diff --git a/docs-site/src/content/docs/fr/guides/claude-code.md b/docs-site/src/content/docs/fr/guides/claude-code.md
index 22f57ac1954..02b1641d58b 100644
--- a/docs-site/src/content/docs/fr/guides/claude-code.md
+++ b/docs-site/src/content/docs/fr/guides/claude-code.md
@@ -423,7 +423,7 @@ d'OpenCodex, de réappliquer ce profil et de démarrer une nouvelle conversation
 dépannage, sans garantie de résolution. OpenCodex ne peut pas observer l'état du sélecteur ; il achemine
 l'identifiant du modèle porté par chaque requête. Vérifiez ce que le client envoie sous **Logs → requestedModel**.
 
-Les modèles dont la fenêtre de contexte de référence atteint 1M obtiennent une ligne supplémentaire `…[1m]` dans le sélecteur.
+Les modèles dont la fenêtre de contexte de référence atteint 1M ou au moins le seuil de compactage par défaut (829 800) obtiennent une ligne supplémentaire `…[1m]` dans le sélecteur.
 Sa sélection indique à Claude Code la fenêtre complète de 1M pour ce modèle, tout en maintenant le compactage automatique ; le proxy retire
 le marqueur avant le routage.
 La sélection est conservée dans le champ `model` de `settings.json` ; pour les requêtes entrantes, l'alias est de nouveau
@@ -458,8 +458,10 @@ conservent leurs identifiants canoniques sur les deux interfaces.
 
 ### Marqueur `[1m]` de variante contextuelle
 
-Les modèles dont la fenêtre de contexte de référence atteint 1M — ou, avec le contexte automatique, dépasse 200k tout en atteignant
-au moins le seuil de compactage — obtiennent une ligne supplémentaire `…[1m]` dans le sélecteur. En la sélectionnant, Claude Code
+Les modèles dont la fenêtre de contexte de référence atteint 1M ou au moins le seuil de compactage par défaut (829 800)
+obtiennent une ligne supplémentaire `…[1m]` dans le sélecteur. Ce plancher est fixe : modifier la valeur de compactage ne l'abaisse pas,
+et un modèle Claude sur une route Anthropic doit réellement atteindre 1M. Sans contexte automatique, seule une fenêtre de 1M compte.
+Au-delà de la fenêtre réelle, l'erreur `prompt is too long` déclenche le compactage automatique de Claude Code. En la sélectionnant, Claude Code
 tient compte d'un contexte complet de 1M. Le proxy supprime le suffixe `[1m]`, sans tenir compte de la casse, avant la
 résolution de l'alias et le routage.
 
@@ -468,8 +470,9 @@ résolution de l'alias et le routage.
 Claude Code attribue une limite de 200k jetons à tout modèle qu'il ne reconnaît pas. Le **contexte automatique**, activé
 par défaut, corrige ce comportement :
 
-1. Les modèles dont la fenêtre réelle dépasse 200k **et** atteint au moins le seuil de compactage automatique obtiennent le
-   marqueur `[1m]` dans les lignes du sélecteur et les variables d'environnement qui les désignent.
+1. Les variables d'environnement de lancement reçoivent le marqueur `[1m]` lorsque la fenêtre réelle dépasse 200k **et**
+   atteint au moins le seuil de compactage automatique configuré. Les lignes de découverte et du sélecteur Desktop ne suivent pas
+   ce seuil : elles utilisent le plancher fixe de 829 800 jetons (1M réel pour un modèle Claude sur une route Anthropic).
 2. `CLAUDE_CODE_AUTO_COMPACT_WINDOW` (`829800` par défaut, plage `100000`–`1000000`) est injecté afin
    que la conversation soit automatiquement résumée à ce seuil.
 
@@ -482,8 +485,9 @@ Trois états de configuration :
 La valeur de compactage est réglable dans **Connexion → Claude**. **Avertissement :** une valeur supérieure à la fenêtre réelle d'un modèle
 rend ce modèle inutilisable : les tours échouent avant que le résumé puisse se déclencher.
 
-Les modèles Anthropic natifs dont le contexte est inférieur à 1M ne sont jamais marqués automatiquement. Les valeurs que vous exportez vous-même
-restent prioritaires ; le proxy s'appuie sur votre valeur pour déterminer les modèles qui peuvent recevoir le marqueur sans risque.
+Les modèles Anthropic natifs dont le contexte est inférieur à 1M ne sont jamais marqués automatiquement. Une fenêtre de compactage que vous exportez vous-même
+reste prioritaire pour le marquage des variables de lancement ; le proxy s'appuie sur votre valeur pour déterminer les modèles qui peuvent recevoir le marqueur sans risque.
+Les lignes de découverte ignorent cet export et conservent le plancher fixe.
 Les valeurs de configuration invalides définies manuellement reviennent à 829,800.
 
 ### Environnement effectif des modèles
@@ -744,7 +748,7 @@ une valeur `ANTHROPIC_BASE_URL` obsolète. Ouvrez un nouveau terminal ou réexé
 
 **Plafond de contexte 200k malgré un grand modèle** — Sélectionnez la variante `[1m]` dans le sélecteur ou activez
 le contexte automatique, activé par défaut. Si le sélecteur n'affiche aucune ligne `[1m]`, la fenêtre de contexte de référence du modèle
-peut être inférieure au seuil de compactage automatique.
+peut être inférieure au plancher fixe de 829 800 jetons.
 
 **Nombre élevé de jetons provenant du chargement des compétences** — La compétence `claude-api` fournie (~136k jetons) se charge automatiquement
 quand un modèle Claude est mentionné. Ce comportement est normal avec le transfert natif ; pour les modèles routés, opencodex la remplace
@@ -768,7 +772,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/docs-site/src/content/docs/guides/claude-code.md b/docs-site/src/content/docs/guides/claude-code.md
index ecdc703d3ce..9922dde3fa5 100644
--- a/docs-site/src/content/docs/guides/claude-code.md
+++ b/docs-site/src/content/docs/guides/claude-code.md
@@ -596,7 +596,8 @@ reapplying the profile, and starting a new conversation. This is a troubleshooti
 guaranteed fix. OpenCodex cannot observe picker state; it routes the model id carried by each
 request. Confirm what the client sends under **Logs → requestedModel**.
 
-Models with an authoritative 1M context window get an extra `…[1m]` picker row: selecting it makes
+Models with a long context window (1M, or at least the 829,800-token default compaction threshold;
+Claude models on an Anthropic route need a genuine 1M) get an extra `…[1m]` picker row: selecting it makes
 Claude Code account a full 1M context for that model (auto-compaction stays on) — the proxy strips
 the marker before routing.
 Selecting one persists it to Claude Code's `settings.json` `model` field; inbound requests resolve
@@ -632,8 +633,10 @@ keep their canonical ids on both surfaces.
 
 ### Context-variant `[1m]` marker
 
-Models with an authoritative context window of 1M (or, under auto-context, above 200k and at
-least the compaction threshold) get an extra `…[1m]` picker row. Selecting it makes Claude Code
+Models with an authoritative context window of 1M, or at least the 829,800-token default
+compaction threshold, get an extra `…[1m]` picker row; the floor is fixed, so a custom compaction
+value does not lower it, and Claude models on an Anthropic route need a genuine 1M. Turning
+auto-context off limits the row to 1M windows. Selecting it makes Claude Code
 account a full 1M context. The proxy strips the case-insensitive `[1m]` suffix before alias
 resolution and routing.
 
@@ -642,8 +645,10 @@ resolution and routing.
 Claude Code accounts 200k tokens for any model it does not recognize. **Auto context** (on by
 default) fixes that:
 
-1. Models whose real window is above 200k **and** at least the auto-compact threshold get the
-   `[1m]` marker on their picker rows and env slots.
+1. Launch env slots get the `[1m]` marker when the model's real window is above 200k **and** at
+   least the configured auto-compact threshold. Discovery and Desktop picker rows do not follow
+   that threshold: they use the fixed 829,800-token floor (a genuine 1M for Claude models on an
+   Anthropic route).
 2. `CLAUDE_CODE_AUTO_COMPACT_WINDOW` (default `829800`, range `100000`–`1000000`) is injected so
    the conversation auto-summarizes at that point.
 
@@ -668,8 +673,9 @@ history and try compacting a fork with fewer enabled tool or MCP schemas, if you
 that workflow. A `[1m]` marker or larger client accounting setting does not raise the upstream
 limit, and OpenCodex does not silently remove history or tools to make it fit.
 
-Sub-1M native Anthropic models are never auto-marked. Values you export yourself always win (the
-proxy uses YOUR value to decide which models are safe to mark). Invalid hand-edited config values
+Sub-1M native Anthropic models are never auto-marked. A compact window you export yourself wins
+for launch-slot marking (the proxy uses YOUR value to decide which models are safe to mark).
+Discovery rows ignore that export and keep the fixed floor. Invalid hand-edited config values
 fall back to 829,800.
 
 ### Effective model environment
@@ -1001,7 +1007,7 @@ set (automatic with `ocx claude`). Run `ocx claude` to refresh the gateway model
 
 **200k context ceiling despite big model** — Select the `[1m]` variant in the picker, or enable
 auto-context (on by default). If the picker shows no `[1m]` row, the model's authoritative context
-window may be below the auto-compact threshold.
+window may be below the fixed 829,800-token floor.
 
 **High token count from skill loads** — The bundled `claude-api` skill (~136k tokens) auto-loads
 on Claude model mentions. This is normal for native passthrough; on routed models, opencodex stubs
@@ -1059,7 +1065,7 @@ the confirmed obsolete token file and applying first-party mode again; never del
 
 ### First-party picker context markers
 
-The Desktop Code-tab picker adds `[1m]` to routed models whose authoritative context window is at least one million tokens, so Claude uses its 1M accounting instead of the smaller custom-model fallback. Labels, profile order, and provider routes stay unchanged. Unknown and sub-million windows remain unmarked, including native long-window opt-ins: the picker cannot guarantee that a Desktop or remote runner receives the matching compaction environment. The paired auto-context setup for `ocx claude` is unchanged. An existing conversation keeps its saved selector until you select the model again from the refreshed picker.
+The Desktop Code-tab picker adds `[1m]` to routed models whose authoritative context window is at least one million tokens or at least the 829,800-token default compaction threshold (for example a native GPT-6 model opted into its 872k window), so Claude uses its 1M accounting instead of the smaller custom-model fallback. Claude models on an Anthropic route still need a genuine 1M. Labels, profile order, and provider routes stay unchanged. Unknown and shorter windows remain unmarked. The picker cannot guarantee that a Desktop or remote runner receives the matching compaction environment, so a long-window model that outgrows its real window gets a `prompt is too long` error, which Claude Code answers by compacting. Desktop 3P `supports1m`/`prefer1m` and generated subagents follow the same rule. The paired auto-context setup for `ocx claude` is unchanged. An existing conversation keeps its saved selector until you select the model again from the refreshed picker.
 
