feat: reserve queue capacity for recent models

README.md

@@ -55,6 +55,9 @@ Optional tuning:
 - `MODELHUB_QUEUE_CLEANUP_INTERVAL_CYCLES` default `120`; cleanup also runs once at startup
 - `MODELHUB_QUEUE_CLEANUP_READ_CONCURRENCY` default `6`
 - `MODELHUB_QUEUE_CLEANUP_REPORT_PATH` default `.modelhub_state/queue_cleanup_latest.json`
+- `MODELHUB_OLD_MODEL_QUEUE_THRESHOLD` default `80` per account
+- `MODELHUB_DYNAMIC_OLD_MODEL_CLEANUP_THRESHOLD` default `95` per account
+- `MODELHUB_RECENT_MODEL_DAYS` default `7`
 
 ## Adaptive GPU Strategy
 
@@ -135,6 +138,25 @@ same model or infer failure from historical similarity. The cleanup repeats
 every 120 poll cycles by default and writes its full evidence report to
 `.modelhub_state/queue_cleanup_latest.json`.
 
+Each account reserves only its first 80 active queue positions for models older
+than seven days. Old-model reservations are made under the same account lock as
+capacity reservations, so concurrent submissions cannot cross position 80.
+Once every account has at least 80 active tasks, discovery is capped at the
+seven-day window and only recent models can fill positions 81–100. Unknown model
+timestamps are treated as old for new submissions.
+
+Queue cleanup applies a two-level version of the policy. On poller cycle 1 it
+first stops deterministic OOM tasks, recalculates each account's surviving task
+order by numeric task ID, and removes older-than-seven-days tasks after position
+80. On later scheduled cleanup cycles, the age threshold relaxes to 95: old
+tasks in positions 81-95 are retained and only old tasks at position 96 or later
+are removed. New old-model submissions are still blocked at position 80, so the
+upper queue remains available to recent models without repeatedly over-pruning
+existing work. Recent overflow tasks are always kept. ModelScope metadata
+failures fail closed and never trigger cancellation. Immediately before
+mutation, task ownership, active status, and post-OOM queue position are checked
+again.
+
 The verified capacities, safe repository-size boundaries, evidence hierarchy,
 and source links are recorded in
 `docs/gpu-memory-capacity-2026-08-10.md`.
@@ -218,12 +240,16 @@ second active-state check and fail-closed handling for incomplete evidence.
 Version `2026.08.11.2` removes LLM inference from every online path; credentials
 alone cannot activate it, and unresolved cases remain available for offline,
 human-reviewed rule development.
+Version `2026.08.11.3` adds atomic per-account 80/7-day admission, prevents
+history expansion once old-model positions are exhausted, performs a strict
+position-80 startup cleanup, and relaxes scheduled age cleanup to position 95
+after OOM cleanup and a second queue check.
 
 ## Deploy
 
 Create a tag and submit the repository URL plus tag in "我的适配智能体".
 
 ```bash
-git tag agent-v12
+git tag agent-v17
-git push origin agent-v12
+git push origin agent-v17
 ```

modelhub_submmit_api/README.md

@@ -120,6 +120,12 @@ bash run_poll.sh --dry-run
   memory rule as new submissions. Only tasks whose own repository size times `1.20` exceeds
   their selected GPU capacity are stopped, after a fresh account-scoped active-state check.
   Incomplete size/capacity evidence is never used for cancellation.
+- Models older than seven days may occupy only positions 1–80 of each account's active queue.
+  The account pool enforces the boundary atomically. Once all accounts reach 80, discovery no
+  longer expands beyond seven days. Cleanup stops OOM tasks first, recalculates the surviving
+  queue order, and on startup stops old tasks still ranked 81 or later. Scheduled cleanup then
+  retains positions 81–95 and stops only old tasks ranked 96 or later. Recent overflow tasks stay.
+  Unknown ModelScope timestamps never authorize a cancellation.
 
 ## Important Flags
 
@@ -153,7 +159,12 @@ Common flags:
 - `--submit-concurrency`: concurrent task submission calls used by each cycle (0 = auto)
 - `--post-cycle-cooldown-seconds`: pause after a successful cycle before next cycle (default 2)
 - `--max-cycles`: optional hard stop for testing or batch windows
-- `--disable-queue-cleanup`: disable automatic deterministic OOM cleanup
+- `--disable-queue-cleanup`: disable automatic OOM and old-overflow queue cleanup
+
+The first automatic cleanup removes models older than seven days after each
+account's position 80. Later scheduled cleanups use position 95, while admission
+continues to reserve positions 81-100 for recent models. Override the later
+threshold with `MODELHUB_DYNAMIC_OLD_MODEL_CLEANUP_THRESHOLD`.
 
 Failure-informed preflight is enabled by default. It rejects deterministic
 missing-file and predicted-OOM cases, clamps unsafe context-length arguments,

modelhub_submmit_api/daily_runner.py

@@ -85,6 +85,18 @@ def build_parser() -> argparse.ArgumentParser:
         default=0,
         help="Maximum tasks to submit in one run (0 means unlimited)",
     )
+    parser.add_argument(
+        "--old-model-queue-threshold",
+        type=int,
+        default=int(os.getenv("MODELHUB_OLD_MODEL_QUEUE_THRESHOLD", "80")),
+        help=argparse.SUPPRESS,
+    )
+    parser.add_argument(
+        "--recent-model-days",
+        type=int,
+        default=int(os.getenv("MODELHUB_RECENT_MODEL_DAYS", "7")),
+        help=argparse.SUPPRESS,
+    )
     parser.add_argument("--disable-candidate-preflight", action="store_true", help=argparse.SUPPRESS)
     parser.add_argument("--llm-classifier-endpoint", default=os.getenv("MODELHUB_LLM_CLASSIFIER_ENDPOINT"), help=argparse.SUPPRESS)
     parser.add_argument("--llm-classifier-model", default=os.getenv("MODELHUB_LLM_CLASSIFIER_MODEL"), help=argparse.SUPPRESS)
@@ -240,6 +252,8 @@ def make_wave_namespace(base_args: argparse.Namespace, wave: WaveSpec) -> argpar
         capacity_probe_interval_cycles=getattr(base_args, "capacity_probe_interval_cycles", 3),
         submit_concurrency=getattr(base_args, "submit_concurrency", 1),
         max_submits_per_run=getattr(base_args, "max_submits_per_run", 0),
+        old_model_queue_threshold=getattr(base_args, "old_model_queue_threshold", 80),
+        recent_model_days=getattr(base_args, "recent_model_days", 7),
         disable_candidate_preflight=getattr(base_args, "disable_candidate_preflight", False),
         llm_classifier_endpoint=getattr(base_args, "llm_classifier_endpoint", None),
         llm_classifier_model=getattr(base_args, "llm_classifier_model", None),

modelhub_submmit_api/hf_discovery.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import os
 import threading
 import time
+from datetime import datetime, timezone
 from pathlib import PurePosixPath
 from typing import Any
 from urllib.parse import quote
@@ -71,6 +72,8 @@ class HuggingFaceDiscovery:
         self._repo_tree_lock = threading.Lock()
         self._model_config_cache: dict[str, tuple[dict[str, Any], str | None]] = {}
         self._model_config_lock = threading.Lock()
+        self._model_last_modified_cache: dict[str, datetime | None] = {}
+        self._model_last_modified_lock = threading.Lock()
         self._model_page_cache: dict[tuple[str, int, int], tuple[float, list[dict[str, Any]]]] = {}
         self._model_page_cache_ttl = max(
             0.0,
@@ -237,6 +240,47 @@ class HuggingFaceDiscovery:
             self._model_config_cache[repo_id] = result
         return dict(result[0]), result[1]
 
