diff --git a/.roo/skills/probe-vscode-lm-api/SKILL.md b/.roo/skills/probe-vscode-lm-api/SKILL.md
new file mode 100644
index 0000000000..0150cd8afa
--- /dev/null
+++ b/.roo/skills/probe-vscode-lm-api/SKILL.md
@@ -0,0 +1,99 @@
+---
+name: probe-vscode-lm-api
+description: How to empirically probe the VS Code Language Model API (`vscode.lm`) with a scratch extension against a real extension host, and the measured findings about Copilot Claude models leaking tool-call markup into text. Use when asked to "test the vscode.lm API", "probe Copilot model behavior", "capture a raw LM transcript", "does the model leak tool-call markup", "verify LanguageModelToolCallPart behavior", or when reasoning about `extractLeakedToolCalls()` in the vscode-lm provider.
+---
+
+# Probing the VS Code LM API Empirically
+
+## When to Use This Skill
+
+- A claim is being made about what a Copilot-backed model _actually_ emits over `vscode.lm` (tool-call parts vs. text), and it needs evidence rather than inference.
+- Changing [`extractLeakedToolCalls()`](../../../src/api/providers/vscode-lm.ts) or its guards, and you need a false-positive corpus.
+- Any question that can only be answered by real `model.sendRequest()` traffic — the mocked unit tests cannot answer it.
+
+## When NOT to Use This Skill
+
+- Ordinary provider work covered by [`src/api/providers/__tests__/vscode-lm.spec.ts`](../../../src/api/providers/__tests__/vscode-lm.spec.ts). Live probing is slow and burns Copilot quota.
+- Anything about non-`vscode-lm` providers. Anthropic-API behavior does not transfer.
+
+## Running the Probe
+
+Scripts live in [`scripts/probe-vscode-lm-api/`](../../../scripts/probe-vscode-lm-api/) at the repo root.
+
+1. Copy `scripts/probe-vscode-lm-api/package.json` and `scripts/probe-vscode-lm-api/extension.js` into a scratch directory, e.g. `<repo>\.tmp\lmprobe\`. No build, no `npm install` — it is plain CommonJS against the `vscode` module.
+2. Adjust `OUT_DIR` at the top of the copied `extension.js` (or set `LM_PROBE_OUT_DIR`) to the transcript output directory.
+3. Launch a **new** extension host window:
+
+```
+code --extensionDevelopmentPath=<repo>\.tmp\lmprobe --new-window <repo>
+```
+
+4. In that new window: `Ctrl+Shift+P` -> **LM Probe: Run** (or click "Run probe" on the toast).
+5. Wait for the completion notification. Transcripts and `summary.json` land in `OUT_DIR`.
+
+Each run writes a `.json` of every stream part and a `.txt` of the exact concatenated text, named `<modelId>__<scenario>__run<N>`.
+
+### Delegate the UI driving
+
+Steps 3-5 involve a live window. Delegate them to **`ui-operator`** mode rather than doing them inline; screenshots and control trees consume large amounts of context.
+
+## Gotchas
+
+### The consent gate needs a real user gesture
+
+**Do not call `model.sendRequest()` from `activate()`.** Every request fails with:
+
+```
+Language model '<model-id>' cannot be used by '<publisher>.<ext-id>'
+```
+
+This is not a quota, auth, or manifest problem — `vscode.lm` grants consent only in response to a genuine user gesture. The probe must therefore be triggered from the Command Palette (or a notification button click). This is the single most expensive trap here; it silently fails 100% of requests and looks like an entitlement bug.
+
+### Never `Stop-Process` filtered on window title
+
+**WARNING:** During this experiment, killing processes matched by window title destroyed the user's unrelated VS Code windows and their unsaved work.
+
+Safe alternative: only ever _launch_ new windows with `--new-window`, and close the probe window by hand. Never bulk-terminate `Code.exe` by title, `MainWindowTitle`, or any other fuzzy match.
+
+### Run tests with pnpm, not npx
+
+```
+pnpm --dir src exec vitest run <path>
+```
+
+Never `npx vitest` — it resolves a wrong hoisted 3.2.4 instead of the pinned 4.1.9 and produces phantom failures.
+
+## Measured Findings
+
+Sample: 210 live requests, 7 Copilot Claude models x 6 scenarios x 5 repeats, 0 errors. The raw transcripts were not retained; the counts below are the retained record of that run, and re-running the probe is the way to regenerate the underlying evidence.
+
+| Scenario | Setup                                     | Runs | `<invoke` in text |
+| -------- | ----------------------------------------- | ---- | ----------------- |
+| A        | tools declared + agent system prompt      | 35   | 0                 |
+| B        | tools declared, no system prompt          | 35   | 0                 |
+| C        | tools declared + ~300KB filler context    | 35   | 0                 |
+| D        | no tools, model asked to emit the markup  | 35   | 14                |
+| E        | asked to quote the markup in prose        | 35   | 23                |
+| F        | asked to quote the markup in a code fence | 35   | 21                |
+
+Observations:
+
+- **The leak did not reproduce.** 105/105 tool-declared runs (A+B+C) emitted a proper `LanguageModelToolCallPart` and leaked nothing into text parts. This bounds the leak rate at a low value; it is **not** proof of absence. 105 runs across 7 models cannot exclude a rare or prompt-specific trigger.
+- **Wrapped vs. bare inverts the intuition.** All 14 genuine emitted invocations (D) were wrapped in `<function_calls>`; 0 were bare. All 44 quoted-in-prose cases (E+F) were bare; 0 were wrapped. In this sample, _bare correlates with quoting and wrapped with genuine invocation_ — so requiring a `<function_calls>` wrapper would not have been the discriminator it appears to be.
+- **No `antml:` prefix appeared** in any of the 210 runs.
+- **Zero false positives.** Replaying `extractLeakedToolCalls()` over all 58 transcripts containing `<invoke` with `validToolNames = {read_file}`: 9 recovered (all genuine wrapped invocations, arguments parsed correctly), 49 passed through as text, including all 44 bare quoted cases. The fenced/quoted guard is what does the work here, not the wrapper requirement.
+
+### Caveats
+
+- Copilot's `vscode.lm` endpoint sits behind its own prompt assembly; results describe that surface, not the raw Anthropic API.
+- The real-world shape of the leak that motivated the recovery code is **inferred** from third-party Anthropic-API reports (anthropics/claude-code#66153, #73808), not captured from `vscode-lm`. No `vscode-lm` transcript of the failure exists.
+
+## Reproducing the False-Positive Replay
+
+[`scripts/probe-vscode-lm-api/probe-false-positives.spec.ts`](../../../scripts/probe-vscode-lm-api/probe-false-positives.spec.ts) replays [`extractLeakedToolCalls()`](../../../src/api/providers/vscode-lm.ts) over a transcript directory and writes a `RECOVERED`/`passthrough` report. It has no committed inputs — run the probe first to produce them. Its `../vscode-lm` import and `TRANSCRIPTS` default are written for the copy destination, not for where it is committed. Drop it into `src/api/providers/__tests__/`, point `TRANSCRIPTS` (or `LM_PROBE_TRANSCRIPTS`) at the probe's `OUT_DIR`, then:
+
+```
+pnpm --dir src exec vitest run api/providers/__tests__/probe-false-positives.spec.ts
+```
+
+It is a scratch harness, not a committed test — remove it afterwards.
diff --git a/scripts/probe-vscode-lm-api/extension.js b/scripts/probe-vscode-lm-api/extension.js
new file mode 100644
index 0000000000..7829dae8aa
--- /dev/null
+++ b/scripts/probe-vscode-lm-api/extension.js
@@ -0,0 +1,208 @@
+const vscode = require("vscode")
+const fs = require("fs")
+const path = require("path")
+
+// Absolute: the extension host's cwd is not the repo, so a relative path would scatter output.