 ### Forced Claude Code subagent model
 
@@ -1067,7 +1073,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/docs-site/src/content/docs/ja/guides/claude-code.md b/docs-site/src/content/docs/ja/guides/claude-code.md
index c6308ab24b8..c5690dc3a67 100644
--- a/docs-site/src/content/docs/ja/guides/claude-code.md
+++ b/docs-site/src/content/docs/ja/guides/claude-code.md
@@ -311,8 +311,9 @@ v2 エイリアスはエスケープを展開します。読みやすい形式
 
 ### コンテキスト変種 `[1m]` 標識
 
-公式コンテキストウィンドウが 1M のモデルには `…[1m]` ピッカー行がもう一つできます。自動コンテキスト使用時は
-コンテキストが 200k を超え、圧縮しきい値以上のモデルも該当します。この行を選ぶと Claude Code が
+公式コンテキストウィンドウが 1M、またはデフォルトの圧縮しきい値（829,800）以上のモデルには `…[1m]` ピッカー行がもう一つできます。
+この基準は固定で、圧縮値を変えても下がりません。Anthropic 経路の Claude モデルは実際に 1M である必要があります。自動コンテキストをオフにすると
+1M のモデルだけが対象です。実際のウィンドウを超えると `prompt is too long` エラーが返り、Claude Code が自動で圧縮します。この行を選ぶと Claude Code が
 全体 1M コンテキストを計算します。プロキシはエイリアス解決とルーティング前に大文字小文字を区別せず `[1m]`
 接尾辞を削除します。
 
@@ -321,8 +322,9 @@ v2 エイリアスはエスケープを展開します。読みやすい形式
 Claude Code は未知モデルのコンテキストを 200k トークンとして計算します。デフォルトでオンの**自動
 コンテキスト**がこの問題を解決します。
 
-1. 実際のコンテキストウィンドウが 200k より大きく自動圧縮しきい値以上のモデルのピッカー行と環境スロットに
-   `[1m]` 標識が付きます。
+1. 起動時の環境スロットには、実際のコンテキストウィンドウが 200k より大きく、設定した自動圧縮しきい値以上のとき
+   `[1m]` 標識が付きます。ディスカバリーと Desktop のピッカー行はこのしきい値に従わず、固定の 829,800 トークン下限
+   （Anthropic 経路の Claude モデルは実際の 1M）を使います。
 2. `CLAUDE_CODE_AUTO_COMPACT_WINDOW`(デフォルト `829800`、範囲 `100000`–`1000000`)を注入し、該当
    地点で会話を自動要約します。
 
@@ -335,7 +337,7 @@ Claude Code は未知モデルのコンテキストを 200k トークンとし
 **接続 → Claude** で圧縮値を調整できます。**警告:** モデルの実際のコンテキストウィンドウより大きく上げると
 要約を開始する前にチャットエラーが発生します。
 
-1M 未満のネイティブ Anthropic モデルには自動で標識を付けません。直接 export した値が常に優先し、プロキシは**ユーザーが指定した**値を基準にどのモデルに安全に標識を付けるか決定します。
+1M 未満のネイティブ Anthropic モデルには自動で標識を付けません。自分で export した圧縮値は起動スロットの標識に優先し、プロキシは**ユーザーが指定した**値を基準にどのモデルに安全に標識を付けるか決定します。ディスカバリー行はその export を無視し、固定の下限を保ちます。
 直接編集した設定値が不正な場合は 829,800 に戻ります。
 
 ### 実モデル環境
@@ -603,7 +605,7 @@ Anthropic バックエンドを明示すると意図的に失敗後停止しま
 
 **大型モデルなのにコンテキストが 200k に制限される** — ピッカーで `[1m]` 変種を選ぶか、デフォルトでオンの
 自動コンテキストを使ってください。ピッカーに `[1m]` 行がない場合はモデルの公式コンテキストウィンドウが
-自動圧縮しきい値より小さい可能性があります。
+固定の 829,800 トークン下限より小さい可能性があります。
 
 **スキル呼び出し時のトークン数が多い** — バンドル `claude-api` スキル(約 136k トークン)は Claude モデルに
 言及すると自動で読み込まれます。ネイティブパススルーでは正常で、ルーティングモデルでは opencodex が
@@ -623,7 +625,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/docs-site/src/content/docs/ko/guides/claude-code.md b/docs-site/src/content/docs/ko/guides/claude-code.md
index 1de6c56ba53..91de7d969bd 100644
--- a/docs-site/src/content/docs/ko/guides/claude-code.md
+++ b/docs-site/src/content/docs/ko/guides/claude-code.md
@@ -283,7 +283,7 @@ import/export는 로컬 설정만 다뤄요. 허브 프로필을 바꾸지 않
 능력 정보(추론 강도 사다리, thinking 타입)를 실어 보냅니다 — Claude Desktop의 서드파티
 게이트웨이 모드가 추론 강도 선택 UI를 열 수 있게 하기 위해서입니다. 실제 Anthropic 모델은
 원래 id를 그대로 유지합니다. 합성된 2026 날짜는 내부 슬롯이며 출시일이 아닙니다. 구버전의
-해시 별칭과 `claude-ocx-<provider>--<model>`, `claude-ocx2-<provider>--<model>` 별칭도 계속 해석됩니다. 저장된 `claude-ocx-`는 `ocx-claude-`로, 이스케이프된 `claude-ocx2-`는 `ocx-claude2-`로 한 번 다시 고르면 실제 컨텍스트 창과 compact가 함께 적용됩니다. 컨텍스트가 1M인 모델에는
+해시 별칭과 `claude-ocx-<provider>--<model>`, `claude-ocx2-<provider>--<model>` 별칭도 계속 해석됩니다. 저장된 `claude-ocx-`는 `ocx-claude-`로, 이스케이프된 `claude-ocx2-`는 `ocx-claude2-`로 한 번 다시 고르면 실제 컨텍스트 창과 compact가 함께 적용됩니다. 컨텍스트가 1M이거나 기본 압축 임곗값(829,800) 이상인 모델에는
 `…[1m]` 행이 하나 더 생깁니다 — 이걸 고르면 Claude Code가 그 모델의 컨텍스트를 1M로 계산합니다
 (자동 요약 유지, 프록시가 표식을 떼고 라우팅). 선택하면 Claude Code의
 `settings.json` `model` 필드에 저장되고, 인바운드 요청에서
@@ -343,8 +343,9 @@ v2 별칭은 이스케이프를 펼쳐요. 읽기 쉬운 형식으로 표현할
 
 ### 컨텍스트 변형 `[1m]` 표식
 
-공식 컨텍스트 창이 1M인 모델에는 `…[1m]` 선택기 행이 하나 더 생겨요. 자동 컨텍스트를 사용할
-때는 컨텍스트가 200k를 넘고 압축 임곗값 이상인 모델도 해당해요. 이 행을 선택하면 Claude Code가
+공식 컨텍스트 창이 1M이거나 기본 압축 임곗값(829,800) 이상인 모델에는 `…[1m]` 선택기 행이 하나 더
+생겨요. 이 기준은 고정이라 압축 값을 직접 바꿔도 낮아지지 않고, Anthropic 경로의 Claude 모델은 실제 1M이어야
+해요. 자동 컨텍스트를 끄면 1M 모델만 해당해요. 실제 창을 넘으면 `prompt is too long` 오류가 오고 Claude Code가 자동으로 압축해요. 이 행을 선택하면 Claude Code가
 전체 1M 컨텍스트를 계산해요. 프록시는 별칭 해석과 라우팅 전에 대소문자를 구분하지 않고 `[1m]`
 접미사를 제거해요.
 
@@ -353,8 +354,9 @@ v2 별칭은 이스케이프를 펼쳐요. 읽기 쉬운 형식으로 표현할
 Claude Code는 알 수 없는 모델의 컨텍스트를 200k 토큰으로 계산해요. 기본으로 켜져 있는 **자동
 컨텍스트**는 이 문제를 해결해요.
 
-1. 실제 컨텍스트 창이 200k보다 크고 자동 압축 임곗값 이상인 모델의 선택기 행과 환경 슬롯에
-   `[1m]` 표식이 붙어요.
+1. 실행 환경 슬롯에는 실제 컨텍스트 창이 200k보다 크고 설정한 자동 압축 임곗값 이상일 때
+   `[1m]` 표식이 붙어요. 디스커버리와 Desktop 선택기 행은 그 임곗값을 따르지 않고, 고정된 829,800 토큰 하한
+   (Anthropic 경로의 Claude 모델은 실제 1M)을 써요.
 2. `CLAUDE_CODE_AUTO_COMPACT_WINDOW`(기본값 `829800`, 범위 `100000`–`1000000`)를 주입해 해당
    지점에서 대화를 자동으로 요약해요.
 
@@ -367,8 +369,9 @@ Claude Code는 알 수 없는 모델의 컨텍스트를 200k 토큰으로 계산
 **연결 → Claude**에서 압축 값을 조절할 수 있어요. **경고:** 모델의 실제 컨텍스트 창보다 크게 올리면
 요약을 시작하기 전에 채팅 오류가 발생해요.
 
-1M 미만인 네이티브 Anthropic 모델에는 자동으로 표식을 붙이지 않아요. 직접 내보낸 값이 항상
+1M 미만인 네이티브 Anthropic 모델에는 자동으로 표식을 붙이지 않아요. 직접 내보낸 압축 값은 실행 슬롯 표식에서 항상
 우선하며, 프록시는 **사용자가 지정한** 값을 기준으로 어떤 모델에 안전하게 표식을 붙일지 결정해요.
+디스커버리 행은 그 값을 무시하고 고정 하한을 유지해요.
 직접 편집한 설정값이 잘못되면 829,800로 돌아가요.
 
 ### 실제 모델 환경
@@ -666,7 +669,7 @@ Anthropic 백엔드를 명시하면 의도적으로 실패 후 중단해요.
 
 **대형 모델인데도 컨텍스트가 200k로 제한됨** — 선택기에서 `[1m]` 변형을 고르거나 기본으로
 켜져 있는 자동 컨텍스트를 사용하세요. 선택기에 `[1m]` 행이 없다면 모델의 공식 컨텍스트 창이
-자동 압축 임곗값보다 작을 수 있어요.
+고정된 829,800 토큰 하한보다 작을 수 있어요.
 
 **스킬을 불러올 때 토큰 수가 많음** — 번들 `claude-api` 스킬(약 136k 토큰)은 Claude 모델을
 언급하면 자동으로 불러와요. 네이티브 패스스루에서는 정상이며, 라우팅 모델에서는 opencodex가
@@ -686,7 +689,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/docs-site/src/content/docs/ru/guides/claude-code.md b/docs-site/src/content/docs/ru/guides/claude-code.md
index 6bb458df309..dce282cfcc9 100644
--- a/docs-site/src/content/docs/ru/guides/claude-code.md
+++ b/docs-site/src/content/docs/ru/guides/claude-code.md
@@ -330,8 +330,10 @@ Anthropic сохраняют канонические id в обоих инте
 
 ### Маркер контекстного варианта `[1m]`
 
-Модели с подтверждённым контекстным окном 1M (или, при автоконтексте, больше 200k и не ниже
-порога сжатия) получают дополнительную строку `…[1m]` в селекторе. Её выбор заставляет Claude Code
+Модели с подтверждённым контекстным окном 1M или не ниже порога сжатия по умолчанию (829 800)
+получают дополнительную строку `…[1m]` в селекторе. Этот порог фиксирован: изменение значения сжатия его
+не понижает, а модель Claude на маршруте Anthropic должна действительно иметь 1M. Без автоконтекста учитываются
+только окна 1M. При выходе за реальное окно приходит ошибка `prompt is too long`, и Claude Code сам сжимает контекст. Её выбор заставляет Claude Code
 учитывать полный контекст 1M. Прокси удаляет суффикс `[1m]` (без учёта регистра) до разрешения
 алиасов и маршрутизации.
 
