fix(context): thread custom_providers to all context-length resolution call sites

agent/agent_init.py

@@ -2756,6 +2756,7 @@ def _parse_prune_int(raw, default):
             proactive_prune_min_reclaim_tokens=compression_proactive_prune_min_reclaim,
             min_tail_user_messages=compression_min_tail_users,
             tail_mode=compression_tail_mode,
+            custom_providers=_custom_providers,
         )
     _bind_session_state = getattr(agent.context_compressor, "bind_session_state", None)
     if callable(_bind_session_state):

agent/auxiliary_client.py

@@ -5592,12 +5592,22 @@ def _candidate_context_window(
     """
     if not model:
         return None
+    # Load custom_providers from config so per-model context_length
+    # overrides (custom_providers[].models.<id>.context_length) are
+    # honored for fallback candidates too — not just the main model.
+    _custom_providers: list | None = None
+    try:
+        from hermes_cli.config import get_compatible_custom_providers, load_config_readonly
+        _custom_providers = get_compatible_custom_providers(load_config_readonly())
+    except Exception:
+        pass
     try:
         ctx = get_model_context_length(
             model,
             base_url=base_url,
             api_key=api_key,
             provider=provider,
+            custom_providers=_custom_providers,
         )
     except Exception as exc:
         logger.debug(

agent/context_compressor.py

@@ -2171,6 +2171,7 @@ def _resolve_context_length(self) -> int:
                 api_key=self.api_key,
                 config_context_length=self._config_context_length,
                 provider=self.provider,
+                custom_providers=self._custom_providers,
             )
             # Small-context threshold floor: models under 512K trigger at
             # >=75% so compaction doesn't fire with half the window still
@@ -2996,6 +2997,7 @@ def __init__(
         proactive_prune_min_reclaim_tokens: int = 4096,
         min_tail_user_messages: int = 1,
         tail_mode: str = "legacy",
+        custom_providers: list | None = None,
     ):
         self.model = model
         self.base_url = base_url
@@ -3006,6 +3008,11 @@ def __init__(
         # tail + verbatim-user-message summary section + recovery pointers;
         # "legacy" = 0.20*window tail (shipping behavior).
         self.tail_mode = tail_mode if tail_mode in ("legacy", "lean") else "legacy"
+        # Per-model context_length overrides from custom_providers config.
+        # Threaded to get_model_context_length() in _resolve_context_length()
+        # so deferred resolution (first context_length property access) honors
+        # the same per-model overrides that startup resolution does (#15779).
+        self._custom_providers = custom_providers
         # Per-model threshold overrides (longest substring match wins).
         # Stored as a plain dict; resolved in _resolve_threshold(), then the
         # small-context floor is applied on top.

agent/moa_loop.py

@@ -658,6 +658,21 @@ def _run_reference(
 _REFERENCE_TRIM_SAFETY_FRACTION = 0.10
 
 
+def _load_custom_providers() -> list | None:
+    """Best-effort load of custom_providers from config.
+
+    Used by _trim_messages_for_reference to honor per-model context_length
+    overrides (custom_providers[].models.<id>.context_length) when resolving
+    reference model context windows. Returns None on any failure so the
+    resolver falls through to probing — never breaks a MoA turn.
+    """
+    try:
+        from hermes_cli.config import get_compatible_custom_providers, load_config_readonly
+        return get_compatible_custom_providers(load_config_readonly())
+    except Exception:
+        return None
+
+
 def _trim_messages_for_reference(
     messages: list[dict[str, Any]],
     slot: dict[str, str],
@@ -723,6 +738,7 @@ def _trim_messages_for_reference(
                 base_url=str(runtime.get("base_url") or ""),
                 api_key=str(runtime.get("api_key") or ""),
                 provider=provider,
+                custom_providers=_load_custom_providers(),
             )
         except Exception:
             logger.debug(

gateway/slash_commands.py

@@ -864,8 +864,35 @@ def _resolve_nonresident_context():
             try:
                 from agent.model_metadata import get_model_context_length
 
+                # Inactive-agent fallback: resolve the configured route
+                # identity + compatible custom_providers so per-model
+                # context_length overrides (custom_providers[].models.<id>)
+                # are honored — passing only model_name falls through to
+                # generic metadata for custom endpoints (#15779).
+                _provider = ""
+                _custom_providers = None
+                try:
+                    from hermes_cli.config import (
+                        get_compatible_custom_providers,
+                        load_config_readonly,
+                    )
+
+                    _cfg = load_config_readonly() or {}
+                    _model_cfg = _cfg.get("model")
+                    if isinstance(_model_cfg, dict):
+                        _provider = str(_model_cfg.get("provider") or "").strip()
+                    _custom_providers = get_compatible_custom_providers(_cfg)
+                except Exception:
+                    _provider = ""
+                    _custom_providers = None
+
                 context_length = _int_value(
-                    await asyncio.to_thread(get_model_context_length, model_name)
+                    await asyncio.to_thread(
+                        get_model_context_length,
+                        model_name,
+                        provider=_provider,
+                        custom_providers=_custom_providers,
+                    )
                 )
             except Exception:
                 context_length = 0

hermes_cli/web_server.py

@@ -6922,11 +6922,14 @@ def get_model_info(profile: Optional[str] = None):
         # purely auto-detected value, then separately report the override)
         try:
             from agent.model_metadata import get_model_context_length
+            from hermes_cli.config import get_compatible_custom_providers
+            _cp = get_compatible_custom_providers(cfg)
             auto_ctx = get_model_context_length(
                 model=model_name,
                 base_url=base_url,
                 provider=provider,
                 config_context_length=None,  # ignore override — we want auto value
+                custom_providers=_cp,
             )
         except Exception:
             auto_ctx = 0

model_tools.py

@@ -720,12 +720,29 @@ def _resolve_active_context_length() -> int:
                     return cached_ctx
             except Exception:
                 pass
+        # Per-model context_length overrides from custom_providers config
+        # (custom_providers[].models.<id>.context_length) must be honored by
+        # the tool-search gate too — otherwise the gate sizes against generic
+        # metadata for custom endpoints (#15779). Best-effort load: config
+        # failure degrades to None so resolution falls through to probing.
+        custom_providers = None
+        try:
+            from hermes_cli.config import (
+                get_compatible_custom_providers,
+                load_config_readonly,
+            )
+            custom_providers = get_compatible_custom_providers(
+                load_config_readonly()
+            )
+        except Exception:
+            custom_providers = None
         return int(get_model_context_length(
             model_id,
             base_url=base_url,
             api_key=api_key,
             config_context_length=config_ctx,
             provider=provider,
+            custom_providers=custom_providers,
         ) or 0)
     except Exception as e:
         logger.debug("Could not resolve active context length: %s", e)