diff --git a/agent/agent_init.py b/agent/agent_init.py
index 68fa17b0dff8..9ae224335a99 100644
--- a/agent/agent_init.py
+++ b/agent/agent_init.py
@@ -432,14 +432,55 @@ def _custom_provider_extra_body_for_agent(
     model: str,
     base_url: str,
     custom_providers: List[Dict[str, Any]],
+    requested_provider: str = "",
 ) -> Optional[Dict[str, Any]]:
     provider_norm = (provider or "").strip().lower()
+    requested_norm = (requested_provider or "").strip().lower()
+
     if provider_norm == "custom":
-        provider_key_filter = ""
+        # ``agent.provider`` is the bare canonical "custom" for every named
+        # providers:/custom_providers: entry AT AGENT-INIT TIME — the runtime
+        # resolver (hermes_cli/runtime_provider.py::_resolve_named_custom_runtime)
+        # never emits "custom:<name>". The entry's actual identity survives
+        # only on ``agent.requested_provider``. Without filtering on it, two
+        # entries sharing (base_url, model) — e.g. a vLLM endpoint listed
+        # twice as "vllm" / "vllm-no-think" with the same model id — are
+        # indistinguishable here and the matcher falls through to "first
+        # entry with a non-empty extra_body", silently applying the WRONG
+        # entry's extra_body to every provider at that endpoint regardless
+        # of which one is actually selected.
+        provider_key_filter = requested_norm if requested_norm not in ("", "custom") else ""
+        if provider_key_filter.startswith("custom:"):
+            provider_key_filter = provider_key_filter.split(":", 1)[1].strip()
     elif provider_norm.startswith("custom:"):
         provider_key_filter = provider_norm.split(":", 1)[1].strip()
     else:
-        return None
+        # A LIVE ``/model`` switch (hermes_cli/model_switch.py's pure
+        # switch_model(), which every gateway/TUI/CLI switch path calls
+        # before mutating the live agent) resolves ``target_provider`` to
+        # the entry's OWN identity ("vllm"), not "custom" — a different
+        # convention than agent-init resolution. So after any live switch,
+        # neither ``provider`` nor ``requested_provider`` ever looks like
+        # "custom"/"custom:<name>" for a named custom provider, and every
+        # branch above misses. Recognize this shape explicitly: only engage
+        # when the bare name actually names a configured custom_providers
+        # entry (by provider_key or display name) — this still can't
+        # false-positive on a builtin provider (openai, anthropic, ...)
+        # since those never appear in ``custom_providers``.
+        known_identities = {
+            str(e.get("provider_key", "") or "").strip().lower()
+            for e in (custom_providers or []) if isinstance(e, dict)
+        } | {
+            str(e.get("name", "") or "").strip().lower()
+            for e in (custom_providers or []) if isinstance(e, dict)
+        }
+        known_identities.discard("")
+        if provider_norm in known_identities:
+            provider_key_filter = provider_norm
+        elif requested_norm in known_identities:
+            provider_key_filter = requested_norm
+        else:
+            return None
 
     target_url = _normalized_custom_base_url(base_url)
     if not target_url:
@@ -472,22 +513,46 @@ def _custom_provider_extra_body_for_agent(
 
 
 def _merge_custom_provider_extra_body(agent, custom_providers: List[Dict[str, Any]]) -> None:
+    """(Re)apply the active custom provider's ``extra_body`` onto the agent.
+
+    Called at agent init AND again on every live ``/model`` switch (see
+    ``agent.agent_runtime_helpers.switch_model``) — nothing else touches
+    ``agent.request_overrides['extra_body']`` on a switch, so without the
+    second call a provider switched-away-from with an ``extra_body`` (e.g.
+    a vLLM endpoint listed twice as "vllm" / "vllm-no-think", the latter
+    with ``chat_template_kwargs.enable_thinking``) leaves its extra_body
+    stuck on every request for the rest of the session, even after
+    switching to a provider with none.
+
+    Idempotent/repeatable: each call first strips whatever keys the
+    *previous* call itself contributed (tracked via
+    ``agent._custom_provider_extra_body_keys``) before merging in the
+    newly resolved ``extra_body``, so keys set by something else entirely
+    (e.g. an explicit fast-mode ``service_tier`` override) are preserved
+    across the switch.
+    """
     extra_body = _custom_provider_extra_body_for_agent(
         provider=agent.provider,
         model=agent.model,
         base_url=agent.base_url,
         custom_providers=custom_providers,
-    )
-    if not extra_body:
-        return
+        requested_provider=getattr(agent, "requested_provider", ""),
+    ) or {}
 