@@ -340,8 +342,9 @@ Anthropic сохраняют канонические id в обоих инте
 Для любой незнакомой модели Claude Code считает контекст равным 200k токенов. **Автоконтекст**
 (включён по умолчанию) исправляет это:
 
-1. Модели, чьё реальное окно больше 200k **и** не меньше порога автосжатия, получают маркер
-   `[1m]` в строках селектора и слотах окружения.
+1. Слоты окружения при запуске получают маркер `[1m]`, когда реальное окно больше 200k **и** не меньше
+   настроенного порога автосжатия. Строки обнаружения и селектора Desktop этому порогу не следуют:
+   для них действует фиксированный пол 829 800 токенов (настоящий 1M для моделей Claude на маршруте Anthropic).
 2. Внедряется `CLAUDE_CODE_AUTO_COMPACT_WINDOW` (по умолчанию `829800`, диапазон
    `100000`–`1000000`), чтобы в этой точке диалог автоматически резюмировался.
 
@@ -355,9 +358,9 @@ Anthropic сохраняют канонические id в обоих инте
 реального окна модели, эта модель ломается — чат завершится ошибкой раньше, чем успеет сработать
 резюмирование.
 
-Нативные модели Anthropic с окном меньше 1M никогда не помечаются автоматически. Экспортированные
-вами значения всегда имеют приоритет (прокси использует ВАШЕ значение, чтобы решить, какие модели
-можно безопасно пометить). Некорректные значения, внесённые в конфигурацию вручную, откатываются
+Нативные модели Anthropic с окном меньше 1M никогда не помечаются автоматически. Экспортированное
+вами окно сжатия имеет приоритет при пометке слотов запуска (прокси использует ВАШЕ значение, чтобы решить, какие модели
+можно безопасно пометить). Строки обнаружения игнорируют этот экспорт и сохраняют фиксированный пол. Некорректные значения, внесённые в конфигурацию вручную, откатываются
 к 829,800.
 
 ### Эффективное окружение моделей
@@ -637,7 +640,7 @@ Responses `web_search_call` в парные блоки Anthropic `server_tool_us
 
 **Потолок контекста 200k, хотя модель больше** — выберите в селекторе вариант `[1m]` или
 включите автоконтекст (включён по умолчанию). Если строки `[1m]` в селекторе нет, подтверждённое
-контекстное окно модели может быть меньше порога автосжатия.
+контекстное окно модели может быть меньше фиксированного пола в 829 800 токенов.
 
 **Большой расход токенов при загрузке скиллов** — встроенный скилл `claude-api` (~136k токенов)
 автоматически загружается при упоминании моделей Claude. Для нативного проброса это нормально;
@@ -658,7 +661,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/docs-site/src/content/docs/tr/guides/claude-code.md b/docs-site/src/content/docs/tr/guides/claude-code.md
index b0015531c14..1a99421d774 100644
--- a/docs-site/src/content/docs/tr/guides/claude-code.md
+++ b/docs-site/src/content/docs/tr/guides/claude-code.md
@@ -470,10 +470,11 @@ giderme adımıdır; kesin çözüm değildir. OpenCodex seçici durumunu gözle
 her isteğin taşıdığı model kimliğini yönlendirir. İstemcinin ne gönderdiğini
 **Logs → requestedModel** altında kontrol edin.
 
-Yetkili 1M bağlam penceresine sahip modeller fazladan bir `…[1m]` seçici satırı
-alır: bunu seçmek Claude Code'un bu model için tam 1M bağlam hesabı yapmasını
-sağlar (otomatik sıkıştırma açık kalır) — proxy yönlendirmeden önce işaretçiyi
-kaldırır.
+Uzun bağlam penceresine sahip modeller (1M, veya en az 829.800 token varsayılan
+sıkıştırma eşiği; Anthropic rotasındaki Claude modellerinin gerçekten 1M olması
+gerekir) fazladan bir `…[1m]` seçici satırı alır: bunu seçmek Claude Code'un bu
+model için tam 1M bağlam hesabı yapmasını sağlar (otomatik sıkıştırma açık kalır)
+— proxy yönlendirmeden önce işaretçiyi kaldırır.
 Birini seçmek, onu Claude Code'un `settings.json` `model` alanına kaydeder;
 gelen istekler takma adı yönlendirilen modele geri çözer. Eski Claude Code
 sürümlerinde seçici yerel kalır — `ANTHROPIC_MODEL` aracılığıyla yuvaları
@@ -512,10 +513,12 @@ kimliklerini korur.
 
 ### Bağlam değişkeni `[1m]` işaretçisi
 
-Yetkili bağlam penceresi 1M olan (veya otomatik bağlam altında 200k üzerinde ve
-en az sıkıştırma eşiğinde olan) modeller fazladan bir `…[1m]` seçici satırı
-alır. Bunu seçmek Claude Code'un tam 1M bağlam hesabı yapmasını sağlar. Proxy,
-takma ad çözümleme ve yönlendirmeden önce büyük/küçük harfe duyarsız `[1m]`
+Yetkili bağlam penceresi 1M olan veya varsayılan sıkıştırma eşiği olan 829.800
+tokene ulaşan modeller fazladan bir `…[1m]` seçici satırı alır. Bu taban sabittir:
+sıkıştırma değerini değiştirmek onu düşürmez ve Anthropic rotasındaki Claude
+modellerinin gerçekten 1M olması gerekir. Otomatik bağlam kapatılırsa satır yalnızca
+1M pencerelere kalır. Bunu seçmek Claude Code'un tam 1M bağlam hesabı yapmasını sağlar.
+Proxy, takma ad çözümleme ve yönlendirmeden önce büyük/küçük harfe duyarsız `[1m]`
 sonekini kaldırır.
 
 ## Otomatik bağlam (200k tavanı olmayan büyük bağlamlı modeller)
@@ -523,9 +526,10 @@ sonekini kaldırır.
 Claude Code, tanımadığı herhangi bir model için 200k token hesabı yapar.
 **Otomatik bağlam** (varsayılan olarak açık) bunu düzeltir:
 
-1. Gerçek penceresi 200k'nın üzerinde **ve** en az otomatik sıkıştırma eşiğinde
-   olan modeller, seçici satırlarında ve ortam yuvalarında `[1m]` işaretçisini
-   alır.
+1. Başlatma ortam yuvaları, gerçek pencere 200k'nın üzerinde **ve** en az yapılandırılmış
+   otomatik sıkıştırma eşiğindeyken `[1m]` işaretçisini alır. Keşif ve Desktop seçici
+   satırları bu eşiği izlemez: sabit 829.800 token tabanını kullanırlar (Anthropic
+   rotasındaki Claude modelleri için gerçek 1M).
 2. Görüşmenin bu noktada otomatik olarak özetlenmesi için
    `CLAUDE_CODE_AUTO_COMPACT_WINDOW` (varsayılan `829800`, aralık
    `100000`–`1000000`) enjekte edilir.
@@ -542,8 +546,9 @@ penceresinin üzerine çıkarmak o modeli bozar — sohbet özetleme tetiklenmed
 önce hata verir.
 
 1M altı yerel Anthropic modelleri hiçbir zaman otomatik olarak işaretlenmez.
-Kendi dışa aktardığınız değerler her zaman kazanır (proxy hangi modellerin
+Kendi dışa aktardığınız sıkıştırma penceresi, başlatma yuvası işaretlemesinde her zaman kazanır (proxy hangi modellerin
 işaretlenmesinin güvenli olduğuna karar vermek için SİZİN değerinizi kullanır).
+Keşif satırları bu dışa aktarımı yok sayar ve sabit tabanı korur.
 Geçersiz elle düzenlenen yapılandırma değerleri 829,800'ya geri döner.
 
 ### Geçerli model ortamı (Effective model environment)
@@ -852,8 +857,8 @@ veya `ocx claude` komutunu yeniden çalıştırın.
 
 **Büyük modele rağmen 200k bağlam tavanı** — Seçicide `[1m]` varyantını seçin
 veya otomatik bağlamı etkinleştirin (varsayılan olarak açık). Seçici hiçbir
-`[1m]` satırı göstermiyorsa, modelin yetkili bağlam penceresi otomatik
-sıkıştırma eşiğinin altında olabilir.
+`[1m]` satırı göstermiyorsa, modelin yetkili bağlam penceresi sabit 829.800
+token tabanının altında olabilir.
 
 **Yetenek yüklemelerinden kaynaklanan yüksek token sayısı** — Paketlenmiş
 `claude-api` yeteneği (~136k token) Claude model adları anıldığında otomatik
@@ -880,7 +885,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/docs-site/src/content/docs/zh-cn/guides/claude-code.md b/docs-site/src/content/docs/zh-cn/guides/claude-code.md
index aafb1f310a7..9f791d54db0 100644
--- a/docs-site/src/content/docs/zh-cn/guides/claude-code.md
+++ b/docs-site/src/content/docs/zh-cn/guides/claude-code.md
@@ -239,7 +239,7 @@ UI。真实 Anthropic 模型保留其原始 id。合成的 2026 日期是内部
 别名和 `claude-ocx-<provider>--<model>` 别名仍可解析，转义的 `claude-ocx2-<provider>--<model>` 也同样可解析。
 已保存的旧 id 仍会路由，但 Claude Code 对它仍按 200k 计算。把已保存的 `claude-ocx-` 重新选一次对应的
 `ocx-claude-`，转义的 `claude-ocx2-` 重新选一次 `ocx-claude2-`，即可同时用上真实上下文窗口和 compact。
-拥有 1M 上下文的模型会多出一行 `…[1m]`：
+上下文为 1M 或不低于默认压缩阈值（829,800）的模型会多出一行 `…[1m]`：
 选中后 Claude Code 会按 1M 计算该模型的上下文（自动压缩保留，代理在路由前去掉该标记）。
 选中后会保存到 Claude Code 的 `settings.json` `model` 字段；入站请求会将别名解析回路由
 模型。旧版 Claude Code 中选择器保持原生 — 通过 `ANTHROPIC_MODEL` 设置槽位，或直接在 `/model`
@@ -293,8 +293,9 @@ v1 别名按字面解码（历史上 model ID 中包含的两字符序列 `~s` /
 
 ### 上下文变体 `[1m]` 标记
 
-权威上下文窗口为 1M 的模型（或者启用自动上下文时，窗口大于 200k 且至少达到压缩阈值的模型）
-会多出一个带 `…[1m]` 的选择器条目。选择它后，Claude Code 会按完整的 1M 上下文计算。
+权威上下文窗口为 1M 或不低于默认压缩阈值（829,800）的模型会多出一个带 `…[1m]` 的选择器条目。
+该下限是固定的，修改压缩值不会降低它；Anthropic 路由上的 Claude 模型必须真正达到 1M；关闭自动上下文后只对 1M 模型生效。
+超出真实窗口时会返回 `prompt is too long` 错误，Claude Code 会自动压缩。选择它后，Claude Code 会按完整的 1M 上下文计算。
 代理会在进行别名解析和路由之前移除不区分大小写的 `[1m]` 后缀。
 