+    def get_model_last_modified(self, repo_id: str) -> datetime | None:
+        with self._model_last_modified_lock:
+            if repo_id in self._model_last_modified_cache:
+                return self._model_last_modified_cache[repo_id]
+
+        encoded_repo_id = "/".join(quote(part, safe="") for part in repo_id.split("/"))
+        try:
+            payload = self.legacy_http_client.request_json(
+                "GET",
+                f"/api/v1/models/{encoded_repo_id}",
+            )
+        except HttpJsonError as exc:
+            print(f"[modelscope] model_metadata_error repo={repo_id} error={exc}", flush=True)
+            payload = None
+
+        data = payload.get("Data") if isinstance(payload, dict) else None
+        if not isinstance(data, dict) and isinstance(payload, dict):
+            data = payload.get("data")
+        result: datetime | None = None
+        if isinstance(data, dict):
+            raw_value = (
+                data.get("LastUpdatedTime")
+                or data.get("lastUpdatedTime")
+                or data.get("last_modified")
+                or data.get("updated_at")
+            )
+            if isinstance(raw_value, (int, float)):
+                timestamp = float(raw_value)
+                if timestamp > 10_000_000_000:
+                    timestamp /= 1000.0
+                try:
+                    result = datetime.fromtimestamp(timestamp, tz=timezone.utc)
+                except (OverflowError, OSError, ValueError):
+                    result = None
+            else:
+                result = parse_datetime(raw_value)
+
+        with self._model_last_modified_lock:
+            self._model_last_modified_cache[repo_id] = result
+        return result
+
     def list_repo_tree(self, repo_id: str) -> list[dict[str, Any]]:
         with self._repo_tree_lock:
             cached = self._repo_tree_cache.get(repo_id)

modelhub_submmit_api/main.py

@@ -33,6 +33,7 @@ from modelhub_client import (
     ModelHubAPIError,
     ModelHubClient,
     ModelHubClientPool,
+    OldModelQueuePolicyError,
     is_duplicate_submission_error,
     is_model_uniqueness_error,
 )
@@ -107,6 +108,18 @@ def build_parser() -> argparse.ArgumentParser:
         default=0,
         help="Maximum tasks to submit in one run (0 means unlimited)",
     )
+    parser.add_argument(
+        "--old-model-queue-threshold",
+        type=int,
+        default=int(os.getenv("MODELHUB_OLD_MODEL_QUEUE_THRESHOLD", "80")),
+        help=argparse.SUPPRESS,
+    )
+    parser.add_argument(
+        "--recent-model-days",
+        type=int,
+        default=int(os.getenv("MODELHUB_RECENT_MODEL_DAYS", "7")),
+        help=argparse.SUPPRESS,
+    )
     parser.add_argument(
         "--disable-candidate-preflight",
         action="store_true",
@@ -548,9 +561,17 @@ def build_adaptive_scan_stages(
     initial_updated_after,
     initial_limit: int,
     max_models: int = ADAPTIVE_SCAN_MAX_MODELS,
+    allow_older_than_recent_window: bool = True,
+    recent_model_days: int = 7,
 ) -> list[dict[str, Any]]:
     max_models = max(1, min(ADAPTIVE_SCAN_MAX_MODELS, int(max_models)))
     initial_limit = min(max_models, max(1, int(initial_limit)))
+    recent_cutoff = now - timedelta(days=max(1, int(recent_model_days)))
+    if not allow_older_than_recent_window and (
+        initial_updated_after is None or initial_updated_after < recent_cutoff
+    ):
+        initial_updated_after = recent_cutoff
+
     stages: list[dict[str, Any]] = [
         {
             "name": "configured_window",
@@ -559,17 +580,16 @@ def build_adaptive_scan_stages(
         }
     ]
 
-    seven_days_ago = now - timedelta(days=7)
     thirty_days_ago = now - timedelta(days=30)
-    if initial_updated_after is not None and initial_updated_after > seven_days_ago:
+    if initial_updated_after is not None and initial_updated_after > recent_cutoff:
         stages.append(
             {
                 "name": "last_7_days",
-                "updatedAfter": seven_days_ago,
+                "updatedAfter": recent_cutoff,
                 "limit": min(max_models, max(initial_limit, ADAPTIVE_SCAN_MIN_FALLBACK_MODELS)),
             }
         )
-    if initial_updated_after is not None and initial_updated_after > thirty_days_ago:
+    if allow_older_than_recent_window and initial_updated_after is not None and initial_updated_after > thirty_days_ago:
         stages.append(
             {
                 "name": "last_30_days",
@@ -577,7 +597,7 @@ def build_adaptive_scan_stages(
                 "limit": min(max_models, max(initial_limit, 1500)),
             }
         )
-    if initial_updated_after is not None or initial_limit < max_models:
+    if allow_older_than_recent_window and (initial_updated_after is not None or initial_limit < max_models):
         stages.append(
             {
                 "name": "all_history",
@@ -598,6 +618,13 @@ def build_adaptive_scan_stages(
     return deduped
 
 
+def is_recent_model(last_modified: Any, *, reference_time, recent_model_days: int = 7) -> bool:
+    parsed = parse_datetime(last_modified)
+    if parsed is None:
+        return False
+    return parsed >= reference_time - timedelta(days=max(1, int(recent_model_days)))
+
+
 def collect_candidates_from_models(
     *,
     models: list[HFModelSummary],
@@ -702,7 +729,14 @@ def submit_candidate(
         payload["strategyId"] = strategy_id
     submit_time = utc_now()
     try:
-        response = modelhub_client.add_task(payload)
+        if hasattr(modelhub_client, "add_task_for_model"):
+            response = modelhub_client.add_task_for_model(
+                payload,
+                model_last_modified=candidate.get("lastModified"),
+                submitted_at=submit_time,
+            )
+        else:
+            response = modelhub_client.add_task(payload)
         task_id = extract_task_id_from_submit_response(response)
         if task_id is None:
             task_id = modelhub_client.find_recent_task_id(candidate["repoId"], candidate["targetGpu"], submit_time)
@@ -714,6 +748,12 @@ def submit_candidate(
             "responseData": response.get("data"),
         }
     except ModelHubAPIError as exc:
+        if isinstance(exc, OldModelQueuePolicyError):
+            return {
+                "outcome": "old_model_policy_skipped",
+                "candidate": candidate,
+                "reason": "all_accounts_at_old_model_queue_threshold",
+            }
         if is_model_uniqueness_error(exc):
             print(
                 f"[submit] skipped repo={candidate['repoId']} gpu={candidate['targetGpu']} "
@@ -799,6 +839,8 @@ def run_submission(
             clients,
             capacity_probe_interval_cycles=max(0, int(getattr(args, "capacity_probe_interval_cycles", 3) or 0)),
             capacity_state_path=Path(getattr(args, "capacity_state_path", DEFAULT_CAPACITY_STATE_PATH)),
+            old_model_queue_threshold=max(1, int(getattr(args, "old_model_queue_threshold", 80) or 80)),
+            recent_model_days=max(1, int(getattr(args, "recent_model_days", 7) or 7)),
         )
 
     if hasattr(modelhub_client, "begin_cycle"):
@@ -1065,11 +1107,25 @@ def run_submission(
 
     pipeline_tags = pipeline_tags_for_task_types(selected_task_types)
     explicit_scan_cap = max(0, int(getattr(args, "max_scan_models", 0) or 0))
+    recent_model_days = max(1, int(getattr(args, "recent_model_days", 7) or 7))
+    old_model_queue_threshold = max(1, int(getattr(args, "old_model_queue_threshold", 80) or 80))
+    old_model_slots_before_scan: int | None = None
+    if hasattr(modelhub_client, "old_model_submit_slots"):
+        old_model_slots_before_scan = int(modelhub_client.old_model_submit_slots())
+    allow_older_models_for_scan = old_model_slots_before_scan is None or old_model_slots_before_scan > 0
+    print(
+        f"[age-policy] recent_days={recent_model_days} account_threshold={old_model_queue_threshold} "
+        f"old_model_slots={old_model_slots_before_scan if old_model_slots_before_scan is not None else 'n/a'} "
+        f"scan_older={'on' if allow_older_models_for_scan else 'off'}",
+        flush=True,
+    )
     for stage in build_adaptive_scan_stages(
         now=now,
         initial_updated_after=updated_after,
         initial_limit=scan_limit,
         max_models=explicit_scan_cap or ADAPTIVE_SCAN_MAX_MODELS,
+        allow_older_than_recent_window=allow_older_models_for_scan,
+        recent_model_days=recent_model_days,
     ):
         if candidate_goal <= 0 or len(candidates) >= candidate_goal:
             break
@@ -1088,6 +1144,16 @@ def run_submission(
             )
         except TypeError:
             models = hf_discovery.list_recent_models(**query_kwargs)
+        if not allow_older_models_for_scan:
+            models = [
+                model
+                for model in models
+                if is_recent_model(
+                    model.last_modified,
+                    reference_time=now,
+                    recent_model_days=recent_model_days,
+                )
+            ]
 