     overrides = dict(getattr(agent, "request_overrides", {}) or {})
+    existing_extra_body = dict(overrides.get("extra_body") or {})
+    for key in getattr(agent, "_custom_provider_extra_body_keys", set()):
+        existing_extra_body.pop(key, None)
+
     merged_extra_body = dict(extra_body)
-    existing_extra_body = overrides.get("extra_body")
-    if isinstance(existing_extra_body, dict):
-        merged_extra_body.update(existing_extra_body)
-    overrides["extra_body"] = merged_extra_body
+    merged_extra_body.update(existing_extra_body)
+
+    if merged_extra_body:
+        overrides["extra_body"] = merged_extra_body
+    elif "extra_body" in overrides:
+        del overrides["extra_body"]
     agent.request_overrides = overrides
+    agent._custom_provider_extra_body_keys = set(extra_body.keys())
 
 
 def init_agent(
diff --git a/agent/agent_runtime_helpers.py b/agent/agent_runtime_helpers.py
index a9bf1a1d38df..a32749097230 100644
--- a/agent/agent_runtime_helpers.py
+++ b/agent/agent_runtime_helpers.py
@@ -2755,6 +2755,22 @@ def _restore_snapshot() -> None:
         _runtime_context_length,
     )
 
+    # ── Re-apply the active custom provider's extra_body ──
+    # Nothing else in this function touches request_overrides['extra_body'],
+    # so switching away from a providers:/custom_providers: entry that had one
+    # (e.g. a vLLM endpoint listed twice as "vllm" / "vllm-no-think", the
+    # latter carrying extra_body.chat_template_kwargs.enable_thinking) would
+    # otherwise leave that extra_body stuck on every request for the rest of
+    # the session, even after switching to a provider with none configured.
+    try:
+        from agent.agent_init import _merge_custom_provider_extra_body
+        _merge_custom_provider_extra_body(agent, _sm_custom_providers or [])
+    except Exception:
+        logger.debug(
+            "switch_model: custom-provider extra_body reconciliation skipped",
+            exc_info=True,
+        )
+
     # ── Re-evaluate prompt caching ──
     # Refresh the custom-provider snapshot from the config just loaded above
     # so the per-model ``prompt_caching`` capability lookup sees the same
diff --git a/hermes_cli/model_switch.py b/hermes_cli/model_switch.py
index d6cd312c9967..d3948e7bb0fd 100644
--- a/hermes_cli/model_switch.py
+++ b/hermes_cli/model_switch.py
@@ -1981,6 +1981,69 @@ def _extra_headers_from_config(entry: Any) -> dict[str, str]:
     return normalize_extra_headers(entry.get("extra_headers"))
 
 
+def _extra_body_identity(entry: Any) -> str:
+    """Stable, hashable identity for an entry's ``extra_body``.
+
+    Two entries can share (api_url, credential, api_mode) yet send different
+    request bodies — e.g. a vLLM endpoint listed twice with only
+    ``extra_body.chat_template_kwargs.enable_thinking`` differing between the
+    "think" and "no-think" variants. Without this in the group identity the
+    picker silently collapses them into one row and one of the two configured
+    ``providers:``/``custom_providers:`` entries becomes unreachable.
+    """
+    if not isinstance(entry, dict):
+        return ""
+    extra_body = entry.get("extra_body")
+    if not isinstance(extra_body, dict) or not extra_body:
+        return ""
+    import json
+
+    try:
+        return json.dumps(extra_body, sort_keys=True, default=str)
+    except TypeError:
+        return repr(sorted(extra_body.items(), key=lambda kv: str(kv[0])))
+
+
+def _group_display_prefix(raw_name: str) -> str:
+    """Version-stripped display prefix used to decide whether two same-
+    endpoint ``providers:`` entries are the same logical provider (several
+    models, one row) or genuinely distinct providers that happen to share a
+    connection.
+
+    Strips a trailing " — " / " - " suffix (Hermes's own config writer uses
+    this to separate a provider label from its model, e.g. "Ollama — GLM
+    5.1"), then drops trailing numeric/version tokens ("Palantir Claude 4.7
+    Opus" → "Palantir Claude") so multi-model entries for the same provider
+    still collapse into one picker row — this mirrors the cosmetic label
+    logic section 4 (``custom_providers:``) already uses for its own display
+    prefix, applied here as part of the *grouping identity* instead of only
+    the label.
+
+    A name with neither a separator nor a trailing digit token is returned
+    unchanged. That is what disambiguates a case like a vLLM endpoint listed
+    twice under distinct names ("vLLM" / "vLLM No-Think", e.g. to toggle
+    ``extra_body.chat_template_kwargs.enable_thinking`` for a hybrid-thinking
+    model) — those names carry no version pattern, so they stay distinct
+    picker rows rather than collapsing into one with only the first entry's
+    identity reachable.
+    """
+    prefix = str(raw_name or "")
+    for sep in ("—", " - "):
+        if sep in prefix:
+            prefix = prefix.split(sep)[0].strip()
+            break
+    tokens = prefix.split()
+    cut_at = None
+    for i, tok in enumerate(tokens):
+        stripped_tok = tok.strip(".,()")
+        if stripped_tok and any(c.isdigit() for c in stripped_tok):
+            cut_at = i
+            break
+    if cut_at is not None and cut_at >= 2:
+        prefix = " ".join(tokens[:cut_at]).strip()
+    return prefix
+
+
 def prewarm_picker_cache_async() -> Optional["_threading.Thread"]:
     """Warm the provider-models disk cache in a background daemon thread.
 