+const OUT_DIR = process.env.LM_PROBE_OUT_DIR || "c:\\git\\<repo>\\.tmp\\transcripts"
+
+function write(name, data) {
+	fs.mkdirSync(OUT_DIR, { recursive: true })
+	fs.writeFileSync(path.join(OUT_DIR, name), typeof data === "string" ? data : JSON.stringify(data, null, 2))
+}
+
+const READ_TOOL = {
+	name: "read_file",
+	description: "Read the contents of a file at the given path.",
+	inputSchema: {
+		type: "object",
+		properties: { path: { type: "string", description: "File path to read" } },
+		required: ["path"],
+	},
+}
+
+const TOOL_SYSTEM_PROMPT = [
+	"You are Zoo Code, an autonomous coding agent.",
+	"You accomplish tasks by calling the tools provided to you.",
+	"You MUST call exactly one tool per message. Never ask the user a question.",
+	"Do not answer from memory; always read the file first using the read_file tool.",
+].join("\n")
+
+async function runOnce(model, scenario) {
+	const record = {
+		scenario: scenario.name,
+		modelId: model.id,
+		modelFamily: model.family,
+		modelVendor: model.vendor,
+		modelVersion: model.version,
+		maxInputTokens: model.maxInputTokens,
+		parts: [],
+		concatenatedText: "",
+		toolCallParts: [],
+		error: null,
+	}
+	const messages = []
+	if (scenario.system) {
+		messages.push(vscode.LanguageModelChatMessage.Assistant(scenario.system))
+	}
+	for (const userText of scenario.userMessages) {
+		messages.push(vscode.LanguageModelChatMessage.User(userText))
+	}
+	const options = { justification: "Empirical probe of leaked tool-call formatting." }
+	if (scenario.tools) {
+		options.tools = [READ_TOOL]
+	}
+	const source = new vscode.CancellationTokenSource()
+	try {
+		const response = await model.sendRequest(messages, options, source.token)
+		for await (const chunk of response.stream) {
+			const typeName = chunk && chunk.constructor ? chunk.constructor.name : typeof chunk
+			if (chunk instanceof vscode.LanguageModelTextPart) {
+				record.parts.push({ type: typeName, value: chunk.value })
+				record.concatenatedText += chunk.value
+			} else if (chunk instanceof vscode.LanguageModelToolCallPart) {
+				const call = { type: typeName, name: chunk.name, callId: chunk.callId, input: chunk.input }
+				record.parts.push(call)
+				record.toolCallParts.push(call)
+			} else {
+				record.parts.push({ type: typeName, raw: String(chunk) })
+			}
+		}
+	} catch (error) {
+		record.error = { name: error && error.name, message: error && error.message, stack: error && error.stack }
+	} finally {
+		source.dispose()
+	}
+
+	const text = record.concatenatedText
+	record.markers = {
+		hasInvoke: /<(?:antml:)?invoke\b/i.test(text),
+		hasFunctionCalls: /<(?:antml:)?function_calls\b/i.test(text),
+		hasAntmlNamespace: /antml:/i.test(text),
+		invokeWrappedInFunctionCalls: /<(?:antml:)?function_calls\s*>[\s\S]*?<(?:antml:)?invoke\b/i.test(text),
+		bareInvokeWithoutWrapper: /<(?:antml:)?invoke\b/i.test(text) && !/<(?:antml:)?function_calls\b/i.test(text),
+		insideFencedCodeBlock: /```[\s\S]*?<(?:antml:)?invoke\b/i.test(text),
+	}
+	return record
+}
+
+function buildScenarios() {
+	const longFiller = "This is filler context line used to grow the prompt toward the context window. ".repeat(4000)
+	return [
+		{
+			name: "A_tools_declared_compelling_prompt",
+			tools: true,
+			system: TOOL_SYSTEM_PROMPT,
+			userMessages: ["Read the file c:/git/Zoo-Code/package.json and tell me the version field."],
+		},
+		{
+			name: "B_tools_declared_no_system_prompt",
+			tools: true,
+			userMessages: ["Read the file c:/git/Zoo-Code/package.json and tell me the version field."],
+		},
+		{
+			name: "C_tools_declared_long_context",
+			tools: true,
+			system: TOOL_SYSTEM_PROMPT,
+			userMessages: [longFiller, "Now read the file c:/git/Zoo-Code/package.json and tell me the version field."],
+		},
+		{
+			name: "D_no_tools_asked_to_emit_markup",
+			tools: false,
+			system: TOOL_SYSTEM_PROMPT,
+			userMessages: [
+				"You have a tool named read_file that takes a path. Emit the tool invocation using Anthropic's internal function-call XML format, exactly as you would internally, to read c:/git/Zoo-Code/package.json. Output only the markup.",
+			],
+		},
+		{
+			name: "E_quoted_markup_in_prose_false_positive_check",
+			tools: true,
+			system: TOOL_SYSTEM_PROMPT,
+			userMessages: [
+				'Do NOT call any tool. Instead, explain in prose what an invoke block looks like, and include a literal example using the tag name "invoke" with a name attribute of read_file and a parameter named path, written as plain text in your answer.',
+			],
+		},
+		{
+			name: "F_quoted_markup_in_fenced_code_block",
+			tools: true,
+			system: TOOL_SYSTEM_PROMPT,
+			userMessages: [
+				"Do NOT call any tool. Show me, inside a fenced markdown code block, an example of an invoke block naming the tool read_file with a parameter named path set to /etc/passwd. Only output the fenced code block.",
+			],
+		},
+	]
+}
+
+async function run() {
+	const summary = { startedAt: new Date().toISOString(), vscodeVersion: vscode.version, models: [], runs: [] }
+	let models = []
+	let selectError = null
+	try {
+		models = (await vscode.lm.selectChatModels({ vendor: "copilot" })) || []
+	} catch (error) {
+		selectError = { name: error && error.name, message: error && error.message }
+	}
+	summary.selectError = selectError
+	summary.allModels = models.map((model) => ({
+		id: model.id,
+		family: model.family,
+		vendor: model.vendor,
+		version: model.version,
+		maxInputTokens: model.maxInputTokens,
+	}))
+
+	const claudeModels = models.filter((model) => /claude/i.test(model.id) || /claude/i.test(model.family))
+	summary.claudeModelIds = claudeModels.map((model) => model.id)
+
+	if (claudeModels.length === 0) {
+		write("summary.json", summary)
+		vscode.window.showErrorMessage(
+			`LM Probe: no Claude models. selectChatModels returned ${models.length}. ${selectError ? selectError.message : ""}`,
+		)
+		return
+	}
+
+	const scenarios = buildScenarios()
+	const REPEATS = 5
+	for (const model of claudeModels) {
+		for (const scenario of scenarios) {
+			for (let iter = 1; iter <= REPEATS; iter++) {
+				const record = await runOnce(model, scenario)
+				record.iteration = iter
+				summary.runs.push({
+					scenario: scenario.name,
+					modelId: model.id,
+					iteration: iter,
+					markers: record.markers,
+					toolCallPartCount: record.toolCallParts.length,
+					textLength: record.concatenatedText.length,
+					error: record.error ? record.error.message : null,
+				})
+				write(`${model.id}__${scenario.name}__run${iter}.json`, record)
+				write(`${model.id}__${scenario.name}__run${iter}.txt`, record.concatenatedText)
+			}
+		}
+	}
+
+	summary.finishedAt = new Date().toISOString()
+	write("summary.json", summary)
+	vscode.window.showInformationMessage(`LM Probe complete: ${summary.runs.length} runs written to ${OUT_DIR}`)
+}
+
+function activate(context) {
+	// MUST be user-initiated: vscode.lm consent is only granted from a real user gesture, so an
+	// activation-time sendRequest is auto-denied with "cannot be used by 'scratch.lmprobe'".
+	context.subscriptions.push(
+		vscode.commands.registerCommand("lmprobe.run", () =>
+			run().catch((error) => {
+				write("fatal.json", { message: String(error && error.message), stack: String(error && error.stack) })
+			}),
+		),
+	)
+	vscode.window.showInformationMessage("LM Probe ready", "Run probe").then((choice) => {
+		if (choice === "Run probe") {
+			vscode.commands.executeCommand("lmprobe.run")
+		}
+	})
+}
+
+module.exports = { activate, deactivate() {} }
diff --git a/scripts/probe-vscode-lm-api/package.json b/scripts/probe-vscode-lm-api/package.json
new file mode 100644
index 0000000000..8ad0174674
--- /dev/null
+++ b/scripts/probe-vscode-lm-api/package.json
@@ -0,0 +1,21 @@
+{
+	"name": "lmprobe",
+	"displayName": "LM Leak Probe",
+	"version": "0.0.1",
+	"publisher": "scratch",
+	"engines": {
+		"vscode": "^1.95.0"
+	},
+	"activationEvents": [
+		"onStartupFinished"
+	],
+	"main": "./extension.js",
+	"contributes": {
+		"commands": [
+			{
+				"command": "lmprobe.run",
+				"title": "LM Probe: Run"
+			}
+		]
+	}
+}
diff --git a/scripts/probe-vscode-lm-api/probe-false-positives.spec.ts b/scripts/probe-vscode-lm-api/probe-false-positives.spec.ts
new file mode 100644
index 0000000000..5c48318ec0
--- /dev/null
+++ b/scripts/probe-vscode-lm-api/probe-false-positives.spec.ts
@@ -0,0 +1,26 @@
+import fs from "fs"
+import path from "path"
+
+// Paths here are written for the copy destination `src/api/providers/__tests__/`, not for this
+// file's committed location — it is a template to be copied there, never run in place.
+import { extractLeakedToolCalls } from "../vscode-lm"
+
+// Point this at the probe's OUT_DIR. Scratch harness: not a committed test.