         stage_candidates, stage_skipped, stage_failed, processed_count = collect_candidates_from_models(
             models=models,
@@ -1165,11 +1231,25 @@ def run_submission(
         # already-scanned pool until the desired number of real submissions is
         # reached or account capacity is genuinely exhausted.
         while len(submitted) < target_submit_count and len(attempted_candidates) < max_submit_attempts:
+            allow_older_models_now = (
+                bool(modelhub_client.can_submit_old_models())
+                if hasattr(modelhub_client, "can_submit_old_models")
+                else True
+            )
+            submission_reference_time = utc_now()
             remaining_candidates = [
                 candidate
                 for candidate in diversified_candidates
                 if candidate_key(candidate) not in attempted_keys
                 and not submission_exclusion_store.is_blocked(candidate["repoId"], candidate["targetGpu"])
+                and (
+                    allow_older_models_now
+                    or is_recent_model(
+                        candidate.get("lastModified"),
+                        reference_time=submission_reference_time,
+                        recent_model_days=recent_model_days,
+                    )
+                )
             ]
             remaining_candidates = one_candidate_per_model(remaining_candidates)
             desired_count = min(
@@ -1209,6 +1289,7 @@ def run_submission(
             batch_submitted_candidates: list[dict[str, Any]] = []
             batch_duplicate_candidates: list[dict[str, Any]] = []
             batch_uniqueness_rejected_candidates: list[dict[str, Any]] = []
+            batch_policy_skipped_candidates: list[dict[str, Any]] = []
             batch_failed_candidates: list[dict[str, Any]] = []
             for index in range(len(batch_candidates)):
                 result = ordered_results.get(index)
@@ -1243,6 +1324,16 @@ def run_submission(
                         }
                     )
                     continue
+                if result["outcome"] == "old_model_policy_skipped":
+                    batch_policy_skipped_candidates.append(candidate)
+                    skipped.append(
+                        {
+                            "repoId": candidate["repoId"],
+                            "targetGpu": candidate["targetGpu"],
+                            "reason": "all_accounts_at_old_model_queue_threshold",
+                        }
+                    )
+                    continue
                 if result["outcome"] in {"failed", "precheck_deferred"}:
                     batch_failed_candidates.append(candidate)
                     failed.append(
@@ -1290,7 +1381,7 @@ def run_submission(
             claim_store.mark_submitted(
                 [*batch_submitted_candidates, *batch_duplicate_candidates, *batch_uniqueness_rejected_candidates]
             )
-            claim_store.release(batch_failed_candidates)
+            claim_store.release([*batch_failed_candidates, *batch_policy_skipped_candidates])
             if strategy_manager is not None:
                 strategy_manager.record_accepted(batch_submitted_candidates)
 
@@ -1332,6 +1423,12 @@ def run_submission(
         "gpuStrategy": strategy_summary,
         "marketIntelligence": market_summary,
         "candidatePreflight": preflight_summary,
+        "agePolicy": {
+            "recentModelDays": recent_model_days,
+            "oldModelSubmitThreshold": old_model_queue_threshold,
+            "oldModelSlotsBeforeScan": old_model_slots_before_scan,
+            "olderHistoryScanEnabled": allow_older_models_for_scan,
+        },
         "scanLimit": scan_limit,
         "candidateGoal": candidate_goal,
         "scanStages": scan_stages,

modelhub_submmit_api/modelhub_client.py

@@ -9,7 +9,7 @@ from datetime import datetime, timedelta
 from pathlib import Path
 from typing import Any
 
-from common import format_modelhub_datetime, parse_datetime, read_json, runtime_instance_id, write_json
+from common import format_modelhub_datetime, parse_datetime, read_json, runtime_instance_id, utc_now, write_json
 from defaults import EMBEDDED_MODELHUB_XC_TOKEN
 from http_json import HttpJsonError, JsonHttpClient
 
@@ -21,6 +21,10 @@ class ModelHubAPIError(RuntimeError):
         self.payload = payload
 
 
+class OldModelQueuePolicyError(ModelHubAPIError):
+    """No account may accept an older model under the per-account queue policy."""
+
+
 def parse_model_submission_precheck(payload: Any) -> dict[str, Any]:
     """Validate the community lookup response instead of failing open."""
     if not isinstance(payload, dict):
@@ -385,12 +389,26 @@ class ModelHubClientPool:
         capacity_probe_interval_cycles: int = 3,
         capacity_probe_cooldown_cycles: int = 3,
         capacity_state_path: Path | str | None = None,
+        old_model_queue_threshold: int | None = None,
+        recent_model_days: int | None = None,
     ) -> None:
         if not clients:
             raise ValueError("At least one ModelHub client is required")
         self.clients = clients
         configured_cap = active_task_cap if active_task_cap is not None else os.getenv("MODELHUB_AGENT_ACTIVE_TASK_CAP", "100")
         self.active_task_cap = max(1, int(configured_cap))
+        configured_old_threshold = (
+            old_model_queue_threshold
+            if old_model_queue_threshold is not None
+            else os.getenv("MODELHUB_OLD_MODEL_QUEUE_THRESHOLD", "80")
+        )
+        configured_recent_days = (
+            recent_model_days
+            if recent_model_days is not None
+            else os.getenv("MODELHUB_RECENT_MODEL_DAYS", "7")
+        )
+        self.old_model_queue_threshold = max(1, int(configured_old_threshold))
+        self.recent_model_days = max(1, int(configured_recent_days))
         self._capacity_state_path = Path(capacity_state_path) if capacity_state_path else None
         self._account_keys = [self._account_key(client, index) for index, client in enumerate(clients)]
         self._account_caps = self._load_account_caps(self.active_task_cap)
@@ -671,12 +689,21 @@ class ModelHubClientPool:
     def processed_gpus_for_model(self, model_id: str) -> set[str]:
         return set(self.model_submission_precheck(model_id)["processedGpus"])
 
-    def _reserve_account(self, excluded: set[int]) -> tuple[int, int] | None:
+    def _reserve_account(
+        self,
+        excluded: set[int],
+        *,
+        effective_count_below: int | None = None,
+    ) -> tuple[int, int] | None:
         with self._state_lock:
             remaining_by_index = {
                 index: self._account_caps[index] - self._effective_count_locked(index)
                 for index in range(len(self.clients))
                 if index not in excluded
+                and (
+                    effective_count_below is None
+                    or self._effective_count_locked(index) < effective_count_below
+                )
             }
             usable = {index: remaining for index, remaining in remaining_by_index.items() if remaining > 0}
             if not usable:
@@ -696,7 +723,12 @@ class ModelHubClientPool:
             }
             return selected_index, reservation_id
 
-    def _reserve_probe_account(self, excluded: set[int]) -> tuple[int, int] | None:
+    def _reserve_probe_account(
+        self,
+        excluded: set[int],
+        *,
+        effective_count_below: int | None = None,
+    ) -> tuple[int, int] | None:
         with self._state_lock:
             if not self._capacity_probe_enabled:
                 return None
@@ -707,6 +739,10 @@ class ModelHubClientPool:
                 and index not in self._capacity_probe_attempted
                 and self._capacity_probe_cycle >= self._capacity_probe_cooldown_until[index]
                 and self._effective_count_locked(index) >= self._account_caps[index]
+                and (
+                    effective_count_below is None
+                    or self._effective_count_locked(index) < effective_count_below
+                )
             ]
             if not eligible:
                 return None
@@ -763,17 +799,58 @@ class ModelHubClientPool:
         self._persist_account_caps()
         print(f"[capacity] account={index + 1:02d} discovered_cap={discovered_cap}", flush=True)
 
