omp: update patches

update
ghostty: fix???
2026-04-27 01:36:08 -04:00 · 2026-04-27 01:27:13 -04:00 · 2026-04-26 01:11:09 -04:00 · 2026-04-26 01:03:11 -04:00 · 2026-04-26 00:12:30 -04:00
9 changed files with 976 additions and 421 deletions
--- a/flake.lock
+++ b/flake.lock
@@ -222,11 +222,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1777138175,
-        "narHash": "sha256-UrexPU1xQ/qB0qCjuTeljQOCDmjeCNuipZMBv3FyoJM=",
+        "lastModified": 1777257791,
+        "narHash": "sha256-KE3+aTLGTIp8OZEI4lq1kvp30lmh3KA8Ru84UocbXyE=",
        "owner": "nix-community",
        "repo": "emacs-overlay",
-        "rev": "d7d0c87d15148472eef847dfe298095ef4298dc1",
+        "rev": "b1f88788b2f0e31cfa42e9dffbc5e9de218369de",
        "type": "github"
      },
      "original": {
@@ -266,11 +266,11 @@
      },
      "locked": {
        "dir": "pkgs/firefox-addons",
-        "lastModified": 1777089773,
-        "narHash": "sha256-ZIlNuebeWTncyl7mcV9VbceSLAaZki+UeXLPQG959xI=",
+        "lastModified": 1777262571,
+        "narHash": "sha256-ni1Cz9BChOXO6C0H4cRAq6bJRQIUV40Yet306ZOEEHs=",
        "owner": "rycee",
        "repo": "nur-expressions",
-        "rev": "402ba229617a12d918c2a887a4c83a9a24f9a36c",
+        "rev": "0827fcbe30e591e79b0554ecc5be9c79ba71a86b",
        "type": "gitlab"
      },
      "original": {
@@ -484,11 +484,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1777151655,
-        "narHash": "sha256-Th3a5OZyEy4kCoyLfefnt+2dwRIrFQqYgMsayF9qzFw=",
+        "lastModified": 1777258755,
+        "narHash": "sha256-EC07KwADRE2LdIk7vEDyAaD3I0ZUq24T9jQF9L0iEPk=",
        "owner": "nix-community",
        "repo": "home-manager",
-        "rev": "6f59831b23d03bbf4fbd13ad167ae25da294cc14",
+        "rev": "7f8bbc93d63401e41368d6ddc46a4f631610fa90",
        "type": "github"
      },
      "original": {
@@ -657,11 +657,11 @@
        "treefmt-nix": "treefmt-nix"
      },
      "locked": {
-        "lastModified": 1777154498,
-        "narHash": "sha256-700kin0o6CoNWkg2w5+2hV1wxECeoMRCQjOerBlWleA=",
+        "lastModified": 1777266861,
+        "narHash": "sha256-cdSr2nIz4I+ysG1gAZxbKQo+f79vCCKfQCdiRYnyPec=",
        "owner": "numtide",
        "repo": "llm-agents.nix",
-        "rev": "013ae4bdac7d0f968174d660aeb0760a025f09d0",
+        "rev": "c8f7c7882804510f2b807021cac0a69c1aeb4829",
        "type": "github"
      },
      "original": {
@@ -704,11 +704,11 @@
        "xwayland-satellite-unstable": "xwayland-satellite-unstable"
      },
      "locked": {
-        "lastModified": 1777130270,
-        "narHash": "sha256-AgOIR3O+hLkTe/spgYjp0knc37iy/A5DqGRY+8DP3LE=",
+        "lastModified": 1777240421,
+        "narHash": "sha256-ooPmu+8tqOGh4kozPW4rJC7Y7WM/FHtEY3OK1PoNW7g=",
        "owner": "sodiboo",
        "repo": "niri-flake",
-        "rev": "e43ef13f23c2c7ae5b10e842745cb345faff4f40",
+        "rev": "2bb22af2985e5f3cfd051b3d977ebfbf81126280",
        "type": "github"
      },
      "original": {
@@ -737,11 +737,11 @@
    "niri-unstable": {
      "flake": false,
      "locked": {
-        "lastModified": 1777115961,
-        "narHash": "sha256-ehSMsSpE+0k8r+2Vseu8kangsYxToZv3vinynsDp9zs=",
+        "lastModified": 1777237919,
+        "narHash": "sha256-bZHBzo4EuW/xLzXnnMKsIMdZYqgY2O0mIMdplwDHB8Y=",
        "owner": "YaLTeR",
        "repo": "niri",
-        "rev": "8ed0da44d974c32c6877d2f4630c314da0717ecb",
+        "rev": "a85b922919815c32a3ae34e0838830fe522d6a1c",
        "type": "github"
      },
      "original": {
@@ -761,11 +761,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1777140538,
-        "narHash": "sha256-2y5SwHxTOwEdr8WZv1IGBVoJM47YcomfoxFnZj9TgN0=",
+        "lastModified": 1777227006,
+        "narHash": "sha256-A7GcOXjfo2xmZ3ERgN0j6GcqaVzqIf5zpYQcdfDaMr0=",
        "owner": "xddxdd",
        "repo": "nix-cachyos-kernel",
-        "rev": "ce6083d35e50516dd6eb6156d0cbda67baed9117",
+        "rev": "0f7e2bea4088227a80502557f6c0e3b74949d6b5",
        "type": "github"
      },
      "original": {
@@ -802,11 +802,11 @@
    },
    "nix-flatpak": {
      "locked": {
-        "lastModified": 1776625032,
-        "narHash": "sha256-edvwHiFhgOiwywt6/Iwe+sSn6ybhU3WZGnIoiGcKjfQ=",
+        "lastModified": 1777229239,
+        "narHash": "sha256-OwSaWqlBdKn8QIa7BrPtJmlrr46U7AuwMc/toDKuMZw=",
        "owner": "gmodena",
        "repo": "nix-flatpak",
-        "rev": "479e19f1decb390aa5b75cae13ddf87d763c74cc",
+        "rev": "3f1d78b63b6af353c0685b8a7411c04d980426e4",
        "type": "github"
      },
      "original": {
@@ -937,11 +937,11 @@
    },
    "nixpkgs-stable": {
      "locked": {
-        "lastModified": 1776734388,
-        "narHash": "sha256-vl3dkhlE5gzsItuHoEMVe+DlonsK+0836LIRDnm6MXQ=",
+        "lastModified": 1777077449,
+        "narHash": "sha256-AIiMJiqvGrN4HyLEbKAoCSRRYn0rnlW5VbKNIMIYqm4=",
        "owner": "NixOS",
        "repo": "nixpkgs",
-        "rev": "10e7ad5bbcb421fe07e3a4ad53a634b0cd57ffac",
+        "rev": "a4bf06618f0b5ee50f14ed8f0da77d34ecc19160",
        "type": "github"
      },
      "original": {
@@ -991,11 +991,11 @@
        "noctalia-qs": "noctalia-qs"
      },
      "locked": {
-        "lastModified": 1777079905,
-        "narHash": "sha256-TvYEXwkZnRFQRuFyyqTNSfPnU2tMdhtiBOXSk2AWLJA=",
+        "lastModified": 1777253304,
+        "narHash": "sha256-XqSHEKEW5pSAx9MoMo8mKPgkjoy4FEhZ4x0a6hGYrSI=",
        "owner": "noctalia-dev",
        "repo": "noctalia-shell",
-        "rev": "a50c92167c8d438000270f7eca36f6eea74f388e",
+        "rev": "6773c4750a12c9e9af9c4ce2365e083f1d0d0ad8",
        "type": "github"
      },
      "original": {
@@ -1014,11 +1014,11 @@
        "treefmt-nix": "treefmt-nix_2"
      },
      "locked": {
-        "lastModified": 1776585574,
-        "narHash": "sha256-j35EWhKoGhKrfcXcAOpoRVgXEPQt41Eukji/h59cnjk=",
+        "lastModified": 1777167795,
+        "narHash": "sha256-VHdtmxVX7oF2+FxYQQPARQmtaHw23FoTBiTaH6ucOEg=",
        "owner": "noctalia-dev",
        "repo": "noctalia-qs",
-        "rev": "75d180c28a9ab4470e980f3d6f706ad6c5213add",
+        "rev": "697db4c14e27d841956ff76887fc312443e6fb17",
        "type": "github"
      },
      "original": {
@@ -1133,11 +1133,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1777086717,
-        "narHash": "sha256-vEl3cGHRxEFdVNuP9PbrhAWnmU98aPOLGy9/1JXzSuM=",
+        "lastModified": 1777259803,
+        "narHash": "sha256-fIb/EoVu/1U0qVrE6qZCJ2WCfprRpywNIAVzKEACIQc=",
        "owner": "oxalica",
        "repo": "rust-overlay",
-        "rev": "3be56bd430bfd65d3c468a50626c3a601c7dee03",
+        "rev": "a6cb2224d975e16b5e67de688c6ad306f7203425",
        "type": "github"
      },
      "original": {
@@ -1356,11 +1356,11 @@
    "trackerlist": {
      "flake": false,
      "locked": {
-        "lastModified": 1777154980,
-        "narHash": "sha256-zEJCVDBjo0SDlYOnkfi9o6lJWpMfmmR6Oh67RPybbqI=",
+        "lastModified": 1777241384,
+        "narHash": "sha256-mzqjBOMvL8951W4qt5VA31rQB+TiOYDRyMXTQ7ScSUY=",
        "owner": "ngosang",
        "repo": "trackerslist",
-        "rev": "9599dfb9be9d899bb5abd40a5dc53e5c5be90fd4",
+        "rev": "50a204edfeb4f5f904a28e20b650966241203edb",
        "type": "github"
      },
      "original": {
@@ -1524,11 +1524,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1777138694,
-        "narHash": "sha256-yjAFuyqQyOtQ5entLYmSRf/1L0kuSDWQndS2QNBLQlc=",
+        "lastModified": 1777218171,
+        "narHash": "sha256-+JGU5Cw6Zm3XVl3xBCkbY7/lTxfLQpjuuhF0IB4dJ8k=",
        "owner": "0xc000022070",
        "repo": "zen-browser-flake",
-        "rev": "5ceb2bfc5671bfca6b1b363669309d6871043d66",
+        "rev": "8a8e30610393c7f1a766a119dea37bf82d0ebcf6",
        "type": "github"
      },
      "original": {
--- a/home/progs/ghostty.nix
+++ b/home/progs/ghostty.nix
@@ -55,6 +55,13 @@
      # cursor
      cursor-style = "underline";

+      # always open new windows at $HOME instead of inheriting whatever cwd the
+      # currently-focused ghostty window has. with gtk-single-instance, the
+      # focused-window inherit rule otherwise sticks the daemon's first cwd to
+      # every subsequent niri Mod+T launch.