+const TRANSCRIPTS = process.env.LM_PROBE_TRANSCRIPTS ?? path.resolve(__dirname, "../../../../.tmp/transcripts")
+
+describe("probe: false-positive check against real transcripts", () => {
+	it("reports which transcripts extractLeakedToolCalls would treat as real calls", () => {
+		const names = new Set(["read_file"])
+		const report: string[] = []
+		for (const file of fs.readdirSync(TRANSCRIPTS).filter((name) => name.endsWith(".txt"))) {
+			const text = fs.readFileSync(path.join(TRANSCRIPTS, file), "utf8")
+			if (!/<(?:antml:)?invoke/i.test(text)) {
+				continue
+			}
+			const { calls } = extractLeakedToolCalls(text, names)
+			report.push(`${calls.length > 0 ? "RECOVERED" : "passthrough"}\t${file}\t${JSON.stringify(calls)}`)
+		}
+		fs.writeFileSync(path.join(TRANSCRIPTS, "..", "false-positive-report.txt"), report.join("\n"))
+		console.log(report.join("\n"))
+	})
+})
diff --git a/src/api/providers/__tests__/vscode-lm.spec.ts b/src/api/providers/__tests__/vscode-lm.spec.ts
index 37fb851720..95be295c3c 100644
--- a/src/api/providers/__tests__/vscode-lm.spec.ts
+++ b/src/api/providers/__tests__/vscode-lm.spec.ts
@@ -60,7 +60,13 @@ vi.mock("vscode", () => {
 })
 
 import * as vscode from "vscode"
-import { VsCodeLmHandler } from "../vscode-lm"
+import {
+	VsCodeLmHandler,
+	extractLeakedToolCalls,
+	trailingPartialToolMarkerLength,
+	middleOutTruncate,
+	truncateToolResultsToFitWindow,
+} from "../vscode-lm"
 import type { ApiHandlerOptions } from "../../../shared/api"
 import type { Anthropic } from "@anthropic-ai/sdk"
 import { openAiModelInfoSaneDefaults, vscodeLlmDefaultModelId, vscodeLlmModels } from "@roo-code/types"
@@ -270,6 +276,229 @@ describe("VsCodeLmHandler", () => {
 			})
 		})
 
+		describe("leaked tool-call recovery during streaming", () => {
+			const salvageTools = [
+				{
+					type: "function" as const,
+					function: {
+						name: "calculator",
+						description: "A simple calculator",
+						parameters: { type: "object", properties: { operation: { type: "string" } } },
+					},
+				},
+			]
+
+			const streamTextParts = (parts: string[]) => {
+				mockLanguageModelChat.sendRequest.mockResolvedValueOnce({
+					stream: (async function* () {
+						for (const part of parts) {
+							yield new vscode.LanguageModelTextPart(part)
+						}
+						return
+					})(),
+					text: (async function* () {
+						yield parts.join("")
+						return
+					})(),
+				})
+			}
+
+			const streamMixedParts = (parts: Array<string | { name: string; input: object }>) => {
+				mockLanguageModelChat.sendRequest.mockResolvedValueOnce({
+					stream: (async function* () {
+						for (const part of parts) {
+							yield typeof part === "string"
+								? new vscode.LanguageModelTextPart(part)
+								: new vscode.LanguageModelToolCallPart("native-1", part.name, part.input)
+						}
+						return
+					})(),
+					text: (async function* () {
+						yield ""
+						return
+					})(),
+				})
+			}
+
+			const drain = async () => {
+				const stream = handler.createMessage("system", [{ role: "user" as const, content: "hi" }], {
+					taskId: "test-task",
+					tools: salvageTools,
+				})
+				const chunks = []
+				for await (const chunk of stream) {
+					chunks.push(chunk)
+				}
+				return chunks
+			}
+
+			const collect = async (parts: string[]) => {
+				streamTextParts(parts)
+				const stream = handler.createMessage("system", [{ role: "user" as const, content: "hi" }], {
+					taskId: "test-task",
+					tools: salvageTools,
+				})
+				const chunks = []
+				for await (const chunk of stream) {
+					chunks.push(chunk)
+				}
+				return chunks
+			}
+
+			it("recovers a tool call the model streamed as raw invoke XML", async () => {
+				const chunks = await collect([
+					"Thinking. ",
+					'<invoke name="calculator"><parameter name="operation">add</parameter></invoke>',
+				])
+
+				expect(chunks.filter((chunk) => chunk.type === "text")).toEqual([{ type: "text", text: "Thinking. " }])
+				expect(chunks.filter((chunk) => chunk.type === "tool_call")).toEqual([
+					{
+						type: "tool_call",
+						id: expect.stringContaining("vscodelm-salvaged-"),
+						name: "calculator",
+						arguments: JSON.stringify({ operation: "add" }),
+					},
+				])
+			})
+
+			it("detects a marker split across stream chunks", async () => {
+				const chunks = await collect([
+					"abc <inv",
+					'oke name="calculator"><parameter name="operation">sub</parameter></invoke>',
+				])
+
+				expect(chunks.filter((chunk) => chunk.type === "text")).toEqual([{ type: "text", text: "abc " }])
+				expect(chunks.filter((chunk) => chunk.type === "tool_call")).toMatchObject([
+					{ name: "calculator", arguments: JSON.stringify({ operation: "sub" }) },
+				])
+			})
+
+			it("emits a carried tail as plain text when it never becomes a marker", async () => {
+				const chunks = await collect(["hello <par"])
+
+				expect(chunks.filter((chunk) => chunk.type === "text")).toEqual([
+					{ type: "text", text: "hello " },
+					{ type: "text", text: "<par" },
+				])
+				expect(chunks.some((chunk) => chunk.type === "tool_call")).toBe(false)
+			})
+
+			it("buffers across chunks that arrive after the marker", async () => {
+				const chunks = await collect([
+					'prose <invoke name="calculator">',
+					'<parameter name="operation">',
+					"mul</parameter>",
+					"</invoke>",
+				])
+
+				expect(chunks.filter((chunk) => chunk.type === "text")).toEqual([{ type: "text", text: "prose " }])
+				expect(chunks.filter((chunk) => chunk.type === "tool_call")).toMatchObject([
+					{ name: "calculator", arguments: JSON.stringify({ operation: "mul" }) },
+				])
+			})
+
+			it("keeps an invoke block for an unknown tool as literal text", async () => {
+				const block = '<invoke name="not_our_tool"><parameter name="a">1</parameter></invoke>'
+				const chunks = await collect([block])
+
+				expect(chunks.filter((chunk) => chunk.type === "text")).toEqual([{ type: "text", text: block }])
+				expect(chunks.some((chunk) => chunk.type === "tool_call")).toBe(false)
+			})
+
+			it("emits prose before the recovered tool call", async () => {
+				const chunks = await collect([
+					"Thinking. ",
+					'<invoke name="calculator"><parameter name="operation">add</parameter></invoke>',
+				])
+
+				expect(chunks.map((chunk) => chunk.type)).toEqual(["text", "tool_call", "usage"])
+			})
+
+			it("flushes buffered text before a native tool call so no text follows a tool_use", async () => {
+				streamMixedParts([
+					'partial <invoke name="calculator">',
+					{ name: "calculator", input: { operation: "div" } },
+				])
+				const chunks = await drain()
+
+				const lastText = chunks.map((chunk) => chunk.type).lastIndexOf("text")
+				const firstToolCall = chunks.map((chunk) => chunk.type).indexOf("tool_call")
+				expect(firstToolCall).toBeGreaterThan(lastText)
+			})
+
+			it("does not latch buffering on prose that merely mentions the tag", async () => {
+				const chunks = await collect(["never emit <invoke> markup as text. ", "Streaming continues."])
+
+				expect(chunks.filter((chunk) => chunk.type === "text")).toEqual([
+					{ type: "text", text: "never emit <invoke> markup as text. " },
+					{ type: "text", text: "Streaming continues." },
+				])
+				expect(chunks.some((chunk) => chunk.type === "tool_call")).toBe(false)
+			})
+
+			it("does not recover an invoke block quoted inside a fenced code block", async () => {
+				const block = '<invoke name="calculator"><parameter name="operation">add</parameter></invoke>'
+				const chunks = await collect(["Do NOT do this:\n```\n" + block + "\n```\n"])
+
+				expect(chunks.some((chunk) => chunk.type === "tool_call")).toBe(false)
+			})
+
+			it("flushes an over-long never-closing invoke as plain text before the stream ends", async () => {
+				// Defect 4: without a cap the buffer is only drained once the stream finishes, so the
+				// user sees nothing until then. Releasing it at the end looks identical in content —
+				// only the timing distinguishes the fix, so track how much of the source has been
+				// produced at the moment each text chunk reaches the consumer.
+				const filler = "x".repeat(5000)
+				const parts = ['<invoke name="calculator">', filler, filler, filler, filler]
+				let partsProduced = 0
+
+				mockLanguageModelChat.sendRequest.mockResolvedValueOnce({
+					stream: (async function* () {
+						for (const part of parts) {
+							partsProduced++
+							yield new vscode.LanguageModelTextPart(part)
+						}
+						return
+					})(),
+					text: (async function* () {
+						yield ""
+						return
+					})(),
+				})
+
+				const stream = handler.createMessage("system", [{ role: "user" as const, content: "hi" }], {
+					taskId: "test-task",
+					tools: salvageTools,
+				})
+
+				let sawTextBeforeStreamEnd = false
+				let streamedText = ""
+				for await (const chunk of stream) {
+					if (chunk.type === "text") {
+						streamedText += chunk.text
+						if (partsProduced < parts.length) {
+							sawTextBeforeStreamEnd = true
+						}
+					}
+				}
+
+				expect(sawTextBeforeStreamEnd).toBe(true)
+				expect(streamedText).toContain('<invoke name="calculator">')
+				expect(streamedText).toContain(filler)
+			})
+
+			it("sanitizes lone surrogates in the system prompt", async () => {
+				streamTextParts(["ok"])
+				const stream = handler.createMessage("sys\uD800tem", [{ role: "user" as const, content: "hi" }])
+				for await (const _chunk of stream) {
+					// drain
+				}
+
+				expect(vscode.LanguageModelChatMessage.Assistant).toHaveBeenCalledWith("sys\uFFFDtem")
+			})
+		})
+
 		it("should handle native tool calls when tools are provided", async () => {
 			const systemPrompt = "You are a helpful assistant"
 			const messages: Anthropic.Messages.MessageParam[] = [
@@ -1075,3 +1304,393 @@ describe("VsCodeLmHandler", () => {
 		})
 	})
 })
+
+describe("leaked tool-call recovery", () => {
+	// Builders keep the XML fixtures readable and prevent this file's own markup from being
+	// mistaken for a real tool call.