 ## 自动上下文（突破 200k 上限的大上下文模型）
@@ -302,8 +303,8 @@ v1 别名按字面解码（历史上 model ID 中包含的两字符序列 `~s` /
 对于任何无法识别的模型，Claude Code 都会按 200k token 计算。默认开启的**自动上下文**可解决
 这一问题：
 
-1. 实际窗口大于 200k **且**至少达到自动压缩阈值的模型，其选择器条目和环境变量槽位会带有
-   `[1m]` 标记。
+1. 启动环境槽位会在实际窗口大于 200k **且**至少达到所配置的自动压缩阈值时带上
+   `[1m]` 标记。发现列表和 Desktop 选择器条目不跟随该阈值：它们使用固定的 829,800 token 下限（Anthropic 路由上的 Claude 模型必须真正达到 1M）。
 2. 系统会注入 `CLAUDE_CODE_AUTO_COMPACT_WINDOW`（默认 `829800`，范围 `100000`–`1000000`），
    使对话在该位置自动进行摘要。
 
@@ -316,8 +317,8 @@ v1 别名按字面解码（历史上 model ID 中包含的两字符序列 `~s` /
 可以在 Claude 页面调整压缩值。**警告：**如果将其提高到超过模型的实际窗口，该模型将无法正常
 工作——聊天会在触发摘要之前报错。
 
-低于 1M 的原生 Anthropic 模型绝不会被自动标记。你自行导出的值始终优先（代理会使用**你的**
-值来判断哪些模型可以安全标记）。手动编辑配置时填入的无效值会回退到 829,800。
+低于 1M 的原生 Anthropic 模型绝不会被自动标记。你自行导出的压缩值在启动槽位标记上始终优先（代理会使用**你的**
+值来判断哪些模型可以安全标记）。发现列表忽略该导出值并保持固定下限。手动编辑配置时填入的无效值会回退到 829,800。
 
 ### 有效模型环境变量
 
@@ -566,7 +567,7 @@ OAuth 活动账户未标记 `needsReauth`。显式选择 Anthropic 却没有可
 `ANTHROPIC_BASE_URL` 可能已经过时。请打开一个新终端，或重新运行 `ocx claude`。
 
 **大模型仍受 200k 上下文上限限制**——在选择器中选择 `[1m]` 变体，或启用自动上下文
-（默认开启）。如果选择器中没有 `[1m]` 条目，该模型的权威上下文窗口可能低于自动压缩阈值。
+（默认开启）。如果选择器中没有 `[1m]` 条目，该模型的权威上下文窗口可能低于固定的 829,800 token 下限。
 
 **技能加载导致 token 数量过高**——内置的 `claude-api` 技能（约 136k token）会在提及
 Claude 模型时自动加载。对于原生透传，这是正常现象；对于已路由模型，opencodex 默认会将其
@@ -581,7 +582,7 @@ Claude 模型时自动加载。对于原生透传，这是正常现象；对于
 
 ### 第一方模型选择器的上下文标记
 
-对于权威上下文窗口至少为一百万 token 的路由模型，Desktop Code 标签页的模型选择器会添加 `[1m]`，使 Claude 按 1M 窗口计量，而不是使用自定义模型较小的默认窗口。标签、配置中的顺序和提供方路由保持不变。窗口未知或小于一百万 token 的模型（包括原生模型的大窗口选项）不添加标记，因为选择器无法保证 Desktop 或远程运行器收到配套的压缩环境设置。`ocx claude` 的自动上下文与压缩配套设置保持不变。现有对话会保留已保存的选择器，直到你在刷新后的模型选择器中重新选择该模型。
+对于权威上下文窗口至少为一百万 token，或不低于默认压缩阈值 829,800（例如选择了 872k 窗口的原生 GPT-6 模型）的路由模型，Desktop Code 标签页的模型选择器会添加 `[1m]`，使 Claude 按 1M 窗口计量，而不是使用自定义模型较小的默认窗口。Anthropic 路由上的 Claude 模型仍必须真正达到 1M。标签、配置中的顺序和提供方路由保持不变。窗口未知或更短的模型不添加标记。选择器无法保证 Desktop 或远程运行器收到配套的压缩环境设置，因此超出真实窗口的长窗口模型会收到 `prompt is too long` 错误，Claude Code 会据此压缩。Desktop 3P 的 `supports1m`/`prefer1m` 和生成的子代理遵循同一规则。`ocx claude` 的自动上下文与压缩配套设置保持不变。现有对话会保留已保存的选择器，直到你在刷新后的模型选择器中重新选择该模型。
 
 ### Forced Claude Code subagent model
 
@@ -589,7 +590,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/docs-site/src/content/docs/zh-tw/guides/claude-code.md b/docs-site/src/content/docs/zh-tw/guides/claude-code.md
index c6ec3d79339..4f033558616 100644
--- a/docs-site/src/content/docs/zh-tw/guides/claude-code.md
+++ b/docs-site/src/content/docs/zh-tw/guides/claude-code.md
@@ -339,7 +339,7 @@ user-agent 會獲得易讀的 CLI 形式，其他用戶端會獲得 Desktop 代
 也同樣可解析。已儲存的舊 id 仍會路由，但 Claude Code 對它仍按 200k 計算。把已儲存的 `claude-ocx-`
 重新選一次對應的 `ocx-claude-`，跳脫的 `claude-ocx2-` 重新選一次 `ocx-claude2-`，即可同時用上真實上下文
 視窗與 compact。
-擁有權威 1M 上下文視窗的模型會多出一個 `…[1m]` 選擇器列：選中後 Claude Code 會按完整 1M 上下文
+權威上下文視窗為 1M 或不低於預設壓縮閾值（829,800）的模型會多出一個 `…[1m]` 選擇器列：選中後 Claude Code 會按完整 1M 上下文
 計算該模型（自動壓縮仍開啟）——代理在路由前會去掉該標記。
 選中後會儲存到 Claude Code 的 `settings.json` `model` 欄位；入站請求會將別名解析回路由
 模型。在較舊的 Claude Code 版本中，選擇器保持原生——可透過 `ANTHROPIC_MODEL` 設定槽位，或在
@@ -369,8 +369,9 @@ v1 別名按字面解碼（歷史上 model ID 中包含的兩字元序列 `~s` /
 
 ### 上下文變體 `[1m]` 標記
 
-權威上下文視窗為 1M 的模型（或者啟用自動上下文時，視窗大於 200k 且至少達到壓縮閾值的模型）
-會多出一個帶 `…[1m]` 的選擇器條目。選擇它後，Claude Code 會按完整的 1M 上下文計算。
+權威上下文視窗為 1M 或不低於預設壓縮閾值（829,800）的模型會多出一個帶 `…[1m]` 的選擇器條目。
+該下限是固定的，修改壓縮值不會降低它；Anthropic 路由上的 Claude 模型必須真正達到 1M；關閉自動上下文後只對 1M 模型生效。
+超出真實視窗時會回傳 `prompt is too long` 錯誤，Claude Code 會自動壓縮。選擇它後，Claude Code 會按完整的 1M 上下文計算。
 代理會在進行別名解析和路由之前移除不區分大小寫的 `[1m]` 字尾。
 
 ## 自動上下文（突破 200k 上限的大上下文模型）
@@ -378,8 +379,8 @@ v1 別名按字面解碼（歷史上 model ID 中包含的兩字元序列 `~s` /
 對於任何無法識別的模型，Claude Code 都會按 200k token 計算。預設開啟的**自動上下文**可解決
 這一問題：
 
-1. 實際視窗大於 200k **且**至少達到自動壓縮閾值的模型，其選擇器條目和環境變數槽位會帶有
-   `[1m]` 標記。
+1. 啟動環境槽位會在實際視窗大於 200k **且**至少達到所設定的自動壓縮閾值時帶上
+   `[1m]` 標記。探索列表和 Desktop 選擇器條目不跟隨該閾值：它們使用固定的 829,800 token 下限（Anthropic 路由上的 Claude 模型必須真正達到 1M）。
 2. 系統會注入 `CLAUDE_CODE_AUTO_COMPACT_WINDOW`（預設 `829800`，範圍 `100000`–`1000000`），
    使對話在該位置自動進行摘要。
 
@@ -392,8 +393,8 @@ v1 別名按字面解碼（歷史上 model ID 中包含的兩字元序列 `~s` /
 可以在 Claude 頁面調整壓縮值。**警告：**如果將其提高到超過模型的實際視窗，該模型將無法正常
 工作——聊天會在觸發摘要之前報錯。
 
-低於 1M 的原生 Anthropic 模型絕不會被自動標記。你自行匯出的值始終優先（代理會使用**你的**
-值來判斷哪些模型可以安全標記）。手動編輯設定時填入的無效值會回退到 829,800。
+低於 1M 的原生 Anthropic 模型絕不會被自動標記。你自行匯出的壓縮值在啟動槽位標記上始終優先（代理會使用**你的**
+值來判斷哪些模型可以安全標記）。探索列表忽略該匯出值並保持固定下限。手動編輯設定時填入的無效值會回退到 829,800。
 
 ### 有效模型環境變數
 
@@ -642,7 +643,7 @@ OAuth 活動帳號未標記 `needsReauth`。顯式選擇 Anthropic 卻沒有可
 `ANTHROPIC_BASE_URL` 可能已經過時。請開啟一個新終端，或重新執行 `ocx claude`。
 
 **大模型仍受 200k 上下文上限限制**——在選擇器中選擇 `[1m]` 變體，或啟用自動上下文
-（預設開啟）。如果選擇器中沒有 `[1m]` 條目，該模型的權威上下文視窗可能低於自動壓縮閾值。
+（預設開啟）。如果選擇器中沒有 `[1m]` 條目，該模型的權威上下文視窗可能低於固定的 829,800 token 下限。
 
 **技能載入導致 token 數量過高**——內建的 `claude-api` 技能（約 136k token）會在提及
 Claude 模型時自動載入。對於原生透傳，這是正常現象；對於已路由模型，opencodex 預設會將其
@@ -665,7 +666,7 @@ The Subagents page offers **Force all subagents onto one model**, off by default
 
 `ocx agent subagents force combo/tev-auto` sets `claudeCode.subagentModelForce`; `ocx agent subagents force -` clears it. `ocx agent status` reports the setting. `GET /api/subagent-models` returns `force`, `forceAvailable`, and `forceStatus`; `PUT` accepts `{ "force": "combo/tev-auto" }` or `{ "force": null }` without changing the roster. Omitting `force` leaves it unchanged. Invalid or unexposed targets are rejected on write; stale targets are reported and skipped at launch.
 
-This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` only for an authoritative million-token window; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
+This takes effect on the **next routed `ocx claude` launch**, injecting `CLAUDE_CODE_SUBAGENT_MODEL` as an explicit proxy alias (with `[1m]` when the authoritative window reaches the fixed 829,800-token floor, and only for a genuine 1M window on Anthropic Claude models and bare `claude-*` selectors; native Claude targets use a reversible native alias) and `CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1`. Each nonempty shell-exported variable independently wins. Native launches inject neither variable; plain `claude` is not affected. No plugin files or `settings.json` are modified by this setting.
 