+      window-inherit-working-directory = false;
+      working-directory = "home";
+
      # keep one daemon alive so subsequent launches (e.g. niri Mod+T) are
      # instant instead of paying GTK + wgpu init each time. relies on the
      # dbus-activated systemd user service that the HM module wires up.
--- a/home/progs/niri.nix
+++ b/home/progs/niri.nix
@@ -115,12 +115,8 @@ in

        "Mod+O".action = toggle-overview;

-        # open a terminal — pass --working-directory=home so the gtk-single-instance
-        # daemon doesn't keep handing back whatever cwd the focused window has.
-        "Mod+T".action = spawn [
-          config.home.sessionVariables.TERMINAL
-          "--working-directory=home"
-        ];
+        # open a terminal
+        "Mod+T".action = spawn config.home.sessionVariables.TERMINAL;

        # lock the screen
        "Mod+X".action = spawn (lib.getExe pkgs.swaylock);
--- a/home/progs/pi.nix
+++ b/home/progs/pi.nix
@@ -41,16 +41,8 @@ in
    # silently ignores the standard `patches` attribute. Apply patches via `prePatch` instead
    # so they actually take effect. Tracking: nothing upstream yet.
    (inputs.llm-agents.packages.${pkgs.stdenv.hostPlatform.system}.omp.overrideAttrs (old: {
-      prePatch =
-        (old.prePatch or "")
-        + ''
-          # 0001 — retry without strict tools when DeepSeek (via OpenRouter) rejects strict-mode
-          #        `anyOf` nullable unions with `Invalid tool parameters schema : field \`anyOf\`:
-          #        missing field \`type\``.
-          patch -p1 < ${../../patches/omp/0001-openai-completions-retry-without-strict-on-deepseek-openrouter.patch}
-          # 0002 — require `reasoning_content` for OpenRouter reasoning models so DeepSeek V4 Pro
-          #        et al. accept follow-up requests in thinking mode.
-          patch -p1 < ${../../patches/omp/0002-openai-completions-stub-reasoning-content-for-openrouter.patch}
+      prePatch = (old.prePatch or "") + ''
+        patch -p1 < ${../../patches/omp/0001-fix-reasoning_content.patch}
      '';
    }))
  ];
--- a/patches/omp/0001-fix-reasoning_content.patch
+++ b/patches/omp/0001-fix-reasoning_content.patch
@@ -0,0 +1,804 @@
+From e145b627cffb6907e6bde348f1318f48acba3801 Mon Sep 17 00:00:00 2001
+From: sonhyrd <son.hong.do@hyrd.ai>
+Date: Mon, 27 Apr 2026 00:00:18 +0700
+Subject: [PATCH 1/5] fix(ai/providers): cover opencode-go reasoning tool-call
+ history
+
+---
+ .../providers/openai-completions-compat.ts    | 12 +++--
+ .../ai/src/providers/openai-completions.ts    |  4 +-
+ .../ai/test/openai-completions-compat.test.ts | 51 +++++++++++++++----
+ 3 files changed, 49 insertions(+), 18 deletions(-)
+
+diff --git a/packages/ai/src/providers/openai-completions-compat.ts b/packages/ai/src/providers/openai-completions-compat.ts
+index 69f4811c8..c777f312b 100644
+--- a/packages/ai/src/providers/openai-completions-compat.ts
+++ b/packages/ai/src/providers/openai-completions-compat.ts
+@@ -107,12 +107,14 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
+ 		reasoningContentField: "reasoning_content",
+ 		// Backends that 400 follow-up requests when prior assistant tool-call turns lack `reasoning_content`:
+ 		//   - Kimi: documented invariant on its native API and via OpenCode-Go.
+-		//   - Any reasoning-capable model reached through OpenRouter: DeepSeek V4 Pro and similar enforce
+-		//     this server-side whenever the request is in thinking mode. We can't translate Anthropic's
+-		//     redacted/encrypted reasoning into DeepSeek's plaintext form, so cross-provider continuations
+-		//     rely on a placeholder — see `convertMessages` for the placeholder injection.
+		//   - Reasoning-capable models reached through OpenRouter or OpenCode-Go: DeepSeek V4 Pro and
+		//     similar enforce this server-side whenever the request is in thinking mode.
+		// We can't translate Anthropic's redacted/encrypted reasoning into DeepSeek's plaintext form, so
+		// cross-provider continuations rely on a placeholder — see `convertMessages` for injection rules.
+ 		requiresReasoningContentForToolCalls:
+-			isKimiModel || ((provider === "openrouter" || baseUrl.includes("openrouter.ai")) && Boolean(model.reasoning)),
+			isKimiModel ||
+			((provider === "openrouter" || baseUrl.includes("openrouter.ai") || provider === "opencode-go" ||
+				baseUrl.includes("opencode.ai/zen/go")) && Boolean(model.reasoning)),
+ 		requiresAssistantContentForToolCalls: isKimiModel,
+ 		openRouterRouting: undefined,
+ 		vercelGatewayRouting: undefined,
+diff --git a/packages/ai/src/providers/openai-completions.ts b/packages/ai/src/providers/openai-completions.ts
+index 3785af106..70f2e3b63 100644
+--- a/packages/ai/src/providers/openai-completions.ts
+++ b/packages/ai/src/providers/openai-completions.ts
+@@ -1213,8 +1213,8 @@ export function convertMessages(
+ 			// Inject a `reasoning_content` placeholder on assistant tool-call turns when the backend
+ 			// rejects history without it. The compat flag captures the rule:
+ 			//   - Kimi (native or via OpenCode-Go): chat completion endpoint demands the field.
+-			//   - Reasoning models reached through OpenRouter (e.g. DeepSeek V4 Pro): the underlying
+-			//     provider's thinking-mode validator demands it on every prior assistant turn. omp
+			//   - Reasoning models reached through OpenRouter or OpenCode-Go (e.g. DeepSeek V4 Pro):
+			//     the upstream thinking-mode validator demands it on every prior assistant turn. omp
+ 			//     cannot synthesize real reasoning when the conversation was warmed up by another
+ 			//     provider whose reasoning is redacted/encrypted (Anthropic) or simply absent, so we
+ 			//     emit a placeholder. Real captured reasoning, when present, is preserved earlier via
+diff --git a/packages/ai/test/openai-completions-compat.test.ts b/packages/ai/test/openai-completions-compat.test.ts
+index 6fc3ca9af..6d60ba5e4 100644
+--- a/packages/ai/test/openai-completions-compat.test.ts
+++ b/packages/ai/test/openai-completions-compat.test.ts
+@@ -283,23 +283,59 @@ describe("openai-completions compatibility", () => {
+ });
+ 
+ describe("kimi model detection via detectCompat", () => {
+-	function kimiOpenCodeModel(id: string): Model<"openai-completions"> {
+	function openCodeGoModel(id: string, reasoning = true): Model<"openai-completions"> {
+ 		return {
+ 			...getBundledModel("openai", "gpt-4o-mini"),
+ 			api: "openai-completions",
+ 			provider: "opencode-go",
+ 			baseUrl: "https://opencode.ai/zen/go/v1",
+ 			id,
+-			reasoning: true,
+			reasoning,
+ 		};
+ 	}
+ 
+	function kimiOpenCodeModel(id: string): Model<"openai-completions"> {
+		return openCodeGoModel(id, true);
+	}
+
+ 	it("requires reasoning_content for tool calls on kimi-k2.5 (opencode-go)", () => {
+ 		const compat = detectCompat(kimiOpenCodeModel("kimi-k2.5"));
+ 		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+ 		expect(compat.requiresAssistantContentForToolCalls).toBe(true);
+ 	});
+ 
+	it("requires reasoning_content for tool calls on reasoning DeepSeek models via opencode-go", () => {
+		const compat = detectCompat(openCodeGoModel("deepseek-v4-pro", true));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+		expect(compat.requiresAssistantContentForToolCalls).toBe(false);
+	});
+
+	it("injects reasoning_content placeholder for reasoning DeepSeek tool-call turns via opencode-go", () => {
+		const model = openCodeGoModel("deepseek-v4-pro", true);
+		const compat = detectCompat(model);
+		const toolCallMessage: AssistantMessage = {
+			role: "assistant",
+			content: [{ type: "toolCall", id: "call_ds_go", name: "web_search", arguments: { query: "hi" } }],
+			api: model.api,
+			provider: model.provider,
+			model: model.id,
+			usage: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				totalTokens: 0,
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+			},
+			stopReason: "toolUse",
+			timestamp: Date.now(),
+		};
+		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+		const assistant = messages.find(m => m.role === "assistant");
+		expect(assistant).toBeDefined();
+		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+	});
+
+ 	it("injects reasoning_content placeholder when assistant with tool calls has no reasoning field", () => {
+ 		const model = kimiOpenCodeModel("kimi-k2.5");
+ 		const compat = detectCompat(model);
+@@ -338,15 +374,8 @@ describe("kimi model detection via detectCompat", () => {
+ 		expect((reasoningContent as string).length).toBeGreaterThan(0);
+ 	});
+ 
+-	it("does not inject reasoning_content when model is not kimi", () => {
+-		const model: Model<"openai-completions"> = {
+-			...getBundledModel("openai", "gpt-4o-mini"),
+-			api: "openai-completions",
+-			provider: "opencode-go",
+-			baseUrl: "https://opencode.ai/zen/go/v1",
+-			id: "some-other-model",
+-		};
+-		const compat = detectCompat(model);
+	it("does not require reasoning_content when opencode-go model is not reasoning-capable", () => {
+		const compat = detectCompat(openCodeGoModel("some-other-model", false));
+ 		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+ 	});
+ 
+
+From 70eda0132d7ff48314cbf2dc9560339f0a765d9e Mon Sep 17 00:00:00 2001
+From: sonhyrd <son.hong.do@hyrd.ai>
+Date: Mon, 27 Apr 2026 00:08:04 +0700
+Subject: [PATCH 2/5] fix(ai/providers): generalize opencode reasoning_content
+ gating
+
+---
+ .../providers/openai-completions-compat.ts    |  14 +-
+ .../ai/src/providers/openai-completions.ts    |   4 +-
+ .../ai/test/openai-completions-compat.test.ts | 160 ++++++++----------
+ 3 files changed, 82 insertions(+), 96 deletions(-)
+
+diff --git a/packages/ai/src/providers/openai-completions-compat.ts b/packages/ai/src/providers/openai-completions-compat.ts
+index c777f312b..b4825a31c 100644
+--- a/packages/ai/src/providers/openai-completions-compat.ts
+++ b/packages/ai/src/providers/openai-completions-compat.ts
+@@ -54,6 +54,8 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
+ 	const isKimiModel = model.id.includes("moonshotai/kimi") || /^kimi[-.]/i.test(model.id);
+ 	const isAlibaba = provider === "alibaba-coding-plan" || baseUrl.includes("dashscope");
+ 	const isQwen = model.id.toLowerCase().includes("qwen");
+	const isOpenRouter = provider === "openrouter" || baseUrl.includes("openrouter.ai");
+	const isOpenCode = provider === "opencode-zen" || provider === "opencode-go" || baseUrl.includes("opencode.ai/zen");
+ 
+ 	const isNonStandard =
+ 		isCerebras ||
+@@ -99,22 +101,20 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
+ 		requiresMistralToolIds: isMistral,
+ 		thinkingFormat: isZai
+ 			? "zai"
+-			: provider === "openrouter" || baseUrl.includes("openrouter.ai")
+			: isOpenRouter
+ 				? "openrouter"
+ 				: isAlibaba || isQwen
+ 					? "qwen"
+ 					: "openai",
+ 		reasoningContentField: "reasoning_content",
+ 		// Backends that 400 follow-up requests when prior assistant tool-call turns lack `reasoning_content`:
+-		//   - Kimi: documented invariant on its native API and via OpenCode-Go.