+	const invoke = (name: string, body: string) => `<in${"voke"} name="${name}">${body}</in${"voke"}>`
+	const param = (name: string, value: string) => `<param${"eter"} name="${name}">${value}</param${"eter"}>`
+
+	describe("extractLeakedToolCalls", () => {
+		it("recovers a known-tool block and strips it from the leftover text", () => {
+			const text = `Working on it.\n${invoke("update_todo_list", param("todos", "[x] one\n[ ] two"))}`
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toEqual([{ name: "update_todo_list", input: { todos: "[x] one\n[ ] two" } }])
+			expect(leftoverText).toBe("Working on it.\n")
+		})
+
+		it("recovers an unwrapped leak preceded by a stray token", () => {
+			const text = `court\n${invoke("update_todo_list", param("todos", "[x] done"))}`
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toEqual([{ name: "update_todo_list", input: { todos: "[x] done" } }])
+			expect(leftoverText).toBe("court\n")
+		})
+
+		it("recovers multiple params and strips function-call wrapper tags", () => {
+			const body = param("mode", "code") + param("message", "go")
+			const text = `<function_calls>${invoke("new_task", body)}</function_calls>`
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["new_task"]))
+
+			expect(calls).toEqual([{ name: "new_task", input: { mode: "code", message: "go" } }])
+			expect(leftoverText).toBe("")
+		})
+
+		it("passes through invoke blocks for tools that were not offered", () => {
+			const text = invoke("some_other_tool", param("x", "1"))
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toEqual([])
+			expect(leftoverText).toBe(text)
+		})
+
+		it("returns no calls for ordinary text", () => {
+			const { calls, leftoverText } = extractLeakedToolCalls("just a normal reply", new Set(["update_todo_list"]))
+
+			expect(calls).toEqual([])
+			expect(leftoverText).toBe("just a normal reply")
+		})
+	})
+
+	describe("trailingPartialToolMarkerLength", () => {
+		it("holds back a split marker prefix at the end of a chunk", () => {
+			expect(trailingPartialToolMarkerLength("some text <in")).toBe(3)
+		})
+
+		it("returns 0 for plain text and complete tags", () => {
+			expect(trailingPartialToolMarkerLength("hello world")).toBe(0)
+			expect(trailingPartialToolMarkerLength("a < b")).toBe(0)
+			expect(trailingPartialToolMarkerLength("text <function_calls>")).toBe(0)
+		})
+
+		it("holds back an invoke tag whose name attribute has not arrived", () => {
+			expect(trailingPartialToolMarkerLength("text <invoke ")).toBe(8)
+		})
+
+		it("does not hold back an over-long trailing fragment", () => {
+			expect(trailingPartialToolMarkerLength("<invoke " + "x".repeat(200))).toBe(0)
+		})
+
+		it("does not hold back an over-long generic tag fragment", () => {
+			expect(trailingPartialToolMarkerLength("text <" + "a".repeat(200))).toBe(0)
+		})
+	})
+
+	describe("quoted markup", () => {
+		it("does not recover an invoke block inside a fenced code block", () => {
+			const text = "```\n" + invoke("update_todo_list", param("todos", "[x] one")) + "\n```"
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toHaveLength(0)
+			expect(leftoverText).toContain("invoke")
+		})
+
+		it("does not recover an invoke block inside an inline code span", () => {
+			const text = "avoid `" + invoke("update_todo_list", param("todos", "x")) + "`"
+
+			const { calls } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toHaveLength(0)
+		})
+
+		it("does not recover an invoke block quoted in unfenced, backtick-free prose", () => {
+			const text = "You must never emit " + invoke("update_todo_list", param("todos", "x")) + " directly."
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toHaveLength(0)
+			expect(leftoverText).toBe(text)
+		})
+
+		it("does not recover a quoted invoke block that ends its line", () => {
+			// Defect 3: an empty rest-of-line previously made this look like a genuine leak.
+			const text = "You must never emit " + invoke("update_todo_list", param("todos", "x"))
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toHaveLength(0)
+			expect(leftoverText).toBe(text)
+		})
+
+		it("does not recover an invoke block inside a tilde fence", () => {
+			const text = "~~~\n" + invoke("update_todo_list", param("todos", "[x] one")) + "\n~~~"
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toHaveLength(0)
+			expect(leftoverText).toBe(text)
+		})
+
+		it("does not recover an invoke block inside a four-backtick fence", () => {
+			const text = "````\n" + invoke("update_todo_list", param("todos", "[x] one")) + "\n````"
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toHaveLength(0)
+			expect(leftoverText).toBe(text)
+		})
+
+		it("does not treat doubled angle brackets as trailing prose after stripping", () => {
+			// Defect 1: a single strip pass turns `<<x>>` into a tag-looking `<x>`, so the
+			// trailing-text check must strip repeatedly until stable.
+			const text = invoke("update_todo_list", param("todos", "x")) + "<<script>>"
+
+			const { calls } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toEqual([{ name: "update_todo_list", input: { todos: "x" } }])
+		})
+
+		it("keeps wrapper tags around a block that was not recovered", () => {
+			const text = `<function_calls>${invoke("some_other_tool", param("x", "1"))}</function_calls>`
+
+			const { calls, leftoverText } = extractLeakedToolCalls(text, new Set(["update_todo_list"]))
+
+			expect(calls).toHaveLength(0)
+			expect(leftoverText).toBe(text)
+		})
+	})
+})
+
+describe("context-window tool_result truncation", () => {
+	describe("middleOutTruncate", () => {
+		it("returns text unchanged when within the limit", () => {
+			expect(middleOutTruncate("hello world", 100)).toBe("hello world")
+		})
+
+		it("keeps the head and tail and inserts a truncation marker", () => {
+			const text = "A".repeat(500) + "B".repeat(500)
+			const result = middleOutTruncate(text, 200)
+
+			expect(result.length).toBeLessThanOrEqual(200)
+			expect(result).toContain("characters truncated to fit the model context window")
+			expect(result.startsWith("A")).toBe(true)
+			expect(result.endsWith("B")).toBe(true)
+		})
+
+		it("returns an empty string for a non-positive limit", () => {
+			expect(middleOutTruncate("anything", 0)).toBe("")
+		})
+
+		// A lone surrogate cannot be encoded as UTF-8 and 400s the whole request.
+		it("never splits a surrogate pair across the removed middle", () => {
+			const pair = "\u{1F600}" // one astral char = high + low surrogate
+			const text = pair.repeat(400)
+			const result = middleOutTruncate(text, 200)
+
+			expect(/[\uD800-\uDBFF](?![\uDC00-\uDFFF])|(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/.test(result)).toBe(false)
+		})
+	})
+
+	describe("truncateToolResultsToFitWindow", () => {
+		const toolUseMessage = (id: string): Anthropic.Messages.MessageParam => ({
+			role: "assistant",
+			content: [
+				{ type: "text", text: "Calling a tool." },
+				{ type: "tool_use", id, name: "some_tool", input: { a: 1 } },
+			],
+		})
+
+		const toolResultMessage = (id: string, content: string): Anthropic.Messages.MessageParam => ({
+			role: "user",
+			content: [
+				{ type: "tool_result", tool_use_id: id, content },
+				{ type: "text", text: "<environment_details>env</environment_details>" },
+			],
+		})
+
+		const findBlock = (message: Anthropic.Messages.MessageParam, type: string) =>
+			(message.content as unknown as Array<{ type: string; [key: string]: unknown }>).find(
+				(block) => block.type === type,
+			)!
+
+		it("is a no-op when the conversation already fits the budget", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", "small result"),
+			]
+			const before = JSON.parse(JSON.stringify(messages))
+
+			truncateToolResultsToFitWindow(messages, 100_000)
+
+			expect(messages).toEqual(before)
+		})
+
+		it("returns messages untouched when the budget is not a usable number", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", "Y".repeat(50_000)),
+			]
+			const before = JSON.parse(JSON.stringify(messages))
+
+			expect(truncateToolResultsToFitWindow(messages, 0)).toBe(messages)
+			expect(truncateToolResultsToFitWindow(messages, Number.NaN)).toBe(messages)
+			expect(messages).toEqual(before)
+		})
+
+		it("truncates array-form tool_result content and preserves non-text parts", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				{
+					role: "user",
+					content: [
+						{
+							type: "tool_result",
+							tool_use_id: "t1",
+							content: [
+								{ type: "text", text: "Z".repeat(50_000) },
+								{ type: "image", source: { type: "base64", media_type: "image/png", data: "abc" } },
+							],
+						},
+					],
+				} as unknown as Anthropic.Messages.MessageParam,
+			]
+
+			truncateToolResultsToFitWindow(messages, 10_000)
+
+			const toolResult = findBlock(messages[1], "tool_result")
+			const parts = toolResult.content as Array<{ type: string; text?: string }>
+			expect(parts[0].type).toBe("text")
+			expect(parts[0].text).toContain("characters truncated")
+			expect(parts.some((part) => part.type === "image")).toBe(true)
+		})
+
+		it("ignores string content and skips messages that cannot hold tool_result blocks", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				{ role: "user", content: "a plain string turn" },
+				toolUseMessage("t1"),
+				{
+					role: "user",
+					content: [
+						{ type: "tool_result", tool_use_id: "t1", content: "W".repeat(50_000) },
+						{ type: "image", source: { type: "base64", media_type: "image/png", data: "abc" } },
+					],
+				} as unknown as Anthropic.Messages.MessageParam,
+			]
+
+			truncateToolResultsToFitWindow(messages, 10_000)
+
+			expect(messages[0].content).toBe("a plain string turn")
+			expect(String(findBlock(messages[2], "tool_result").content)).toContain("characters truncated")
+		})
+
+		it("ignores a tool_result whose content is neither string nor array", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", "V".repeat(50_000)),
+				{
+					role: "user",
+					content: [{ type: "tool_result", tool_use_id: "t2", content: undefined }],
+				} as unknown as Anthropic.Messages.MessageParam,
+			]
+
+			truncateToolResultsToFitWindow(messages, 10_000)
+
+			expect(findBlock(messages[2], "tool_result").content).toBeUndefined()
+			expect(String(findBlock(messages[1], "tool_result").content)).toContain("characters truncated")
+		})
+
+		it("skips a tool_result already small enough to need no trimming", () => {
+			// Overage is tiny, so the largest block's target lands at its current length.
