From 1add4dbd0ed1c87c2228d7aa15fde60e89b76065 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 27 May 2026 14:32:11 +0000
Subject: [PATCH] Tighten comments across PR 5711 (no behaviour change)

Audited every comment added by this PR; condensed multi-paragraph
docstrings and inline blocks to 1-2 lines where the WHY survives.

Files touched: 6 backend + 7 frontend. Net -148 lines.
- external_provider.py: -80 (docstring + stop_sequences + compaction)
- chat-settings-sheet.tsx: -70 (InfoHint tooltips + section headers)
- routes/inference.py: -54 (parallel_tool_calls cap comments)
- provider-capabilities.ts: -108 (stop-cap table, gemini bucket,
  service_tier resolver, max-output cap header)
- chat-adapter.ts: -27 (sampling forwarding stanza)
- providers.py: -24 (kimi reasoning class, deepseek aliases)
- llama_cpp.py: -24 (payload builder shared header)
- models/inference.py: -28 (Pydantic Field descriptions)
- chat-settings-storage.ts: -23 (nullable handling comments)
- stop-sequences-input.tsx: -23 (JSDoc + chip-key + draft-commit)
- anthropic_compat.py: -9 (serial tool-call gate)
- types/runtime.ts: -12 (DRY JSDoc, null convention header)
- types/api.ts: -6 (service_tier JSDoc)

Tightening rules applied:
- Removed em-dashes everywhere a comment was rewritten.
- Kept every external doc URL (Anthropic, OpenAI, vLLM, llama.cpp,
  Ollama, OpenRouter, Mistral, DeepSeek, Gemini, Kimi).
- Collapsed "silently dropped on unsupported routes" prose since the
  bucket-comment at the top of each capability table already states it.
- Removed restated rationale prose in comment blocks where the field
  name plus the master rule already encodes the WHY.

393/393 backend tests pass (test_sampling_params_routing 65, plus
anthropic / openai / gemini / llama-server suites). Frontend tsc +
vite build clean.
---
 .../core/inference/anthropic_compat.py        |   9 +-
 .../core/inference/external_provider.py       |  80 +++++--------
 studio/backend/core/inference/llama_cpp.py    |  24 ++--
 studio/backend/core/inference/providers.py    |  24 ++--
 studio/backend/models/inference.py            |  28 ++---
 studio/backend/routes/inference.py            |  54 +++------
 .../components/ui/stop-sequences-input.tsx    |  23 +---
 .../src/features/chat/api/chat-adapter.ts     |  27 ++---
 .../src/features/chat/chat-settings-sheet.tsx |  70 ++++--------
 .../features/chat/provider-capabilities.ts    | 108 +++++++-----------
 .../frontend/src/features/chat/types/api.ts   |   6 +-
 .../src/features/chat/types/runtime.ts        |  12 +-
 .../chat/utils/chat-settings-storage.ts       |  23 ++--
 13 files changed, 170 insertions(+), 318 deletions(-)

diff --git a/studio/backend/core/inference/anthropic_compat.py b/studio/backend/core/inference/anthropic_compat.py
index c2c1094806..cb2736a525 100644
--- a/studio/backend/core/inference/anthropic_compat.py
+++ b/studio/backend/core/inference/anthropic_compat.py
@@ -410,11 +410,10 @@ class AnthropicPassthroughEmitter:
         self._tool_call_states: dict = {}  # delta index -> {block_index, id, name}
         self._usage: dict = {}
         self._stop_reason: str = "end_turn"
-        # When the caller opted out of parallel tool calls (Anthropic
-        # `tool_choice.disable_parallel_tool_use=true` mapped to local
-        # `parallel_tool_calls=false`), only emit the first tool-call
-        # index and drop later siblings. llama.cpp may not enforce the
-        # flag on every jinja template (ggml-org/llama.cpp#22043).
+        # parallel_tool_calls=False (Anthropic
+        # tool_choice.disable_parallel_tool_use=true): emit only the first
+        # tool-call index; llama.cpp's flag isn't enforced by every jinja
+        # template (ggml-org/llama.cpp#22043).
         self._serial_tool_calls: bool = parallel_tool_calls is False
         self._first_tool_call_idx: Optional[int] = None
 
diff --git a/studio/backend/core/inference/external_provider.py b/studio/backend/core/inference/external_provider.py
index c38a077627..bdb56c33ab 100644
--- a/studio/backend/core/inference/external_provider.py
+++ b/studio/backend/core/inference/external_provider.py
@@ -81,15 +81,10 @@ def _normalize_stop_for_provider(
 def _is_openai_family_cloud(base_url: Optional[str]) -> bool:
     """True iff ``base_url`` points at OpenAI cloud or Azure OpenAI Foundry.
 
-    Host-anchored to avoid subdomain-injection bypass
-    (https://evil.com/api.openai.com/v1, https://api.openai.com.attacker.com/v1).
-    Used to gate cloud-only Responses-API extensions
-    (prompt_cache_retention, context_management compaction, container
-    shell tool) that 400 on non-cloud OAI-compat servers.
-
-    Azure Foundry uses <resource>.openai.azure.com; match via endswith
-    with the leading dot so the apex `openai.azure.com` can't slip
-    through (no apex Foundry endpoint exists).
+    Host-anchored against subdomain-injection (api.openai.com.attacker.com).
+    Gates Responses-API extensions (prompt_cache_retention, context_management,
+    container shell) that 400 on non-cloud OAI-compat servers. Azure Foundry
+    matches via .openai.azure.com suffix; leading dot blocks the apex.
     """
     if not base_url:
         return False
@@ -917,19 +912,14 @@ class ExternalProviderClient:
         For OpenAI-compatible providers, lines are forwarded verbatim.
         For Anthropic, the native Messages API SSE is translated to OpenAI format.
 
-        ``top_k`` and ``presence_penalty`` are forwarded only when the caller
-        supplies a value the provider accepts — the frontend's
-        provider-capability map already filters these per provider, so we
-        treat them as opt-in here.
-
+        Optional sampling extras (``top_k``, ``presence_penalty``,
         ``frequency_penalty``, ``seed``, ``stop``, ``service_tier``,
-        ``parallel_tool_calls`` follow the same rule: the per-provider
-        stream helpers silently drop fields the upstream API does not
-        accept (e.g. Responses rejects all of seed / frequency / stop;
-        Anthropic does not implement seed / frequency / logprobs).
+        ``parallel_tool_calls``) are opt-in. Per-provider helpers silently
+        drop fields the upstream rejects (Responses: seed/freq/stop;
+        Anthropic: seed/freq/logprobs).
 