+-		//   - Reasoning-capable models reached through OpenRouter or OpenCode-Go: DeepSeek V4 Pro and
+-		//     similar enforce this server-side whenever the request is in thinking mode.
+		//   - Kimi: documented invariant on its native API and via OpenCode.
+		//   - Reasoning-capable models reached through OpenRouter or OpenCode (Zen/Go): DeepSeek V4 Pro,
+		//     Kimi, and similar models can enforce this server-side whenever the request is in thinking mode.
+ 		// We can't translate Anthropic's redacted/encrypted reasoning into DeepSeek's plaintext form, so
+ 		// cross-provider continuations rely on a placeholder — see `convertMessages` for injection rules.
+ 		requiresReasoningContentForToolCalls:
+-			isKimiModel ||
+-			((provider === "openrouter" || baseUrl.includes("openrouter.ai") || provider === "opencode-go" ||
+-				baseUrl.includes("opencode.ai/zen/go")) && Boolean(model.reasoning)),
+			isKimiModel || ((isOpenRouter || isOpenCode) && Boolean(model.reasoning)),
+ 		requiresAssistantContentForToolCalls: isKimiModel,
+ 		openRouterRouting: undefined,
+ 		vercelGatewayRouting: undefined,
+diff --git a/packages/ai/src/providers/openai-completions.ts b/packages/ai/src/providers/openai-completions.ts
+index 70f2e3b63..e25aeffb3 100644
+--- a/packages/ai/src/providers/openai-completions.ts
+++ b/packages/ai/src/providers/openai-completions.ts
+@@ -1212,8 +1212,8 @@ export function convertMessages(
+ 				(assistantMsg as any).reasoning_text !== undefined;
+ 			// Inject a `reasoning_content` placeholder on assistant tool-call turns when the backend
+ 			// rejects history without it. The compat flag captures the rule:
+-			//   - Kimi (native or via OpenCode-Go): chat completion endpoint demands the field.
+-			//   - Reasoning models reached through OpenRouter or OpenCode-Go (e.g. DeepSeek V4 Pro):
+			//   - Kimi (native or via OpenCode Zen/Go): chat completion endpoint demands the field.
+			//   - Reasoning models reached through OpenRouter or OpenCode Zen/Go (e.g. DeepSeek V4 Pro):
+ 			//     the upstream thinking-mode validator demands it on every prior assistant turn. omp
+ 			//     cannot synthesize real reasoning when the conversation was warmed up by another
+ 			//     provider whose reasoning is redacted/encrypted (Anthropic) or simply absent, so we
+diff --git a/packages/ai/test/openai-completions-compat.test.ts b/packages/ai/test/openai-completions-compat.test.ts
+index 6d60ba5e4..c743dd246 100644
+--- a/packages/ai/test/openai-completions-compat.test.ts
+++ b/packages/ai/test/openai-completions-compat.test.ts
+@@ -282,105 +282,91 @@ describe("openai-completions compatibility", () => {
+ 	});
+ });
+ 
+-describe("kimi model detection via detectCompat", () => {
+-	function openCodeGoModel(id: string, reasoning = true): Model<"openai-completions"> {
+describe("opencode reasoning-content compatibility via detectCompat", () => {
+	type OpenCodeProvider = "opencode-go" | "opencode-zen";
+
+	function openCodeModel(provider: OpenCodeProvider, id: string, reasoning = true): Model<"openai-completions"> {
+		const baseUrl = provider === "opencode-go" ? "https://opencode.ai/zen/go/v1" : "https://opencode.ai/zen/v1";
+ 		return {
+ 			...getBundledModel("openai", "gpt-4o-mini"),
+ 			api: "openai-completions",
+-			provider: "opencode-go",
+-			baseUrl: "https://opencode.ai/zen/go/v1",
+			provider,
+			baseUrl,
+ 			id,
+ 			reasoning,
+ 		};
+ 	}
+ 
+-	function kimiOpenCodeModel(id: string): Model<"openai-completions"> {
+-		return openCodeGoModel(id, true);
+-	}
+-
+-	it("requires reasoning_content for tool calls on kimi-k2.5 (opencode-go)", () => {
+-		const compat = detectCompat(kimiOpenCodeModel("kimi-k2.5"));
+-		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+-		expect(compat.requiresAssistantContentForToolCalls).toBe(true);
+-	});
+-
+-	it("requires reasoning_content for tool calls on reasoning DeepSeek models via opencode-go", () => {
+-		const compat = detectCompat(openCodeGoModel("deepseek-v4-pro", true));
+-		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+-		expect(compat.requiresAssistantContentForToolCalls).toBe(false);
+-	});
+-
+-	it("injects reasoning_content placeholder for reasoning DeepSeek tool-call turns via opencode-go", () => {
+-		const model = openCodeGoModel("deepseek-v4-pro", true);
+-		const compat = detectCompat(model);
+-		const toolCallMessage: AssistantMessage = {
+-			role: "assistant",
+-			content: [{ type: "toolCall", id: "call_ds_go", name: "web_search", arguments: { query: "hi" } }],
+-			api: model.api,
+-			provider: model.provider,
+-			model: model.id,
+-			usage: {
+-				input: 0,
+-				output: 0,
+-				cacheRead: 0,
+-				cacheWrite: 0,
+-				totalTokens: 0,
+-				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+-			},
+-			stopReason: "toolUse",
+-			timestamp: Date.now(),
+	it.each(["opencode-go", "opencode-zen"] as const)(
+		"requires reasoning_content for tool calls on kimi-k2.5 via %s",
+		provider => {
+			const compat = detectCompat(openCodeModel(provider, "kimi-k2.5", true));
+			expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+			expect(compat.requiresAssistantContentForToolCalls).toBe(true);
+		},
+	);
+
+	it.each(["opencode-go", "opencode-zen"] as const)(
+		"requires reasoning_content for tool calls on reasoning DeepSeek models via %s",
+		provider => {
+			const compat = detectCompat(openCodeModel(provider, "deepseek-v4-pro", true));
+			expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+			expect(compat.requiresAssistantContentForToolCalls).toBe(false);
+		},
+	);
+
+	it("requires reasoning_content when custom openai provider targets opencode zen baseUrl", () => {
+		const model: Model<"openai-completions"> = {
+			...getBundledModel("openai", "gpt-4o-mini"),
+			api: "openai-completions",
+			provider: "openai",
+			baseUrl: "https://opencode.ai/zen/v1",
+			id: "deepseek-v4-pro",
+			reasoning: true,
+ 		};
+-		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+-		const assistant = messages.find(m => m.role === "assistant");
+-		expect(assistant).toBeDefined();
+-		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+-	});
+-
+-	it("injects reasoning_content placeholder when assistant with tool calls has no reasoning field", () => {
+-		const model = kimiOpenCodeModel("kimi-k2.5");
+ 		const compat = detectCompat(model);
+-		const toolCallMessage: AssistantMessage = {
+-			role: "assistant",
+-			content: [
+-				// Thinking returned as plain text (as kimi-k2.5 on opencode-go does)
+-				{ type: "text", text: "Let me research this." },
+-				{
+-					type: "toolCall",
+-					id: "call_abc123",
+-					name: "web_search",
+-					arguments: { query: "beads gastownhall" },
+-				},
+-			],
+-			api: model.api,
+-			provider: model.provider,
+-			model: model.id,
+-			usage: {
+-				input: 0,
+-				output: 0,
+-				cacheRead: 0,
+-				cacheWrite: 0,
+-				totalTokens: 0,
+-				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+-			},
+-			stopReason: "toolUse",
+-			timestamp: Date.now(),
+-		};
+-		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+-		const assistant = messages.find(m => m.role === "assistant");
+-		expect(assistant).toBeDefined();
+-		const reasoningContent = Reflect.get(assistant as object, "reasoning_content");
+-		expect(reasoningContent).toBeDefined();
+-		expect(typeof reasoningContent).toBe("string");
+-		expect((reasoningContent as string).length).toBeGreaterThan(0);
+-	});
+-
+-	it("does not require reasoning_content when opencode-go model is not reasoning-capable", () => {
+-		const compat = detectCompat(openCodeGoModel("some-other-model", false));
+-		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+ 	});
+ 
+-	it.each(["kimi-k2.5", "kimi-k1.5", "kimi-k2-5"])("matches kimi model id: %s", id => {
+-		const compat = detectCompat(kimiOpenCodeModel(id));
+	it.each(["opencode-go", "opencode-zen"] as const)(
+		"injects reasoning_content placeholder for reasoning DeepSeek tool-call turns via %s",
+		provider => {
+			const model = openCodeModel(provider, "deepseek-v4-pro", true);
+			const compat = detectCompat(model);
+			const toolCallMessage: AssistantMessage = {
+				role: "assistant",
+				content: [{ type: "toolCall", id: `call_ds_${provider}`, name: "web_search", arguments: { query: "hi" } }],
+				api: model.api,
+				provider: model.provider,
+				model: model.id,
+				usage: {
+					input: 0,
+					output: 0,
+					cacheRead: 0,
+					cacheWrite: 0,
+					totalTokens: 0,
+					cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+				},
+				stopReason: "toolUse",
+				timestamp: Date.now(),
+			};
+			const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+			const assistant = messages.find(m => m.role === "assistant");
+			expect(assistant).toBeDefined();
+			expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+		},
+	);
+
+	it.each(["opencode-go", "opencode-zen"] as const)(
+		"does not require reasoning_content when %s model is not reasoning-capable",
+		provider => {
+			const compat = detectCompat(openCodeModel(provider, "some-other-model", false));
+			expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+		},
+	);
+
+	it.each(["kimi-k2.5", "kimi-k1.5", "kimi-k2-5"])("matches kimi model id pattern via opencode-zen: %s", id => {
+		const compat = detectCompat(openCodeModel("opencode-zen", id, true));
+ 		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+ 	});
+ 
+
+From 76c1fe9ee083836ecca43900fefc458c8cf4c4fb Mon Sep 17 00:00:00 2001
+From: sonhyrd <son.hong.do@hyrd.ai>
+Date: Mon, 27 Apr 2026 00:14:27 +0700
+Subject: [PATCH 3/5] test(ai): restore non-kimi coverage while adding
+ opencode-zen cases
+
+---
+ .../ai/test/openai-completions-compat.test.ts | 215 +++++++++++++-----
+ 1 file changed, 154 insertions(+), 61 deletions(-)
+
+diff --git a/packages/ai/test/openai-completions-compat.test.ts b/packages/ai/test/openai-completions-compat.test.ts
+index c743dd246..8b8cef393 100644
+--- a/packages/ai/test/openai-completions-compat.test.ts
+++ b/packages/ai/test/openai-completions-compat.test.ts
+@@ -282,38 +282,56 @@ describe("openai-completions compatibility", () => {
+ 	});
+ });
+ 
+-describe("opencode reasoning-content compatibility via detectCompat", () => {
+-	type OpenCodeProvider = "opencode-go" | "opencode-zen";
+describe("kimi model detection via detectCompat", () => {
+	function openCodeGoModel(id: string, reasoning = true): Model<"openai-completions"> {
+		return {