 Claude Code **2.1.257 or newer** is required for FORCE. Plugin and built-in agents (including Explore/Plan) and per-call model arguments are overridden. Forks and subagent skills with `model: inherit` keep the main conversation model. The main loop and Haiku/small-fast sidecars are unaffected. Existing roster files remain available.
 
diff --git a/scripts/test-layout/layout.json b/scripts/test-layout/layout.json
index adb7db7b2b6..9e9d2939b1d 100644
--- a/scripts/test-layout/layout.json
+++ b/scripts/test-layout/layout.json
@@ -418,7 +418,7 @@
     "messages-native-decline-trace.test.ts": "claude-integration",
     "messages-native-oauth.test.ts": "claude-integration",
     "messages-native-opaque-state.test.ts": "claude-integration",
-    "messages-surface-matrix.test.ts": "claude-integration",
+    "messages-surface-matrix.test.ts": "claude-integration", "claude-long-context-eligibility.test.ts": "claude-integration",
     "claude-model-info.test.ts": "claude-integration",
     "claude-models-discovery.test.ts": "claude-integration",
     "claude-native-affinity.test.ts": "claude-integration",
diff --git a/src/claude/context-windows.ts b/src/claude/context-windows.ts
index 858f62d1818..5c3caeda824 100644
--- a/src/claude/context-windows.ts
+++ b/src/claude/context-windows.ts
@@ -15,8 +15,9 @@ import { desktop3pAlias } from "./desktop-3p";
 import { nativeOpenAiContextWindow, type CatalogModel, type NativeContextLimitsInput } from "../codex/catalog";
 import { ANTHROPIC_MODEL_CONTEXT_WINDOWS } from "../providers/registry/model-seeds";
 import type { OcxClaudeCodeConfig } from "../types";
+import { AUTO_COMPACT_WINDOW_DEFAULT, AUTO_CONTEXT_FLOOR, ONE_MILLION } from "./long-context";
 
-const ONE_MILLION = 1_000_000;
+export { AUTO_COMPACT_WINDOW_DEFAULT, AUTO_CONTEXT_FLOOR } from "./long-context";
 
 /**
  * The native id each Claude Code tier alias resolves to (2.1.282: `--model opus` sends
@@ -27,21 +28,6 @@ const ONE_MILLION = 1_000_000;
  */
 const CLAUDE_CODE_NATIVE_TIERS = { opus: "claude-opus-5-5", sonnet: "claude-sonnet-5", fable: "claude-fable-5-1" } as const;
 
-/**
- * Auto-context defaults (devlog 260712 020, user-approved).
- *
- * The compact window is the token count at which Claude Code starts compacting, and it is
- * also the floor `shouldMarkOneMillion` uses — a model may only carry the marker if it can
- * host this window. 350,000 was chosen when the widest native row advertised 372,000.
- *
- * It now matches the auto-compaction limit the Codex catalog ships for the same models
- * (`nativeAutoCompactLimit`: 829,800 against the 922,000 native window). Leaving the two
- * apart meant one model compacting at 350k under Claude Code and at 829,800 under Codex.
- * The value stays clear of the measured 922,000 ceiling by ~92k, so compaction still has
- * room to run before the upstream refuses.
- */
-export const AUTO_COMPACT_WINDOW_DEFAULT = 829_800;
-export const AUTO_CONTEXT_FLOOR = 200_000;
 /** Binary-verified accepted range for CLAUDE_CODE_AUTO_COMPACT_WINDOW (2.1.207: pSo=1e5, yDs=1e6). */
 export const AUTO_COMPACT_WINDOW_MIN = 100_000;
 export const AUTO_COMPACT_WINDOW_MAX = ONE_MILLION;
@@ -63,6 +49,13 @@ export interface AutoContextMode {
 
 export const AUTO_CONTEXT_OFF: AutoContextMode = { enabled: false, compactWindow: AUTO_COMPACT_WINDOW_DEFAULT };
 
+/**
+ * Marking mode for surfaces whose runner may not inherit the compaction env (Desktop pickers,
+ * discovery, generated subagents). shouldMarkOneMillion under it equals isLongContextWindow
+ * (long-context.ts); kept as a mode so the existing marker helpers serve every surface.
+ */
+export const UNPAIRED_AUTO_CONTEXT: AutoContextMode = { enabled: true, compactWindow: AUTO_COMPACT_WINDOW_DEFAULT };
+
 interface AutoContextConfigSlice {
   autoContext?: boolean;
   autoCompactWindow?: number;
diff --git a/src/claude/desktop-3p.ts b/src/claude/desktop-3p.ts
index e2f0eec30e6..26eb3fbfb2c 100644
--- a/src/claude/desktop-3p.ts
+++ b/src/claude/desktop-3p.ts
@@ -25,6 +25,7 @@ import {
 import { nativeOpenAiContextWindow, type NativeContextLimitsInput } from "../codex/catalog";
 import { localAdmissionToken, localInferenceDestination } from "../lib/local-destinations";
 import { assertDesktop3pModelsValid } from "./desktop-3p-guard";
+import { claudeSurfaceSupportsOneMillion, ONE_MILLION } from "./long-context";
 
 export interface Desktop3pModelEntry {
   name: string;
@@ -56,12 +57,12 @@ export interface Desktop3pRoutedModel {
 }
 
 /**
- * 1M-context eligibility, shared with the Desktop DTO so the dashboard's 1M chip can
- * never disagree with what the writer emits. The DTO imports this from here — keeping
- * the constant in this module avoids a cycle, since shared.ts already reads
- * claude/desktop-profile.
+ * The genuine-1M threshold real Anthropic rows still need. Every other row follows the
+ * long-window rule; the writer, the profile renderer and the dashboard DTO all decide through
+ * `claudeSurfaceSupportsOneMillion` (long-context.ts), so the 1M chip cannot disagree with the
+ * written config.
  */
-export const DESKTOP_SUPPORTS_1M_THRESHOLD = 1_000_000;
+export const DESKTOP_SUPPORTS_1M_THRESHOLD = ONE_MILLION;
 
 /** CLI arg parsing for `ocx claude desktop` mode flags (mutually exclusive). */
 export function parseDesktop3pModeArgs(flags: string[]): { mode: Desktop3pConfigMode } | { error: string } {
@@ -252,7 +253,7 @@ function collectDesktop3pModels(
   for (const { provider, id, contextWindow } of candidates) {
     const route = `${provider}/${id}`;
     const alias = desktop3pAlias(provider, id);
-    const supports1m = typeof contextWindow === "number" && contextWindow >= DESKTOP_SUPPORTS_1M_THRESHOLD
+    const supports1m = claudeSurfaceSupportsOneMillion(provider, id, contextWindow)
       ? { supports1m: true as const }
       : {};
     if (alias === id) {
diff --git a/src/claude/desktop-profile.ts b/src/claude/desktop-profile.ts
index 9f617f59181..f160cda9430 100644
--- a/src/claude/desktop-profile.ts
+++ b/src/claude/desktop-profile.ts
@@ -1,4 +1,5 @@
 import { createHash } from "node:crypto";
+import { routeSupportsOneMillion } from "./long-context";
 import type {
   OcxClaudeDesktopAssignment,
   OcxClaudeDesktopFamily,
@@ -359,7 +360,7 @@ export function renderDesktopProfile(
       name: isRealAnthropicRoute(route) ? assignment.alias : desktopProfileWireAlias(assignment.alias),
       family: assignment.family,
       isFamilyDefault: effectiveDefaults[assignment.family] === route,
-      supports1m: typeof model.contextWindow === "number" && model.contextWindow >= 1_000_000,
+      supports1m: routeSupportsOneMillion(route, model.contextWindow),
     };
   });
 }
diff --git a/src/claude/intercept/picker-models.ts b/src/claude/intercept/picker-models.ts
index 79c85ca8f39..65d3ca93e13 100644
--- a/src/claude/intercept/picker-models.ts
+++ b/src/claude/intercept/picker-models.ts
@@ -7,7 +7,8 @@ import { dirname } from "node:path";
 import { nativeOpenAiContextWindow, type NativeContextLimitsInput } from "../../codex/catalog";
 import type { OcxClaudeDesktopProfile } from "../../types";
 import { aliasForRoute, claudeCodeNativeAlias } from "../alias";
-import { AUTO_CONTEXT_OFF, stripOneMillionMarker, withOneMillionMarker } from "../context-windows";
+import { AUTO_CONTEXT_OFF, stripOneMillionMarker, UNPAIRED_AUTO_CONTEXT, withOneMillionMarker } from "../context-windows";
+import { isAnthropicClaudeRoute } from "../long-context";
 import { activeDesktop3pAlias, displayModelId, resolveDesktop3pAlias, type Desktop3pRoutedModel } from "../desktop-3p";
 import { reconcileDesktopProfile, renderDesktopProfile, type DesktopProfileModel } from "../desktop-profile";
 import type { PickerModelEntry } from "./picker-bootstrap";
@@ -46,9 +47,14 @@ function splitRoute(route: string): { provider: string; id: string } {
   return { provider: route.slice(0, slash), id: route.slice(slash + 1) };
 }
 
-/** Desktop runners do not inherit the proxy's compaction env: mark only real >=1M windows. */
-function pickerSelector(alias: string, contextWindow: number | undefined): string {
-  return withOneMillionMarker(alias, contextWindow === undefined ? {} : { [alias]: contextWindow }, AUTO_CONTEXT_OFF)!;
+/**
+ * Desktop runners do not inherit the proxy's compaction env, so a long window (>= the default
+ * compact window) is marked on the strength of the `prompt is too long` recovery instead
+ * (devlog/_plan/261009_claude_1m_default/020). Real Anthropic models need a genuine 1M.
+ */
+function pickerSelector(alias: string, contextWindow: number | undefined, provider: string, id: string): string {
+  const mode = isAnthropicClaudeRoute(provider, id) ? AUTO_CONTEXT_OFF : UNPAIRED_AUTO_CONTEXT;
+  return withOneMillionMarker(alias, contextWindow === undefined ? {} : { [alias]: contextWindow }, mode)!;
 }
 