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", "U".repeat(3000)),
+				toolUseMessage("t2"),
+				toolResultMessage("t2", "T".repeat(2500)),
+			]
+
+			truncateToolResultsToFitWindow(messages, 5600)
+
+			const first = String(findBlock(messages[1], "tool_result").content)
+			const second = String(findBlock(messages[3], "tool_result").content)
+			expect(first.length + second.length).toBeLessThanOrEqual(5600)
+		})
+
+		it("leaves a tool_result at or below the minimum size alone", () => {
+			const shortResult = "S".repeat(1500)
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", shortResult),
+				toolUseMessage("t2"),
+				toolResultMessage("t2", shortResult),
+			]
+
+			truncateToolResultsToFitWindow(messages, 100)
+
+			expect(findBlock(messages[1], "tool_result").content).toBe(shortResult)
+			expect(findBlock(messages[3], "tool_result").content).toBe(shortResult)
+		})
+
+		it("shrinks an oversized tool_result so the conversation fits the budget", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", "X".repeat(50_000)),
+			]
+
+			truncateToolResultsToFitWindow(messages, 10_000)
+
+			const toolResult = findBlock(messages[1], "tool_result")
+			expect(toolResult.tool_use_id).toBe("t1") // pairing preserved
+			expect(String(toolResult.content).length).toBeLessThanOrEqual(10_000)
+			expect(String(toolResult.content)).toContain("characters truncated")
+		})
+
+		it("truncates the largest tool_result first and leaves small ones intact", () => {
+			const small = "small but real result"
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", "B".repeat(40_000)),
+				toolUseMessage("t2"),
+				toolResultMessage("t2", small),
+			]
+
+			truncateToolResultsToFitWindow(messages, 12_000)
+
+			expect(String(findBlock(messages[1], "tool_result").content)).toContain("characters truncated")
+			expect(findBlock(messages[3], "tool_result").content).toBe(small) // untouched
+		})
+
+		it("never truncates tool_use blocks, assistant text, or environment details", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				toolResultMessage("t1", "X".repeat(50_000)),
+			]
+
+			truncateToolResultsToFitWindow(messages, 8_000)
+
+			expect(findBlock(messages[0], "text").text).toBe("Calling a tool.")
+			expect(findBlock(messages[0], "tool_use")).toMatchObject({ id: "t1", name: "some_tool" })
+			expect(findBlock(messages[1], "text").text).toBe("<environment_details>env</environment_details>")
+		})
+
+		it("handles array-form tool_result content and keeps it valid", () => {
+			const messages: Anthropic.Messages.MessageParam[] = [
+				toolUseMessage("t1"),
+				{
+					role: "user",
+					content: [
+						{
+							type: "tool_result",
+							tool_use_id: "t1",
+							content: [{ type: "text", text: "Y".repeat(40_000) }],
+						},
+					],
+				},
+			]
+
+			truncateToolResultsToFitWindow(messages, 8_000)
+
+			const toolResult = findBlock(messages[1], "tool_result")
+			expect(toolResult.tool_use_id).toBe("t1")
+			expect(Array.isArray(toolResult.content)).toBe(true)
+			const parts = toolResult.content as Array<{ type: string; text?: string }>
+			expect(parts[0].type).toBe("text")
+			expect(String(parts[0].text)).toContain("characters truncated")
+		})
+	})
+})
diff --git a/src/api/providers/vscode-lm.ts b/src/api/providers/vscode-lm.ts
index c657e6c0d6..0339bab720 100644
--- a/src/api/providers/vscode-lm.ts
+++ b/src/api/providers/vscode-lm.ts
@@ -8,8 +8,12 @@ import type { ApiHandlerOptions } from "../../shared/api"
 import { SELECTOR_SEPARATOR, stringifyVsCodeLmModelSelector } from "../../shared/vsCodeSelectorUtils"
 import { normalizeToolSchema } from "../../utils/json-schema"
 
-import { ApiStream } from "../transform/stream"
-import { convertToVsCodeLmMessages, extractTextCountFromMessage } from "../transform/vscode-lm-format"
+import { ApiStream, ApiStreamChunk } from "../transform/stream"
+import {
+	convertToVsCodeLmMessages,
+	extractTextCountFromMessage,
+	sanitizeSurrogates,
+} from "../transform/vscode-lm-format"
 
 import { BaseProvider } from "./base-provider"
 import type { SingleCompletionHandler, ApiHandlerCreateMessageMetadata, CompletePromptOptions } from "../index"
@@ -60,6 +64,381 @@ function convertToVsCodeLmTools(tools: OpenAI.Chat.ChatCompletionTool[]): vscode
  * }
  * ```
  */
+/**
+ * Recovery for leaked tool calls
+ * ------------------------------
+ * Some VS Code LM backends — notably GitHub Copilot serving Anthropic Claude models —
+ * intermittently stream a tool call as PLAIN TEXT using Anthropic's internal function-call
+ * XML instead of emitting a structured `LanguageModelToolCallPart`. When this happens the
+ * assistant turn contains no tool_use block, so Zoo reports "no tools used" and the task stalls
+ * in a retry loop. The helpers below detect the leaked markup mid-stream and replay it as a real
+ * tool call. Recovery is deliberately conservative: only `<invoke>` blocks whose name matches a
+ * tool we actually offered this turn are treated as calls; everything else is passed through
+ * unchanged as text.
+ */
+// Latching on a bare `<invoke` would let any prose mentioning the tag stop real-time streaming for
+// the rest of the response, so the marker only counts once the `name="` attribute has arrived.
+const LEAKED_TOOL_CALL_START = /<(?:antml:)?(?:function_calls\s*>|invoke\s+name=")/i
+const LEAKED_INVOKE_BLOCK = /<(?:antml:)?invoke\s+name="([^"]+)"\s*>([\s\S]*?)<\/(?:antml:)?invoke\s*>/gi
+const LEAKED_INVOKE_PARAM = /<(?:antml:)?parameter\s+name="([^"]+)"\s*>([\s\S]*?)<\/(?:antml:)?parameter\s*>/gi
+
+/** Upper bound on an incomplete `<invoke ...` tail held back between chunks. */
+const MAX_PARTIAL_INVOKE_CARRY = 64
+
+/**
+ * Upper bound on buffered text held while waiting for a leaked `<invoke>` block to close. Markup
+ * that never closes would otherwise withhold the whole response from the user until the stream
+ * ended, so past this size the buffer is released as ordinary text.
+ */
+const MAX_SALVAGE_BUFFER_CHARS = 64 * MAX_PARTIAL_INVOKE_CARRY
+
+/**
+ * Same-line prose that introduces markup as an example rather than invoking it. This is a
+ * deliberately narrow lexical cue: a quoted invoke that ENDS its line is otherwise
+ * indistinguishable from a genuine leak, which is just as often preceded by prose.
+ */
+const QUOTING_CUE =
+	/\b(?:never|not|do not|don't|does not|doesn't|must not|mustn't|avoid|instead of|rather than|for example|e\.g\.|such as|like this|as follows)\b[^.!?\n]*$/i
+
+/**
+ * True when `before` ends inside an open Markdown code fence. Tracks the fence character and its
+ * width so tilde fences and fences of 4+ backticks (which may legally contain shorter fences) are
+ * recognized, rather than counting three-backtick runs for parity.
+ */
+function isInsideCodeFence(before: string): boolean {
+	let openFence: { marker: string; width: number } | null = null
+	for (const line of before.split("\n")) {
+		const fenceMatch = line.match(/^ {0,3}(`{3,}|~{3,})/)
+		if (!fenceMatch) {
+			continue
+		}
+		const marker = fenceMatch[1][0]
+		const width = fenceMatch[1].length
+		if (!openFence) {
+			openFence = { marker, width }
+		} else if (marker === openFence.marker && width >= openFence.width) {
+			openFence = null
+		}
+	}
+	return openFence !== null
+}
+
+/**
+ * Strips well-formed tags repeatedly until the result stops changing. A single pass is unsafe:
+ * `<<invoke>>` reassembles into a live-looking tag after one replacement.
+ */
+/** True when `text` already contains a closed `<invoke>` block, so buffering is still productive. */
+function hasCompleteInvokeBlock(text: string): boolean {
+	LEAKED_INVOKE_BLOCK.lastIndex = 0
+	const found = LEAKED_INVOKE_BLOCK.test(text)
+	LEAKED_INVOKE_BLOCK.lastIndex = 0
+	return found
+}
+
+function stripTagsCompletely(text: string): string {
+	let current = text
+	for (;;) {
+		const stripped = current.replace(/<[^<>]*>/g, "")
+		if (stripped === current) {
+			return stripped
+		}
+		current = stripped
+	}
+}
+
+/**
+ * Returns the length of a trailing fragment that might be the start of a leaked tool-call marker
+ * split across stream chunks. Such a tail is held back until more text arrives so the marker can
+ * be detected intact.
