["wire_protocol"],
+            "key_name": provider_meta[pid]["key_name"],
-                "key_name": provider_meta[pid]["key_name"],
+        }
-            }
+
-        )
+    plan_models = [to_plan_entry(r) for r in requested]
+    fallback_models = [to_plan_entry(r) for r in fallback_requested]
 
     secrets_path = out_path.with_suffix(".secrets.json")
     secrets_path.write_text(
@@ -306,6 +335,7 @@ def main() -> None:
         "max_tokens": int(cfg.get("max_tokens", 4096)),
         "language": language,
         "models": plan_models,
+        "fallback_models": fallback_models,
         "secrets_file": str(secrets_path),
         "providers_used": sorted(provider_meta.keys()),
     }
@@ -315,6 +345,7 @@ def main() -> None:
             {
                 "providers": plan["providers_used"],
                 "models": [f"{m['provider']}:{m['model']}" for m in plan_models],
+                "fallbacks": [f"{m['provider']}:{m['model']}" for m in fallback_models],
                 "keys": {m["provider"]: m["key_name"] for m in plan_models},
                 "agreement_min": plan["agreement_min"],
                 "language": plan["language"],

scripts/review.sh

@@ -151,7 +151,7 @@ json.dump({"system": system, "user": json.dumps(user, indent=2), "language": lan
 print(f"Prompt ready; floor={floor}; language={language}; plan={len(plan.get('models',[]))} models", flush=True)
 PY
 
-# Fire models in parallel (may be mixed providers)
+# Fire primary models in parallel (may be mixed providers)
 pids=()
 i=0
 while IFS= read -r line; do
@@ -173,6 +173,52 @@ for pid in "${pids[@]}"; do
 done
 echo "Parallel model calls finished (wait_ec=$ec)" >&2
 
+# On partial failure, try fallback_models from providers.yaml (beyond the first 3)
+python3 - "$MODELS_DIR" "$PLAN" <<'PY' >"$OUT/fallback-needed.txt"
+import json, pathlib, sys
+models_dir = pathlib.Path(sys.argv[1])
+plan = json.load(open(sys.argv[2]))
+failed = ok = 0
+for p in models_dir.glob("model-*.json"):
+    try:
+        j = json.loads(p.read_text())
+    except Exception:
+        failed += 1
+        continue
+    if j.get("error") or j.get("parse_error"):
+        failed += 1
+    else:
+        ok += 1
+fallbacks = plan.get("fallback_models") or []
+# Aim for ≥2 usable responses; if zero ok, try up to 3 reserves
+if ok == 0:
+    need = min(len(fallbacks), 3)
+elif ok < 2:
+    need = min(len(fallbacks), 2 - ok)
+else:
+    need = 0
+print(need)
+print(f"gatekeep: primary ok={ok} failed={failed} fallbacks={len(fallbacks)} need={need}", file=sys.stderr)
+PY
+NEED_FB=$(head -1 "$OUT/fallback-needed.txt" | tr -d '[:space:]')
+NEED_FB=${NEED_FB:-0}
+if [[ "$NEED_FB" -gt 0 ]]; then
+  echo "gatekeep: trying $NEED_FB fallback model(s) after partial failure" >&2
+  fb_i=0
+  while IFS= read -r line; do
+    fb_i=$((fb_i + 1))
+    [[ "$fb_i" -gt "$NEED_FB" ]] && break
+    provider=$(printf '%s' "$line" | python3 -c 'import json,sys; print(json.load(sys.stdin)["provider"])')
+    model=$(printf '%s' "$line" | python3 -c 'import json,sys; print(json.load(sys.stdin)["model"])')
+    protocol=$(printf '%s' "$line" | python3 -c 'import json,sys; print(json.load(sys.stdin)["protocol"])')
+    base_url=$(printf '%s' "$line" | python3 -c 'import json,sys; print(json.load(sys.stdin)["base_url"])')
+    safe=$(echo "fb_${provider}_${model}" | tr '/:' '__')
+    bash "$ACTION/scripts/call-model.sh" \
+      "$provider" "$model" "$protocol" "$base_url" \
+      "$OUT/prompt.json" "$MODELS_DIR/model-fb$fb_i-$safe.json" "$SECRETS_FILE" || true
+  done < <(python3 -c 'import json,sys; [print(json.dumps(m)) for m in json.load(open(sys.argv[1])).get("fallback_models") or []]' "$PLAN")
+fi
+
 # Scrub secrets file from workspace artifact path if present
 rm -f "$SECRETS_FILE"