 export function buildPickerModels(input: PickerRouteInput): PickerModelEntry[] {
@@ -61,7 +67,7 @@ export function buildPickerModels(input: PickerRouteInput): PickerModelEntry[] {
     const alias = provider === "native" ? claudeCodeNativeAlias(id) : aliasForRoute(provider, id);
     if (!alias || seen.has(alias)) continue;
     seen.add(alias);
-    out.push({ id: pickerSelector(alias, model.contextWindow), name: model.label,
+    out.push({ id: pickerSelector(alias, model.contextWindow, provider, id), name: model.label,
       ...(model.contextWindow === undefined ? {} : { contextWindow: model.contextWindow }) });
   }
   return out;
@@ -81,7 +87,7 @@ export function buildCliPickerModels(input: PickerRouteInput): PickerModelEntry[
     const alias = activeDesktop3pAlias(provider, id);
     if (seen.has(alias) || resolveDesktop3pAlias(alias) !== model.route) continue;
     seen.add(alias);
-    out.push({ id: pickerSelector(alias, model.contextWindow), name: model.label,
+    out.push({ id: pickerSelector(alias, model.contextWindow, provider, id), name: model.label,
       description: `opencodex · ${model.route}`, route: model.route,
       ...(model.contextWindow === undefined ? {} : { contextWindow: model.contextWindow }) });
   }
diff --git a/src/claude/long-context.ts b/src/claude/long-context.ts
new file mode 100644
index 00000000000..9b1a922d23e
--- /dev/null
+++ b/src/claude/long-context.ts
@@ -0,0 +1,66 @@
+/**
+ * Long-context thresholds shared by every Claude surface
+ * (devlog/_plan/261009_claude_1m_default/020_long_context_eligibility.md).
+ *
+ * A leaf module: desktop-3p.ts and desktop-profile.ts need the predicate, and context-windows.ts
+ * already imports desktop-3p.ts, so the predicate cannot live there.
+ */
+import { isAnthropicInstanceId } from "../providers/anthropic-instance-id";
+
+export const ONE_MILLION = 1_000_000;
+
+/** A window at or below this is Claude Code's own default accounting; widening it gains nothing. */
+export const AUTO_CONTEXT_FLOOR = 200_000;
+
+/**
+ * Auto-context defaults (devlog 260712 020, user-approved).
+ *
+ * The compact window is the token count at which Claude Code starts compacting, and it is
+ * also the floor `shouldMarkOneMillion` uses — a model may only carry the marker if it can
+ * host this window. 350,000 was chosen when the widest native row advertised 372,000.
+ *
+ * It now matches the auto-compaction limit the Codex catalog ships for the same models
+ * (`nativeAutoCompactLimit`: 829,800 against the 922,000 native window). Leaving the two
+ * apart meant one model compacting at 350k under Claude Code and at 829,800 under Codex.
+ * The value stays clear of the measured 922,000 ceiling by ~92k, so compaction still has
+ * room to run before the upstream refuses.
+ */
+export const AUTO_COMPACT_WINDOW_DEFAULT = 829_800;
+
+/**
+ * Whether a window earns the 1M selection on a surface whose Claude Code runner may not inherit
+ * CLAUDE_CODE_AUTO_COMPACT_WINDOW (Desktop pickers, discovery, Desktop 3P, generated subagents).
+ * A window >= 1M always does. A smaller one does when it can host the default compact window, so
+ * a runner that has the variable compacts in time, and one that lacks it overflows into the
+ * `prompt is too long` envelope (anthropicErrorBody) and compacts reactively. The floor is fixed
+ * on purpose: a custom compact window must not re-admit a 372k route (#854).
+ */
+export function isLongContextWindow(window: number | undefined): boolean {
+  if (typeof window !== "number" || window <= 0) return false;
+  return window >= ONE_MILLION || (window > AUTO_CONTEXT_FLOOR && window >= AUTO_COMPACT_WINDOW_DEFAULT);
+}
+
+/**
+ * A Claude model on an Anthropic-named route. Claude windows are 200k or a genuine 1M and a bare
+ * id rides the native passthrough, so the long-window rule never widens one: a Claude row listed
+ * under 1M is a capped one, and on a runner without the compaction env it keeps the 200k
+ * accounting it had before long windows were widened. Keyed on the model id as well, so a
+ * configured gateway that only shares the name `anthropic2` and serves other models keeps the
+ * long-window rule. A custom `anthropic2` gateway serving Claude below 1M is treated like a pool
+ * row on these surfaces, which is the pre-existing (unwidened) behavior.
+ */
+export function isAnthropicClaudeRoute(provider: string, modelId: string): boolean {
+  return isAnthropicInstanceId(provider) && modelId.startsWith("claude-");
+}
+
+/** Desktop 3P / picker / discovery / dashboard 1M eligibility for one provider route. */
+export function claudeSurfaceSupportsOneMillion(provider: string, modelId: string, window: number | undefined): boolean {
+  if (isAnthropicClaudeRoute(provider, modelId)) return typeof window === "number" && window >= ONE_MILLION;
+  return isLongContextWindow(window);
+}
+
+/** Same, for a `provider/id` route string. */
+export function routeSupportsOneMillion(route: string, window: number | undefined): boolean {
+  const slash = route.indexOf("/");
+  return claudeSurfaceSupportsOneMillion(route.slice(0, slash), route.slice(slash + 1), window);
+}
diff --git a/src/claude/model-info.ts b/src/claude/model-info.ts
index 7f5fe0f7266..21162fc269c 100644
--- a/src/claude/model-info.ts
+++ b/src/claude/model-info.ts
@@ -19,7 +19,8 @@ import { orderForModelPicker, catalogModelEfforts, nativeEffortClamp, nativeOpen
 import { claudeCodeAlias, claudeCodeNativeAlias } from "./alias";
 import { cursorFastIdFor } from "../adapters/cursor/catalog";
 import { desktop3pAlias } from "./desktop-3p";
-import { AUTO_CONTEXT_OFF, type AutoContextMode } from "./context-windows";
+import { AUTO_CONTEXT_OFF, shouldMarkOneMillion, UNPAIRED_AUTO_CONTEXT, type AutoContextMode } from "./context-windows";
+import { isAnthropicClaudeRoute } from "./long-context";
 
 const MODEL_INFO_CREATED_AT = "2026-01-01T00:00:00Z";
 const ANTHROPIC_EFFORT_RUNGS = new Set(["low", "medium", "high", "xhigh", "max"]);
@@ -144,23 +145,25 @@ export function buildAnthropicModelInfos(
         : aliasForRoute(m.provider, m.id);
     }),
   ]);
-  // [1m] picker variant (devlog 260712 B1): Claude Code accounts exactly 1M for ids
-  // carrying the marker (2.1.207 binary: /\[1m\]/i → 1e6, compaction preserved), so
-  // ONLY models with an authoritative >=1M window get a second selectable row —
-  // the auto-context widening that let a 372K route carry the marker (and be
-  // over-filled) is the #854 defect and does not come back. Guards (audit R1#11):
-  // same dedupe set, never double-suffix.
+  // [1m] picker variant (devlog 260712 B1; widened in devlog/_plan/261009_claude_1m_default/020):
+  // Claude Code accounts exactly 1M for ids carrying the marker (2.1.207 binary: /\[1m\]/i -> 1e6,
+  // compaction preserved). A discovery client may lack CLAUDE_CODE_AUTO_COMPACT_WINDOW, so the
+  // rule is the fixed unpaired one (long-context.ts), not the configurable launch compact window:
+  // a custom 350k window must not re-admit a 372k route (#854). The server's auto mode only turns
+  // widening off (autoContext: false / maxContextTokens), falling back to >= 1M. Base and Fast
+  // rows stay unmarked: this list is a choice list. Guards (audit R1#11): same dedupe set, never
+  // double-suffix.
+  const variantMode = auto.enabled ? UNPAIRED_AUTO_CONTEXT : AUTO_CONTEXT_OFF;
   const push1mVariant = (
     base: AnthropicModelInfo,
     contextWindow: number | undefined,
     maxInputTokens?: number,
     selectorId?: string,
+    mode: AutoContextMode = variantMode,
   ) => {
-    // The [1m] marker makes Claude Code account 1e6 tokens for the row, so it
-    // may only name models whose AUTHORITATIVE effective window is >= 1M —
-    // never the auto-context widening, which would mark a 372K route and have
-    // Claude Code over-fill it (the #854 defect).
-    if (contextWindow === undefined || contextWindow < ONE_MILLION) return;
+    // A window >= 1M always earns the row; a long window earns it under the unpaired rule,
+    // because an overflow is answered as `prompt is too long`, which Claude Code compacts on.
+    if (!shouldMarkOneMillion(contextWindow, mode)) return;
     if (base.id.includes("[1m]")) return;
     const id = selectorId ?? `${base.id}[1m]`;
     if (seen.has(id)) return;
@@ -170,9 +173,9 @@ export function buildAnthropicModelInfos(
     // (measured — see devlog/_plan/260817_native_gpt56_1m_context/001_measurement_evidence.md).
     // Advertising the flat 1e6 there would invite mid-session context_length_exceeded, so the
     // variant reports whichever of the two is smaller.
-    const advertised = typeof maxInputTokens === "number" && maxInputTokens > 0
-      ? Math.min(ONE_MILLION, maxInputTokens)
-      : ONE_MILLION;
+    // A long window under 1M must not be advertised as 1e6 input either.
+    const ceiling = typeof maxInputTokens === "number" && maxInputTokens > 0 ? maxInputTokens : contextWindow;
+    const advertised = typeof ceiling === "number" && ceiling > 0 ? Math.min(ONE_MILLION, ceiling) : ONE_MILLION;
     out.push({ ...base, id, display_name: `${base.display_name} · 1M`, max_input_tokens: advertised });
   };
   /**
@@ -251,7 +254,7 @@ export function buildAnthropicModelInfos(
       && listedModelId.startsWith("claude-fable-")
       ? `${claudeCodeNativeAlias(listedModelId)}[1m]`
       : undefined;
-    push1mVariant(info, m.contextWindow, routedMaxInput, oneMillionSelector);
+    push1mVariant(info, m.contextWindow, routedMaxInput, oneMillionSelector, isAnthropicClaudeRoute(m.provider, m.id) ? AUTO_CONTEXT_OFF : variantMode);
     // The whole model is passed, not a (provider, id) pair: a combo row lives in its own
     // namespace with no config.providers entry, so the caller classifies it from the
     // aggregated supportsServiceTier the row already carries.
diff --git a/src/claude/subagent-model.ts b/src/claude/subagent-model.ts
index db0ee655455..f5efe44f4d4 100644
--- a/src/claude/subagent-model.ts
+++ b/src/claude/subagent-model.ts
@@ -1,6 +1,7 @@
 import type { OcxConfig } from "../types";
 import { claudeCodeAlias, claudeCodeNativeAlias, resolveAlias } from "./alias";
-import { AUTO_CONTEXT_OFF, shouldMarkOneMillion, stripOneMillionMarker, withOneMillionMarker } from "./context-windows";
+import { AUTO_CONTEXT_OFF, shouldMarkOneMillion, stripOneMillionMarker, UNPAIRED_AUTO_CONTEXT, withOneMillionMarker, type AutoContextMode } from "./context-windows";
+import { isAnthropicClaudeRoute } from "./long-context";
 import { hasOwnProvider } from "../config";
 import { knownModelIdsForProvider } from "../router";
 import { decodeRoutedModelIdOrThrow } from "../providers/slug-codec";
@@ -14,20 +15,38 @@ export { SAFE_AGENT_MODEL_ID } from "../config/subagent-models";
 