+ */
+export function trailingPartialToolMarkerLength(text: string): number {
+	const partialTag = text.match(/<(?:antml:)?[a-zA-Z_]*$/)
+	if (partialTag) {
+		return partialTag[0].length <= MAX_PARTIAL_INVOKE_CARRY ? partialTag[0].length : 0
+	}
+	// An `<invoke` whose `name="` attribute hasn't arrived yet: hold it back so the marker can
+	// latch on the next chunk, bounded so ordinary prose is never swallowed.
+	const partialInvoke = text.match(/<(?:antml:)?invoke\b[^<>]*$/i)
+	return partialInvoke && partialInvoke[0].length <= MAX_PARTIAL_INVOKE_CARRY ? partialInvoke[0].length : 0
+}
+
+/**
+ * True when the block spanning `[index, endIndex)` is being quoted — inside a fenced code block,
+ * inside an inline code span, or embedded mid-sentence in plain prose — rather than invoked.
+ */
+function isQuotedAsCode(text: string, index: number, endIndex: number): boolean {
+	const before = text.slice(0, index)
+	if (isInsideCodeFence(before)) {
+		return true
+	}
+	const lineStart = before.lastIndexOf("\n") + 1
+	const sameLineBefore = before.slice(lineStart)
+	if ((sameLineBefore.match(/`/g)?.length ?? 0) % 2 === 1) {
+		return true
+	}
+	// Narrative words after the block on the same line mean the markup is being talked about
+	// (e.g. "never emit <invoke ...> directly"), which must not be replayed as a live call.
+	const after = text.slice(endIndex)
+	const lineEnd = after.indexOf("\n")
+	const restOfLine = lineEnd === -1 ? after : after.slice(0, lineEnd)
+	if (stripTagsCompletely(restOfLine).trim().length > 0) {
+		return true
+	}
+	// A quoted invoke that ENDS its line leaves no trailing text to judge. Keying off the mere
+	// presence of leading prose was tried and regressed genuine recoveries, because a real leak is
+	// commonly preceded by narration too ("Working on it.\n"), so only an explicit quoting cue
+	// immediately before the markup suppresses it. Heuristic: prose that quotes markup without
+	// such a cue still reads as a live call.
+	return QUOTING_CUE.test(stripTagsCompletely(sameLineBefore))
+}
+
+function parseLeakedInvokeParams(body: string): Record<string, string> {
+	const input: Record<string, string> = {}
+	LEAKED_INVOKE_PARAM.lastIndex = 0
+	let match: RegExpExecArray | null
+	while ((match = LEAKED_INVOKE_PARAM.exec(body)) !== null) {
+		input[match[1]] = match[2].trim()
+	}
+	return input
+}
+
+/**
+ * Extracts complete leaked `<invoke>` tool-call blocks from `text`. Only blocks whose name
+ * is present in `validToolNames` are returned as calls; all other text (including `<invoke>`
+ * blocks for unknown names) is returned as `leftoverText` so legitimate prose is preserved.
+ */
+export function extractLeakedToolCalls(
+	text: string,
+	validToolNames: ReadonlySet<string>,
+	precedingText = "",
+): { calls: Array<{ name: string; input: Record<string, string> }>; leftoverText: string } {
+	const calls: Array<{ name: string; input: Record<string, string> }> = []
+	// Text between recovered/passed-through blocks, kept as segments so the wrapper cleanup below
+	// only touches segments adjacent to a block that was actually recovered.
+	const segments: Array<{ text: string; nearRecovery: boolean }> = []
+	let pending = ""
+	let lastIndex = 0
+
+	LEAKED_INVOKE_BLOCK.lastIndex = 0
+	let match: RegExpExecArray | null
+	while ((match = LEAKED_INVOKE_BLOCK.exec(text)) !== null) {
+		pending += text.slice(lastIndex, match.index)
+		const name = match[1]
+		// Quote detection needs the text streamed before the buffer, since a fence may have opened there.
+		if (
+			validToolNames.has(name) &&
+			!isQuotedAsCode(
+				precedingText + text,
+				precedingText.length + match.index,
+				precedingText.length + match.index + match[0].length,
+			)
+		) {
+			calls.push({ name, input: parseLeakedInvokeParams(match[2]) })
+			segments.push({ text: pending, nearRecovery: true })
+			pending = ""
+			// The segment that follows a recovery also holds that call's closing wrapper.
+			segments.push({ text: "", nearRecovery: true })
+		} else {
+			// Not one of our tools, or quoted as code — keep the block as literal text.
+			pending += match[0]
+		}
+		lastIndex = match.index + match[0].length
+	}
+	pending += text.slice(lastIndex)
+	const trailing =
+		segments.length > 0 && segments[segments.length - 1].nearRecovery && segments[segments.length - 1].text === ""
+	if (trailing) {
+		segments[segments.length - 1].text = pending
+	} else {
+		segments.push({ text: pending, nearRecovery: false })
+	}
+
+	// Remove the wrapper tags belonging to a recovered call (cosmetic; also avoids re-teaching the
+	// model this format when the turn is sent back as history). Wrappers around blocks that were
+	// NOT recovered are user-visible text and must survive verbatim.
+	const leftover = segments
+		.map((segment) =>
+			segment.nearRecovery ? segment.text.replace(/<\/?(?:antml:)?function_calls\s*>/gi, "") : segment.text,
+		)
+		.join("")
+
+	return { calls, leftoverText: leftover }
+}
+
+/**
+ * Context-window safety for Copilot's backend
+ * -------------------------------------------
+ * Copilot's backend enforces its own context window and, for third-party `sendRequest` callers,
+ * trims an over-window request in a way that is NOT tool-pair-aware: it can drop the assistant
+ * message holding a `tool_use` while keeping the matching `tool_result`, after which Anthropic
+ * rejects the request with "unexpected tool_use_id". To keep trimming on OUR side — where
+ * pairing is preserved — we shrink oversized `tool_result` payloads before sending. Only
+ * `tool_result` text is truncated (never `tool_use`, assistant text, summaries, or environment
+ * details), and only when the request would otherwise exceed the budget.
+ */
+
+/**
+ * Conservative characters-per-token ratio used to turn a token window into a character budget.
+ * The token-dense JSON, logs, and code that dominate oversized tool results tokenize to fewer
+ * characters per token than prose, so we intentionally under-count (3, not the ~4 typical of
+ * English) to keep the resulting budget on the safe side of the enforced window.
+ */
+const VSCODE_LM_BUDGET_CHARS_PER_TOKEN = 3
+
+/**
+ * Fraction of the context window the *entire* input (system prompt + tool schemas + conversation)
+ * is allowed to occupy. The remaining headroom absorbs char/token estimation variance and any
+ * output/overhead the backend reserves.
+ */
+const VSCODE_LM_INPUT_BUDGET_FRACTION = 0.8
+
+/** A tool_result is never shrunk below this many characters, so a truncated result stays useful. */
+const MIN_TOOL_RESULT_CHARS = 2000
+
+function readToolResultText(block: Anthropic.Messages.ContentBlockParam): string | undefined {
+	if (!block || (block as { type?: string }).type !== "tool_result") {
+		return undefined
+	}
+	const content = (block as Anthropic.Messages.ToolResultBlockParam).content
+	if (typeof content === "string") {
+		return content
+	}
+	if (Array.isArray(content)) {
+		return content
+			.filter((part): part is Anthropic.Messages.TextBlockParam => (part as { type?: string })?.type === "text")
+			.map((part) => part.text ?? "")
+			.join("")
+	}
+	return undefined
+}
+
+function writeToolResultText(block: Anthropic.Messages.ContentBlockParam, text: string): void {
+	const toolResult = block as Anthropic.Messages.ToolResultBlockParam
+	const content = toolResult.content
+	if (Array.isArray(content)) {
+		// Preserve any non-text parts (e.g. images) and collapse the text into one truncated part.
+		const nonText = content.filter((part) => (part as { type?: string })?.type !== "text")
+		toolResult.content = [{ type: "text", text }, ...nonText] as typeof content
+		return
+	}
+	toolResult.content = text
+}
+
+/**
+ * Middle-out truncate `text` to at most `maxChars`, keeping the head and tail and replacing the
+ * middle with a marker noting how many characters were removed. Head/tail are preserved because
+ * logs and file dumps carry the most signal at their start (structure) and end (recent output).
+ */
+export function middleOutTruncate(text: string, maxChars: number): string {
+	if (maxChars <= 0) {
+		return ""
+	}
+	if (text.length <= maxChars) {
+		return text
+	}
+
+	const buildMarker = (removed: number) =>
+		`\n\n[... ${removed.toLocaleString("en-US")} characters truncated to fit the model context window ...]\n\n`
+
+	// Reserve room for the marker, sized against the original length so the result never grows.
+	const reservedMarkerLength = buildMarker(text.length).length
+	const keep = Math.max(0, maxChars - reservedMarkerLength)
+	const headLength = Math.ceil(keep / 2)
+	const tailLength = keep - headLength
+	let head = text.slice(0, headLength)
+	// Don't end the head on a lone high surrogate — its low half is in the removed middle, and a lone
+	// surrogate cannot be encoded as UTF-8 (the backend 400s the whole request). Drop the split half.