+    def old_model_submit_slots(self) -> int:
+        """Return queue positions where models older than the recent window are allowed."""
+        self._refresh_active_counts()
+        with self._state_lock:
+            return sum(
+                max(
+                    0,
+                    min(self._account_caps[index], self.old_model_queue_threshold)
+                    - self._effective_count_locked(index),
+                )
+                for index in range(len(self.clients))
+            )
+
+    def can_submit_old_models(self) -> bool:
+        return self.old_model_submit_slots() > 0
+
+    def add_task_for_model(
+        self,
+        payload: dict[str, Any],
+        *,
+        model_last_modified: datetime | str | None,
+        submitted_at: datetime | str | None = None,
+    ) -> dict[str, Any]:
+        reference_time = parse_datetime(submitted_at) or utc_now()
+        last_modified = parse_datetime(model_last_modified)
+        is_old_model = (
+            last_modified is None
+            or last_modified < reference_time - timedelta(days=self.recent_model_days)
+        )
+        return self._add_task(payload, old_model_only=is_old_model)
+
     def add_task(self, payload: dict[str, Any]) -> dict[str, Any]:
+        return self._add_task(payload, old_model_only=False)
+
+    def _add_task(self, payload: dict[str, Any], *, old_model_only: bool) -> dict[str, Any]:
         self._refresh_active_counts()
         attempted_accounts: set[int] = set()
         forced_refresh_done = False
         last_capacity_error: ModelHubAPIError | None = None
+        account_count_limit = self.old_model_queue_threshold if old_model_only else None
 
         while True:
-            reservation = self._reserve_account(attempted_accounts)
+            reservation = self._reserve_account(
+                attempted_accounts,
+                effective_count_below=account_count_limit,
+            )
             capacity_probe = False
             if reservation is None:
-                reservation = self._reserve_probe_account(attempted_accounts)
+                reservation = self._reserve_probe_account(
+                    attempted_accounts,
+                    effective_count_below=account_count_limit,
+                )
                 capacity_probe = reservation is not None
             if reservation is None:
                 if not forced_refresh_done:
@@ -782,6 +859,11 @@ class ModelHubClientPool:
                     continue
                 if last_capacity_error is not None:
                     raise last_capacity_error
+                if old_model_only:
+                    raise OldModelQueuePolicyError(
+                        f"所有账号活动队列均已达到 {self.old_model_queue_threshold}，"
+                        f"仅允许提交最近 {self.recent_model_days} 天内更新的模型"
+                    )
                 raise ModelHubAPIError(
                     "当前等待中或运行中的异步模型验证任务数量已达已知上限"
                 )

modelhub_submmit_api/poll_runner.py

@@ -58,6 +58,24 @@ def build_parser() -> argparse.ArgumentParser:
         default=0,
         help="Maximum tasks to submit in one cycle (0 means unlimited)",
     )
+    parser.add_argument(
+        "--old-model-queue-threshold",
+        type=int,
+        default=int(os.getenv("MODELHUB_OLD_MODEL_QUEUE_THRESHOLD", "80")),
+        help=argparse.SUPPRESS,
+    )
+    parser.add_argument(
+        "--recent-model-days",
+        type=int,
+        default=int(os.getenv("MODELHUB_RECENT_MODEL_DAYS", "7")),
+        help=argparse.SUPPRESS,
+    )
+    parser.add_argument(
+        "--dynamic-old-model-cleanup-threshold",
+        type=int,
+        default=int(os.getenv("MODELHUB_DYNAMIC_OLD_MODEL_CLEANUP_THRESHOLD", "95")),
+        help=argparse.SUPPRESS,
+    )
     parser.add_argument("--disable-candidate-preflight", action="store_true", help=argparse.SUPPRESS)
     parser.add_argument("--llm-classifier-endpoint", default=os.getenv("MODELHUB_LLM_CLASSIFIER_ENDPOINT"), help=argparse.SUPPRESS)
     parser.add_argument("--llm-classifier-model", default=os.getenv("MODELHUB_LLM_CLASSIFIER_MODEL"), help=argparse.SUPPRESS)
@@ -206,9 +224,30 @@ def _build_modelhub_client(base_args: argparse.Namespace) -> ModelHubClientPool:
         clients,
         capacity_probe_interval_cycles=max(0, int(getattr(base_args, "capacity_probe_interval_cycles", 3) or 0)),
         capacity_state_path=Path(getattr(base_args, "capacity_state_path", DEFAULT_CAPACITY_STATE_PATH)),
+        old_model_queue_threshold=max(1, int(getattr(base_args, "old_model_queue_threshold", 80) or 80)),
+        recent_model_days=max(1, int(getattr(base_args, "recent_model_days", 7) or 7)),
     )
 
 
+def resolve_age_cleanup_policy(
+    base_args: argparse.Namespace,
+    *,
+    initial_cleanup_pending: bool,
+) -> tuple[str, int]:
+    """Use the strict admission boundary once, then retain a 15-slot buffer."""
+    initial_threshold = max(
+        1,
+        int(getattr(base_args, "old_model_queue_threshold", 80) or 80),
+    )
+    dynamic_threshold = max(
+        initial_threshold,
+        int(getattr(base_args, "dynamic_old_model_cleanup_threshold", 95) or 95),
+    )
+    if initial_cleanup_pending:
+        return "initial", initial_threshold
+    return "dynamic", dynamic_threshold
+
+
 def run_poll_loop(
     *,
     base_args: argparse.Namespace,
@@ -243,6 +282,7 @@ def run_poll_loop(
     submitted_total = 0
     cycles = 0
     stopped_reason = "max_cycles_reached"
+    initial_age_cleanup_pending = True
 
     while True:
         if base_args.max_cycles and cycles >= base_args.max_cycles:
@@ -263,12 +303,22 @@ def run_poll_loop(
             try:
                 cleanup_feedback = outcome_tracker.get_stats_report()
                 cleanup_gpu_memory = cleanup_feedback.get("observedGpuMemoryGiB") or {}
+                age_cleanup_mode, age_cleanup_threshold = resolve_age_cleanup_policy(
+                    base_args,
+                    initial_cleanup_pending=initial_age_cleanup_pending,
+                )
+                log(
+                    f"[queue-cleanup] mode={age_cleanup_mode} "
+                    f"old_model_threshold={age_cleanup_threshold} "
+                    f"recent_days={max(1, int(getattr(base_args, 'recent_model_days', 7) or 7))}"
+                )
                 cleanup_summary = cleanup_certain_oom_tasks(
                     modelhub_client,
                     hf_discovery,
                     dry_run=bool(base_args.dry_run),
                     read_concurrency=max(1, int(getattr(base_args, "queue_cleanup_read_concurrency", 6) or 6)),
                     gpu_memory_gib=cleanup_gpu_memory if isinstance(cleanup_gpu_memory, dict) else None,
+                    age_queue_threshold=age_cleanup_threshold,
                     log=log,
                 )
                 write_json(
@@ -284,12 +334,16 @@ def run_poll_loop(
                 queue_cleanup_runs.append(
                     {
                         "cycle": cycles,
+                        "mode": age_cleanup_mode,
+                        "ageQueueThreshold": age_cleanup_threshold,
                         "activeScanned": cleanup_summary["activeScanned"],
                         "certainOomCount": cleanup_summary["certainOomCount"],
+                        "oldOverflowCount": cleanup_summary["oldOverflowCount"],
                         "cancelledCount": cleanup_summary["cancelledCount"],
                         "stopErrorCount": len(cleanup_summary["stopErrors"]),
                     }
                 )
+                initial_age_cleanup_pending = False
             except Exception as exc:
                 log(f"[queue-cleanup] error={type(exc).__name__}: {exc} continue_polling=true")
 

modelhub_submmit_api/queue_cleanup.py

@@ -3,11 +3,12 @@ from __future__ import annotations
 import argparse
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from dataclasses import dataclass
+from datetime import datetime, timedelta
 from pathlib import Path
 from typing import Any, Callable, Iterable
 
 from candidate_preflight import CandidatePreflightAdvisor, MODEL_LOAD_OVERHEAD
-from common import write_json
+from common import utc_now, write_json
 from hf_discovery import HuggingFaceDiscovery, inspect_repo_tree
 from modelhub_client import ModelHubClient, ModelHubClientPool
 from runner_common import DEFAULT_KEY_PATH, ensure_tokens
@@ -159,6 +160,41 @@ def _load_repository_sizes(
     return sizes, errors
 