+			...getBundledModel("openai", "gpt-4o-mini"),
+			api: "openai-completions",
+			provider: "opencode-go",
+			baseUrl: "https://opencode.ai/zen/go/v1",
+			id,
+			reasoning,
+		};
+	}
+ 
+-	function openCodeModel(provider: OpenCodeProvider, id: string, reasoning = true): Model<"openai-completions"> {
+-		const baseUrl = provider === "opencode-go" ? "https://opencode.ai/zen/go/v1" : "https://opencode.ai/zen/v1";
+	function openCodeZenModel(id: string, reasoning = true): Model<"openai-completions"> {
+ 		return {
+ 			...getBundledModel("openai", "gpt-4o-mini"),
+ 			api: "openai-completions",
+-			provider,
+-			baseUrl,
+			provider: "opencode-zen",
+			baseUrl: "https://opencode.ai/zen/v1",
+ 			id,
+ 			reasoning,
+ 		};
+ 	}
+ 
+-	it.each(["opencode-go", "opencode-zen"] as const)(
+-		"requires reasoning_content for tool calls on kimi-k2.5 via %s",
+-		provider => {
+-			const compat = detectCompat(openCodeModel(provider, "kimi-k2.5", true));
+-			expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+-			expect(compat.requiresAssistantContentForToolCalls).toBe(true);
+-		},
+-	);
+-
+-	it.each(["opencode-go", "opencode-zen"] as const)(
+-		"requires reasoning_content for tool calls on reasoning DeepSeek models via %s",
+-		provider => {
+-			const compat = detectCompat(openCodeModel(provider, "deepseek-v4-pro", true));
+-			expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+-			expect(compat.requiresAssistantContentForToolCalls).toBe(false);
+-		},
+-	);
+	function kimiOpenCodeModel(id: string): Model<"openai-completions"> {
+		return openCodeGoModel(id, true);
+	}
+
+	it("requires reasoning_content for tool calls on kimi-k2.5 (opencode-go)", () => {
+		const compat = detectCompat(kimiOpenCodeModel("kimi-k2.5"));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+		expect(compat.requiresAssistantContentForToolCalls).toBe(true);
+	});
+
+	it("requires reasoning_content for tool calls on kimi-k2.5 (opencode-zen)", () => {
+		const compat = detectCompat(openCodeZenModel("kimi-k2.5", true));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+		expect(compat.requiresAssistantContentForToolCalls).toBe(true);
+	});
+
+	it("requires reasoning_content for tool calls on reasoning DeepSeek models via opencode-go", () => {
+		const compat = detectCompat(openCodeGoModel("deepseek-v4-pro", true));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+		expect(compat.requiresAssistantContentForToolCalls).toBe(false);
+	});
+
+	it("requires reasoning_content for tool calls on reasoning DeepSeek models via opencode-zen", () => {
+		const compat = detectCompat(openCodeZenModel("deepseek-v4-pro", true));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+		expect(compat.requiresAssistantContentForToolCalls).toBe(false);
+	});
+ 
+ 	it("requires reasoning_content when custom openai provider targets opencode zen baseUrl", () => {
+ 		const model: Model<"openai-completions"> = {
+@@ -328,45 +346,120 @@ describe("opencode reasoning-content compatibility via detectCompat", () => {
+ 		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+ 	});
+ 
+-	it.each(["opencode-go", "opencode-zen"] as const)(
+-		"injects reasoning_content placeholder for reasoning DeepSeek tool-call turns via %s",
+-		provider => {
+-			const model = openCodeModel(provider, "deepseek-v4-pro", true);
+-			const compat = detectCompat(model);
+-			const toolCallMessage: AssistantMessage = {
+-				role: "assistant",
+-				content: [{ type: "toolCall", id: `call_ds_${provider}`, name: "web_search", arguments: { query: "hi" } }],
+-				api: model.api,
+-				provider: model.provider,
+-				model: model.id,
+-				usage: {
+-					input: 0,
+-					output: 0,
+-					cacheRead: 0,
+-					cacheWrite: 0,
+-					totalTokens: 0,
+-					cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+	it("injects reasoning_content placeholder for reasoning DeepSeek tool-call turns via opencode-go", () => {
+		const model = openCodeGoModel("deepseek-v4-pro", true);
+		const compat = detectCompat(model);
+		const toolCallMessage: AssistantMessage = {
+			role: "assistant",
+			content: [{ type: "toolCall", id: "call_ds_go", name: "web_search", arguments: { query: "hi" } }],
+			api: model.api,
+			provider: model.provider,
+			model: model.id,
+			usage: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				totalTokens: 0,
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+			},
+			stopReason: "toolUse",
+			timestamp: Date.now(),
+		};
+		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+		const assistant = messages.find(m => m.role === "assistant");
+		expect(assistant).toBeDefined();
+		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+	});
+
+	it("injects reasoning_content placeholder for reasoning DeepSeek tool-call turns via opencode-zen", () => {
+		const model = openCodeZenModel("deepseek-v4-pro", true);
+		const compat = detectCompat(model);
+		const toolCallMessage: AssistantMessage = {
+			role: "assistant",
+			content: [{ type: "toolCall", id: "call_ds_zen", name: "web_search", arguments: { query: "hi" } }],
+			api: model.api,
+			provider: model.provider,
+			model: model.id,
+			usage: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				totalTokens: 0,
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+			},
+			stopReason: "toolUse",
+			timestamp: Date.now(),
+		};
+		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+		const assistant = messages.find(m => m.role === "assistant");
+		expect(assistant).toBeDefined();
+		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+	});
+
+	it("injects reasoning_content placeholder when assistant with tool calls has no reasoning field", () => {
+		const model = kimiOpenCodeModel("kimi-k2.5");
+		const compat = detectCompat(model);
+		const toolCallMessage: AssistantMessage = {
+			role: "assistant",
+			content: [
+				// Thinking returned as plain text (as kimi-k2.5 on opencode-go does)
+				{ type: "text", text: "Let me research this." },
+				{
+					type: "toolCall",
+					id: "call_abc123",
+					name: "web_search",
+					arguments: { query: "beads gastownhall" },
+ 				},
+-				stopReason: "toolUse",
+-				timestamp: Date.now(),
+-			};
+-			const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+-			const assistant = messages.find(m => m.role === "assistant");
+-			expect(assistant).toBeDefined();
+-			expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+-		},
+-	);
+-
+-	it.each(["opencode-go", "opencode-zen"] as const)(
+-		"does not require reasoning_content when %s model is not reasoning-capable",
+-		provider => {
+-			const compat = detectCompat(openCodeModel(provider, "some-other-model", false));
+-			expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+-		},
+-	);
+-
+-	it.each(["kimi-k2.5", "kimi-k1.5", "kimi-k2-5"])("matches kimi model id pattern via opencode-zen: %s", id => {
+-		const compat = detectCompat(openCodeModel("opencode-zen", id, true));
+			],
+			api: model.api,
+			provider: model.provider,
+			model: model.id,
+			usage: {
+				input: 0,
+				output: 0,
+				cacheRead: 0,
+				cacheWrite: 0,
+				totalTokens: 0,
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+			},
+			stopReason: "toolUse",
+			timestamp: Date.now(),
+		};
+		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
+		const assistant = messages.find(m => m.role === "assistant");
+		expect(assistant).toBeDefined();
+		const reasoningContent = Reflect.get(assistant as object, "reasoning_content");
+		expect(reasoningContent).toBeDefined();
+		expect(typeof reasoningContent).toBe("string");
+		expect((reasoningContent as string).length).toBeGreaterThan(0);
+	});
+
+	it("does not inject reasoning_content when model is not kimi", () => {
+		const model: Model<"openai-completions"> = {
+			...getBundledModel("openai", "gpt-4o-mini"),
+			api: "openai-completions",
+			provider: "opencode-go",
+			baseUrl: "https://opencode.ai/zen/go/v1",
+			id: "some-other-model",
+		};
+		const compat = detectCompat(model);
+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+	});
+
+	it("does not require reasoning_content when opencode-go model is not reasoning-capable", () => {
+		const compat = detectCompat(openCodeGoModel("some-other-model", false));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+	});
+
+	it("does not require reasoning_content when opencode-zen model is not reasoning-capable", () => {
+		const compat = detectCompat(openCodeZenModel("some-other-model", false));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+	});
+
+	it.each(["kimi-k2.5", "kimi-k1.5", "kimi-k2-5"])("matches kimi model id: %s", id => {
+		const compat = detectCompat(kimiOpenCodeModel(id));
+ 		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
+ 	});
+ 
+
+From 9c7a8958c682b16990504500551827320508087d Mon Sep 17 00:00:00 2001
+From: sonhyrd <son.hong.do@hyrd.ai>
+Date: Mon, 27 Apr 2026 00:29:48 +0700
+Subject: [PATCH 4/5] fix(ai/providers): gate reasoning_content stubs on
+ deepseek models
+
+---
+ .../providers/openai-completions-compat.ts    |  7 ++--
+ .../ai/src/providers/openai-completions.ts    |  4 +--
+ .../ai/test/openai-completions-compat.test.ts | 36 +++++++++++++++++++
+ 3 files changed, 42 insertions(+), 5 deletions(-)
+
+diff --git a/packages/ai/src/providers/openai-completions-compat.ts b/packages/ai/src/providers/openai-completions-compat.ts
+index b4825a31c..bba1cef70 100644
+--- a/packages/ai/src/providers/openai-completions-compat.ts
+++ b/packages/ai/src/providers/openai-completions-compat.ts
+@@ -54,6 +54,7 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
+ 	const isKimiModel = model.id.includes("moonshotai/kimi") || /^kimi[-.]/i.test(model.id);
+ 	const isAlibaba = provider === "alibaba-coding-plan" || baseUrl.includes("dashscope");
+ 	const isQwen = model.id.toLowerCase().includes("qwen");
+	const isDeepSeekModel = model.id.toLowerCase().includes("deepseek");
+ 	const isOpenRouter = provider === "openrouter" || baseUrl.includes("openrouter.ai");
+ 	const isOpenCode = provider === "opencode-zen" || provider === "opencode-go" || baseUrl.includes("opencode.ai/zen");
+ 
+@@ -109,12 +110,12 @@ export function detectOpenAICompat(model: Model<"openai-completions">, resolvedB
+ 		reasoningContentField: "reasoning_content",
+ 		// Backends that 400 follow-up requests when prior assistant tool-call turns lack `reasoning_content`:
+ 		//   - Kimi: documented invariant on its native API and via OpenCode.