+	if (head.length > 0 && (head.charCodeAt(head.length - 1) & 0xfc00) === 0xd800) {
+		head = head.slice(0, -1)
+	}
+	let tail = tailLength > 0 ? text.slice(text.length - tailLength) : ""
+	// Likewise, don't start the tail on a lone low surrogate (its high half is in the removed middle).
+	if (tail.length > 0 && (tail.charCodeAt(0) & 0xfc00) === 0xdc00) {
+		tail = tail.slice(1)
+	}
+	const removed = text.length - head.length - tail.length
+	return `${head}${buildMarker(removed)}${tail}`
+}
+
+function estimateContentChars(content: Anthropic.Messages.MessageParam["content"]): number {
+	if (typeof content === "string") {
+		return content.length
+	}
+	if (!Array.isArray(content)) {
+		return 0
+	}
+	let total = 0
+	for (const block of content) {
+		const type = (block as { type?: string })?.type
+		if (type === "text") {
+			total += (block as Anthropic.Messages.TextBlockParam).text?.length ?? 0
+		} else if (type === "tool_result") {
+			total += readToolResultText(block)?.length ?? 0
+		} else if (type === "tool_use") {
+			total += JSON.stringify((block as Anthropic.Messages.ToolUseBlockParam).input ?? {}).length
+		} else if (type === "image") {
+			total += 8 // "[IMAGE]" placeholder — VS Code LM drops image data anyway.
+		}
+	}
+	return total
+}
+
+/**
+ * Shrinks oversized `tool_result` payloads (largest first, middle-out) until the conversation fits
+ * `budgetChars`. Mutates the tool_result blocks of the supplied messages in place — callers pass a
+ * cloned array (see `createMessage`) so stored history is never mutated. A no-op when the
+ * conversation already fits.
+ */
+export function truncateToolResultsToFitWindow(
+	messages: Anthropic.Messages.MessageParam[],
+	budgetChars: number,
+): Anthropic.Messages.MessageParam[] {
+	if (!Number.isFinite(budgetChars) || budgetChars <= 0) {
+		return messages
+	}
+
+	let total = messages.reduce((sum, message) => sum + estimateContentChars(message.content), 0)
+	if (total <= budgetChars) {
+		return messages
+	}
+
+	// Collect every truncatable tool_result block, largest first.
+	const toolResultBlocks: Anthropic.Messages.ContentBlockParam[] = []
+	for (const message of messages) {
+		if (!Array.isArray(message.content)) {
+			continue
+		}
+		for (const block of message.content) {
+			if (readToolResultText(block) !== undefined) {
+				toolResultBlocks.push(block)
+			}
+		}
+	}
+	toolResultBlocks.sort((a, b) => (readToolResultText(b)?.length ?? 0) - (readToolResultText(a)?.length ?? 0))
+
+	for (const block of toolResultBlocks) {
+		if (total <= budgetChars) {
+			break
+		}
+		const text = readToolResultText(block)
+		if (text === undefined || text.length <= MIN_TOOL_RESULT_CHARS) {
+			continue
+		}
+
+		const overage = total - budgetChars
+		const target = Math.max(MIN_TOOL_RESULT_CHARS, text.length - overage)
+		if (target >= text.length) {
+			continue
+		}
+
+		const truncated = middleOutTruncate(text, target)
+		total -= text.length - truncated.length
+		writeToolResultText(block, truncated)
+	}
+
+	return messages
+}
+
 export class VsCodeLmHandler extends BaseProvider implements SingleCompletionHandler {
 	protected options: ApiHandlerOptions
 	private client: vscode.LanguageModelChat | null
@@ -383,9 +762,22 @@ export class VsCodeLmHandler extends BaseProvider implements SingleCompletionHan
 			content: this.cleanMessageContent(msg.content),
 		}))
 
+		// Keep context-window trimming on OUR side. Copilot's backend trims an over-window request
+		// without preserving tool_use/tool_result pairing, which orphans a tool_result and triggers a
+		// 400 ("unexpected tool_use_id"). See truncateToolResultsToFitWindow.
+		const contextWindowTokens = this.getCondenseContextWindow()
+		if (Number.isFinite(contextWindowTokens) && contextWindowTokens > 0) {
+			const toolSchemaChars = metadata?.tools ? JSON.stringify(metadata.tools).length : 0
+			const messagesBudgetChars =
+				contextWindowTokens * VSCODE_LM_INPUT_BUDGET_FRACTION * VSCODE_LM_BUDGET_CHARS_PER_TOKEN -
+				systemPrompt.length -
+				toolSchemaChars
+			truncateToolResultsToFitWindow(cleanedMessages, messagesBudgetChars)
+		}
+
 		// Convert Anthropic messages to VS Code LM messages
 		const vsCodeLmMessages: vscode.LanguageModelChatMessage[] = [
-			vscode.LanguageModelChatMessage.Assistant(systemPrompt),
+			vscode.LanguageModelChatMessage.Assistant(sanitizeSurrogates(systemPrompt)),
 			...convertToVsCodeLmMessages(cleanedMessages),
 		]
 
@@ -398,6 +790,65 @@ export class VsCodeLmHandler extends BaseProvider implements SingleCompletionHan
 		// Accumulate the text and count at the end of the stream to reduce token counting overhead.
 		let accumulatedText: string = ""
 
+		// Leaked tool-call recovery state (see `extractLeakedToolCalls`). Only enabled when we
+		// actually offered tools this turn, so it can never misfire on plain conversations.
+		const providedToolNames = new Set(
+			(metadata?.tools ?? [])
+				.filter((tool) => tool.type === "function")
+				.map((tool) => tool.function.name)
+				.filter((name) => name.length > 0),
+		)
+		const salvageLeakedToolCalls = providedToolNames.size > 0
+		let salvageBuffering = false
+		let salvageBuffer = ""
+		let salvageCarry = ""
+		let salvageEmittedText = ""
+		let salvagedToolCallIndex = 0
+
+		// Drains the salvage state into ordered chunks: prose first, then any recovered calls. Must
+		// run before a native tool_call is yielded — text after a tool_use block is rejected by
+		// Anthropic once the turn is serialized back into history.
+		const flushSalvage = (): ApiStreamChunk[] => {
+			const flushed: ApiStreamChunk[] = []
+			if (!salvageLeakedToolCalls) {
+				return flushed
+			}
+
+			if (!salvageBuffering) {
+				if (salvageCarry) {
+					flushed.push({ type: "text", text: salvageCarry })
+					salvageCarry = ""
+				}
+				return flushed
+			}
+
+			const buffered = salvageBuffer
+			salvageBuffering = false
+			salvageBuffer = ""
+			if (!buffered) {
+				return flushed
+			}
+
+			const { calls, leftoverText } = extractLeakedToolCalls(buffered, providedToolNames, salvageEmittedText)
+			salvageEmittedText += buffered
+			if (leftoverText) {
+				flushed.push({ type: "text", text: leftoverText })
+			}
+			for (const call of calls) {
+				console.warn(
+					"Zoo Code <Language Model API>: Recovered a tool call the model emitted as text instead of a structured tool call:",
+					{ name: call.name, params: Object.keys(call.input) },
+				)
+				flushed.push({
+					type: "tool_call",
+					id: `vscodelm-salvaged-${Date.now()}-${salvagedToolCallIndex++}`,
+					name: call.name,
+					arguments: JSON.stringify(call.input),
+				})
+			}
+			return flushed
+		}
+
 		try {
 			// Create the response stream with required options
 			const requestOptions: vscode.LanguageModelChatRequestOptions = {
@@ -421,11 +872,54 @@ export class VsCodeLmHandler extends BaseProvider implements SingleCompletionHan
 					}
 
 					accumulatedText += chunk.value
-					yield {
-						type: "text",
-						text: chunk.value,
+
+					// Fast path: when we didn't offer any tools there is nothing to salvage, so
+					// stream the text straight through exactly as before.
+					if (!salvageLeakedToolCalls) {
+						yield { type: "text", text: chunk.value }
+						continue
+					}
+
+					// Once we've seen the start of a leaked tool-call block, buffer the rest of the
+					// stream so the full markup can be parsed and replayed as a structured call.
+					if (salvageBuffering) {
+						salvageBuffer += chunk.value
+						// Markup that never closes must not withhold the response indefinitely; past the
+						// cap the buffer is released as plain text and buffering stops for the turn.
+						if (salvageBuffer.length > MAX_SALVAGE_BUFFER_CHARS && !hasCompleteInvokeBlock(salvageBuffer)) {
+							const overflowed = salvageBuffer
+							salvageBuffering = false
+							salvageBuffer = ""
+							salvageEmittedText += overflowed
+							yield { type: "text", text: overflowed }
+						}
+						continue
+					}
+
+					// Watch for the start of a leaked tool-call block, carrying a small tail across
+					// chunks so a marker split across chunk boundaries is still detected.