@@ -2658,19 +2721,39 @@ def _has_aws_sdk_creds_for_listing(slug: str) -> bool:
     # and one "custom:openrouter" from section 4, both labelled identically.
     _section3_emitted_pairs: set = set()
     if user_providers and isinstance(user_providers, dict):
-        # Group ``providers:`` entries by (api_url, key_env, api_mode) so that
-        # multiple keyed providers pointing at the same endpoint with the
-        # same credential and wire-protocol collapse into one picker row.
-        # Mirrors section-4's grouping for ``custom_providers:`` lists.
-        # Concrete case: a Palantir Foundry Anthropic-proxy with two
-        # configured models (claude-4.6 + claude-4.7) — both share the same
-        # api/key_env/api_mode and used to produce two near-duplicate rows
-        # labelled "Palantir Claude 4.6 Opus" and "Palantir Claude 4.7 Opus";
-        # now they appear as a single "Palantir Claude" row with both models
-        # in the dropdown. Same-host entries with different ``key_env`` or
-        # ``api_mode`` (e.g. an OpenAI-compat gpt-5.4 alongside the Anthropic
-        # claude-4.7 on the same Palantir host) keep distinct rows since
-        # the wire protocol differs.
+        # Group ``providers:`` entries by (api_url, credential, api_mode,
+        # headers, extra_body, name-prefix) so that multiple keyed providers
+        # pointing at the same endpoint with the same credential and wire
+        # protocol collapse into one picker row — but ONLY when their names
+        # also identify them as the same logical provider. Mirrors (and now
+        # matches) section-4's grouping for ``custom_providers:`` lists.
+        #
+        # Concrete "same provider" case: a Palantir Foundry Anthropic-proxy
+        # with two configured models (claude-4.6 + claude-4.7) — both share
+        # api/key_env/api_mode and are named "Palantir Claude 4.6 Opus" /
+        # "Palantir Claude 4.7 Opus". The trailing version token is stripped
+        # (_group_display_prefix) so both fold into a single "Palantir
+        # Claude" row with both models in the dropdown, instead of two
+        # near-duplicate rows.
+        #
+        # Concrete "distinct providers" case: a vLLM endpoint listed twice —
+        # same base_url/model, no credential — under deliberately different
+        # names ("vLLM" / "vLLM No-Think") so the second entry can pin
+        # ``extra_body: {chat_template_kwargs: {enable_thinking: false}}``
+        # for a hybrid-thinking model. Neither name carries a version
+        # pattern, so the prefix heuristic leaves them unchanged and they
+        # stay distinct rows. Without this — and without extra_body/headers
+        # also participating in the identity below — the group_key collapsed
+        # to one row keyed on the FIRST entry's slug, and the second entry's
+        # extra_body became permanently unreachable from the picker even
+        # though it resolved correctly via ``--provider vllm-no-think`` or a
+        # direct config edit (request-time resolution matches by slug, not
+        # by this grouping).
+        #
+        # Same-host entries with different ``key_env`` or ``api_mode`` (e.g.
+        # an OpenAI-compat gpt-5.4 alongside the Anthropic claude-4.7 on the
+        # same Palantir host) keep distinct rows since the wire protocol
+        # differs — unaffected by any of the above.
         from collections import OrderedDict as _OD3
 
         from hermes_cli.config import is_provider_enabled
@@ -2712,7 +2795,27 @@ def _has_aws_sdk_creds_for_listing(slug: str) -> bool:
             # URL, routed by header) and must keep distinct picker rows.
             entry_extra_headers = _extra_headers_from_config(ep_cfg)
             headers_identity = tuple(sorted(entry_extra_headers.items()))
-            group_key = (api_url_norm, credential_identity, api_mode, headers_identity)
+            # Per-provider extra_body participates in the group identity for
+            # the same reason as headers: two entries sharing
+            # (api_url, credential, api_mode, headers) but declaring different
+            # extra_body (e.g. chat_template_kwargs.enable_thinking) send
+            # different requests and must stay distinct picker rows.
+            body_identity = _extra_body_identity(ep_cfg)
+            # The version-stripped name prefix participates in the group
+            # identity too (see the module-level docstring on
+            # _group_display_prefix and the block comment above this loop):
+            # this is what tells "same provider, more models" (Palantir)
+            # apart from "distinct provider sharing a connection" (vLLM /
+            # vLLM No-Think).
+            group_prefix = _group_display_prefix(display_name)
+            group_key = (
+                api_url_norm,
+                credential_identity,
+                api_mode,
+                headers_identity,
+                body_identity,
+                group_prefix.lower(),
+            )
 