 
+def _load_model_last_modified(
+    model_ids: set[str],
+    *,
+    discovery: HuggingFaceDiscovery,
+    read_concurrency: int,
+    log: Callable[[str], None],
+) -> tuple[dict[str, datetime], dict[str, str]]:
+    values: dict[str, datetime] = {}
+    errors: dict[str, str] = {}
+    completed = 0
+    workers = min(max(1, int(read_concurrency)), max(1, len(model_ids)))
+    with ThreadPoolExecutor(max_workers=workers) as executor:
+        futures = {
+            executor.submit(discovery.get_model_last_modified, model_id): model_id
+            for model_id in sorted(model_ids)
+        }
+        for future in as_completed(futures):
+            model_id = futures[future]
+            try:
+                value = future.result()
+                if value is None:
+                    errors[model_id] = "model_last_modified_unknown"
+                else:
+                    values[model_id] = value
+            except Exception as exc:
+                errors[model_id] = f"{type(exc).__name__}: {exc}"
+            completed += 1
+            if completed == len(model_ids) or completed % 50 == 0:
+                log(
+                    f"[queue-cleanup] age_scan={completed}/{len(model_ids)} "
+                    f"complete={len(values)} unknown={len(errors)}"
+                )
+    return values, errors
+
+
 def find_certain_oom_tasks(
     tasks: list[OwnedTask],
     *,
@@ -202,6 +238,62 @@ def find_certain_oom_tasks(
     return decisions, skipped
 
 
+def find_old_overflow_tasks(
+    tasks: list[OwnedTask],
+    *,
+    model_last_modified: dict[str, datetime],
+    queue_threshold: int = 80,
+    recent_model_days: int = 7,
+    reference_time: datetime | None = None,
+    incomplete_accounts: set[int] | None = None,
+) -> tuple[list[dict[str, Any]], dict[str, int]]:
+    threshold = max(1, int(queue_threshold))
+    recent_days = max(1, int(recent_model_days))
+    cutoff = (reference_time or utc_now()) - timedelta(days=recent_days)
+    incomplete_accounts = incomplete_accounts or set()
+    grouped: dict[int, list[OwnedTask]] = {}
+    for task in tasks:
+        grouped.setdefault(task.account_index, []).append(task)
+
+    decisions: list[dict[str, Any]] = []
+    skipped = {
+        "accountsWithIncompleteListing": len(incomplete_accounts),
+        "withinFirstQueuePositions": 0,
+        "recentOverflowTasks": 0,
+        "modelAgeUnknown": 0,
+    }
+    for account_index, account_tasks in sorted(grouped.items()):
+        if account_index in incomplete_accounts:
+            continue
+        ordered = sorted(account_tasks, key=lambda item: item.task_id)
+        skipped["withinFirstQueuePositions"] += min(threshold, len(ordered))
+        for position, task in enumerate(ordered, start=1):
+            if position <= threshold:
+                continue
+            last_modified = model_last_modified.get(task.model_id)
+            if last_modified is None:
+                skipped["modelAgeUnknown"] += 1
+                continue
+            if last_modified >= cutoff:
+                skipped["recentOverflowTasks"] += 1
+                continue
+            decisions.append(
+                {
+                    "accountIndex": account_index + 1,
+                    "taskId": task.task_id,
+                    "modelId": task.model_id,
+                    "gpuType": task.gpu_type,
+                    "status": task.status,
+                    "queuePosition": position,
+                    "queueThreshold": threshold,
+                    "modelLastModified": last_modified.isoformat(),
+                    "recentCutoff": cutoff.isoformat(),
+                    "reason": "old_model_beyond_account_queue_threshold",
+                }
+            )
+    return decisions, skipped
+
+
 def _chunks(values: list[int], size: int) -> Iterable[list[int]]:
     for offset in range(0, len(values), size):
         yield values[offset : offset + size]
@@ -215,10 +307,20 @@ def cleanup_certain_oom_tasks(
     read_concurrency: int = 6,
     stop_batch_size: int = DEFAULT_STOP_BATCH_SIZE,
     gpu_memory_gib: dict[str, float] | None = None,
+    age_queue_threshold: int | None = None,
+    reference_time: datetime | None = None,
     log: Callable[[str], None] = print,
 ) -> dict[str, Any]:
-    """Stop only active tasks that cannot fit the selected GPU by known capacity."""
+    """Stop deterministic OOM tasks and old tasks beyond each account's protected prefix."""
     clients = list(modelhub.clients)
+    reference_time = reference_time or utc_now()
+    configured_queue_threshold = (
+        age_queue_threshold
+        if age_queue_threshold is not None
+        else getattr(modelhub, "old_model_queue_threshold", 80)
+    )
+    queue_threshold = max(1, int(configured_queue_threshold or 80))
+    recent_model_days = max(1, int(getattr(modelhub, "recent_model_days", 7) or 7))
     tasks, listing_errors = collect_active_tasks(clients, read_concurrency=read_concurrency)
     log(
         f"[queue-cleanup] active_scanned={len(tasks)} accounts={len(clients)} "
@@ -232,27 +334,104 @@ def cleanup_certain_oom_tasks(
         read_concurrency=read_concurrency,
         log=log,
     )
-    decisions, skipped = find_certain_oom_tasks(
+    oom_decisions, skipped = find_certain_oom_tasks(
         tasks,
         repository_sizes=repository_sizes,
         gpu_memory_gib=gpu_memory_gib,
     )
     log(
-        f"[queue-cleanup] certain_oom={len(decisions)} "
+        f"[queue-cleanup] certain_oom={len(oom_decisions)} "
         f"fits={skipped['fitsKnownCapacity']} size_unknown={skipped['repositorySizeUnknown']} "
         f"gpu_unknown={skipped['gpuCapacityUnknown']} dry_run={str(bool(dry_run)).lower()}"
     )
 
+    oom_task_keys = {
+        (int(decision["accountIndex"]) - 1, int(decision["taskId"]))
+        for decision in oom_decisions
+    }
+    # OOM tasks are stopped first. Rank the age-policy queue as it will look
+    # after those certain failures are gone, so an old task moving into the
+    # protected first N positions is not over-cancelled.
+    age_rank_tasks = [
+        task for task in tasks if (task.account_index, task.task_id) not in oom_task_keys
+    ]
+    overflow_model_ids = {
+        task.model_id
+        for account_index in range(len(clients))
+        if account_index not in listing_errors
+        for task in sorted(
+            (item for item in age_rank_tasks if item.account_index == account_index),
+            key=lambda item: item.task_id,
+        )[queue_threshold:]
+    }
+    model_last_modified, age_errors = _load_model_last_modified(
+        overflow_model_ids,
+        discovery=discovery,
+        read_concurrency=read_concurrency,
+        log=log,
+    )
+    old_overflow_decisions, age_skipped = find_old_overflow_tasks(
+        age_rank_tasks,
+        model_last_modified=model_last_modified,
+        queue_threshold=queue_threshold,
+        recent_model_days=recent_model_days,
+        reference_time=reference_time,
+        incomplete_accounts=set(listing_errors),
+    )
+    log(
+        f"[queue-cleanup] old_overflow={len(old_overflow_decisions)} "
+        f"threshold={queue_threshold} recent_days={recent_model_days} "
+        f"recent_overflow={age_skipped['recentOverflowTasks']} "
+        f"age_unknown={age_skipped['modelAgeUnknown']}"
+    )
+
+    decisions_by_key: dict[tuple[int, int], dict[str, Any]] = {}
+    for decision in oom_decisions:
+        enriched = dict(decision)
+        enriched["cleanupReasons"] = [decision["reason"]]
+        decisions_by_key[(int(decision["accountIndex"]), int(decision["taskId"]))] = enriched
+    for decision in old_overflow_decisions:
+        key = (int(decision["accountIndex"]), int(decision["taskId"]))
+        existing = decisions_by_key.get(key)
+        if existing is None:
+            enriched = dict(decision)
+            enriched["cleanupReasons"] = [decision["reason"]]
+            decisions_by_key[key] = enriched
+            continue
+        existing["cleanupReasons"].append(decision["reason"])
+        for field in ("queuePosition", "queueThreshold", "modelLastModified", "recentCutoff"):
+            existing[field] = decision[field]
+    decisions = sorted(
+        decisions_by_key.values(),
+        key=lambda item: (int(item["accountIndex"]), int(item["taskId"])),
+    )
+
     cancelled: list[dict[str, Any]] = []
     disappeared: list[dict[str, Any]] = []
+    policy_no_longer_applies: list[dict[str, Any]] = []
     stop_errors: list[dict[str, Any]] = []
     if decisions and not dry_run:
         # Re-read each account immediately before mutation. If any active-state
         # query fails for that account, fail closed and do not terminate its tasks.
         refreshed_tasks, refresh_errors = collect_active_tasks(clients, read_concurrency=read_concurrency)
         active_ids_by_account: dict[int, set[int]] = {}
+        active_positions_by_account: dict[int, dict[int, int]] = {}
         for task in refreshed_tasks:
             active_ids_by_account.setdefault(task.account_index, set()).add(task.task_id)
+        for account_index in range(len(clients)):
+            planned_oom_ids = {
+                int(decision["taskId"])
+                for decision in oom_decisions
+                if int(decision["accountIndex"]) - 1 == account_index
+            }
+            ordered_ids = sorted(
+                task.task_id
+                for task in refreshed_tasks
+                if task.account_index == account_index and task.task_id not in planned_oom_ids
+            )
+            active_positions_by_account[account_index] = {
+                task_id: position for position, task_id in enumerate(ordered_ids, start=1)
+            }
 