 /**
  * Generated subagent defs cannot rely on the parent's auto-context compaction
- * pairing, so their [1m] marker follows the AUTHORITATIVE window only: mark when
- * the effective window (exact selector, then the canonical [1m] form, then bare)
- * is genuinely >= 1M; strip an inherited unsafe marker back to the bare selector;
- * with no window information, keep the selector as it was. Genuine routed [1m]
- * ids are preserved through the canonical-exact lookup. (#854)
+ * pairing, so their [1m] marker follows the unpaired long-context rule
+ * (long-context.ts): mark when the effective window (exact selector, then the
+ * canonical [1m] form, then bare) is >= 1M or can host the default compact window,
+ * whose overflow Claude Code compacts on as `prompt is too long`; strip an inherited
+ * marker the window cannot carry; with no window information, keep the selector as it
+ * was. Genuine routed [1m] ids are preserved through the canonical-exact lookup.
+ * (#854; devlog/_plan/261009_claude_1m_default/020)
  */
+/**
+ * The marking mode a subagent selector gets: real Anthropic models (a bare `claude-*` passthrough
+ * id, or a Claude row on either Anthropic pool) need a genuine 1M; everything else follows the
+ * unpaired long-context rule. Connected launches read windows from the published catalog, which
+ * does not drop capped Anthropic rows, so this cannot rely on the window map alone.
+ */
+function subagentSelectorMode(bare: string): AutoContextMode {
+  const route = resolveAlias(bare) ?? resolveDesktop3pAlias(bare) ?? bare;
+  const slash = route.indexOf("/");
+  const anthropic = slash < 0
+    ? route.startsWith("claude-")
+    : isAnthropicClaudeRoute(route.slice(0, slash), route.slice(slash + 1));
+  return anthropic ? AUTO_CONTEXT_OFF : UNPAIRED_AUTO_CONTEXT;
+}
+
 export function withSubagentContextMarker(selector: string, windows: Record<string, number>): string {
   const bare = stripOneMillionMarker(selector);
   const wasMarked = selector !== bare;
   const canonicalExact = wasMarked ? `${bare}[1m]` : selector;
   const authoritativeWindow = windows[selector] ?? windows[canonicalExact] ?? windows[bare];
   if (typeof authoritativeWindow === "number" && authoritativeWindow > 0) {
-    return shouldMarkOneMillion(authoritativeWindow, AUTO_CONTEXT_OFF)
-      ? (withOneMillionMarker(selector, windows) ?? selector)
+    const mode = subagentSelectorMode(bare);
+    return shouldMarkOneMillion(authoritativeWindow, mode)
+      ? (withOneMillionMarker(selector, windows, mode) ?? selector)
       : bare;
   }
   return wasMarked ? selector : bare;
@@ -77,7 +96,7 @@ export function resolveSubagentForceModel(config: OcxConfig, windows: Record<str
     const window = windows[parts.alias] ?? windows[`${bare}[1m]`] ?? windows[bare];
     // Identity admission must not let an appended context promise ride on an
     // ordinary exposed model. Exact catalog ids such as kimi/k3[1m] remain literal.
-    if (requestedMarker && !exactMarkedExposure && !(typeof window === "number" && Number.isFinite(window) && window >= 1_000_000)) return null;
+    if (requestedMarker && !exactMarkedExposure && !(typeof window === "number" && Number.isFinite(window) && shouldMarkOneMillion(window, subagentSelectorMode(bare)))) return null;
     const marked = exactMarkedExposure ? parts.alias : withSubagentContextMarker(parts.alias, windows);
     // Bare Anthropic names are indistinguishable from the old CLI's fallback.
     // Encode only the force selector; handlers restore native identity before
diff --git a/src/server/management/shared.ts b/src/server/management/shared.ts
index d3ca423f4d7..f5a0621922c 100644
--- a/src/server/management/shared.ts
+++ b/src/server/management/shared.ts
@@ -364,7 +364,7 @@ export function stripRegistryOnlyStaticHeaders(name: string, provider: OcxProvid
 /** Shared Desktop profile DTO builder for the management API and CLI. */
 export async function buildClaudeDesktopState(config: OcxConfig, stored?: OcxClaudeDesktopProfile) {
   const { filterCatalogVisibleModels, nativeContextLimits, nativeOpenAiContextWindow, desktopVisibleNativeSlugs } = await import("../../codex/catalog");
-  const { DESKTOP_SUPPORTS_1M_THRESHOLD } = await import("../../claude/desktop-3p");
+  const { routeSupportsOneMillion } = await import("../../claude/long-context");
   const { reconcileDesktopProfile, renderDesktopProfile } = await import("../../claude/desktop-profile");
   const routed = filterCatalogVisibleModels(await fetchAllModels(config), config);
   const profileModels: DesktopProfileModel[] = [
@@ -413,8 +413,8 @@ export async function buildClaudeDesktopState(config: OcxConfig, stored?: OcxCla
     ...(modelByRoute.get(route)?.contextWindow ? { contextWindow: modelByRoute.get(route)!.contextWindow } : {}),
     effortSupported: effortByRoute.get(route) ?? false,
     // Read-only view of the 1M capability the written Desktop config already emits,
-    // derived from the SAME threshold so the dashboard chip can never disagree.
-    supports1m: (modelByRoute.get(route)?.contextWindow ?? 0) >= DESKTOP_SUPPORTS_1M_THRESHOLD,
+    // derived from the SAME predicate so the dashboard chip can never disagree.
+    supports1m: routeSupportsOneMillion(route, modelByRoute.get(route)?.contextWindow),
     assignment: profile.assignments[route]!,
   }));
   return {
diff --git a/structure/clients/claude-desktop.md b/structure/clients/claude-desktop.md
index dbc87ee40ae..a27fdadbcc3 100644
--- a/structure/clients/claude-desktop.md
+++ b/structure/clients/claude-desktop.md
@@ -287,7 +287,7 @@ intent and cannot repair the URL a running Desktop already pinned. The `claude.a
 certificate, streams every body and upgrade unchanged, and rewrites only the bootstrap response's
 local Code picker surfaces, `ccd` (what the Desktop Code tab reads) and its `code` fallback, never the
 remote `ccr` (`picker-bootstrap.ts`), failing open to the original bytes; the model list
-comes from a persisted snapshot (`picker-models.ts`), so a bootstrap never waits on discovery. Picker aliases carry `[1m]` only for authoritative windows of at least 1M, using the shared context marker helper with auto-context disabled. Sub-million opt-ins remain unmarked because the picker cannot guarantee the Desktop runner's compaction environment. A
+comes from a persisted snapshot (`picker-models.ts`), so a bootstrap never waits on discovery. Picker aliases carry `[1m]` for long windows (`src/claude/long-context.ts`: at least 1M, or at least the 829,800 default compact window), and Claude models on either Anthropic pool only at a genuine 1M. The picker cannot guarantee the Desktop runner's compaction environment, so an 872k route that outgrows its window relies on the `prompt is too long` envelope, which Claude Code compacts on; Desktop 3P `supports1m`/`prefer1m`, discovery `· 1M` rows and generated subagent markers use the same rule. A
 CONNECT to claude.ai that arrives before the first refresh waits at most 3 s, then goes blind. A
 picker proxy bind failure only disables picker mode; a picker construction or start failure closes
 every socket the start had bound before rethrowing. Ordinary session cookies within the header
diff --git a/structure/subagents.md b/structure/subagents.md
index bdaa57cae19..906ce2149e4 100644
--- a/structure/subagents.md
+++ b/structure/subagents.md
@@ -546,6 +546,6 @@ Automatic account exhaustion and recovery use the [spendable Codex credit eviden
 
 ### Forced Claude Code subagent model
 
-`src/claude/subagent-model.ts` shares the roster alias and authoritative context-marker resolver with routed launch force. `claudeCode.subagentModelForce` is opt-in and default-off. Force availability excludes retained unavailable roster entries. Caller-added `[1m]` suffixes require finite authoritative million-token capacity or an exact advertised marked identity; exact upstream ids remain literal, while legacy roster marker handling is unchanged. The generated-agent legacy directive cannot replace a wire selector matching the configured forced alias. `src/claude/subagent-force-status.ts` performs bounded, read-only server-local version and settings-key inspection; unknown is not supported.
+`src/claude/subagent-model.ts` shares the roster alias and authoritative context-marker resolver with routed launch force. `claudeCode.subagentModelForce` is opt-in and default-off. Force availability excludes retained unavailable roster entries. Caller-added `[1m]` suffixes require a finite authoritative long window (`src/claude/long-context.ts`: 1M, or at least the 829,800 default compact window; a genuine 1M for Claude models on either Anthropic pool or a bare `claude-*` id) or an exact advertised marked identity; generated markers follow the same rule; exact upstream ids remain literal, while legacy roster marker handling is unchanged. The generated-agent legacy directive cannot replace a wire selector matching the configured forced alias. `src/claude/subagent-force-status.ts` performs bounded, read-only server-local version and settings-key inspection; unknown is not supported.
 