             # ``default_model`` is the legacy key; ``model`` matches what
             # custom_providers entries use, so accept either.
@@ -2730,33 +2833,10 @@ def _has_aws_sdk_creds_for_listing(slug: str) -> bool:
                     entry_models.append(model_id)
 
             if group_key not in ep_groups:
-                # Strip per-model suffix so "Palantir Claude 4.7 Opus" becomes
-                # "Palantir Claude". Em dash and " - " are the separators
-                # Hermes's own writer uses (mirrors section-4 grouping).
-                grp_display = display_name
-                for sep in ("—", " - "):
-                    if sep in grp_display:
-                        grp_display = grp_display.split(sep)[0].strip()
-                        break
-                # Drop trailing numeric/version tokens that distinguish per-model
-                # entries ("Palantir Claude 4.7 Opus" → "Palantir Claude").
-                # Keeps the row label short; the model dropdown carries the
-                # per-version detail. Heuristic: split at the first token whose
-                # stripped form contains a digit; keep the prefix only if it
-                # is at least 2 words (avoids over-trimming single-word names).
-                _toks = grp_display.split()
-                _cut_at = None
-                for _i, _t in enumerate(_toks):
-                    _tl = _t.strip(".,()")
-                    if _tl and any(c.isdigit() for c in _tl):
-                        _cut_at = _i
-                        break
-                if _cut_at is not None and _cut_at >= 2:
-                    grp_display = " ".join(_toks[:_cut_at]).strip()
                 grp_slug = ep_name  # primary slug is the first ep_name encountered
                 ep_groups[group_key] = {
                     "slug": grp_slug,
-                    "name": grp_display or display_name,
+                    "name": group_prefix or display_name,
                     "api_url": api_url,
                     "models": [],
                     "has_explicit_models": False,
@@ -3022,6 +3102,15 @@ def _has_aws_sdk_creds_for_listing(slug: str) -> bool:
             entry_extra_headers = _extra_headers_from_config(entry)
             headers_identity = tuple(sorted(entry_extra_headers.items()))
 
+            # Per-provider extra_body participates in the group identity for
+            # the same reason as headers: two entries sharing
+            # (api_url, credential, api_mode, headers) but declaring different
+            # extra_body (e.g. a vLLM endpoint listed twice with only
+            # extra_body.chat_template_kwargs.enable_thinking differing) send
+            # different requests and must stay distinct picker rows rather
+            # than collapsing with only one of the two ``models:`` surviving.
+            body_identity = _extra_body_identity(entry)
+
             # Display-name prefix (text before " — " / " - "), used both
             # as a grouping dimension and to derive the row's display name.
             _display_prefix = raw_name
@@ -3030,7 +3119,14 @@ def _has_aws_sdk_creds_for_listing(slug: str) -> bool:
                     _display_prefix = _display_prefix.split(sep)[0].strip()
                     break
 
-            group_key = (api_url, credential_identity, api_mode, headers_identity, _display_prefix.lower())
+            group_key = (
+                api_url,
+                credential_identity,
+                api_mode,
+                headers_identity,
+                body_identity,
+                _display_prefix.lower(),
+            )
             if group_key not in groups:
                 # Reuse the prefix computed above as the row display name;
                 # fall back to the raw name if stripping left it empty.
diff --git a/tests/agent/test_custom_provider_extra_body.py b/tests/agent/test_custom_provider_extra_body.py
index ea94c104f2a2..f0e2c22b1709 100644
--- a/tests/agent/test_custom_provider_extra_body.py
+++ b/tests/agent/test_custom_provider_extra_body.py
@@ -71,3 +71,208 @@ def test_named_custom_provider_extra_body_matches_provider_key():
     )
 