-        ``fast_mode`` only applies to Anthropic Opus 4.6 / 4.7 (silently
-        dropped elsewhere); adds the beta header and ``speed: "fast"``.
+        ``fast_mode``: Anthropic Opus 4.6/4.7 only; adds the beta header
+        and ``speed: "fast"``.
         """
         # tool_choice="none" hard-disables hosted/builtin tools across
         # every provider so enabled_tools cannot accidentally bill or leak.
@@ -1054,9 +1044,8 @@ class ExternalProviderClient:
             else:
                 body["max_tokens"] = max_tokens
 
-        # Optional sampling extensions. Only forwarded when the caller
-        # passed a value. Per-provider rename / cap via `seed_field` /
-        # `stop_max`; `body_omit` strips fields the upstream rejects.
+        # Optional sampling extras; `seed_field` renames seed (Mistral),
+        # `body_omit` strips upstream-rejected fields.
         from core.inference.providers import get_provider_info
 
         provider_info = get_provider_info(self.provider_type) or {}
@@ -1069,9 +1058,8 @@ class ExternalProviderClient:
         normalized_stop = _normalize_stop_for_provider(stop, provider_info)
         if normalized_stop:
             body["stop"] = normalized_stop
-        # service_tier is OpenAI Chat-only on the generic OAI-compat
-        # branch; opt-in via `accepts_service_tier=True` on the registry
-        # entry. Anthropic and Responses handle it in their own helpers.
+        # service_tier is OAI-Chat-only here (accepts_service_tier registry
+        # opt-in); Anthropic/Responses branches set it themselves.
         if service_tier is not None and provider_info.get(
             "accepts_service_tier", False
         ):
@@ -1468,10 +1456,8 @@ class ExternalProviderClient:
         if max_tokens is not None:
             body["max_tokens"] = max_tokens
 
-        # The default OAI-compat body construction is skipped because
-        # this helper returns early. Apply the same provider-aware
-        # sampling / stop logic here so kimi-with-search matches
-        # kimi-without-search.
+        # Kimi-with-search returns early before the default OAI-compat
+        # body build; re-apply provider-aware sampling/stop here.
         from core.inference.providers import get_provider_info
 
         provider_info = get_provider_info(self.provider_type) or {}
@@ -2141,27 +2127,17 @@ class ExternalProviderClient:
         if top_k is not None and top_k > 0 and not sampling_removed:
             body["top_k"] = top_k
 
-        # Optional sampling extensions. Anthropic has no
-        # frequency_penalty / seed / logprobs equivalents so they are
-        # never forwarded here. The two body-level knobs Anthropic
-        # accepts land here:
-        #   stop          -> stop_sequences (renamed, ws-stripped)
-        #   service_tier  -> service_tier (auto|standard_only only)
-        # parallel_tool_calls inversion is applied after the tools
-        # wiring below because Anthropic requires it nested under
-        # tool_choice.
+        # Anthropic body-knob mapping: stop -> stop_sequences (ws-stripped,
+        # dedup), service_tier (auto|standard_only). parallel_tool_calls is
+        # handled after tools wiring (Anthropic nests under tool_choice).
         if stop:
             sequences: list[str]
             if isinstance(stop, str):
                 sequences = [stop] if stop.strip() else []
             else:
-                # Dedupe + drop whitespace-only entries. Anthropic
-                # rejects any sequence with no non-whitespace char
-                # ("stop_sequences: each stop sequence must contain
-                # non-whitespace"), so "", " ", "\n", "\n\n" are all
-                # filtered. The 16-cap is a client-side guard; the
-                # docs do not publish a max, but every SDK treats 16
-                # as a sane ceiling (Bedrock is the outlier at 8191).
+                # Anthropic rejects whitespace-only stop_sequences ("must
+                # contain non-whitespace"); 16-cap is a client-side guard
+                # (undocumented max; every SDK uses 16, Bedrock at 8191).
                 sequences = list(
                     dict.fromkeys(s for s in stop if isinstance(s, str) and s.strip())
                 )
@@ -2371,9 +2347,8 @@ class ExternalProviderClient:
             if anthropic_code_exec_container_id:
                 body["container"] = anthropic_code_exec_container_id
 
-        # parallel_tool_calls=False maps to disable_parallel_tool_use=True
-        # nested under tool_choice (top-level placement is rejected).
-        # Without tools the flag is a no-op so we skip the block.
+        # parallel_tool_calls=False -> tool_choice.disable_parallel_tool_use=True
+        # (top-level 400s; no-op without tools).
         # https://platform.claude.com/docs/en/agents-and-tools/tool-use/implement-tool-use
         if parallel_tool_calls is False and body.get("tools"):
             tc = body.get("tool_choice")
@@ -5435,11 +5410,8 @@ class ExternalProviderClient:
             "input": input_items,
             "stream": True,
         }
-        # Responses accepts auto|default|flex|priority per the live
-        # docs. The openai-python SDK type happens to include "scale"
-        # too but the public Responses reference does not, so drop it
-        # here to avoid a 400. Scale Tier is still selectable on Chat
-        # Completions backends. parallel_tool_calls default is true.
+        # Responses: auto|default|flex|priority (SDK type lists "scale"
+        # but server 400s; Scale Tier stays on Chat Completions).
         if service_tier in ("auto", "default", "flex", "priority"):
             body["service_tier"] = service_tier
         if parallel_tool_calls is not None:
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 7fd6c44d83..0b3e426dc1 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -4316,10 +4316,8 @@ class LlamaCppBackend:
             else (self._effective_context_length or _DEFAULT_MAX_TOKENS_FLOOR)
         )
         payload["t_max_predict_ms"] = _DEFAULT_T_MAX_PREDICT_MS
-        # Strip empty / non-string stop entries before forwarding to
-        # llama-server (matches `_normalize_stop_for_provider` for the
-        # external path). Without this a stale `stop=["", "END"]` can
-        # 400 the upstream.
+        # Strip empty/non-string stop entries (mirrors
+        # `_normalize_stop_for_provider`); stale `stop=["", "END"]` 400s.
         if stop:
             if isinstance(stop, str):
                 payload["stop"] = stop
@@ -4327,9 +4325,8 @@ class LlamaCppBackend:
                 _cleaned = [s for s in stop if isinstance(s, str) and s]
                 if _cleaned:
                     payload["stop"] = _cleaned
-        # Each field gated `is not None` so explicit 0 / 0.0 / False
-        # values reach the wire. llama-server silently ignores fields
-        # it doesn't recognise.
+        # `is not None` gate so explicit 0/False reach the wire;
+        # llama-server ignores unknown fields.
         if frequency_penalty is not None:
             payload["frequency_penalty"] = frequency_penalty
         if seed is not None:
@@ -5195,12 +5192,9 @@ class LlamaCppBackend:
                 _accumulated_predicted_ms += _it.get("predicted_ms", 0)
                 _accumulated_predicted_n += _it.get("predicted_n", 0)
 