         for account_index, details in sorted(refresh_errors.items()):
             stop_errors.append(
@@ -271,6 +450,16 @@ def cleanup_certain_oom_tasks(
             if int(decision["taskId"]) not in active_ids_by_account.get(account_index, set()):
                 disappeared.append(decision)
                 continue
+            cleanup_reasons = set(decision.get("cleanupReasons") or [decision.get("reason")])
+            age_only = cleanup_reasons == {"old_model_beyond_account_queue_threshold"}
+            current_position = active_positions_by_account.get(account_index, {}).get(int(decision["taskId"]))
+            if age_only and (current_position is None or current_position <= queue_threshold):
+                policy_no_longer_applies.append(
+                    {**decision, "recheckedQueuePosition": current_position}
+                )
+                continue
+            if current_position is not None:
+                decision["recheckedQueuePosition"] = current_position
             by_account.setdefault(account_index, []).append(decision)
 
         batch_size = max(1, min(100, int(stop_batch_size)))
@@ -279,25 +468,36 @@ def cleanup_certain_oom_tasks(
             if stop_failed:
                 break
             decisions_by_id = {int(item["taskId"]): item for item in by_account[account_index]}
-            task_ids = sorted(decisions_by_id)
+            phase_ids = [
-            for batch in _chunks(task_ids, batch_size):
+                sorted(
-                try:
+                    task_id
-                    clients[account_index].stop_tasks(batch)
+                    for task_id, decision in decisions_by_id.items()
-                except Exception as exc:
+                    if ("certain_oom_repository_size_exceeds_gpu_capacity" in decision["cleanupReasons"])
-                    stop_errors.append(
+                    == is_oom_phase
-                        {
-                            "accountIndex": account_index + 1,
-                            "taskIds": batch,
-                            "error": f"{type(exc).__name__}: {exc}",
-                        }
-                    )
-                    stop_failed = True
-                    break
-                cancelled.extend(decisions_by_id[task_id] for task_id in batch)
-                log(
-                    f"[queue-cleanup] account={account_index + 1:02d} "
-                    f"cancelled_batch={len(batch)} cancelled_total={len(cancelled)}"
                 )
+                for is_oom_phase in (True, False)
+            ]
+            for task_ids in phase_ids:
+                for batch in _chunks(task_ids, batch_size):
+                    try:
+                        clients[account_index].stop_tasks(batch)
+                    except Exception as exc:
+                        stop_errors.append(
+                            {
+                                "accountIndex": account_index + 1,
+                                "taskIds": batch,
+                                "error": f"{type(exc).__name__}: {exc}",
+                            }
+                        )
+                        stop_failed = True
+                        break
+                    cancelled.extend(decisions_by_id[task_id] for task_id in batch)
+                    log(
+                        f"[queue-cleanup] account={account_index + 1:02d} "
+                        f"cancelled_batch={len(batch)} cancelled_total={len(cancelled)}"
+                    )
+                if stop_failed:
+                    break
 
         if cancelled and hasattr(modelhub, "refresh_active_counts"):
             modelhub.refresh_active_counts()
@@ -309,19 +509,29 @@ def cleanup_certain_oom_tasks(
         "uniqueModels": len(model_ids),
         "repositorySizesComplete": len(repository_sizes),
         "repositorySizeErrors": size_errors,
+        "modelAgeMetadataComplete": len(model_last_modified),
+        "modelAgeErrors": age_errors,
         "listingErrors": {str(index + 1): values for index, values in listing_errors.items()},
-        "certainOomCount": len(decisions),
+        "certainOomCount": len(oom_decisions),
-        "certainOomTasks": decisions,
+        "certainOomTasks": oom_decisions,
+        "oldOverflowCount": len(old_overflow_decisions),
+        "oldOverflowTasks": old_overflow_decisions,
+        "oldModelQueueThreshold": queue_threshold,
+        "recentModelDays": recent_model_days,
+        "agePolicySkipped": age_skipped,
+        "cleanupCandidateCount": len(decisions),
         "skipped": skipped,
         "cancelledCount": len(cancelled),
         "cancelledTasks": cancelled,
         "noLongerActiveCount": len(disappeared),
+        "policyNoLongerAppliesCount": len(policy_no_longer_applies),
+        "policyNoLongerAppliesTasks": policy_no_longer_applies,
         "stopErrors": stop_errors,
     }
 
 
 def build_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser(description="Safely stop active ModelHub tasks that are certain to run out of GPU memory.")
+    parser = argparse.ArgumentParser(description="Safely stop deterministic OOM and over-threshold old ModelHub tasks.")
     parser.add_argument("--execute", action="store_true", help="Actually terminate selected tasks; otherwise only print a preview")
     parser.add_argument("--read-concurrency", type=int, default=6)
     parser.add_argument("--stop-batch-size", type=int, default=DEFAULT_STOP_BATCH_SIZE)
@@ -353,6 +563,7 @@ def main(argv: list[str] | None = None) -> int:
     write_json(report_path, summary)
     print(
         f"[queue-cleanup] finished certain_oom={summary['certainOomCount']} "
+        f"old_overflow={summary['oldOverflowCount']} "
         f"cancelled={summary['cancelledCount']} stop_errors={len(summary['stopErrors'])} "
         f"report={report_path}",
         flush=True,

modelhub_submmit_api/version.py

@@ -1 +1 @@
-AGENT_VERSION = "2026.08.11.2"
+AGENT_VERSION = "2026.08.11.3"

tests/test_concurrency.py

@@ -17,8 +17,13 @@ if str(PACKAGE_DIR) in sys.path:
 sys.path.insert(0, str(PACKAGE_DIR))
 
 from common import read_json, read_jsonl  # noqa: E402
-from main import build_parser, make_run_dir, run_submission, submit_candidate  # noqa: E402
+from main import build_adaptive_scan_stages, build_parser, make_run_dir, run_submission, submit_candidate  # noqa: E402
-from modelhub_client import ModelHubAPIError, ModelHubClientPool, is_duplicate_submission_error  # noqa: E402
+from modelhub_client import (  # noqa: E402
+    ModelHubAPIError,
+    ModelHubClientPool,
+    OldModelQueuePolicyError,
+    is_duplicate_submission_error,
+)
 from models import HFModelSummary, ModelInspection  # noqa: E402
 from outcome_tracker import OutcomeTracker  # noqa: E402
 from submission_claims import SubmissionClaimStore, candidate_key  # noqa: E402
@@ -210,6 +215,109 @@ def make_candidate(index: int) -> dict:
 