     assert agent.request_overrides == {"extra_body": {"enable_thinking": False}}
+
+
+def test_bare_custom_provider_disambiguates_via_requested_provider():
+    """Two providers: entries at the SAME (base_url, model) — one with
+    extra_body, one without — must not cross-contaminate.
+
+    ``agent.provider`` is always the bare canonical "custom" for named
+    providers:/custom_providers: entries (the runtime resolver never emits
+    "custom:<name>"); only ``agent.requested_provider`` carries the actual
+    selected entry's identity ("vllm" vs "vllm-no-think"). Without filtering
+    on it, the matcher fell through to "first entry with a non-empty
+    extra_body" and applied the WRONG entry's extra_body regardless of which
+    provider was actually selected (repro: a vLLM endpoint listed twice with
+    the same model id, only "vllm-no-think" carrying
+    extra_body.chat_template_kwargs.enable_thinking)."""
+    custom_providers = [
+        {
+            "name": "vLLM",
+            "provider_key": "vllm",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        },
+        {
+            "name": "vLLM No-Think",
+            "provider_key": "vllm-no-think",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+            "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+        },
+    ]
+
+    agent_plain = SimpleNamespace(
+        provider="custom",
+        requested_provider="vllm",
+        model="unsloth/Qwen3.6-35B-A3B-NVFP4",
+        base_url="http://192.168.15.115:8000/v1",
+        request_overrides={},
+    )
+    _merge_custom_provider_extra_body(agent_plain, custom_providers)
+    assert agent_plain.request_overrides == {}
+
+    agent_no_think = SimpleNamespace(
+        provider="custom",
+        requested_provider="vllm-no-think",
+        model="unsloth/Qwen3.6-35B-A3B-NVFP4",
+        base_url="http://192.168.15.115:8000/v1",
+        request_overrides={},
+    )
+    _merge_custom_provider_extra_body(agent_no_think, custom_providers)
+    assert agent_no_think.request_overrides == {
+        "extra_body": {"chat_template_kwargs": {"enable_thinking": False}}
+    }
+
+
+def test_switching_away_from_extra_body_provider_clears_stale_value():
+    """A second call for a provider WITHOUT extra_body must clear whatever
+    the previous call (e.g. at agent init, before a live /model switch)
+    applied — mirrors what switch_model() now does on every live switch."""
+    custom_providers = [
+        {
+            "name": "vLLM",
+            "provider_key": "vllm",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        },
+        {
+            "name": "vLLM No-Think",
+            "provider_key": "vllm-no-think",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+            "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+        },
+    ]
+    agent = SimpleNamespace(
+        provider="custom",
+        requested_provider="vllm-no-think",
+        model="unsloth/Qwen3.6-35B-A3B-NVFP4",
+        base_url="http://192.168.15.115:8000/v1",
+        request_overrides={},
+    )
+    _merge_custom_provider_extra_body(agent, custom_providers)
+    assert agent.request_overrides == {
+        "extra_body": {"chat_template_kwargs": {"enable_thinking": False}}
+    }
+
+    # Simulate a live /model switch to "vllm" (no extra_body).
+    agent.requested_provider = "vllm"
+    _merge_custom_provider_extra_body(agent, custom_providers)
+    assert agent.request_overrides == {}
+
+
+def test_bare_raw_entry_name_as_provider_resolves_via_known_identities():
+    """hermes_cli/model_switch.py's live /model-switch path sets
+    agent.provider to the entry's OWN raw identity (e.g. "vllm-no-think"),
+    not "custom" or "custom:vllm-no-think" — a different convention than
+    agent-init resolution. Matching must still work by recognizing the bare
+    name against the configured custom_providers identities."""
+    custom_providers = [
+        {
+            "name": "vLLM",
+            "provider_key": "vllm",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        },
+        {
+            "name": "vLLM No-Think",
+            "provider_key": "vllm-no-think",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+            "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+        },
+    ]
+
+    agent_vllm = SimpleNamespace(
+        provider="vllm",
+        requested_provider="vllm",
+        model="unsloth/Qwen3.6-35B-A3B-NVFP4",
+        base_url="http://192.168.15.115:8000/v1",
+        request_overrides={},
+    )
+    _merge_custom_provider_extra_body(agent_vllm, custom_providers)
+    assert agent_vllm.request_overrides == {}
+
+    agent_no_think = SimpleNamespace(
+        provider="vllm-no-think",
+        requested_provider="vllm-no-think",
+        model="unsloth/Qwen3.6-35B-A3B-NVFP4",
+        base_url="http://192.168.15.115:8000/v1",
+        request_overrides={},
+    )
+    _merge_custom_provider_extra_body(agent_no_think, custom_providers)
+    assert agent_no_think.request_overrides == {
+        "extra_body": {"chat_template_kwargs": {"enable_thinking": False}}
+    }
+
+
+def test_bare_provider_name_matching_builtin_never_leaks_custom_extra_body():
+    """A builtin provider name (never present in custom_providers) must not
+    accidentally match — the known-identities fallback only engages for
+    names that actually belong to a configured custom_providers entry."""
+    agent = SimpleNamespace(
+        provider="openai",
+        requested_provider="openai",
+        model="gpt-5.5",
+        base_url="https://api.openai.com/v1",
+        request_overrides={},
+    )
+    _merge_custom_provider_extra_body(
+        agent,
+        [
+            {
+                "name": "vLLM No-Think",
+                "provider_key": "vllm-no-think",
+                "base_url": "https://api.openai.com/v1",
+                "model": "gpt-5.5",
+                "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+            }
+        ],
+    )
+    assert agent.request_overrides == {}
+
+
+def test_switching_preserves_unrelated_caller_override_across_switch():
+    """A caller-set override key unrelated to the custom-provider config
+    (e.g. an explicit fast-mode service_tier) must survive a switch that
+    also changes the custom-provider-derived extra_body."""
+    custom_providers = [
+        {
+            "name": "A",
+            "provider_key": "a",
+            "base_url": "https://proxy.example.com/v1",
+            "model": "m",
+            "extra_body": {"chat_template_kwargs": {"enable_thinking": True}},
+        },
+        {
+            "name": "B",
+            "provider_key": "b",
+            "base_url": "https://proxy.example.com/v1",
+            "model": "m",
+            "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+        },
+    ]
+    agent = SimpleNamespace(
+        provider="custom",
+        requested_provider="a",
+        model="m",
+        base_url="https://proxy.example.com/v1",
+        request_overrides={"extra_body": {"service_tier": "flex"}},
+    )
+    _merge_custom_provider_extra_body(agent, custom_providers)
+    assert agent.request_overrides == {
+        "extra_body": {
+            "chat_template_kwargs": {"enable_thinking": True},
+            "service_tier": "flex",
+        }
+    }
+
+    agent.requested_provider = "b"
+    _merge_custom_provider_extra_body(agent, custom_providers)
+    assert agent.request_overrides == {
+        "extra_body": {
+            "chat_template_kwargs": {"enable_thinking": False},
+            "service_tier": "flex",
+        }
+    }
diff --git a/tests/hermes_cli/test_model_switch_custom_providers.py b/tests/hermes_cli/test_model_switch_custom_providers.py
index 381c5870c9bb..8f5986d3d6fa 100644
--- a/tests/hermes_cli/test_model_switch_custom_providers.py
+++ b/tests/hermes_cli/test_model_switch_custom_providers.py
@@ -715,6 +715,42 @@ def fake_fetch_api_models(api_key, base_url, **kwargs):
     assert models_by_row == {("model-a",), ("model-b",)}
 