-                # When the caller opted out of parallel tool calls
-                # (parallel_tool_calls=False), enforce at most one call
-                # per assistant turn even if llama-server emitted more.
-                # llama.cpp's parallel_tool_calls flag isn't enforced by
-                # every jinja template (see ggml-org/llama.cpp#22043),
-                # so this client-side cap is the only guarantee.
+                # parallel_tool_calls=False: client-side cap to 1
+                # (llama.cpp flag isn't enforced by every jinja template,
+                # ggml-org/llama.cpp#22043).
                 if parallel_tool_calls is False and tool_calls:
                     tool_calls = tool_calls[:1]
 
@@ -5414,8 +5408,8 @@ class LlamaCppBackend:
                 _cleaned = [s for s in stop if isinstance(s, str) and s]
                 if _cleaned:
                     stream_payload["stop"] = _cleaned
-        # Match the per-iteration tool loop above so sampling behavior
-        # stays consistent when the cap-exhausted final-answer pass runs.
+        # Match per-iteration tool-loop sampling for the cap-exhausted
+        # final-answer pass.
         if frequency_penalty is not None:
             stream_payload["frequency_penalty"] = frequency_penalty
         if seed is not None:
diff --git a/studio/backend/core/inference/providers.py b/studio/backend/core/inference/providers.py
index 777d11b702..ed0047d8da 100644
--- a/studio/backend/core/inference/providers.py
+++ b/studio/backend/core/inference/providers.py
@@ -137,19 +137,17 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
             r"gemini-pro-latest|gemini-flash-latest|gemini-flash-lite-latest"
             r")$"
         ),
-        # Gemini's OpenAI-compatible layer inherits OpenAI's 4-stop cap
-        # (https://ai.google.dev/gemini-api/docs/openai). Without the
-        # explicit cap the default 16 leaks through and the upstream
-        # silently drops the overflow.
+        # Gemini OAI-compat inherits OpenAI's 4-stop cap; default 16
+        # silently truncates upstream.
+        # https://ai.google.dev/gemini-api/docs/openai
         "stop_max": 4,
     },
     "deepseek": {
         "display_name": "DeepSeek",
         "base_url": "https://api.deepseek.com/v1",
-        # Legacy aliases (deepseek-chat / deepseek-reasoner) retire
-        # 2026-07-24 per https://api-docs.deepseek.com/updates. Surface
-        # the new canonical ids (deepseek-v4-flash / deepseek-v4-pro)
-        # alongside so the picker keeps working on cutover.
+        # deepseek-chat / deepseek-reasoner retire 2026-07-24; list
+        # v4-pro / v4-flash alongside for cutover.
+        # https://api-docs.deepseek.com/updates
         "default_models": [
             "deepseek-v4-pro",
             "deepseek-v4-flash",
@@ -217,13 +215,9 @@ PROVIDER_REGISTRY: dict[str, dict[str, Any]] = {
         "auth_prefix": "Bearer ",
         "notes": "Moonshot API key. China: use base URL https://api.moonshot.cn/v1",
         "model_id_allowlist": re.compile(r"^kimi-k2\.[56]$"),
-        # Both k2.6 and k2.5 are reasoning-class. The API rejects
-        # custom sampling ("invalid temperature: only 1 is allowed for
-        # this model", same for top_p). frequency_penalty follows the
-        # same lock on those models. seed and parallel_tool_calls are
-        # not in Kimi's documented chat schema; strip them too so a
-        # stale client or direct API caller cannot smuggle them onto
-        # the wire and 400 the request.
+        # k2.5/k2.6 are reasoning-class: API locks temperature=1, top_p=1,
+        # frequency_penalty; seed and parallel_tool_calls are undocumented
+        # and 400.
         "body_omit": (
             "temperature",
             "top_p",
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 04daba30d0..e16a074a85 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -876,10 +876,8 @@ class ChatCompletionRequest(BaseModel):
     seed: Optional[int] = Field(
         None,
         description = (
-            "Best-effort determinism seed. Forwarded to OpenAI Chat "
-            "Completions and OpenAI-compatible local backends. The "
-            "Responses family rejects it server-side and Anthropic does "
-            "not implement it, so it is silently dropped on those routes."
+            "Best-effort determinism seed. Forwarded to OpenAI Chat and "
+            "OAI-compat local backends; dropped on Anthropic and OpenAI Responses."
         ),
     )
     service_tier: Optional[
@@ -887,22 +885,18 @@ class ChatCompletionRequest(BaseModel):
     ] = Field(
         None,
         description = (
-            "Provider service tier. Anthropic accepts only `auto` and "
-            "`standard_only`; OpenAI Chat accepts "
-            "`auto|default|flex|priority|scale`; OpenAI Responses accepts "
-            "`auto|default|flex|priority`. Unsupported values per provider "
-            "are dropped in the per-provider stream helper instead of "
-            "422'ing here so a stale frontend never breaks a request."
+            "Provider service tier. Anthropic: auto|standard_only. "
+            "OpenAI Chat: auto|default|flex|priority|scale. "
+            "OpenAI Responses: auto|default|flex|priority. "
+            "Unsupported values are dropped per provider rather than 422'd here."
         ),
     )
     parallel_tool_calls: Optional[bool] = Field(
         None,
         description = (
-            "Whether the provider may dispatch tool calls in parallel. "
-            "OpenAI: forwarded as `parallel_tool_calls`. Anthropic: "
-            "inverted into `disable_parallel_tool_use` on the Messages "
-            "body. Default `None` preserves each provider's upstream "
-            "default (which is `true` everywhere today)."
+            "Allow parallel tool calls. Forwarded as `parallel_tool_calls` "
+            "on OpenAI; inverted to `disable_parallel_tool_use` on Anthropic. "
+            "None preserves upstream default (currently true everywhere)."
         ),
     )
     typical_p: Optional[float] = Field(
@@ -958,8 +952,8 @@ class ChatCompletionRequest(BaseModel):
         None,
         ge = 0.0,
         description = (
-            "llama.cpp DRY multiplier. 0 disables the 4-field chain "
-            "(dry_base / dry_allowed_length / dry_penalty_last_n). Local only."
+            "llama.cpp DRY multiplier. 0 disables the dry_base / "
+            "dry_allowed_length / dry_penalty_last_n chain. Local only."
         ),
     )
     dry_base: Optional[float] = Field(
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 3dfaf398da..cc0e236b49 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -240,12 +240,9 @@ studio_router = APIRouter()
 
 
 def _clean_local_stop_list(stop) -> Optional[list[str]]:
-    """Strip empty / non-string entries from a stop sequence input.
-
-    Mirrors `_normalize_stop_for_provider` (external_provider.py) so
-    local llama-server callers cannot ship `stop=["", "END"]` and
-    get a 400. Returns `None` when nothing survives, so the caller
-    can omit the field entirely.
+    """Strip empty/non-string stop entries; returns None when empty so
+    callers can omit. Mirrors `_normalize_stop_for_provider` so
+    `stop=["", "END"]` cannot 400 llama-server.
     """
     if isinstance(stop, str):
         return [stop] if stop else None
@@ -4185,9 +4182,7 @@ def _build_chat_request(
         chat_kwargs["top_p"] = payload.top_p
     if payload.max_output_tokens is not None:
         chat_kwargs["max_tokens"] = payload.max_output_tokens
-    # parallel_tool_calls is first-class on ChatCompletionRequest and
-    # the OpenAI-compat passthrough builder forwards it. Translate it
-    # so a Responses API caller's preference reaches llama-server.
+    # Forward parallel_tool_calls from Responses caller through to llama-server.
     if payload.parallel_tool_calls is not None:
         chat_kwargs["parallel_tool_calls"] = payload.parallel_tool_calls
 