 
 class ClientPoolConcurrencyTests(unittest.TestCase):
+    def test_old_models_use_only_each_accounts_first_eighty_queue_positions(self) -> None:
+        below_threshold = FakeClient(active_count=79)
+        at_threshold = FakeClient(active_count=80)
+        pool = ModelHubClientPool(
+            [below_threshold, at_threshold],  # type: ignore[list-item]
+            active_task_cap=100,
+            active_counts_ttl=60,
+            old_model_queue_threshold=80,
+            recent_model_days=7,
+            instance_id="old-model-threshold-test",
+        )
+        submitted_at = datetime(2026, 8, 11, tzinfo=timezone.utc)
+
+        pool.add_task_for_model(
+            {"model": "old-allowed-as-position-80"},
+            model_last_modified=datetime(2026, 7, 1, tzinfo=timezone.utc),
+            submitted_at=submitted_at,
+        )
+
+        self.assertEqual(1, len(below_threshold.submitted))
+        self.assertEqual(0, len(at_threshold.submitted))
+        self.assertEqual(0, pool.old_model_submit_slots())
+        with self.assertRaises(OldModelQueuePolicyError):
+            pool.add_task_for_model(
+                {"model": "old-rejected"},
+                model_last_modified=datetime(2026, 7, 1, tzinfo=timezone.utc),
+                submitted_at=submitted_at,
+            )
+
+        pool.add_task_for_model(
+            {"model": "recent-allowed"},
+            model_last_modified=datetime(2026, 8, 10, tzinfo=timezone.utc),
+            submitted_at=submitted_at,
+        )
+        self.assertEqual(2, len(below_threshold.submitted) + len(at_threshold.submitted))
+
+    def test_concurrent_old_model_submissions_cannot_cross_eighty(self) -> None:
+        client = FakeClient(active_count=78)
+        pool = ModelHubClientPool(
+            [client],  # type: ignore[list-item]
+            active_task_cap=100,
+            active_counts_ttl=60,
+            old_model_queue_threshold=80,
+            recent_model_days=7,
+        )
+        submitted_at = datetime(2026, 8, 11, tzinfo=timezone.utc)
+
+        def submit(index: int) -> str:
+            try:
+                pool.add_task_for_model(
+                    {"model": f"old-{index}"},
+                    model_last_modified=datetime(2026, 7, 1, tzinfo=timezone.utc),
+                    submitted_at=submitted_at,
+                )
+                return "submitted"
+            except OldModelQueuePolicyError:
+                return "policy_skipped"
+
+        with ThreadPoolExecutor(max_workers=8) as executor:
+            outcomes = list(executor.map(submit, range(8)))
+
+        self.assertEqual(2, outcomes.count("submitted"))
+        self.assertEqual(6, outcomes.count("policy_skipped"))
+        self.assertEqual(2, len(client.submitted))
+        self.assertEqual(0, pool.old_model_submit_slots())
+
+    def test_scan_does_not_expand_beyond_recent_window_when_old_slots_are_full(self) -> None:
+        now = datetime(2026, 8, 11, tzinfo=timezone.utc)
+        stages = build_adaptive_scan_stages(
+            now=now,
+            initial_updated_after=now - timedelta(hours=48),
+            initial_limit=100,
+            allow_older_than_recent_window=False,
+            recent_model_days=7,
+        )
+
+        self.assertEqual(["configured_window", "last_7_days"], [stage["name"] for stage in stages])
+        self.assertTrue(all(stage["updatedAfter"] >= now - timedelta(days=7) for stage in stages))
+
+    def test_submit_candidate_reports_old_model_policy_skip_at_eighty(self) -> None:
+        client = FakeClient(active_count=80)
+        pool = ModelHubClientPool(
+            [client],  # type: ignore[list-item]
+            active_task_cap=100,
+            old_model_queue_threshold=80,
+            recent_model_days=7,
+        )
+        result = submit_candidate(
+            {
+                "repoId": "owner/old",
+                "modelAddress": "https://modelscope.cn/models/owner/old",
+                "targetGpu": "Iluvatar_bi-150",
+                "taskType": "text-generation",
+                "framework": "vllm",
+                "configParams": "framework: vllm",
+                "lastModified": "2025-01-01T00:00:00+00:00",
+            },
+            pool,  # type: ignore[arg-type]
+        )
+
+        self.assertEqual("old_model_policy_skipped", result["outcome"])
+        self.assertEqual([], client.submitted)
+
     def test_online_submission_does_not_construct_llm_even_when_key_is_present(self) -> None:
         with tempfile.TemporaryDirectory() as temporary_dir:
             root = Path(temporary_dir)

tests/test_modelscope_discovery.py

@@ -47,7 +47,27 @@ class RateLimitedPageClient:
         return {"success": True, "data": {"models": items}}
 
 
+class ModelMetadataClient:
+    def __init__(self) -> None:
+        self.calls = 0
+
+    def request_json(self, _method: str, _path: str, **_kwargs) -> dict:  # noqa: ANN003
+        self.calls += 1
+        return {"Code": 200, "Data": {"LastUpdatedTime": 1786294389}}
+
+
 class ModelScopeDiscoveryTests(unittest.TestCase):
+    def test_model_last_modified_uses_detail_api_and_is_cached(self) -> None:
+        metadata_client = ModelMetadataClient()
+        discovery = HuggingFaceDiscovery(legacy_http_client=metadata_client)  # type: ignore[arg-type]
+
+        first = discovery.get_model_last_modified("owner/model")
+        second = discovery.get_model_last_modified("owner/model")
+
+        self.assertEqual(first, second)
+        self.assertEqual(1786294389, int(first.timestamp()))  # type: ignore[union-attr]
+        self.assertEqual(1, metadata_client.calls)
+
     def test_openapi_page_size_never_exceeds_platform_limit(self) -> None:
         http_client = RecordingHttpClient()
         discovery = HuggingFaceDiscovery(http_client=http_client)  # type: ignore[arg-type]

tests/test_poll_policy.py

@@ -0,0 +1,46 @@
+from __future__ import annotations
+
+import argparse
+import sys
+import unittest
+from pathlib import Path
+
+
+PACKAGE_DIR = Path(__file__).resolve().parents[1] / "modelhub_submmit_api"
+if str(PACKAGE_DIR) in sys.path:
+    sys.path.remove(str(PACKAGE_DIR))
+sys.path.insert(0, str(PACKAGE_DIR))
+
+from poll_runner import resolve_age_cleanup_policy  # noqa: E402
+
+
+class PollPolicyTests(unittest.TestCase):
+    def test_age_cleanup_uses_eighty_once_then_ninety_five(self) -> None:
+        args = argparse.Namespace(
+            old_model_queue_threshold=80,
+            dynamic_old_model_cleanup_threshold=95,
+        )
+
+        self.assertEqual(
+            ("initial", 80),
+            resolve_age_cleanup_policy(args, initial_cleanup_pending=True),
+        )
+        self.assertEqual(
+            ("dynamic", 95),
+            resolve_age_cleanup_policy(args, initial_cleanup_pending=False),
+        )
+
+    def test_dynamic_cleanup_cannot_be_stricter_than_admission(self) -> None:
+        args = argparse.Namespace(
+            old_model_queue_threshold=80,
+            dynamic_old_model_cleanup_threshold=70,
+        )
+
+        self.assertEqual(
+            ("dynamic", 80),
+            resolve_age_cleanup_policy(args, initial_cleanup_pending=False),
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_queue_cleanup.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import unittest
 import sys
+from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any
 
@@ -12,16 +13,28 @@ if str(PACKAGE_DIR) in sys.path:
 sys.path.insert(0, str(PACKAGE_DIR))
 
 from modelhub_client import ModelHubClient, ModelHubClientPool  # noqa: E402
-from queue_cleanup import OwnedTask, cleanup_certain_oom_tasks, find_certain_oom_tasks  # noqa: E402
+from queue_cleanup import (  # noqa: E402
+    OwnedTask,
+    cleanup_certain_oom_tasks,
+    find_certain_oom_tasks,
+    find_old_overflow_tasks,
+)
 