 
+def test_same_endpoint_different_extra_body_not_collapsed(monkeypatch):
+    """Entries sharing (api_url, credential, api_mode, headers) but declaring
+    different extra_body must NOT collapse into one picker row — e.g. a vLLM
+    endpoint listed twice where only extra_body.chat_template_kwargs
+    .enable_thinking differs between "think" and "no-think" variants."""
+    monkeypatch.setattr("agent.models_dev.fetch_models_dev", lambda: {})
+    monkeypatch.setattr("hermes_cli.providers.HERMES_OVERLAYS", {})
+    monkeypatch.setattr("hermes_cli.models.fetch_api_models", lambda *a, **k: [])
+
+    providers = list_authenticated_providers(
+        current_provider="openrouter",
+        current_base_url="https://openrouter.ai/api/v1",
+        custom_providers=[
+            {
+                "name": "vLLM",
+                "base_url": "http://192.168.15.115:8000/v1",
+                "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+                "discover_models": False,
+                "models": {"unsloth/Qwen3.6-35B-A3B-NVFP4": {}},
+            },
+            {
+                "name": "vLLM",
+                "base_url": "http://192.168.15.115:8000/v1",
+                "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+                "discover_models": False,
+                "models": {"unsloth/Qwen3.6-35B-A3B-NVFP4": {}},
+                "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+            },
+        ],
+        max_models=50,
+    )
+
+    rows = [
+        p for p in providers if p.get("api_url") == "http://192.168.15.115:8000/v1"
+    ]
+    assert len(rows) == 2, f"expected 2 rows, got {len(rows)}: {rows}"
 
 
 
diff --git a/tests/hermes_cli/test_provider_section3_grouping.py b/tests/hermes_cli/test_provider_section3_grouping.py
index 2889fc723c63..0449337607b9 100644
--- a/tests/hermes_cli/test_provider_section3_grouping.py
+++ b/tests/hermes_cli/test_provider_section3_grouping.py
@@ -6,6 +6,16 @@
 mirroring section 4's grouping for ``custom_providers:``. These are invariant
 tests — grouping identity, header-routed separation, list-of-dict model
 declarations, and display-only RID stripping.
+
+Extended for the "distinct name" invariant: two entries at the same endpoint
+whose names carry no shared version pattern (unlike "Palantir Claude 4.6/4.7
+Opus") are genuinely distinct providers, not "one provider, several models",
+and must stay separate picker rows regardless of whether anything else in
+their config differs — extra_body, extra_headers, or nothing at all. Before
+this, section 3 grouped purely on connection identity (never on name), so a
+second same-named-endpoint entry could become permanently unreachable from
+the picker even when perfectly valid via ``--provider <name>`` or a direct
+config edit.
 """
 
 import hermes_cli.providers as providers_mod
@@ -80,6 +90,66 @@ def test_different_extra_headers_keep_distinct_rows(monkeypatch):
     assert len(rows) == 2
 