+-		//   - Reasoning-capable models reached through OpenRouter or OpenCode (Zen/Go): DeepSeek V4 Pro,
+-		//     Kimi, and similar models can enforce this server-side whenever the request is in thinking mode.
+		//   - DeepSeek reasoning models reached through OpenRouter or OpenCode (Zen/Go): enforced when
+		//     thinking mode is enabled on those model families.
+ 		// We can't translate Anthropic's redacted/encrypted reasoning into DeepSeek's plaintext form, so
+ 		// cross-provider continuations rely on a placeholder — see `convertMessages` for injection rules.
+ 		requiresReasoningContentForToolCalls:
+-			isKimiModel || ((isOpenRouter || isOpenCode) && Boolean(model.reasoning)),
+			isKimiModel || (isDeepSeekModel && (isOpenRouter || isOpenCode) && Boolean(model.reasoning)),
+ 		requiresAssistantContentForToolCalls: isKimiModel,
+ 		openRouterRouting: undefined,
+ 		vercelGatewayRouting: undefined,
+diff --git a/packages/ai/src/providers/openai-completions.ts b/packages/ai/src/providers/openai-completions.ts
+index e25aeffb3..89a997a0f 100644
+--- a/packages/ai/src/providers/openai-completions.ts
+++ b/packages/ai/src/providers/openai-completions.ts
+@@ -1213,8 +1213,8 @@ export function convertMessages(
+ 			// Inject a `reasoning_content` placeholder on assistant tool-call turns when the backend
+ 			// rejects history without it. The compat flag captures the rule:
+ 			//   - Kimi (native or via OpenCode Zen/Go): chat completion endpoint demands the field.
+-			//   - Reasoning models reached through OpenRouter or OpenCode Zen/Go (e.g. DeepSeek V4 Pro):
+-			//     the upstream thinking-mode validator demands it on every prior assistant turn. omp
+			//   - DeepSeek reasoning models reached through OpenRouter or OpenCode Zen/Go: the upstream
+			//     thinking-mode validator demands it on every prior assistant turn. omp
+ 			//     cannot synthesize real reasoning when the conversation was warmed up by another
+ 			//     provider whose reasoning is redacted/encrypted (Anthropic) or simply absent, so we
+ 			//     emit a placeholder. Real captured reasoning, when present, is preserved earlier via
+diff --git a/packages/ai/test/openai-completions-compat.test.ts b/packages/ai/test/openai-completions-compat.test.ts
+index 8b8cef393..c083c2151 100644
+--- a/packages/ai/test/openai-completions-compat.test.ts
+++ b/packages/ai/test/openai-completions-compat.test.ts
+@@ -333,6 +333,29 @@ describe("kimi model detection via detectCompat", () => {
+ 		expect(compat.requiresAssistantContentForToolCalls).toBe(false);
+ 	});
+ 
+	it("does not require reasoning_content for non-DeepSeek reasoning models via opencode-go", () => {
+		const compat = detectCompat(openCodeGoModel("glm-5", true));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+	});
+
+	it("does not require reasoning_content for non-DeepSeek reasoning models via opencode-zen", () => {
+		const compat = detectCompat(openCodeZenModel("glm-5", true));
+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+	});
+
+	it("does not require reasoning_content when custom openai provider targets opencode zen baseUrl with non-DeepSeek model", () => {
+		const model: Model<"openai-completions"> = {
+			...getBundledModel("openai", "gpt-4o-mini"),
+			api: "openai-completions",
+			provider: "openai",
+			baseUrl: "https://opencode.ai/zen/v1",
+			id: "glm-5",
+			reasoning: true,
+		};
+		const compat = detectCompat(model);
+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+	});
+
+ 	it("requires reasoning_content when custom openai provider targets opencode zen baseUrl", () => {
+ 		const model: Model<"openai-completions"> = {
+ 			...getBundledModel("openai", "gpt-4o-mini"),
+@@ -453,6 +476,19 @@ describe("kimi model detection via detectCompat", () => {
+ 		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+ 	});
+ 
+	it("does not require reasoning_content for non-DeepSeek reasoning models via openrouter", () => {
+		const model: Model<"openai-completions"> = {
+			...getBundledModel("openai", "gpt-4o-mini"),
+			api: "openai-completions",
+			provider: "openrouter",
+			baseUrl: "https://openrouter.ai/api/v1",
+			id: "openai/gpt-4.1-mini",
+			reasoning: true,
+		};
+		const compat = detectCompat(model);
+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+	});
+
+ 	it("does not require reasoning_content when opencode-zen model is not reasoning-capable", () => {
+ 		const compat = detectCompat(openCodeZenModel("some-other-model", false));
+ 		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
+
+From 53a03286cf658bb4aeab67dad3246b7ba80cf244 Mon Sep 17 00:00:00 2001
+From: sonhyrd <son.hong.do@hyrd.ai>
+Date: Mon, 27 Apr 2026 00:52:22 +0700
+Subject: [PATCH 5/5] fix(ai/providers): set content when reasoning placeholder
+ is injected
+
+---
+ packages/ai/src/providers/openai-completions.ts    | 3 ++-
+ packages/ai/test/openai-completions-compat.test.ts | 2 ++
+ 2 files changed, 4 insertions(+), 1 deletion(-)
+
+diff --git a/packages/ai/src/providers/openai-completions.ts b/packages/ai/src/providers/openai-completions.ts
+index 89a997a0f..b490e254e 100644
+--- a/packages/ai/src/providers/openai-completions.ts
+++ b/packages/ai/src/providers/openai-completions.ts
+@@ -1206,7 +1206,7 @@ export function convertMessages(
+ 			}
+ 
+ 			const toolCalls = msg.content.filter(b => b.type === "toolCall") as ToolCall[];
+-			const hasReasoningField =
+			let hasReasoningField =
+ 				(assistantMsg as any).reasoning_content !== undefined ||
+ 				(assistantMsg as any).reasoning !== undefined ||
+ 				(assistantMsg as any).reasoning_text !== undefined;
+@@ -1227,6 +1227,7 @@ export function convertMessages(
+ 			if (toolCalls.length > 0 && stubsReasoningContent && !hasReasoningField) {
+ 				const reasoningField = compat.reasoningContentField ?? "reasoning_content";
+ 				(assistantMsg as any)[reasoningField] = ".";
+				hasReasoningField = true;
+ 			}
+ 			if (toolCalls.length > 0) {
+ 				assistantMsg.tool_calls = toolCalls.map((tc, toolCallIndex) => {
+diff --git a/packages/ai/test/openai-completions-compat.test.ts b/packages/ai/test/openai-completions-compat.test.ts
+index c083c2151..8efae899a 100644
+--- a/packages/ai/test/openai-completions-compat.test.ts
+++ b/packages/ai/test/openai-completions-compat.test.ts
+@@ -393,6 +393,7 @@ describe("kimi model detection via detectCompat", () => {
+ 		const assistant = messages.find(m => m.role === "assistant");
+ 		expect(assistant).toBeDefined();
+ 		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+		expect(Reflect.get(assistant as object, "content")).toBe("");
+ 	});
+ 
+ 	it("injects reasoning_content placeholder for reasoning DeepSeek tool-call turns via opencode-zen", () => {
+@@ -419,6 +420,7 @@ describe("kimi model detection via detectCompat", () => {
+ 		const assistant = messages.find(m => m.role === "assistant");
+ 		expect(assistant).toBeDefined();
+ 		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
+		expect(Reflect.get(assistant as object, "content")).toBe("");
+ 	});
+ 
+ 	it("injects reasoning_content placeholder when assistant with tool calls has no reasoning field", () => {
--- a/patches/omp/0001-openai-completions-retry-without-strict-on-deepseek-openrouter.patch
+++ b/patches/omp/0001-openai-completions-retry-without-strict-on-deepseek-openrouter.patch
@@ -1,126 +0,0 @@
-Subject: [PATCH] fix(openai-completions): retry without strict tools for DeepSeek-via-OpenRouter anyOf rejections
-
-The retry-on-strict-tool-error path in openai-completions failed to recover when
-DeepSeek (and similar backends fronted by OpenRouter) reject strict-mode tool
-schemas with errors of the form:
-
-  Invalid tool parameters schema : field `anyOf`: missing field `type`
-
-Two reasons:
-
-1. Retry only triggered in "all_strict" mode. OpenRouter defaults to "mixed"
-   (per-tool strict), so the early return prevented retry.