 Explicit gateway selectors outrank generated-agent `ocx-route` fallback independently of saved force state. `src/claude/inbound-model-options.ts` shares this precedence across Messages and count-tokens; bare Claude fallback and requests without directives retain their existing behavior. Native Claude force targets use reversible native aliases, restored before existing credential/model-map checks. Connected launch exposure comes from fresh authenticated gateway rows, not cached context-window keys; acquisition is independent of cache-write success.
diff --git a/tests/claude-integration/claude-long-context-eligibility.test.ts b/tests/claude-integration/claude-long-context-eligibility.test.ts
new file mode 100644
index 00000000000..ed41dc2e72f
--- /dev/null
+++ b/tests/claude-integration/claude-long-context-eligibility.test.ts
@@ -0,0 +1,89 @@
+/**
+ * Long windows (>= the default compact window) are 1M on every Claude surface whose runner may
+ * lack CLAUDE_CODE_AUTO_COMPACT_WINDOW; Claude models on either Anthropic pool still need a genuine
+ * 1M (devlog/_plan/261009_claude_1m_default/020_long_context_eligibility.md).
+ */
+import { describe, expect, test } from "bun:test";
+import { claudeSurfaceSupportsOneMillion, isLongContextWindow, routeSupportsOneMillion } from "../../src/claude/long-context";
+import { AUTO_CONTEXT_OFF, UNPAIRED_AUTO_CONTEXT } from "../../src/claude/context-windows";
+import { buildAnthropicModelInfos } from "../../src/claude/model-info";
+import { generateDesktop3pModels } from "../../src/claude/desktop-3p";
+import { withSubagentContextMarker } from "../../src/claude/subagent-model";
+
+const ASTRA_872K = { modelWindows: { "gpt-6-astra": 872_000 } };
+
+describe("isLongContextWindow / claudeSurfaceSupportsOneMillion", () => {
+  test("the floor is the default compact window, never a configurable one", () => {
+    for (const window of [undefined, 0, 200_000, 262_144, 400_000, 829_799]) expect(isLongContextWindow(window)).toBe(false);
+    for (const window of [829_800, 872_000, 922_000, 1_000_000, 2_000_000]) expect(isLongContextWindow(window)).toBe(true);
+  });
+
+  test("Claude models on either Anthropic pool need a genuine 1M; a gateway that shares the name does not", () => {
+    expect(claudeSurfaceSupportsOneMillion("anthropic", "claude-sonnet-5", 872_000)).toBe(false);
+    expect(claudeSurfaceSupportsOneMillion("anthropic2", "claude-sonnet-5", 872_000)).toBe(false);
+    expect(claudeSurfaceSupportsOneMillion("anthropic2", "claude-sonnet-5", 1_000_000)).toBe(true);
+    // A configured gateway named anthropic2 serving another vendor's model is not an Anthropic pool.
+    expect(claudeSurfaceSupportsOneMillion("anthropic2", "kimi-k3", 900_000)).toBe(true);
+    expect(routeSupportsOneMillion("native/gpt-6-astra", 872_000)).toBe(true);
+    expect(routeSupportsOneMillion("kimi/k3", 262_144)).toBe(false);
+  });
+});
+
+describe("Anthropic discovery 1M variants", () => {
+  const variants = (infos: ReturnType<typeof buildAnthropicModelInfos>) => infos.filter(info => info.id.endsWith("[1m]"));
+
+  test("an 872k native gets a variant that reports its real input ceiling", () => {
+    const infos = buildAnthropicModelInfos(["gpt-6-astra"], [], UNPAIRED_AUTO_CONTEXT, "readable", undefined, ASTRA_872K);
+    const [variant] = variants(infos);
+    expect(variant?.display_name).toBe("gpt-6-astra (native) · 1M");
+    expect(variant?.max_input_tokens).toBe(872_000);
+    expect(variants(buildAnthropicModelInfos(["gpt-6-astra"], [], AUTO_CONTEXT_OFF, "readable", undefined, ASTRA_872K))).toHaveLength(0);
+  });
+
+  test("a custom compact window does not lower the floor (#854)", () => {
+    const routed = [{ provider: "example", id: "mid", contextWindow: 400_000 }];
+    expect(variants(buildAnthropicModelInfos([], routed, { enabled: true, compactWindow: 350_000 }, "readable"))).toHaveLength(0);
+  });
+
+  test("a long routed row without an input ceiling advertises its window, not 1e6", () => {
+    const infos = buildAnthropicModelInfos([], [{ provider: "example", id: "long", contextWindow: 900_000 }], UNPAIRED_AUTO_CONTEXT, "readable");
+    expect(variants(infos).map(info => info.max_input_tokens)).toEqual([900_000]);
+  });
+
+  test("Anthropic rows and short windows get no widened variant", () => {
+    const routed = [
+      { provider: "anthropic", id: "claude-x", contextWindow: 872_000 },
+      { provider: "anthropic2", id: "claude-y", contextWindow: 872_000 },
+      { provider: "kimi", id: "k3", contextWindow: 262_144 },
+    ];
+    expect(variants(buildAnthropicModelInfos([], routed, UNPAIRED_AUTO_CONTEXT, "readable"))).toHaveLength(0);
+  });
+});
+
+describe("Desktop 3P", () => {
+  test("an 872k native offers and prefers 1M; a 262k route does not", () => {
+    const [astra] = generateDesktop3pModels(["gpt-6-astra"], [], undefined, ASTRA_872K);
+    expect(astra).toMatchObject({ supports1m: true, prefer1m: true });
+    const [k3] = generateDesktop3pModels([], [{ provider: "kimi", id: "k3", contextWindow: 262_144 }]);
+    expect(k3!.supports1m).toBeUndefined();
+    expect(k3!.prefer1m).toBeUndefined();
+  });
+});
+
+describe("generated subagent markers", () => {
+  const alias = "ocx-claude-native--gpt-6-astra";
+
+  test("an 872k window is marked, a 262k window is not, and an unsafe marker is stripped", () => {
+    expect(withSubagentContextMarker(alias, { [alias]: 872_000 })).toBe(`${alias}[1m]`);
+    expect(withSubagentContextMarker(alias, { [alias]: 262_144 })).toBe(alias);
+    expect(withSubagentContextMarker(`${alias}[1m]`, { [alias]: 262_144 })).toBe(alias);
+  });
+
+  test("a capped Claude row on Pool 2 stays unmarked even when a connected window map lists it", () => {
+    // Connected launches read the published catalog, which keeps capped Anthropic rows.
+    const pool2 = "ocx-claude-anthropic2--claude-sonnet-5";
+    expect(withSubagentContextMarker(pool2, { [pool2]: 900_000 })).toBe(pool2);
+    expect(withSubagentContextMarker(`${pool2}[1m]`, { [pool2]: 900_000 })).toBe(pool2);
+    expect(withSubagentContextMarker(pool2, { [pool2]: 1_000_000 })).toBe(`${pool2}[1m]`);
+  });
+});
diff --git a/tests/claude-integration/claude-picker-models.test.ts b/tests/claude-integration/claude-picker-models.test.ts
index fd1ad3445c6..7fee7abe587 100644
--- a/tests/claude-integration/claude-picker-models.test.ts
+++ b/tests/claude-integration/claude-picker-models.test.ts
@@ -79,25 +79,30 @@ test("picker preserves real million-token routed windows without promoting small
   expect(rows[1]?.contextWindow).toBe(1_048_576);
 });
 
-test("picker leaves sub-million native opt-ins unmarked without a runner compaction guarantee", () => {
+test("picker marks long native opt-ins on the prompt-too-long recovery and leaves 272k unmarked", () => {
+  // Desktop runners lack CLAUDE_CODE_AUTO_COMPACT_WINDOW; an 872k window is marked anyway because an
+  // overflow comes back as `prompt is too long`, which Claude Code compacts on (261009 020).
   const rows = buildPickerModels({
     nativeSlugs: ["gpt-5.5", "gpt-6-astra", "gpt-6-sol", "gpt-6-luna", "gpt-5.6-terra"],
     routedModels: [], nativeContextCap: { modelWindows: {
       "gpt-6-astra": 872_000, "gpt-6-sol": 872_000, "gpt-6-luna": 872_000, "gpt-5.6-terra": 922_000,
     } },
   });
-  expect(rows.every(row => !row.id.endsWith("[1m]"))).toBe(true);
+  expect(rows.map(row => row.id.endsWith("[1m]"))).toEqual([false, true, true, true, true]);
   expect(rows.map(row => row.contextWindow)).toEqual([272_000, 872_000, 872_000, 872_000, 922_000]);
 });
 
-test("picker marks the million-token boundary and leaves smaller routed windows unmarked", () => {
+test("picker marks windows from the default compact floor up and leaves smaller routed windows unmarked", () => {
   const rows = buildPickerModels({ nativeSlugs: [], routedModels: [
-    { provider: "example", id: "below", contextWindow: 999_999 },
+    { provider: "example", id: "short", contextWindow: 262_144 },
+    { provider: "example", id: "below", contextWindow: 829_799 },
+    { provider: "example", id: "floor", contextWindow: 829_800 },
     { provider: "example", id: "exact", contextWindow: 1_000_000 },
     { provider: "example", id: "larger", contextWindow: 2_000_000 },
+    { provider: "anthropic2", id: "claude-pool-two", contextWindow: 872_000 },
   ] });
   expect(rows.map(row => row.id)).toEqual([
-    "ocx-claude-example--below", "ocx-claude-example--exact[1m]", "ocx-claude-example--larger[1m]",
+    "ocx-claude-example--short", "ocx-claude-example--below", "ocx-claude-example--floor[1m]",
+    "ocx-claude-example--exact[1m]", "ocx-claude-example--larger[1m]", "ocx-claude-anthropic2--claude-pool-two",
   ]);
-  expect(rows.map(row => row.contextWindow)).toEqual([999_999, 1_000_000, 2_000_000]);
 });
diff --git a/tests/claude-integration/claude-subagent-model-force.test.ts b/tests/claude-integration/claude-subagent-model-force.test.ts
index be3c07403ca..670beb458ca 100644
--- a/tests/claude-integration/claude-subagent-model-force.test.ts
+++ b/tests/claude-integration/claude-subagent-model-force.test.ts
@@ -39,11 +39,12 @@ test("exported values independently win and native launches retain exports", ()
   }
 });
 
-test("alias resolution reuses roster provider/native semantics and authoritative million boundary", () => {
+test("alias resolution reuses roster provider/native semantics and the long-context boundary", () => {
   for (const entry of deps.forceAvailable) {
     const c = config(entry);
     const { alias } = entryParts(entry, c);
-    expect(resolveSubagentForceModel(c, { [alias]: 999999 }, { entries: deps.forceAvailable })).toBe(alias);
+    expect(resolveSubagentForceModel(c, { [alias]: 829_799 }, { entries: deps.forceAvailable })).toBe(alias);
+    expect(resolveSubagentForceModel(c, { [alias]: 829_800 }, { entries: deps.forceAvailable })).toBe(`${alias}[1m]`);
     expect(resolveSubagentForceModel(c, { [alias]: 1000000 }, { entries: deps.forceAvailable })).toBe(`${alias}[1m]`);
   }
 });
diff --git a/tests/clients/desktop-3p.test.ts b/tests/clients/desktop-3p.test.ts
index 8e61aeaf48a..0156ee1ab12 100644
--- a/tests/clients/desktop-3p.test.ts
+++ b/tests/clients/desktop-3p.test.ts
@@ -212,19 +212,18 @@ describe("Claude Desktop 3P models", () => {
   });
 
   test("an openai context cap reaches the Desktop writer, not just the dashboard", () => {
-    // No surviving native advertises a 1M window (gpt-5.4 was the last). Sol's
-    // opt-in ceiling is 922k, so even a 1M provider cap must not invent
-    // supports1m — nativeOpenAiContextWindow clamps it under the threshold.
-    // A 272k cap has to take the same path, or the written Desktop config
-    // would promise a window the proxy will not serve (#854's effective-window
+    // Sol's default window is 272k: no 1M. A 1M provider cap opts it into its 922k
+    // ceiling, a long window, so the written config now offers and prefers 1M; an
+    // overflow comes back as `prompt is too long`, which Claude compacts on (261009 020).
+    // A 272k cap has to take the same path as the default (#854's effective-window
     // contract).
     const uncapped = generateDesktop3pModels(["gpt-5.6-sol"], []);
     expect(uncapped[0]!.supports1m).toBeUndefined();
     expect(uncapped[0]!.prefer1m).toBeUndefined();
 
     const optedIn = generateDesktop3pModels(["gpt-5.6-sol"], [], undefined, 1_000_000);
-    expect(optedIn[0]!.supports1m).toBeUndefined();
-    expect(optedIn[0]!.prefer1m).toBeUndefined();
+    expect(optedIn[0]!.supports1m).toBe(true);
+    expect(optedIn[0]!.prefer1m).toBe(true);
 
     const capped = generateDesktop3pModels(["gpt-5.6-sol"], [], undefined, 272_000);
     expect(capped[0]!.supports1m).toBeUndefined();
diff --git a/tests/fixtures/test-layout-expected.json b/tests/fixtures/test-layout-expected.json
index 2ec78f200e9..a3c47be575c 100644
--- a/tests/fixtures/test-layout-expected.json
+++ b/tests/fixtures/test-layout-expected.json
@@ -303,6 +303,7 @@
   "claude-model-info.test.ts": "claude-integration", "claude-models-discovery.test.ts": "claude-integration",
   "claude-native-affinity.test.ts": "claude-integration", "claude-native-passthrough.test.ts": "claude-integration",
   "claude-native-rate-limit-headers.test.ts": "claude-integration", "claude-outbound.test.ts": "claude-integration",
+  "claude-long-context-eligibility.test.ts": "claude-integration",
   "claude-prompt-too-long.test.ts": "claude-integration",
   "claude-shell-hook.test.ts": "claude-integration", "claude-sidecar-override.test.ts": "claude-integration",
   "claude-source-envelope.test.ts": "claude-integration", "claude-system-env-auto.test.ts": "claude-integration",