@@ -4250,9 +4245,8 @@ async def _responses_non_streaming(
         msg = choices[0].get("message", {}) or {}
         text = msg.get("content", "") or ""
         tool_calls = msg.get("tool_calls") or []
-    # Match the cap applied on GGUF / Anthropic / safetensors tool
-    # paths: when the caller opted out of parallel tool calls, surface
-    # at most one. llama.cpp may not enforce the flag.
+    # parallel_tool_calls=False -> cap to 1 (llama.cpp flag isn't enforced;
+    # mirrors GGUF/Anthropic/safetensors paths).
     if payload.parallel_tool_calls is False and tool_calls:
         tool_calls = tool_calls[:1]
 
@@ -4376,10 +4370,9 @@ async def _responses_stream(
         tool_call_state: dict[int, dict] = {}
         # Text message lives at output_index 0; tool calls claim 1, 2, ...
         next_output_index = 1
-        # When the caller opted out of parallel tool calls, latch the
-        # first index we see and drop subsequent siblings — mirrors the
-        # GGUF agentic-loop / Anthropic-passthrough caps; llama.cpp may
-        # not enforce the upstream flag (ggml-org/llama.cpp#22043).
+        # parallel_tool_calls=False: latch the first tc index, drop the
+        # rest; llama.cpp flag isn't enforced by every jinja template
+        # (ggml-org/llama.cpp#22043).
         serial_tool_calls = payload.parallel_tool_calls is False
         first_serial_idx: Optional[int] = None
 
@@ -4872,10 +4865,9 @@ async def anthropic_messages(
     if openai_tool_choice is None:
         openai_tool_choice = "auto"
 
-    # Anthropic nests `disable_parallel_tool_use` inside `tool_choice`
-    # (https://docs.claude.com/en/docs/agents-and-tools/tool-use/implement-tool-use).
-    # Flip it into the OpenAI-shaped `parallel_tool_calls` toggle so the
-    # local GGUF tool loop respects clients that opt out of parallel calls.
+    # Anthropic nests `disable_parallel_tool_use` under `tool_choice`;
+    # flip to OAI `parallel_tool_calls` so the local GGUF tool loop honors it.
+    # https://docs.claude.com/en/docs/agents-and-tools/tool-use/implement-tool-use
     anthropic_parallel_tool_calls: Optional[bool] = None
     if isinstance(payload.tool_choice, dict):
         _disable = payload.tool_choice.get("disable_parallel_tool_use")
@@ -5393,10 +5385,8 @@ def _build_passthrough_payload(
         else (backend_ctx or _DEFAULT_MAX_TOKENS_FLOOR)
     )
     body["t_max_predict_ms"] = _DEFAULT_T_MAX_PREDICT_MS
-    # Strip empty / non-string stop entries before forwarding to
-    # llama-server; the external-provider helper does this via
-    # `_normalize_stop_for_provider`, and the local path needs the same
-    # defensive shape so a stale `stop=["", "END"]` cannot 400 upstream.
+    # Strip empty stop entries (mirrors `_normalize_stop_for_provider`);
+    # stale `stop=["", "END"]` would 400 llama-server.
     if stop:
         if isinstance(stop, str):
             if stop:
@@ -5412,12 +5402,9 @@ def _build_passthrough_payload(
         body["repeat_penalty"] = repetition_penalty
     if presence_penalty is not None:
         body["presence_penalty"] = presence_penalty
-    # llama-server's /v1/chat/completions accepts the standard OpenAI
-    # fields. parallel_tool_calls is a no-op on llama-server today but
-    # forwarded so a future release picks it up automatically.
-    # Each field below gated `is not None` so explicit 0 / False reach
-    # the wire; llama-server silently ignores unknown fields, Ollama's
-    # OAI translator drops everything outside the OAI subset.
+    # parallel_tool_calls is a no-op on llama-server today but forwarded
+    # for future support. `is not None` gate lets explicit 0/False through;
+    # llama-server ignores unknowns, Ollama drops non-OAI fields.
     if frequency_penalty is not None:
         body["frequency_penalty"] = frequency_penalty
     if seed is not None:
@@ -5705,11 +5692,8 @@ async def _anthropic_passthrough_non_streaming(
             content_blocks.append(AnthropicResponseTextBlock(text = text))
 
     tool_calls = message.get("tool_calls") or []
-    # Mirror the GGUF agentic-loop client-side cap: when the caller
-    # opted out of parallel tool calls, surface at most one tool_use
-    # block even if llama-server returned more than one. llama.cpp may
-    # not enforce the flag on every jinja template (see
-    # ggml-org/llama.cpp#22043).
+    # parallel_tool_calls=False: cap to 1 tool_use block; llama.cpp flag
+    # isn't enforced by every jinja template (ggml-org/llama.cpp#22043).
     if parallel_tool_calls is False and tool_calls:
         tool_calls = tool_calls[:1]
     for tc in tool_calls:
diff --git a/studio/frontend/src/components/ui/stop-sequences-input.tsx b/studio/frontend/src/components/ui/stop-sequences-input.tsx
index 539571313a..e9adbc894f 100644
--- a/studio/frontend/src/components/ui/stop-sequences-input.tsx
+++ b/studio/frontend/src/components/ui/stop-sequences-input.tsx
@@ -7,14 +7,8 @@ import { cn } from "@/lib/utils";
 import { XIcon } from "lucide-react";
 import { type KeyboardEvent, useState } from "react";
 