-2. The error-message regex required "strict" near "tool". DeepSeek's message
-   never mentions "strict".
-
-Fix:
- Allow retry whenever any tool was sent with strict (i.e. mode != "none").
- Recognize "Invalid tool parameters" in the regex.
-
-Includes a regression test reproducing the exact DeepSeek error body via
-OpenRouter mixed-strict mode.
-
-Applies cleanly against v14.2.1.
-
---
-diff --git a/packages/ai/src/providers/openai-completions.ts b/packages/ai/src/providers/openai-completions.ts
-index e58189607..3c20631c1 100644
--- a/packages/ai/src/providers/openai-completions.ts
-+++ b/packages/ai/src/providers/openai-completions.ts
-@@ -1245,7 +1245,10 @@ function shouldRetryWithoutStrictTools(
- 	toolStrictMode: AppliedToolStrictMode,
- 	tools: Tool[] | undefined,
- ): boolean {
-	if (!tools || tools.length === 0 || toolStrictMode !== "all_strict") {
-+	// Retry whenever any tool was sent with `strict: true`. OpenRouter routes to underlying
-+	// providers (e.g. DeepSeek) whose schema validators reject the strict-mode `anyOf` shape
-+	// even when omp emitted strict per-tool ("mixed"), not just provider-wide ("all_strict").
-+	if (!tools || tools.length === 0 || toolStrictMode === "none") {
- 		return false;
- 	}
- 	const status = extractHttpStatusFromError(error) ?? capturedErrorResponse?.status;
-@@ -1255,7 +1258,14 @@ function shouldRetryWithoutStrictTools(
- 	const messageParts = [error instanceof Error ? error.message : undefined, capturedErrorResponse?.bodyText]
- 		.filter((value): value is string => typeof value === "string" && value.trim().length > 0)
- 		.join("\n");
-	return /wrong_api_format|mixed values for 'strict'|tool[s]?\b.*strict|\bstrict\b.*tool/i.test(messageParts);
-+	// Patterns:
-+	// - `wrong_api_format`, `mixed values for 'strict'`: OpenAI rejecting mixed strict flags.
-+	// - `tool ... strict` / `strict ... tool`: generic strict-tool complaints.
-+	// - `Invalid tool parameters schema`: DeepSeek (via OpenRouter) rejecting strict-mode
-+	//   nullable unions because their validator demands `type` alongside `anyOf`.
-+	return /wrong_api_format|mixed values for 'strict'|tool[s]?\b.*strict|\bstrict\b.*tool|invalid tool parameters/i.test(
-+		messageParts,
-+	);
- }
- 
- function mapStopReason(reason: ChatCompletionChunk.Choice["finish_reason"] | string): {
-diff --git a/packages/ai/test/openai-tool-strict-mode.test.ts b/packages/ai/test/openai-tool-strict-mode.test.ts
-index 2bf17e6d8..24d5a09d5 100644
--- a/packages/ai/test/openai-tool-strict-mode.test.ts
-+++ b/packages/ai/test/openai-tool-strict-mode.test.ts
-@@ -231,6 +231,64 @@ describe("OpenAI tool strict mode", () => {
- 		expect(result.content).toContainEqual({ type: "text", text: "Hello" });
- 		expect(strictFlags).toEqual([[true], [false]]);
- 	});
-+	it("retries with non-strict tool schemas when OpenRouter backend rejects strict anyOf nullable unions", async () => {
-+		// Reproduces deepseek/deepseek-v4-pro via OpenRouter rejecting the strict-mode schema with:
-+		//   400 Provider returned error
-+		//   {"error":{"message":"Invalid tool parameters schema : field `anyOf`: missing field `type`",...}}
-+		// OpenRouter is in mixed-strict mode by default (per-tool strict), so the original retry condition
-+		// (only "all_strict") prevented recovery. The retry now triggers whenever any tool sent strict=true.
-+		const model = getBundledModel("openrouter", "deepseek/deepseek-v3.2") as Model<"openai-completions">;
-+		const strictFlags: boolean[][] = [];
-+		global.fetch = Object.assign(
-+			async (_input: string | URL | Request, init?: RequestInit): Promise<Response> => {
-+				const bodyText = typeof init?.body === "string" ? init.body : "";
-+				const payload = JSON.parse(bodyText) as {
-+					tools?: Array<{ function?: { strict?: boolean } }>;
-+				};
-+				strictFlags.push((payload.tools ?? []).map(tool => tool.function?.strict === true));
-+				if (strictFlags.length === 1) {
-+					return new Response(
-+						JSON.stringify({
-+							error: {
-+								message: "Invalid tool parameters schema : field `anyOf`: missing field `type`",
-+								type: "invalid_request_error",
-+								param: null,
-+								code: "invalid_request_error",
-+							},
-+						}),
-+						{
-+							status: 400,
-+							headers: { "content-type": "application/json" },
-+						},
-+					);
-+				}
-+				return createSseResponse([
-+					{
-+						id: "chatcmpl-or",
-+						object: "chat.completion.chunk",
-+						created: 0,
-+						model: model.id,
-+						choices: [{ index: 0, delta: { content: "Hello" } }],
-+					},
-+					{
-+						id: "chatcmpl-or",
-+						object: "chat.completion.chunk",
-+						created: 0,
-+						model: model.id,
-+						choices: [{ index: 0, delta: {}, finish_reason: "stop" }],
-+					},
-+					"[DONE]",
-+				]);
-+			},
-+			{ preconnect: originalFetch.preconnect },
-+		);
-+
-+		const result = await streamOpenAICompletions(model, testContext, { apiKey: "test-key" }).result();
-+		expect(result.stopReason).toBe("stop");
-+		expect(result.content).toContainEqual({ type: "text", text: "Hello" });
-+		expect(strictFlags).toEqual([[true], [false]]);
-+	});
-+
- 
- 	it("sends strict=true for openai-responses tool schemas on OpenAI", async () => {
- 		const model = getBundledModel("openai", "gpt-5-mini") as Model<"openai-responses">;
--- a/patches/omp/0002-openai-completions-stub-reasoning-content-for-openrouter.patch
+++ b/patches/omp/0002-openai-completions-stub-reasoning-content-for-openrouter.patch
@@ -1,233 +0,0 @@
-Subject: [PATCH] fix(openai-completions): require `reasoning_content` for OpenRouter reasoning models
-
-DeepSeek V4 Pro (and similar reasoning models reached via OpenRouter) reject
-multi-turn requests in thinking mode with:
-
-  400 The `reasoning_content` in the thinking mode must be passed back to
-  the API.
-
-omp's existing kimi placeholder injection (`requiresReasoningContentForToolCalls`)
-covered this requirement only for `thinkingFormat === "openai"`. OpenRouter
-sets `thinkingFormat === "openrouter"`, so the gate never fired even though
-the underlying providers behind OpenRouter (DeepSeek, Kimi, etc.) all enforce
-the same invariant.
-
-This patch:
-
-1. Extends `requiresReasoningContentForToolCalls` detection: any
-   reasoning-capable model fronted by OpenRouter now sets the flag.
-2. Extends the placeholder gate in `convertMessages` to accept
-   `thinkingFormat === "openrouter"` alongside `"openai"`.
-
-Cross-provider continuations are the dominant trigger: a conversation warmed
-up by Anthropic Claude (whose reasoning is redacted/encrypted on the wire)
-followed by a switch to DeepSeek V4 Pro via OpenRouter. omp cannot
-synthesize plaintext `reasoning_content` from Anthropic's encrypted blocks,
-so the placeholder satisfies DeepSeek's validator without fabricating a
-reasoning trace. Real captured reasoning, when present, short-circuits the
-placeholder via `hasReasoningField` and survives intact.
-
-Side benefit: also closes a latent gap where Kimi-via-OpenRouter
-(`thinkingFormat === "openrouter"`) had the compat flag set but the
-placeholder gate silently rejected it.
-
-Regression tests cover:
-  - compat flag detection on OpenRouter reasoning models
-  - opt-out for non-reasoning OpenRouter models
-  - cross-provider redacted-thinking placeholder
-  - Kimi-via-OpenRouter placeholder firing
-  - real reasoning preserved over the placeholder
-
-Applies cleanly on top of patch 0001.
-
---
-diff --git a/packages/ai/src/providers/openai-completions-compat.ts b/packages/ai/src/providers/openai-completions-compat.ts
--- a/packages/ai/src/providers/openai-completions-compat.ts
-+++ b/packages/ai/src/providers/openai-completions-compat.ts
-@@ -105,7 +105,14 @@
- 					? "qwen"
- 					: "openai",
- 		reasoningContentField: "reasoning_content",
-		requiresReasoningContentForToolCalls: isKimiModel,
-+		// Backends that 400 follow-up requests when prior assistant tool-call turns lack `reasoning_content`:
-+		//   - Kimi: documented invariant on its native API and via OpenCode-Go.
-+		//   - Any reasoning-capable model reached through OpenRouter: DeepSeek V4 Pro and similar enforce
-+		//     this server-side whenever the request is in thinking mode. We can't translate Anthropic's
-+		//     redacted/encrypted reasoning into DeepSeek's plaintext form, so cross-provider continuations
-+		//     rely on a placeholder — see `convertMessages` for the placeholder injection.