 
+def test_different_extra_body_keeps_distinct_rows(monkeypatch):
+    """Two providers: entries sharing (api_url, credential, api_mode, headers)
+    but declaring different extra_body must NOT collapse — e.g. a vLLM
+    endpoint listed twice where only extra_body.chat_template_kwargs
+    .enable_thinking differs between the "think" and "no-think" variants.
+    Regression test: previously these merged into one row and one of the two
+    ``models:`` entries silently disappeared from the picker."""
+    rows = _user_rows(_providers(monkeypatch, {
+        "vllm": {
+            "name": "vLLM",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+            "discover_models": False,
+            "models": {"unsloth/Qwen3.6-35B-A3B-NVFP4": {}},
+        },
+        "vllm-no-think": {
+            "name": "vLLM No-Think",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+            "discover_models": False,
+            "models": {"unsloth/Qwen3.6-35B-A3B-NVFP4": {}},
+            "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+        },
+    }))
+    assert len(rows) == 2, f"expected 2 rows, got {len(rows)}: {rows}"
+    slugs = {row["slug"] for row in rows}
+    assert slugs == {"vllm", "vllm-no-think"}
+
+
+def test_distinctly_named_entries_stay_separate_even_with_identical_config(monkeypatch):
+    """Two providers: entries at the same endpoint, with IDENTICAL config in
+    every other respect (no extra_body, no extra_headers, same credential),
+    but deliberately different names must still stay separate rows — name is
+    the primary signal, not a fallback used only when something else also
+    differs. Neither "vLLM" nor "vLLM No-Think" carries a version-number
+    pattern, so the Palantir-style suffix-stripping heuristic leaves both
+    unchanged and they don't collapse."""
+    rows = _user_rows(_providers(monkeypatch, {
+        "vllm": {
+            "name": "vLLM",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+            "discover_models": False,
+            "models": {"unsloth/Qwen3.6-35B-A3B-NVFP4": {}},
+        },
+        "vllm-no-think": {
+            "name": "vLLM No-Think",
+            "base_url": "http://192.168.15.115:8000/v1",
+            "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+            "discover_models": False,
+            "models": {"unsloth/Qwen3.6-35B-A3B-NVFP4": {}},
+        },
+    }))
+    assert len(rows) == 2, f"expected 2 rows, got {len(rows)}: {rows}"
+    slugs = {row["slug"] for row in rows}
+    assert slugs == {"vllm", "vllm-no-think"}
+    names = {row["name"] for row in rows}
+    assert names == {"vLLM", "vLLM No-Think"}
+
+
 class TestFormatModelForDisplay:
     def test_palantir_rid_stripped_to_trailing_slug(self):
         rid = "ri.language-model-service..language-model.anthropic-claude-4-7-opus"
diff --git a/tests/run_agent/test_switch_model_extra_body.py b/tests/run_agent/test_switch_model_extra_body.py
new file mode 100644
index 000000000000..1a8599c365e4
--- /dev/null
+++ b/tests/run_agent/test_switch_model_extra_body.py
@@ -0,0 +1,158 @@
+"""Regression tests: switch_model() must reconcile request_overrides['extra_body']
+on every live /model switch, for both agent-init and live-switch provider
+identity conventions.
+
+Root-cause bugs fixed together:
+
+1. ``agent.provider`` carries the entry's identity DIFFERENTLY depending on
+   when it's resolved. At agent-init (hermes_cli/runtime_provider.py::
+   _resolve_named_custom_runtime) it's always the bare canonical "custom",
+   with the real identity only on ``agent.requested_provider``. On a live
+   ``/model`` switch (hermes_cli/model_switch.py's pure switch_model(), which
+   every gateway/TUI/CLI switch path calls before mutating the live agent)
+   ``target_provider`` — and so ``agent.provider`` after the swap — is the
+   entry's OWN raw identity instead (e.g. "vllm"), never "custom". Matching
+   extra_body only recognized the first convention, so two providers:
+   entries sharing (base_url, model) — e.g. a vLLM endpoint listed twice as
+   "vllm" / "vllm-no-think", the latter pinning
+   extra_body.chat_template_kwargs.enable_thinking for a hybrid-thinking
+   model — were indistinguishable and the WRONG entry's extra_body could be
+   applied regardless of which was actually selected, or (for the live-
+   switch convention) extra_body silently stopped resolving at all.
+
+2. switch_model() never touched request_overrides['extra_body'] at all, so a
+   provider switched away from would leave its extra_body stuck on every
+   request for the rest of the session even after switching to one with none
+   configured.
+"""
+
+from unittest.mock import MagicMock, patch
+
+
+def _make_agent(provider="vllm", requested_provider=None, extra_body=None):
+    """Minimal AIAgent with just enough surface for switch_model()'s
+    non-anthropic (openai-client) branch, mirroring
+    tests/run_agent/test_switch_model_reapplies_headers.py."""
+    from run_agent import AIAgent
+
+    agent = AIAgent.__new__(AIAgent)