-/**
- * Chips editor for the `stop` / `stop_sequences` array.
- *
- * Commit a chip with Enter or comma; press Backspace on an empty input
- * to delete the most recent chip. Caps at `maxEntries` -- OpenAI Chat
- * Completions documents a hard cap of 4 stop sequences; Anthropic
- * Messages accepts arbitrarily many. Pass `Infinity` to disable the cap.
- */
+/** Chips editor for stop/stop_sequences. Enter or comma commits; Backspace
+ *  on empty deletes the last. OpenAI Chat caps at 4; pass Infinity to disable. */
 export interface StopSequencesInputProps {
   value: string[];
   onChange: (next: string[]) => void;
@@ -38,12 +32,9 @@ export function StopSequencesInput({
   const atCap = value.length >= maxEntries;
 
   function commitDraft() {
-    // Reject only the empty draft; preserve whitespace exactly (stop
-    // matching is byte-exact). OpenAI-compat / llama-server backends
-    // accept whitespace-only stops like "\n\n" for blank-line halts;
-    // pasting such a value into the input should round-trip rather
-    // than be silently dropped. Anthropic's helper strips whitespace
-    // entries on the wire so a chip that's invalid there cannot 400.
+    // Preserve whitespace exactly (stops are byte-exact); llama-server
+    // accepts "\n\n" for blank-line halts. Anthropic strips whitespace
+    // entries on the wire.
     if (!draft) return;
     if (atCap) return;
     if (value.includes(draft)) {
@@ -85,9 +76,7 @@ export function StopSequencesInput({
     >
       {value.map((entry, index) => (
         <Badge
-          // Stop sequences are not guaranteed unique across edits (the
-          // user could enter "END" twice if the previous one was just
-          // deleted), so combine value + index for a stable key.
+          // Not unique across edits (re-typed value); combine value+index for a stable key.
           key={`${entry}-${index}`}
           variant="secondary"
           className="gap-1 pl-2 pr-1"
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index 8d16d14204..5ec6cf40a9 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -1991,10 +1991,8 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               ...(externalCapabilities?.presencePenalty
                 ? { presence_penalty: params.presencePenalty }
                 : {}),
-              // Optional sampling extensions. Each gate is per-provider
-              // (see provider-capabilities.ts); the backend additionally
-              // drops fields the upstream API does not accept, so a
-              // stale frontend cannot 400 the request.
+              // Optional sampling extras. Per-provider gates live in
+              // provider-capabilities.ts; backend drops unknown fields.
               ...(externalCapabilities?.frequencyPenalty
                 ? { frequency_penalty: params.frequencyPenalty }
                 : {}),
@@ -2007,16 +2005,14 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               ...(externalCapabilities?.serviceTier && params.serviceTier
                 ? { service_tier: params.serviceTier }
                 : {}),
-              // Forward parallel_tool_calls when the user explicitly
-              // turned it off (the upstream default is `true` on every
-              // provider we ship, so default true is a no-op).
+              // Upstream default is true on every shipped provider;
+              // forward only on explicit opt-out.
               ...(externalCapabilities?.parallelToolCalls &&
               params.parallelToolCalls === false
                 ? { parallel_tool_calls: false }
                 : {}),
-              // llama.cpp / vLLM / OpenRouter extras. Each is gated by
-              // (a) the active provider's capability flag and (b) a
-              // non-default value, so only meaningful knobs hit the wire.
+              // llama.cpp / vLLM / OpenRouter extras: cap-flag plus
+              // non-default value gates so only meaningful knobs hit wire.
               ...(externalCapabilities?.typicalP &&
               params.typicalP !== null &&
               params.typicalP !== 1
@@ -2244,10 +2240,9 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
             min_p: params.minP,
             repetition_penalty: params.repetitionPenalty,
             presence_penalty: params.presencePenalty,
-            // llama-server accepts the standard OAI extensions via
-            // _build_passthrough_payload and silently ignores unknown
-            // fields. parallel_tool_calls defaults to false upstream so
-            // we forward unconditionally to honour the default-on UI.
+            // llama-server (_build_passthrough_payload) accepts OAI
+            // extras and ignores unknown; parallel_tool_calls default
+            // false upstream so forward unconditionally.
             ...(params.frequencyPenalty !== 0
               ? { frequency_penalty: params.frequencyPenalty }
               : {}),
@@ -2312,9 +2307,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
             ...(params.minTokens !== null && params.minTokens > 0
               ? { min_tokens: params.minTokens }
               : {}),
-            // Forward only when value diverges from upstream default;
-            // per-backend capability gating decides whether the wire
-            // even sees these.
+            // Forward only on non-default; per-backend cap-gates wire visibility.
             ...(params.skipSpecialTokens === false
               ? { skip_special_tokens: false }
               : {}),
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index 90246ed901..2abc423ae5 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -413,15 +413,11 @@ export function ChatSettingsPanel({
 }: ChatSettingsPanelProps) {
   const isMobile = useIsMobile();
   const isGguf = useChatRuntimeStore((s) => s.activeGgufVariant) != null;
-  // For non-external (local) models we show every knob — providerCapabilities
-  // is only consulted when `isExternalModel` is true. An external model with an
-  // unknown provider falls back to the OpenAI-compat shape via
-  // getProviderCapabilities, so these flags never undercount support.
-  // GGUF (llama-server) honours frequency_penalty/seed/stop/parallel_tool_calls;
-  // the safetensors / HF transformers path does not, so we hide those toggles
-  // on local non-GGUF backends to keep the UI honest. Anything still set in
-  // params from a prior GGUF session is harmlessly ignored by the safetensors
-  // worker, so this is purely a presentation gate.
+  // Local models show every knob (providerCapabilities only consulted
+  // when isExternalModel; unknown providers fall back to OPENAI_COMPAT_BASE).
+  // GGUF llama-server honours frequency_penalty/seed/stop/parallel_tool_calls;
+  // HF transformers path doesn't, so hide on local non-GGUF (stale params
+  // are harmlessly ignored by the safetensors worker).
   const localSamplerSupportsExtras = !isExternalModel ? isGguf : true;
   const showTemperature =
     !isExternalModel || Boolean(providerCapabilities?.temperature);
@@ -446,9 +442,7 @@ export function ChatSettingsPanel({
   const showParallelToolCalls = isExternalModel
     ? Boolean(providerCapabilities?.parallelToolCalls)
     : localSamplerSupportsExtras;
-  // Extended llama.cpp / vLLM / OpenRouter samplers. Same gate as the
-  // core knobs above: external → providerCapabilities flag; local →
-  // GGUF-only (safetensors transformers ignores these).
+  // Extended samplers: external uses cap flag, local is GGUF-only.
   const capAdv = (k: keyof ProviderCapabilities): boolean =>
     isExternalModel
       ? Boolean(providerCapabilities?.[k])
@@ -484,9 +478,8 @@ export function ChatSettingsPanel({
     postSamplingProbs: capAdv("postSamplingProbs"),
   };
   const showAdvancedSamplingSection = Object.values(advCaps).some(Boolean);
-  // Per-provider stop cap from provider-capabilities.ts; backend
-  // re-trims on the wire if a stale UI sends more than the upstream
-  // accepts.
+  // Per-provider stop cap; backend re-trims on the wire if a stale
+  // UI sends more than the upstream accepts.
   const stopMaxEntries = getProviderStopMax(externalProviderType);
   const serviceTierOptions = getServiceTierOptions(externalProviderType);
   const hasModelContent =
@@ -1379,10 +1372,9 @@ export function ChatSettingsPanel({
                     Seed
                   </span>
                   <InfoHint>
-                    Best-effort determinism seed. Same seed + prompt =
-                    same output (approximately). OpenAI Chat Completions
-                    and OpenAI-compat local backends honor it; OpenAI
-                    Responses and Anthropic silently drop it.
+                    Best-effort determinism. OpenAI Chat and OAI-compat
+                    local backends honor it; OpenAI Responses and Anthropic
+                    silently drop it.
                   </InfoHint>
                 </div>
                 <Input
@@ -1413,11 +1405,9 @@ export function ChatSettingsPanel({
                     Stop sequences
                   </span>
                   <InfoHint>
-                    Strings that halt generation as soon as the model
-                    emits them. Enter a value and press Enter or comma
-                    to commit a chip. Backend translates to
-                    `stop_sequences` on Anthropic and `stop` on OpenAI
-                    Chat (capped at 4 entries).
+                    Strings that halt generation. Enter or comma to commit.
+                    Maps to `stop_sequences` (Anthropic) / `stop` (OpenAI,
+                    cap 4).
                   </InfoHint>
                 </div>
                 <StopSequencesInput
@@ -1435,22 +1425,16 @@ export function ChatSettingsPanel({
                     Service tier
                   </span>
                   <InfoHint>
-                    Provider routing tier. `auto` (default) lets the
-                    provider choose. `flex` / `priority` / `scale` route
-                    to higher-latency-tolerant or premium queues on
-                    OpenAI; `standard_only` opts out of Anthropic's
-                    priority tier.
+                    Provider routing tier. `auto` = provider default.
+                    OpenAI: flex / priority / scale. Anthropic:
+                    `standard_only` opts out of Priority Tier.
                   </InfoHint>
                 </div>
                 <Select
                   value={params.serviceTier ?? "auto"}
                   onValueChange={(value) => {
-                    // Store every selected tier verbatim, including the
-                    // explicit "auto". On Anthropic the docs distinguish
-                    // omitting service_tier (provider default) from
-                    // setting "auto" (opt into Priority Tier when
-                    // available); preserve the user's choice through to
-                    // the adapter so the wire reflects it.
+                    // Store "auto" verbatim: Anthropic distinguishes
+                    // omitted (provider default) from auto (Priority Tier opt-in).
                     const allowed: readonly ServiceTier[] = serviceTierOptions;
                     if (allowed.includes(value as ServiceTier)) {
                       set("serviceTier")(value as ServiceTier);
@@ -1480,11 +1464,8 @@ export function ChatSettingsPanel({
                     Parallel tool calls
                   </span>
                   <InfoHint>
-                    When on, the model may dispatch multiple tool calls
-                    in a single turn (default). Turn off to force one
-                    tool call at a time. Anthropic implements this as
-                    `disable_parallel_tool_use`; OpenAI as
-                    `parallel_tool_calls`.
+                    Allow multiple tool calls per turn (default).
+                    Anthropic uses inverse `disable_parallel_tool_use`.
                   </InfoHint>
                 </div>
                 <Switch
@@ -1623,7 +1604,7 @@ export function ChatSettingsPanel({
                   max={5}
                   step={0.1}
                   onChange={(v) => set("dynatempExponent")(v)}
-                  info="llama.cpp `dynatemp_exponent`. Curve exponent for dynamic temperature. Paired with Dynatemp Range."
+                  info="llama.cpp `dynatemp_exponent`. Curve exponent, pairs with Dynatemp Range."
                 />
               ) : null}
               {advCaps.mirostat ? (
@@ -1699,7 +1680,7 @@ export function ChatSettingsPanel({
                       ? "Off"
                       : undefined
                   }
-                  info="llama.cpp DRY sampler. Master switch for the 4-field DRY chain (base / allowed length / penalty last N). 0 = off."
+                  info="llama.cpp DRY master switch (unlocks base / allowed length / penalty last N). 0 = off."
                 />
               ) : null}
               {advCaps.dryBase && (params.dryMultiplier ?? 0) > 0 ? (
@@ -1892,9 +1873,8 @@ export function ChatSettingsPanel({
                       Skip Special Tokens
                     </span>
                     <InfoHint>
-                      vLLM `skip_special_tokens`. Default on. Turn off to keep
-                      chat-template markers (e.g. `&lt;|im_end|&gt;`) in the
-                      decoded output.
+                      vLLM `skip_special_tokens` (default on). Off keeps
+                      chat-template markers like `&lt;|im_end|&gt;` in output.
                     </InfoHint>
                   </div>
                   <Switch
diff --git a/studio/frontend/src/features/chat/provider-capabilities.ts b/studio/frontend/src/features/chat/provider-capabilities.ts
index 2c1f35e197..0707ee6895 100644
--- a/studio/frontend/src/features/chat/provider-capabilities.ts
+++ b/studio/frontend/src/features/chat/provider-capabilities.ts
@@ -1,13 +1,11 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-// Per-provider sampling capability matrix. Sourced from each
-// provider's chat-completion docs (2026-05). The panel hides params
-// the active provider does not accept so users never move a knob that
-// would be silently dropped or rejected.
-// When adding a new knob: default it to false on every SaaS bucket;
-// only local backends + the permissive openrouter bucket should
-// expose llama.cpp-specific samplers.
+// Per-provider sampling capability matrix sourced from each provider's
+// chat-completion docs (2026-05); panel hides params the active
+// provider would silently drop or reject.
+// New knobs: default to false on every SaaS bucket; only local +
+// openrouter expose llama.cpp samplers.
 export interface ProviderCapabilities {
   /** OpenAI gpt-5.x / o-series reject via /v1/responses. */
   temperature: boolean;
@@ -76,17 +74,12 @@ export interface ProviderCapabilities {
 }
 