-+		requiresReasoningContentForToolCalls:
-+			isKimiModel || ((provider === "openrouter" || baseUrl.includes("openrouter.ai")) && Boolean(model.reasoning)),
- 		requiresAssistantContentForToolCalls: isKimiModel,
- 		openRouterRouting: undefined,
- 		vercelGatewayRouting: undefined,
-diff --git a/packages/ai/src/providers/openai-completions.ts b/packages/ai/src/providers/openai-completions.ts
--- a/packages/ai/src/providers/openai-completions.ts
-+++ b/packages/ai/src/providers/openai-completions.ts
-@@ -1059,12 +1059,21 @@
- 				(assistantMsg as any).reasoning_content !== undefined ||
- 				(assistantMsg as any).reasoning !== undefined ||
- 				(assistantMsg as any).reasoning_text !== undefined;
-			if (
-				toolCalls.length > 0 &&
-+			// Inject a `reasoning_content` placeholder on assistant tool-call turns when the backend
-+			// rejects history without it. The compat flag captures the rule:
-+			//   - Kimi (native or via OpenCode-Go): chat completion endpoint demands the field.
-+			//   - Reasoning models reached through OpenRouter (e.g. DeepSeek V4 Pro): the underlying
-+			//     provider's thinking-mode validator demands it on every prior assistant turn. omp
-+			//     cannot synthesize real reasoning when the conversation was warmed up by another
-+			//     provider whose reasoning is redacted/encrypted (Anthropic) or simply absent, so we
-+			//     emit a placeholder. Real captured reasoning, when present, is preserved earlier via
-+			//     the `thinkingSignature` echo path and short-circuits via `hasReasoningField`.
-+			// `thinkingFormat` is gated to formats that consume the field (openai/openrouter chat
-+			// completions); formats with their own conventions (zai, qwen) are excluded.
-+			const stubsReasoningContent =
- 				compat.requiresReasoningContentForToolCalls &&
-				compat.thinkingFormat === "openai" &&
-				!hasReasoningField
-			) {
-+				(compat.thinkingFormat === "openai" || compat.thinkingFormat === "openrouter");
-+			if (toolCalls.length > 0 && stubsReasoningContent && !hasReasoningField) {
- 				const reasoningField = compat.reasoningContentField ?? "reasoning_content";
- 				(assistantMsg as any)[reasoningField] = ".";
- 			}
-diff --git a/packages/ai/test/openai-completions-compat.test.ts b/packages/ai/test/openai-completions-compat.test.ts
--- a/packages/ai/test/openai-completions-compat.test.ts
-+++ b/packages/ai/test/openai-completions-compat.test.ts
-@@ -367,4 +367,137 @@
- 		const compat = detectCompat(model);
- 		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
- 	});
-+
-+	it("requires reasoning_content for tool calls on reasoning-capable models via OpenRouter", () => {
-+		const model: Model<"openai-completions"> = {
-+			...(getBundledModel("openrouter", "deepseek/deepseek-v3.2") as Model<"openai-completions">),
-+			reasoning: true,
-+		};
-+		const compat = detectCompat(model);
-+		expect(compat.thinkingFormat).toBe("openrouter");
-+		expect(compat.requiresReasoningContentForToolCalls).toBe(true);
-+	});
-+
-+	it("does not require reasoning_content for non-reasoning OpenRouter models", () => {
-+		const model: Model<"openai-completions"> = {
-+			...(getBundledModel("openrouter", "deepseek/deepseek-v3.2") as Model<"openai-completions">),
-+			reasoning: false,
-+		};
-+		const compat = detectCompat(model);
-+		expect(compat.requiresReasoningContentForToolCalls).toBe(false);
-+	});
-+
-+	it("injects reasoning_content placeholder for OpenRouter reasoning models lacking captured reasoning", () => {
-+		// Reproduces the failing path from real usage: a conversation generated under Anthropic Claude (whose
-+		// reasoning is redacted/encrypted) is continued with deepseek/deepseek-v4-pro via OpenRouter. The
-+		// prior assistant turns persist as ThinkingContent blocks with empty `thinking` text plus an opaque
-+		// Anthropic signature cookie. omp cannot translate that into DeepSeek's plain-text `reasoning_content`,
-+		// so the empty thinking block is filtered out and the placeholder fires — satisfying DeepSeek's
-+		// thinking-mode validator without fabricating a reasoning trace.
-+		const model: Model<"openai-completions"> = {
-+			...(getBundledModel("openrouter", "deepseek/deepseek-v3.2") as Model<"openai-completions">),
-+			reasoning: true,
-+		};
-+		const compat = detectCompat(model);
-+		const toolCallMessage: AssistantMessage = {
-+			role: "assistant",
-+			content: [
-+				// Anthropic-style redacted thinking block: empty text plus opaque signature.
-+				// `thinking.trim().length === 0` filters this out before the signature echo can fire.
-+				{ type: "thinking", thinking: "", thinkingSignature: "Ep4CClkIDRgCKkDOpaqueAnthropicCookie" },
-+				{ type: "toolCall", id: "call_anth_to_ds", name: "web_search", arguments: { query: "hi" } },
-+			],
-+			api: model.api,
-+			provider: model.provider,
-+			model: model.id,
-+			usage: {
-+				input: 0,
-+				output: 0,
-+				cacheRead: 0,
-+				cacheWrite: 0,
-+				totalTokens: 0,
-+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
-+			},
-+			stopReason: "toolUse",
-+			timestamp: Date.now(),
-+		};
-+		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
-+		const assistant = messages.find(m => m.role === "assistant");
-+		expect(assistant).toBeDefined();
-+		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
-+	});
-+
-+	it("injects reasoning_content placeholder for kimi-k2-5 via OpenRouter (closes the kimi-via-openrouter gap)", () => {
-+		// Before this fix, `requiresReasoningContentForToolCalls` was true for Kimi via OpenRouter but the
-+		// stub gate only fired when `thinkingFormat === "openai"`. OpenRouter sets thinkingFormat="openrouter",
-+		// so the stub silently never fired and Kimi-via-OpenRouter conversations 400'd the same way.
-+		const model: Model<"openai-completions"> = {
-+			...getBundledModel("openai", "gpt-4o-mini"),
-+			api: "openai-completions",
-+			provider: "openrouter",
-+			baseUrl: "https://openrouter.ai/api/v1",
-+			id: "moonshotai/kimi-k2-5",
-+			reasoning: true,
-+		};
-+		const compat = detectCompat(model);
-+		const toolCallMessage: AssistantMessage = {
-+			role: "assistant",
-+			content: [
-+				{ type: "toolCall", id: "call_kimi_or", name: "web_search", arguments: { query: "hi" } },
-+			],
-+			api: model.api,
-+			provider: model.provider,
-+			model: model.id,
-+			usage: {
-+				input: 0,
-+				output: 0,
-+				cacheRead: 0,
-+				cacheWrite: 0,
-+				totalTokens: 0,
-+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
-+			},
-+			stopReason: "toolUse",
-+			timestamp: Date.now(),
-+		};
-+		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
-+		const assistant = messages.find(m => m.role === "assistant");
-+		expect(assistant).toBeDefined();
-+		expect(Reflect.get(assistant as object, "reasoning_content")).toBe(".");
-+	});
-+
-+	it("preserves real captured reasoning over the placeholder when the assistant has non-empty thinking", () => {
-+		// Sanity check: the placeholder must not overwrite real reasoning. When the prior assistant turn was
-+		// generated by the same provider and surfaces plaintext reasoning, the existing thinkingSignature
-+		// echo path sets `reasoning_content` first, and `hasReasoningField` short-circuits the stub.
-+		const model: Model<"openai-completions"> = {
-+			...(getBundledModel("openrouter", "deepseek/deepseek-v3.2") as Model<"openai-completions">),
-+			reasoning: true,
-+		};
-+		const compat = detectCompat(model);
-+		const toolCallMessage: AssistantMessage = {
-+			role: "assistant",
-+			content: [
-+				{ type: "thinking", thinking: "Step 1: read the file. Step 2: search.", thinkingSignature: "reasoning_content" },
-+				{ type: "toolCall", id: "call_real", name: "web_search", arguments: { query: "hi" } },
-+			],
-+			api: model.api,
-+			provider: model.provider,
-+			model: model.id,
-+			usage: {
-+				input: 0,
-+				output: 0,
-+				cacheRead: 0,
-+				cacheWrite: 0,
-+				totalTokens: 0,
-+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
-+			},
-+			stopReason: "toolUse",
-+			timestamp: Date.now(),
-+		};
-+		const messages = convertMessages(model, { messages: [toolCallMessage] }, compat);
-+		const assistant = messages.find(m => m.role === "assistant");
-+		expect(assistant).toBeDefined();
-+		expect(Reflect.get(assistant as object, "reasoning_content")).toBe("Step 1: read the file. Step 2: search.");
-+	});
-+
- });
--- a/services/jellyfin/jellyfin-qbittorrent-monitor.py
+++ b/services/jellyfin/jellyfin-qbittorrent-monitor.py
@@ -38,6 +38,7 @@ class JellyfinQBittorrentMonitor:
        stream_bitrate_headroom=1.1,
        webhook_port=0,
        webhook_bind="127.0.0.1",
+        gateway_ip=None,
    ):
        self.jellyfin_url = jellyfin_url
        self.qbittorrent_url = qbittorrent_url
@@ -77,6 +78,15 @@ class JellyfinQBittorrentMonitor:
            ipaddress.ip_network("fe80::/10"),  # IPv6 link-local
        ]

+        # Hairpin marker. When a LAN client reaches Jellyfin via the public
+        # hostname, the router NAT-loopbacks the packet and SNATs the source
+        # to itself — the session arrives looking local but still costs WAN
+        # bandwidth. Sessions whose source equals the gateway must therefore
+        # NOT be skipped. None disables the check (pre-hairpin-aware behavior).
+        if gateway_ip is None:
+            gateway_ip = self._discover_default_gateway()
+        self.gateway_ip = gateway_ip
+
    def is_local_ip(self, ip_address: str) -> bool:
        """Check if an IP address is from a local network"""
        try:
@@ -86,6 +96,39 @@ class JellyfinQBittorrentMonitor:
            logger.warning(f"Invalid IP address format: {ip_address}")
            return True  # Treat invalid IPs as local for safety

+    def _discover_default_gateway(self) -> str | None:
+        """Read the IPv4 default gateway from /proc/net/route, or None."""