 
 GIB = 1024**3
 
 
 class FakeQueueClient:
-    def __init__(self, records: list[dict[str, Any]], *, disappear_on_recheck: bool = False) -> None:
+    def __init__(
+        self,
+        records: list[dict[str, Any]],
+        *,
+        disappear_on_recheck: bool = False,
+        drop_first_on_recheck: bool = False,
+    ) -> None:
         self.records = list(records)
         self.disappear_on_recheck = disappear_on_recheck
+        self.drop_first_on_recheck = drop_first_on_recheck
         self.waiting_reads = 0
         self.stopped: list[list[int]] = []
 
@@ -32,6 +45,8 @@ class FakeQueueClient:
                 records: list[dict[str, Any]] = []
             else:
                 records = [record for record in self.records if record["status"] == "waiting"]
+                if self.drop_first_on_recheck and self.waiting_reads >= 2:
+                    records = sorted(records, key=lambda item: int(item["taskId"]))[1:]
         else:
             records = [record for record in self.records if record["status"] == status]
         return {"code": 0, "data": {"records": records, "pages": 1}}
@@ -47,8 +62,13 @@ class FakeQueueClient:
 
 
 class FakeDiscovery:
-    def __init__(self, sizes: dict[str, int | None]) -> None:
+    def __init__(
+        self,
+        sizes: dict[str, int | None],
+        last_modified: dict[str, datetime | None] | None = None,
+    ) -> None:
         self.sizes = sizes
+        self.last_modified = last_modified or {}
 
     def list_repo_tree(self, repo_id: str) -> list[dict[str, Any]]:
         size = self.sizes[repo_id]
@@ -56,6 +76,9 @@ class FakeDiscovery:
             return [{"Path": "model.safetensors"}]
         return [{"Path": "model.safetensors", "Size": size}]
 
+    def get_model_last_modified(self, repo_id: str) -> datetime | None:
+        return self.last_modified.get(repo_id)
+
 
 class RecordingHttpClient:
     def __init__(self) -> None:
@@ -74,6 +97,151 @@ class RecordingHttpClient:
 
 
 class QueueCleanupTests(unittest.TestCase):
+    def test_old_models_are_selected_only_after_each_accounts_first_eighty_tasks(self) -> None:
+        now = datetime(2026, 8, 11, tzinfo=timezone.utc)
+        tasks = [
+            OwnedTask(0, index, "owner/old", "Iluvatar_bi-100", "waiting")
+            for index in range(1, 82)
+        ]
+        tasks.append(OwnedTask(0, 82, "owner/recent", "Iluvatar_bi-100", "waiting"))
+
+        selected, skipped = find_old_overflow_tasks(
+            tasks,
+            model_last_modified={
+                "owner/old": datetime(2026, 7, 1, tzinfo=timezone.utc),
+                "owner/recent": datetime(2026, 8, 10, tzinfo=timezone.utc),
+            },
+            queue_threshold=80,
+            recent_model_days=7,
+            reference_time=now,
+        )
+
+        self.assertEqual([81], [item["taskId"] for item in selected])
+        self.assertEqual(81, selected[0]["queuePosition"])
+        self.assertEqual(1, skipped["recentOverflowTasks"])
+
+    def test_old_overflow_task_is_not_stopped_if_it_moves_into_first_eighty(self) -> None:
+        records = [
+            {
+                "taskId": index,
+                "modelId": "owner/old",
+                "gpuType": "Iluvatar_bi-100",
+                "status": "waiting",
+            }
+            for index in range(1, 82)
+        ]
+        client = FakeQueueClient(records, drop_first_on_recheck=True)
+        pool = ModelHubClientPool(
+            [client],  # type: ignore[list-item]
+            active_task_cap=100,
+            old_model_queue_threshold=80,
+        )
+        summary = cleanup_certain_oom_tasks(
+            pool,
+            FakeDiscovery(
+                {"owner/old": 1 * GIB},
+                {"owner/old": datetime(2026, 7, 1, tzinfo=timezone.utc)},
+            ),  # type: ignore[arg-type]
+            reference_time=datetime(2026, 8, 11, tzinfo=timezone.utc),
+            log=lambda _message: None,
+        )
+
+        self.assertEqual(1, summary["oldOverflowCount"])
+        self.assertEqual(0, summary["cancelledCount"])
+        self.assertEqual(1, summary["policyNoLongerAppliesCount"])
+        self.assertEqual([], client.stopped)
+
+    def test_cleanup_stops_old_task_beyond_eightieth_position(self) -> None:
+        records = [
+            {
+                "taskId": index,
+                "modelId": "owner/old",
+                "gpuType": "Iluvatar_bi-100",
+                "status": "waiting",
+            }
+            for index in range(1, 82)
+        ]
+        client = FakeQueueClient(records)
+        pool = ModelHubClientPool(
+            [client],  # type: ignore[list-item]
+            active_task_cap=100,
+            old_model_queue_threshold=80,
+        )
+        summary = cleanup_certain_oom_tasks(
+            pool,
+            FakeDiscovery(
+                {"owner/old": 1 * GIB},
+                {"owner/old": datetime(2026, 7, 1, tzinfo=timezone.utc)},
+            ),  # type: ignore[arg-type]
+            reference_time=datetime(2026, 8, 11, tzinfo=timezone.utc),
+            log=lambda _message: None,
+        )
+
+        self.assertEqual(1, summary["oldOverflowCount"])
+        self.assertEqual(1, summary["cancelledCount"])
+        self.assertEqual([[81]], client.stopped)
+
+    def test_dynamic_cleanup_keeps_positions_through_ninety_five(self) -> None:
+        records = [
+            {
+                "taskId": index,
+                "modelId": "owner/old",
+                "gpuType": "Iluvatar_bi-100",
+                "status": "waiting",
+            }
+            for index in range(1, 97)
+        ]
+        client = FakeQueueClient(records)
+        pool = ModelHubClientPool(
+            [client],  # type: ignore[list-item]
+            active_task_cap=100,
+            old_model_queue_threshold=80,
+        )
+        summary = cleanup_certain_oom_tasks(
+            pool,
+            FakeDiscovery(
+                {"owner/old": 1 * GIB},
+                {"owner/old": datetime(2026, 7, 1, tzinfo=timezone.utc)},
+            ),  # type: ignore[arg-type]
+            age_queue_threshold=95,
+            reference_time=datetime(2026, 8, 11, tzinfo=timezone.utc),
+            log=lambda _message: None,
+        )
+
+        self.assertEqual(95, summary["oldModelQueueThreshold"])
+        self.assertEqual([96], [item["taskId"] for item in summary["oldOverflowTasks"]])
+        self.assertEqual([[96]], client.stopped)
+
+    def test_oom_is_removed_before_recalculating_old_overflow_positions(self) -> None:
+        records = [
+            {
+                "taskId": index,
+                "modelId": "owner/large" if index == 1 else "owner/old",
+                "gpuType": "Iluvatar_bi-100",
+                "status": "waiting",
+            }
+            for index in range(1, 83)
+        ]
+        client = FakeQueueClient(records)
+        pool = ModelHubClientPool(
+            [client],  # type: ignore[list-item]
+            active_task_cap=100,
+            old_model_queue_threshold=80,
+        )
+        summary = cleanup_certain_oom_tasks(
+            pool,
+            FakeDiscovery(
+                {"owner/large": 40 * GIB, "owner/old": 1 * GIB},
+                {"owner/old": datetime(2026, 7, 1, tzinfo=timezone.utc)},
+            ),  # type: ignore[arg-type]
+            reference_time=datetime(2026, 8, 11, tzinfo=timezone.utc),
+            log=lambda _message: None,
+        )
+
+        self.assertEqual(1, summary["certainOomCount"])
+        self.assertEqual([82], [item["taskId"] for item in summary["oldOverflowTasks"]])
+        self.assertEqual([[1], [82]], client.stopped)
+
     def test_stop_tasks_uses_documented_put_endpoint_and_integer_ids(self) -> None:
         http = RecordingHttpClient()
         client = ModelHubClient(http_client=http)  # type: ignore[arg-type]