 /**
- * Per-provider stop-sequence max count. Resolved by
- * `getProviderStopMax(providerType)`. Mirrors the backend's
- * `provider_info.stop_max` for the same provider type.
- *   - openai:     4   (Chat Completions hard cap; Responses drops stop)
- *   - anthropic:  16  (client-side guard; docs publish no max)
- *   - kimi:       5   (https://platform.kimi.ai/docs/api/chat)
- *   - deepseek:   16  (https://api-docs.deepseek.com/api/create-chat-completion)
- *   - mistral:    16  (no documented max; widen to permissive default)
- *   - gemini:     4   (https://ai.google.dev/gemini-api/docs/openai inherits OAI cap)
- *   - openrouter: 4   (normalises to OpenAI's chat schema)
- *   - default:    16  (covers ollama, vllm, llama.cpp, custom)
+ * Per-provider stop-sequence max count. Mirrors backend `stop_max`.
+ *   openai 4 (Chat hard cap; Responses drops stop)
+ *   anthropic 16 (client-side guard, docs no max)
+ *   kimi 5 (https://platform.kimi.ai/docs/api/chat)
+ *   deepseek 16 (https://api-docs.deepseek.com/api/create-chat-completion)
+ *   mistral 16, gemini 4, openrouter 4, default 16 (ollama/vllm/llama.cpp/custom)
  */
 const PROVIDER_STOP_MAX: Record<string, number> = {
   openai: 4,
@@ -114,13 +107,9 @@ export type ServiceTierOption =
   | "standard_only";
 