+    agent.model = "unsloth/Qwen3.6-35B-A3B-NVFP4"
+    agent.provider = provider
+    agent.requested_provider = requested_provider if requested_provider is not None else provider
+    agent.base_url = "http://192.168.15.115:8000/v1"
+    agent.api_key = "no-key-required"
+    agent.api_mode = "chat_completions"
+    agent.client = MagicMock()
+    agent.quiet_mode = True
+    agent._config_context_length = None
+    agent._client_kwargs = {"api_key": "no-key-required", "base_url": agent.base_url}
+    agent.request_overrides = dict(extra_body and {"extra_body": extra_body} or {})
+    agent._custom_provider_extra_body_keys = set(extra_body or {})
+    agent._credential_pool = None
+    agent._credential_pool_entry_id = None
+    agent._transport_cache = {}
+    agent.context_compressor = None
+    agent._primary_runtime = {}
+    agent._cached_system_prompt = None
+    agent._anthropic_api_key = ""
+    agent._anthropic_base_url = None
+    agent._is_anthropic_oauth = False
+    agent._anthropic_prompt_cache_policy = MagicMock(return_value=(False, False))
+    agent._ensure_lmstudio_runtime_loaded = MagicMock(return_value=None)
+    agent._lmstudio_load_was_unverified = MagicMock(return_value=False)
+    agent._effective_lmstudio_context_length = MagicMock(return_value=None)
+    agent._create_openai_client = MagicMock(return_value=MagicMock())
+    agent._apply_client_headers_for_base_url = MagicMock()
+    return agent
+
+
+_CUSTOM_PROVIDERS = [
+    {
+        "name": "vLLM",
+        "provider_key": "vllm",
+        "base_url": "http://192.168.15.115:8000/v1",
+        "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+    },
+    {
+        "name": "vLLM No-Think",
+        "provider_key": "vllm-no-think",
+        "base_url": "http://192.168.15.115:8000/v1",
+        "model": "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        "extra_body": {"chat_template_kwargs": {"enable_thinking": False}},
+    },
+]
+
+
+@patch("agent.model_metadata.get_model_context_length", return_value=131_072)
+@patch("hermes_cli.config.get_custom_provider_context_length", return_value=None)
+@patch("hermes_cli.config.get_compatible_custom_providers", return_value=_CUSTOM_PROVIDERS)
+@patch("hermes_cli.config.load_config", return_value={})
+def test_switch_to_no_think_variant_picks_up_its_extra_body(
+    mock_cfg, mock_cps, mock_ctx_len_cp, mock_ctx_len
+):
+    """Real invocation convention: new_provider is the entry's OWN identity
+    ("vllm-no-think"), matching what hermes_cli/model_switch.py's
+    target_provider actually resolves to for a named providers: entry — NOT
+    the "custom" canonical form used at agent-init."""
+    agent = _make_agent(provider="vllm")
+
+    agent.switch_model(
+        "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        "vllm-no-think",
+        base_url="http://192.168.15.115:8000/v1",
+    )
+
+    assert agent.request_overrides.get("extra_body") == {
+        "chat_template_kwargs": {"enable_thinking": False}
+    }
+
+
+@patch("agent.model_metadata.get_model_context_length", return_value=131_072)
+@patch("hermes_cli.config.get_custom_provider_context_length", return_value=None)
+@patch("hermes_cli.config.get_compatible_custom_providers", return_value=_CUSTOM_PROVIDERS)
+@patch("hermes_cli.config.load_config", return_value={})
+def test_switch_away_from_no_think_variant_clears_stale_extra_body(
+    mock_cfg, mock_cps, mock_ctx_len_cp, mock_ctx_len
+):
+    agent = _make_agent(
+        provider="vllm-no-think",
+        extra_body={"chat_template_kwargs": {"enable_thinking": False}},
+    )
+
+    agent.switch_model(
+        "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        "vllm",
+        base_url="http://192.168.15.115:8000/v1",
+    )
+
+    assert agent.request_overrides.get("extra_body") is None
+
+
+@patch("agent.model_metadata.get_model_context_length", return_value=131_072)
+@patch("hermes_cli.config.get_custom_provider_context_length", return_value=None)
+@patch("hermes_cli.config.get_compatible_custom_providers", return_value=_CUSTOM_PROVIDERS)
+@patch("hermes_cli.config.load_config", return_value={})
+def test_agent_init_time_custom_convention_still_resolves_correctly(
+    mock_cfg, mock_cps, mock_ctx_len_cp, mock_ctx_len
+):
+    """The OTHER convention — agent.provider == "custom" (bare canonical,
+    as set at agent-init by hermes_cli/runtime_provider.py) with the real
+    identity only on agent.requested_provider — must keep working too."""
+    agent = _make_agent(provider="custom", requested_provider="vllm")
+
+    agent.switch_model(
+        "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        "custom",
+        base_url="http://192.168.15.115:8000/v1",
+    )
+    agent.requested_provider = "vllm-no-think"
+
+    agent.switch_model(
+        "unsloth/Qwen3.6-35B-A3B-NVFP4",
+        "custom",
+        base_url="http://192.168.15.115:8000/v1",
+    )
+
+    assert agent.request_overrides.get("extra_body") == {
+        "chat_template_kwargs": {"enable_thinking": False}
+    }