+        try:
+            with open("/proc/net/route") as f:
+                next(f)  # skip header
+                for line in f:
+                    fields = line.split()
+                    if len(fields) < 8 or fields[1] != "00000000":
+                        continue
+                    flags = int(fields[3], 16)
+                    if not flags & 0x2:  # RTF_GATEWAY
+                        continue
+                    gw_bytes = bytes.fromhex(fields[2])[::-1]  # little-endian
+                    if len(gw_bytes) != 4:
+                        continue
+                    return ".".join(str(b) for b in gw_bytes)
+        except (OSError, ValueError) as e:
+            logger.warning(f"Could not autodetect default gateway: {e}")
+        return None
+
+    def is_skippable(self, ip_address: str) -> bool:
+        """True iff this source IP can be ignored when deciding to throttle.
+
+        Truly LAN-direct sessions are skippable (no WAN cost). Hairpin-NAT'd
+        LAN sessions arrive with the LAN gateway as their source — those still
+        cost WAN bandwidth and must NOT be skipped.
+        """
+        if not self.is_local_ip(ip_address):
+            return False
+        if self.gateway_ip and ip_address == self.gateway_ip:
+            return False
+        return True
+
    def signal_handler(self, signum, frame):
        logger.info("Received shutdown signal, cleaning up...")
        self.running = False
@@ -164,7 +207,7 @@ class JellyfinQBittorrentMonitor:
            if (
                "NowPlayingItem" in session
                and not session.get("PlayState", {}).get("IsPaused", True)
-                and not self.is_local_ip(session.get("RemoteEndPoint", ""))
+                and not self.is_skippable(session.get("RemoteEndPoint", ""))
            ):
                item = session["NowPlayingItem"]
                item_type = item.get("Type", "").lower()
@@ -354,6 +397,9 @@ class JellyfinQBittorrentMonitor:
        logger.info(f"Default stream bitrate: {self.default_stream_bitrate} bps")
        logger.info(f"Minimum torrent speed: {self.min_torrent_speed} KB/s")
        logger.info(f"Stream bitrate headroom: {self.stream_bitrate_headroom}x")
+        logger.info(
+            f"LAN gateway (hairpin marker): {self.gateway_ip or 'none / autodetect failed'}"
+        )
        if self.webhook_port:
            logger.info(f"Webhook receiver: {self.webhook_bind}:{self.webhook_port}")

@@ -484,6 +530,7 @@ if __name__ == "__main__":
    stream_bitrate_headroom = float(os.getenv("STREAM_BITRATE_HEADROOM", "1.1"))
    webhook_port = int(os.getenv("WEBHOOK_PORT", "0"))
    webhook_bind = os.getenv("WEBHOOK_BIND", "127.0.0.1")
+    gateway_ip = os.getenv("LAN_GATEWAY_IP") or None

    monitor = JellyfinQBittorrentMonitor(
        jellyfin_url=jellyfin_url,
@@ -499,6 +546,7 @@ if __name__ == "__main__":
        stream_bitrate_headroom=stream_bitrate_headroom,
        webhook_port=webhook_port,
        webhook_bind=webhook_bind,
+        gateway_ip=gateway_ip,
    )

    monitor.run()
--- a/tests/jellyfin-qbittorrent-monitor.nix
+++ b/tests/jellyfin-qbittorrent-monitor.nix
@@ -428,6 +428,73 @@ pkgs.testers.runNixOSTest {
        local_playback["PositionTicks"] = 50000000
        server.succeed(f"curl -sf -X POST 'http://localhost:8096/Sessions/Playing/Stopped' -d '{json.dumps(local_playback)}' -H 'Content-Type:application/json' -H 'X-Emby-Authorization:{local_auth}, Token={local_token}'")

+    with subtest("Hairpin'd LAN session (source IP = configured gateway) DOES throttle"):
+        # Simulates a LAN client reaching Jellyfin via the public hostname:
+        # the router SNATs the source to itself, so Jellyfin sees the gateway
+        # IP and IsInLocalNetwork=True even though WAN bandwidth is in play.
+        # We use 127.0.0.1 as the "gateway" in this VM because the localhost
+        # curl below produces source 127.0.0.1 from Jellyfin's view.
+        server.succeed("systemctl stop monitor-test || true")
+        time.sleep(1)
+        server.succeed(f"""
+          systemd-run --unit=monitor-hairpin \
+            --setenv=JELLYFIN_URL=http://localhost:8096 \
+            --setenv=JELLYFIN_API_KEY={token} \
+            --setenv=QBITTORRENT_URL=http://localhost:8080 \
+            --setenv=CHECK_INTERVAL=1 \
+            --setenv=STREAMING_START_DELAY=1 \
+            --setenv=STREAMING_STOP_DELAY=1 \
+            --setenv=TOTAL_BANDWIDTH_BUDGET=50000000 \
+            --setenv=SERVICE_BUFFER=2000000 \
+            --setenv=DEFAULT_STREAM_BITRATE=10000000 \
+            --setenv=MIN_TORRENT_SPEED=100 \
+            --setenv=LAN_GATEWAY_IP=127.0.0.1 \
+            {python} {monitor}
+        """)
+        time.sleep(2)
+        assert not is_throttled(), "Should start unthrottled (no streams yet)"
+
+        hairpin_auth = 'MediaBrowser Client="Hairpin Client", DeviceId="hairpin-2222", Device="HairpinDevice", Version="1.0"'
+        hairpin_auth_result = json.loads(server.succeed(
+            f"curl -sf -X POST 'http://localhost:8096/Users/AuthenticateByName' -d '@${jfLib.payloads.auth}' -H 'Content-Type:application/json' -H 'X-Emby-Authorization:{hairpin_auth}'"
+        ))
+        hairpin_token = hairpin_auth_result["AccessToken"]
+
+        hairpin_playback = {
+            "ItemId": movie_id,
+            "MediaSourceId": media_source_id,
+            "PlaySessionId": "test-play-session-hairpin",
+            "CanSeek": True,
+            "IsPaused": False,
+        }
+        server.succeed(f"curl -sf -X POST 'http://localhost:8096/Sessions/Playing' -d '{json.dumps(hairpin_playback)}' -H 'Content-Type:application/json' -H 'X-Emby-Authorization:{hairpin_auth}, Token={hairpin_token}'")
+        time.sleep(3)
+        assert is_throttled(), "Hairpin'd session (source=gateway) should throttle even though source is RFC1918"
+
+        # Cleanup: stop the playback and the override-monitor, restore the normal one.
+        hairpin_playback["PositionTicks"] = 50000000
+        server.succeed(f"curl -sf -X POST 'http://localhost:8096/Sessions/Playing/Stopped' -d '{json.dumps(hairpin_playback)}' -H 'Content-Type:application/json' -H 'X-Emby-Authorization:{hairpin_auth}, Token={hairpin_token}'")
+        time.sleep(2)
+        assert not is_throttled(), "Should unthrottle after hairpin'd playback stops"
+
+        server.succeed("systemctl stop monitor-hairpin || true")
+        time.sleep(1)
+        server.succeed(f"""
+          systemd-run --unit=monitor-test \
+            --setenv=JELLYFIN_URL=http://localhost:8096 \
+            --setenv=JELLYFIN_API_KEY={token} \
+            --setenv=QBITTORRENT_URL=http://localhost:8080 \
+            --setenv=CHECK_INTERVAL=1 \
+            --setenv=STREAMING_START_DELAY=1 \
+            --setenv=STREAMING_STOP_DELAY=1 \
+            --setenv=TOTAL_BANDWIDTH_BUDGET=50000000 \
+            --setenv=SERVICE_BUFFER=2000000 \
+            --setenv=DEFAULT_STREAM_BITRATE=10000000 \
+            --setenv=MIN_TORRENT_SPEED=100 \
+            {python} {monitor}
+        """)
+        time.sleep(2)
+
    # === WEBHOOK TESTS ===
    #
    # Configure the Jellyfin Webhook plugin to target the monitor, then verify
@@ -589,7 +656,7 @@ pkgs.testers.runNixOSTest {
        server.succeed("systemctl restart jellyfin.service")
        server.wait_for_unit("jellyfin.service")
        server.wait_for_open_port(8096)
-        server.wait_until_succeeds("curl -sf http://localhost:8096/health | grep -q Healthy", timeout=60)
+        server.wait_until_succeeds("curl -sf http://localhost:8096/health | grep -q Healthy", timeout=180)

        # During Jellyfin restart, monitor can't reach Jellyfin
        # After restart, sessions are cleared - monitor should eventually unthrottle
@@ -645,7 +712,7 @@ pkgs.testers.runNixOSTest {
        server.succeed("systemctl start jellyfin.service")
        server.wait_for_unit("jellyfin.service")
        server.wait_for_open_port(8096)
-        server.wait_until_succeeds("curl -sf http://localhost:8096/health | grep -q Healthy", timeout=60)
+        server.wait_until_succeeds("curl -sf http://localhost:8096/health | grep -q Healthy", timeout=180)

        # After Jellyfin comes back, sessions are gone - should unthrottle
        time.sleep(3)
Author	SHA1	Message	Date
Simon Gardling	bbdc478e84	omp: update patches All checks were successful Build and Deploy / mreow (push) Successful in 13m8s Details Build and Deploy / yarn (push) Successful in 1m11s Details Build and Deploy / muffin (push) Successful in 7m15s Details	2026-04-27 01:36:08 -04:00
Simon Gardling	675fc7f805	update Some checks failed Build and Deploy / mreow (push) Failing after 5m10s Details Build and Deploy / yarn (push) Failing after 1m1s Details Build and Deploy / muffin (push) Has been cancelled Details	2026-04-27 01:27:13 -04:00
Simon Gardling	141754ca39	ghostty: fix??? All checks were successful Build and Deploy / mreow (push) Successful in 1m20s Details Build and Deploy / yarn (push) Successful in 54s Details Build and Deploy / muffin (push) Successful in 1m14s Details	2026-04-26 01:11:09 -04:00
Simon Gardling	4b173ef164	jellyfin-qbittorrent-monitor: fix hairpin handling	2026-04-26 01:03:11 -04:00
Simon Gardling	3201b5726e	update Some checks failed Build and Deploy / mreow (push) Successful in 1m44s Details Build and Deploy / yarn (push) Successful in 1m3s Details Build and Deploy / muffin (push) Failing after 27s Details	2026-04-26 00:12:30 -04:00