 /**
- * Legal `service_tier` values per provider. Anthropic exposes only
- * `auto` and `standard_only`. OpenAI in Studio is routed through
- * `/v1/responses`, which the live docs list as
- * `auto|default|flex|priority`; `scale` is excluded here even though
- * the openai-python SDK type happens to include it. Other providers
- * fall through to a permissive `auto` / `default` pair so the picker
- * stays usable for OpenAI-compat backends.
+ * Legal `service_tier` per provider. anthropic=auto|standard_only;
+ * openai (/v1/responses)=auto|default|flex|priority (scale excluded
+ * though SDK lists it); others fall through to auto|default.
  */
 export function getServiceTierOptions(
   providerType: string | null | undefined,
@@ -179,13 +168,12 @@ export function clampReasoningEffortToLevels(
  */
 export const EXTERNAL_MAX_OUTPUT_TOKENS = 32768;
 
-// Per-model max-output caps from each provider's docs (verified May 2026):
+// Per-model max-output caps (verified May 2026). Longer prefixes first
+// so .startsWith() picks the specific id over the family root.
 //   OpenAI:    developers.openai.com/api/docs/models/<model>
 //   Anthropic: platform.claude.com/docs/en/about-claude/models/overview
 //   Gemini:    ai.google.dev/gemini-api/docs/models
 //   DeepSeek:  api-docs.deepseek.com/quick_start/pricing
-// Order matters: list specific chat-class ids before broader gpt-5 /
-// claude-opus-4 entries so the longer prefix wins via .startsWith().
 const EXTERNAL_MAX_OUTPUT_TOKENS_BY_MODEL: Array<{
   providerType: string;
   prefixes: readonly string[];
@@ -262,16 +250,13 @@ function _inferProviderFromOpenrouterId(
   return null;
 }
 
-// Gates the composer's Search button. Backend translates
-// enable_tools:["web_search"] into each provider's tool schema:
-//   OpenAI:     tools:[{type:"web_search"}] on /v1/responses
-//   Anthropic:  tools:[{type:"web_search_20250305", max_uses:5}] on /v1/messages
-//   OpenRouter: plugins:[{id:"web"}] (router's universal shape)
-//   Kimi:       $web_search builtin (two-call round trip via
-//               _stream_kimi_web_search)
-// Mistral excluded: their web_search is on the Agents API, not chat
-// completions, and 400s if injected. Gemini grounded-search needs
-// matching backend translation first.
+// Gates the composer's Search button. Backend maps
+// enable_tools:["web_search"] to each provider's tool schema:
+//   OpenAI: tools:[{type:"web_search"}] on /v1/responses
+//   Anthropic: tools:[{type:"web_search_20250305", max_uses:5}]
+//   OpenRouter: plugins:[{id:"web"}]
+//   Kimi: $web_search builtin (2-call via _stream_kimi_web_search)
+// Mistral excluded (web_search lives on Agents API, 400s on /v1/chat).
 export function providerSupportsBuiltinWebSearch(
   providerType: string | null | undefined,
   modelId?: string | null | undefined,
@@ -418,9 +403,8 @@ export function providerSupportsBuiltinCodeExecution(
   return false;
 }
 
-// OpenAI Responses-API image_generation tool. OpenAI cloud +
-// Responses-family ids only; backend mirrors via is_openai_cloud so
-// custom OAI-compat servers reporting provider_type="openai" don't 400.
+// OpenAI Responses-API image_generation. OpenAI cloud +
+// Responses-family ids only; backend mirrors via is_openai_cloud.
 const OPENAI_IMAGE_GENERATION_MODEL_PREFIXES = [
   "gpt-5.5-pro",
   "gpt-5.5",
@@ -510,10 +494,8 @@ function geminiImageModelAllowsGoogleSearch(modelId: string): boolean {
   );
 }
 
-// Per-provider min on outbound max_tokens. Kimi thinking models need
-// >=16000 or the response truncates mid-stream. Other providers fall
-// through to the generic 64. chat-adapter bumps the user's stored
-// maxTokens up to the floor on send; the slider min mirrors the same.
+// Per-provider min on outbound max_tokens. Kimi thinking needs >=16000
+// (truncates mid-stream below); chat-adapter bumps on send.
 const EXTERNAL_MIN_OUTPUT_TOKENS_BY_PROVIDER: Record<string, number> = {
   kimi: 16000,
 };
@@ -902,12 +884,9 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
     postSamplingProbs: false,
   },
   mistral: OPENAI_COMPAT_BASE,
-  // Gemini's native generationConfig accepts temperature, topP, topK,
-  // presencePenalty, frequencyPenalty (not surfaced today), seed and
-  // stopSequences. minP and repetitionPenalty are not part of the
-  // contract -- see https://ai.google.dev/api/rest/v1beta/GenerationConfig.
-  // Backend request shaping lives in _stream_gemini in
-  // studio/backend/core/inference/external_provider.py.
+  // Gemini generationConfig: temperature/topP/topK/presencePenalty/
+  // frequencyPenalty/seed/stopSequences. No minP/repetitionPenalty.
+  // https://ai.google.dev/api/rest/v1beta/GenerationConfig
   gemini: {
     temperature: true,
     topP: true,
@@ -1053,14 +1032,9 @@ const PROVIDER_CAPABILITIES: Record<string, ProviderCapabilities> = {
 
 const DEFAULT_EXTERNAL_CAPABILITIES = OPENAI_COMPAT_BASE;
 
-// Per-model specialisations:
-//   openai + chat-class (gpt-4o, gpt-4-turbo, gpt-4, gpt-3.5):
-//     full sampling surface (OPENAI_CHAT_CAPABILITIES).
-//   openai + reasoning (gpt-5.x, o1, o3, o4): OPENAI_REASONING_CAPABILITIES.
-//   anthropic + claude-opus-4-7: strips temp/top_p/top_k (Opus only
-//     in 4.7; Sonnet/Haiku don't ship).
-//   deepseek reasoner: hides temp/top_p (silently ignored upstream).
-// Returns null for local models (caller treats as "every knob applies").
+// Per-model overrides: openai+chat-class -> OPENAI_CHAT_CAPABILITIES;
+// anthropic claude-opus-4-7 strips temp/top_p/top_k; deepseek reasoner
+// hides temp/top_p. Local (no providerType) returns null = every knob.
 export function getProviderCapabilities(
   providerType: string | null | undefined,
   modelId?: string | null | undefined,
@@ -1114,12 +1088,9 @@ const NO_REASONING_CAPS: ReasoningCaps = {
   reasoningEffortLevels: DEFAULT_EFFORT_LEVELS,
 };
 