+					const combined = salvageCarry + chunk.value
+					const markerMatch = combined.match(LEAKED_TOOL_CALL_START)
+					if (markerMatch) {
+						const before = combined.slice(0, markerMatch.index)
+						if (before) {
+							salvageEmittedText += before
+							yield { type: "text", text: before }
+						}
+						salvageBuffering = true
+						salvageBuffer = combined.slice(markerMatch.index)
+						salvageCarry = ""
+					} else {
+						const carryLength = trailingPartialToolMarkerLength(combined)
+						const emit = carryLength > 0 ? combined.slice(0, combined.length - carryLength) : combined
+						salvageCarry = carryLength > 0 ? combined.slice(combined.length - carryLength) : ""
+						if (emit) {
+							salvageEmittedText += emit
+							yield { type: "text", text: emit }
+						}
 					}
 				} else if (chunk instanceof vscode.LanguageModelToolCallPart) {
+					yield* flushSalvage()
 					try {
 						// Validate tool call parameters
 						if (!chunk.name || typeof chunk.name !== "string") {
@@ -472,6 +966,9 @@ export class VsCodeLmHandler extends BaseProvider implements SingleCompletionHan
 				}
 			}
 
+			// Flush any leaked tool-call recovery state accumulated during streaming.
+			yield* flushSalvage()
+
 			// Count tokens in the accumulated text after stream completion
 			const totalOutputTokens: number = await this.internalCountTokens(accumulatedText)
 
diff --git a/src/api/transform/__tests__/vscode-lm-format.spec.ts b/src/api/transform/__tests__/vscode-lm-format.spec.ts
index 3265f2745b..275126b5ce 100644
--- a/src/api/transform/__tests__/vscode-lm-format.spec.ts
+++ b/src/api/transform/__tests__/vscode-lm-format.spec.ts
@@ -3,7 +3,12 @@
 import { Anthropic } from "@anthropic-ai/sdk"
 import * as vscode from "vscode"
 
-import { convertToVsCodeLmMessages, convertToAnthropicRole, extractTextCountFromMessage } from "../vscode-lm-format"
+import {
+	convertToVsCodeLmMessages,
+	convertToAnthropicRole,
+	extractTextCountFromMessage,
+	sanitizeSurrogates,
+} from "../vscode-lm-format"
 
 // Mock crypto using Vitest
 vitest.stubGlobal("crypto", {
@@ -325,6 +330,79 @@ describe("convertToVsCodeLmMessages", () => {
 	})
 })
 
+describe("sanitizeSurrogates", () => {
+	it("leaves plain ASCII unchanged", () => {
+		expect(sanitizeSurrogates("hello world")).toBe("hello world")
+	})
+
+	it("leaves valid surrogate pairs unchanged", () => {
+		// 😀 U+1F600 and 𐀀 U+10000 are astral-plane code points encoded as surrogate pairs.
+		expect(sanitizeSurrogates("a\uD83D\uDE00b\uD800\uDC00c")).toBe("a\uD83D\uDE00b\uD800\uDC00c")
+	})
+
+	it("replaces a lone high surrogate with U+FFFD", () => {
+		expect(sanitizeSurrogates("a\uD800b")).toBe("a\uFFFDb")
+	})
+
+	it("replaces a lone low surrogate with U+FFFD", () => {
+		expect(sanitizeSurrogates("a\uDC00b")).toBe("a\uFFFDb")
+	})
+
+	it("replaces a trailing lone high surrogate", () => {
+		expect(sanitizeSurrogates("abc\uD800")).toBe("abc\uFFFD")
+	})
+
+	it("replaces a reversed (low-then-high) pair as two lone surrogates", () => {
+		expect(sanitizeSurrogates("\uDC00\uD800")).toBe("\uFFFD\uFFFD")
+	})
+
+	it("returns empty input unchanged", () => {
+		expect(sanitizeSurrogates("")).toBe("")
+	})
+})
+
+describe("convertToVsCodeLmMessages surrogate sanitization", () => {
+	const lone = "bad\uD800end"
+	const sanitized = "bad\uFFFDend"
+
+	const textValues = (message: { content: unknown }) =>
+		(message.content as MockLanguageModelTextPart[]).map((part) => part.value)
+
+	it("sanitizes a simple string message", () => {
+		const result = convertToVsCodeLmMessages([{ role: "user", content: lone }])
+		expect(textValues(result[0])).toEqual([sanitized])
+	})
+
+	it("sanitizes string tool_result content", () => {
+		const result = convertToVsCodeLmMessages([
+			{ role: "user", content: [{ type: "tool_result", tool_use_id: "tool-1", content: lone }] },
+		])
+		const toolResult = result[0].content[0] as MockLanguageModelToolResultPart
+		expect(toolResult.content[0].value).toBe(sanitized)
+	})
+
+	it("sanitizes tool_result text blocks", () => {
+		const result = convertToVsCodeLmMessages([
+			{
+				role: "user",
+				content: [{ type: "tool_result", tool_use_id: "tool-1", content: [{ type: "text", text: lone }] }],
+			},
+		])
+		const toolResult = result[0].content[0] as MockLanguageModelToolResultPart
+		expect(toolResult.content[0].value).toBe(sanitized)
+	})
+
+	it("sanitizes user text blocks", () => {
+		const result = convertToVsCodeLmMessages([{ role: "user", content: [{ type: "text", text: lone }] }])
+		expect(textValues(result[0])).toContain(sanitized)
+	})
+
+	it("sanitizes assistant text blocks", () => {
+		const result = convertToVsCodeLmMessages([{ role: "assistant", content: [{ type: "text", text: lone }] }])
+		expect(textValues(result[0])).toContain(sanitized)
+	})
+})
+
 describe("convertToAnthropicRole", () => {
 	it("should convert assistant role correctly", () => {
 		const result = convertToAnthropicRole(vscode.LanguageModelChatMessageRole.Assistant)
diff --git a/src/api/transform/vscode-lm-format.ts b/src/api/transform/vscode-lm-format.ts
index 7ac51e024f..a03cba257c 100644
--- a/src/api/transform/vscode-lm-format.ts
+++ b/src/api/transform/vscode-lm-format.ts
@@ -28,6 +28,23 @@ function asObjectSafe(value: unknown): object {
 	}
 }
 
+/**
+ * Replaces unpaired UTF-16 surrogate code units with the Unicode replacement character (U+FFFD).
+ *
+ * The VS Code LM backend forwards requests to model APIs that require valid UTF-8. A lone surrogate
+ * — e.g. left behind when some upstream step slices a string through an astral-plane character
+ * (emoji, CJK extension, etc.) — cannot be encoded as UTF-8, so the backend rejects the entire
+ * request with a 400 ("string contains an unpaired UTF-16 surrogate code point and cannot be
+ * encoded as valid UTF-8"). Valid surrogate pairs are matched by the lookahead/lookbehind and left
+ * untouched. The regex intentionally omits the `u` flag so it operates on UTF-16 code units.
+ */
+export function sanitizeSurrogates(text: string): string {
+	if (!text) {
+		return text
+	}
+	return text.replace(/[\uD800-\uDBFF](?![\uDC00-\uDFFF])|(?<![\uD800-\uDBFF])[\uDC00-\uDFFF]/g, "\uFFFD")
+}
+
 export function convertToVsCodeLmMessages(
 	anthropicMessages: Anthropic.Messages.MessageParam[],
 ): vscode.LanguageModelChatMessage[] {
@@ -36,10 +53,11 @@ export function convertToVsCodeLmMessages(
 	for (const anthropicMessage of anthropicMessages) {
 		// Handle simple string messages
 		if (typeof anthropicMessage.content === "string") {
+			const safeContent = sanitizeSurrogates(anthropicMessage.content)
 			vsCodeLmMessages.push(
 				anthropicMessage.role === "assistant"
-					? vscode.LanguageModelChatMessage.Assistant(anthropicMessage.content)
-					: vscode.LanguageModelChatMessage.User(anthropicMessage.content),
+					? vscode.LanguageModelChatMessage.Assistant(safeContent)
+					: vscode.LanguageModelChatMessage.User(safeContent),
 			)
 			continue
 		}
@@ -69,7 +87,7 @@ export function convertToVsCodeLmMessages(
 						// Process tool result content into TextParts
 						const toolContentParts: vscode.LanguageModelTextPart[] =
 							typeof toolMessage.content === "string"
-								? [new vscode.LanguageModelTextPart(toolMessage.content)]
+								? [new vscode.LanguageModelTextPart(sanitizeSurrogates(toolMessage.content))]
 								: (toolMessage.content?.map((part) => {
 										if (part.type === "image") {
 											if (part.source.type === "base64") {
@@ -82,7 +100,7 @@ export function convertToVsCodeLmMessages(
 											)
 										}
 										if (part.type === "text") {
-											return new vscode.LanguageModelTextPart(part.text)
+											return new vscode.LanguageModelTextPart(sanitizeSurrogates(part.text))
 										}
 										return new vscode.LanguageModelTextPart("")
 									}) ?? [new vscode.LanguageModelTextPart("")])
@@ -102,7 +120,7 @@ export function convertToVsCodeLmMessages(
 								`[Image (${part.source.type}): not supported by VSCode LM API]`,
 							)
 						}
-						return new vscode.LanguageModelTextPart(part.text)
+						return new vscode.LanguageModelTextPart(sanitizeSurrogates(part.text))
 					}),
 				]
 
@@ -135,7 +153,7 @@ export function convertToVsCodeLmMessages(
 						if (part.type === "image") {
 							return new vscode.LanguageModelTextPart("[Image generation not supported by VSCode LM API]")
 						}
-						return new vscode.LanguageModelTextPart(part.text)
+						return new vscode.LanguageModelTextPart(sanitizeSurrogates(part.text))
 					}),
 
 					// Convert tool messages to ToolCallParts after text