-// Order matters: longest prefixes first so find() picks the right
-// bucket before the bare-family fallback ("claude-opus-4") sweeps.
-// Levels per platform.claude.com/docs/en/about-claude/models/overview;
-// 4.5 line uses budget_tokens mapped by the backend. Legacy 4.x
-// (opus-4-1 / opus-4 / sonnet-4) supports Extended thinking per the
-// overview table; sonnet-4 / opus-4 retire 2026-06-15.
+// Longest prefixes first (find() must match before the bare-family
+// fallback). Levels per platform.claude.com/docs/en/about-claude/models/overview.
+// 4.5 line maps to budget_tokens; sonnet-4/opus-4 retire 2026-06-15.
 const ANTHROPIC_REASONING_MODELS = [
   {
     prefixes: ["claude-opus-4-7"],
@@ -1373,10 +1344,9 @@ function resolveGeminiReasoningCapabilities(
 }
 
 function resolveMistralReasoningCapabilities(modelId: string): ExternalReasoningCapabilities {
-  // magistral-* is native always-on (no reasoning_effort param; 422 if
-  // injected). mistral-{small,medium,vibe-cli}-latest is adjustable
-  // none/low/medium/high. See docs.mistral.ai/studio-api/conversations/
-  // reasoning + mistral.ai/news/magistral.
+  // magistral-*: native always-on (422 on reasoning_effort).
+  // mistral-{small,medium,vibe-cli}-latest: none/low/medium/high.
+  // https://docs.mistral.ai/studio-api/conversations/reasoning
   if (
     modelId === "magistral-medium-latest" ||
     modelId === "magistral-small-latest"
diff --git a/studio/frontend/src/features/chat/types/api.ts b/studio/frontend/src/features/chat/types/api.ts
index b3c254ce43..520acea25c 100644
--- a/studio/frontend/src/features/chat/types/api.ts
+++ b/studio/frontend/src/features/chat/types/api.ts
@@ -320,11 +320,7 @@ export interface OpenAIChatCompletionsRequest {
   seed?: number;
   /** OAI Chat caps at 4; Anthropic mapped to `stop_sequences`. */
   stop?: string[];
-  /**
-   * Per-provider enum (see getServiceTierOptions). Union stays
-   * permissive; external_provider.py drops values the active provider
-   * doesn't accept.
-   */
+  /** Per-provider enum (see getServiceTierOptions); external_provider.py drops unsupported values. */
   service_tier?:
     | "auto"
     | "default"
diff --git a/studio/frontend/src/features/chat/types/runtime.ts b/studio/frontend/src/features/chat/types/runtime.ts
index ccdddbbd37..0e88ac0e0c 100644
--- a/studio/frontend/src/features/chat/types/runtime.ts
+++ b/studio/frontend/src/features/chat/types/runtime.ts
@@ -9,11 +9,8 @@ export type ServiceTier =
   | "scale"
   | "standard_only";
 
-// All `number | null` / `boolean | null` fields below follow the same
-// convention: `null` = field omitted from the wire request (provider
-// uses its own default). Per-provider capability gating lives in
-// provider-capabilities.ts; the chat-adapter forwards only when the
-// active provider's bucket has the matching flag set true.
+// null = field omitted from wire (provider default).
+// Per-provider gates in provider-capabilities.ts.
 export interface InferenceParams {
   temperature: number;
   topP: number;
@@ -47,10 +44,7 @@ export interface InferenceParams {
   mirostatEta: number | null;
   /** OpenRouter `top_a`. Range [0, 1]. */
   topA: number | null;
-  /**
-   * llama.cpp DRY sampler — multiplier is the master switch (0 disables
-   * the 4-field chain). See llama.cpp/tools/server/README.md.
-   */
+  /** llama.cpp DRY: multiplier is master switch (0 disables 4-field chain). See llama.cpp/tools/server/README.md. */
   dryMultiplier: number | null;
   /** Default 1.75. */
   dryBase: number | null;
diff --git a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
index 7c31d2caa7..dd20d812f2 100644
--- a/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
+++ b/studio/frontend/src/features/chat/utils/chat-settings-storage.ts
@@ -45,12 +45,9 @@ const NUMERIC_INFERENCE_FIELDS = [
   "maxTokens",
 ] as const satisfies readonly (keyof PersistedInferenceParams)[];
 
-// `seed` is numeric but nullable (null = no seed on the wire) so it
-// can't go through the NUMERIC_INFERENCE_FIELDS finite-number filter.
-// Keep this set in sync with `ServiceTier` in ../types/runtime.ts and
-// `getServiceTierOptions` in ../provider-capabilities.ts. `scale` is
-// kept in the allowlist for forward-compat with legacy persisted data
-// even though the resolver no longer surfaces it for OpenAI Responses.
+// `seed` is nullable so it skips NUMERIC_INFERENCE_FIELDS' finite-number filter.
+// Keep in sync with ServiceTier (../types/runtime.ts) and getServiceTierOptions.
+// "scale" stays for forward-compat with legacy persisted data.
 const VALID_SERVICE_TIERS = new Set([
   "auto",
   "default",
@@ -162,15 +159,12 @@ function sanitizeInferenceParams(
   } else if (typeof value.seed === "number" && Number.isInteger(value.seed)) {
     params.seed = value.seed;
   }
-  // stop: cap at 16 (Anthropic's widest supported). The backend's
-  // per-provider stream helper re-truncates to the wire cap. An EMPTY
-  // array must persist (not be sanitized away) so clearing the last
-  // chip actually saves; otherwise the old stops come back on reload.
+  // stop: cap at 16 (Anthropic widest); backend re-truncates per provider.
+  // Empty array MUST persist or clearing the last chip is reverted on reload.
   if (Array.isArray(value.stop)) {
     const stops = value.stop.filter((s): s is string => typeof s === "string");
     params.stop = stops.slice(0, 16);
   }
-  // serviceTier: nullable enum string.
   if (value.serviceTier === null) {
     params.serviceTier = null;
   } else if (
@@ -182,8 +176,8 @@ function sanitizeInferenceParams(
   if (typeof value.parallelToolCalls === "boolean") {
     params.parallelToolCalls = value.parallelToolCalls;
   }
-  // typicalP: nullable float (null = no typ_p on the wire, matching
-  // llama-server's default 1.0). Mirror seed's nullable-float handling.
+  // typicalP: nullable float (null = no typ_p on wire, matches
+  // llama-server default 1.0). Mirrors seed handling.
   if (value.typicalP === null) {
     params.typicalP = null;
   } else if (
@@ -192,8 +186,7 @@ function sanitizeInferenceParams(
   ) {
     params.typicalP = value.typicalP;
   }
-  // New llama.cpp / OpenRouter samplers — all nullable numbers with
-  // the same handling as `typicalP` / `seed`.
+  // Nullable numeric samplers (same handling as typicalP/seed).
   for (const key of [
     "topNSigma",
     "repeatLastN",