Skip to content

Claude Code. Any Model. The most powerful AI coding agent now speaks every language.

fix(tools): an empty-string required argument is present, not missing (item 1)

Jack Rudenko417ef5c49c3a
diff --git a/packages/cli/src/format-translation.test.ts b/packages/cli/src/format-translation.test.ts
index 21e0fb68..10c1b8aa 100644
--- a/packages/cli/src/format-translation.test.ts
+++ b/packages/cli/src/format-translation.test.ts
@@ -2709,3 +2709,132 @@ describe("Regression: finalize teardown survives callback failures", () => {
expect(result.events?.filter((event) => event.data?.type === "message_stop")).toHaveLength(1);
});
});
+
+// ─── Item 1 — an empty-string required argument is PRESENT ──────────────────
+
+/**
+ * Regression from a LIVE `[email protected]` run, captured 2026-09-16.
+ *
+ * The model was asked to delete a line and emitted exactly the right call:
+ *
+ * Edit{"file_path":"…/sample.ts","old_string":"const dead = 1;\n","new_string":""}
+ *
+ * `finish_reason: "tool_calls"`, nothing truncated, the JSON complete. An empty
+ * `new_string` IS the deletion — it is the whole point of the call. claudish's
+ * presence filter tested `parsedArgs[param] === ""`, declared `new_string`
+ * missing, suppressed the tool call and emitted
+ * `⚠️ Tool call "Edit" failed: missing required parameters: new_string`
+ * instead. The file was left unchanged and the model was told its own correct
+ * call was malformed.
+ *
+ * The fixture is that response, verbatim. The capture logged two concurrent
+ * upstream requests into one file, so its events were separated by response id;
+ * nothing else was changed.
+ */
+describe("OpenAI SSE: a required argument whose value is an empty string is present", () => {
+ async function getParser() {
+ const mod = await import("./handlers/shared/openai-compat.js");
+ return mod.createStreamingResponseHandler;
+ }
+
+ async function getDefaultAdapter() {
+ const mod = await import("./adapters/base-api-format.js");
+ return new mod.DefaultAPIFormat("test-model");
+ }
+
+ /** Claude Code's own Edit schema: all three of these are required. */
+ const EDIT_SCHEMA = [
+ {
+ name: "Edit",
+ input_schema: {
+ type: "object",
+ properties: {
+ file_path: { type: "string" },
+ old_string: { type: "string" },
+ new_string: { type: "string" },
+ replace_all: { type: "boolean" },
+ },
+ required: ["file_path", "old_string", "new_string"],
+ },
+ },
+ ];
+
+ const CAPTURE = "grok-4.6-openai-edit-empty-new-string.sse";
+
+ async function replayEditCapture(toolSchemas: any[] = EDIT_SCHEMA): Promise<ClaudeEvent[]> {
+ const createStreamingResponseHandler = await getParser();
+ const adapter = await getDefaultAdapter();
+ const response = createStreamingResponseHandler(
+ createMockContext(),
+ fixtureToResponse(join(FIXTURES_DIR, CAPTURE)),
+ adapter,
+ "grok-4.6",
+ null,
+ undefined,
+ toolSchemas
+ );
+ return parseClaudeSseStream(response);
+ }
+
+ /** The complete input JSON of the first tool_use block, as the client sees it. */
+ function toolInput(events: ClaudeEvent[]): Record<string, unknown> {
+ const start = events.find(
+ (e) => e.data?.type === "content_block_start" && e.data?.content_block?.type === "tool_use"
+ );
+ expect(start).toBeDefined();
+ const index = start?.data.index;
+ const json = events
+ .filter(
+ (e) =>
+ e.data?.type === "content_block_delta" &&
+ e.data?.index === index &&
+ e.data?.delta?.type === "input_json_delta"
+ )
+ .map((e) => e.data.delta.partial_json)
+ .join("");
+ return JSON.parse(json);
+ }
+
+ test('the Edit call reaches the client with new_string preserved as ""', async () => {
+ const events = await replayEditCapture();
+
+ expect(extractToolNames(events)).toContain("Edit");
+
+ const input = toolInput(events);
+ expect(Object.hasOwn(input, "new_string")).toBe(true);
+ expect(input.new_string).toBe("");
+ expect(input.old_string).toBe("const dead = 1;\n");
+ expect(input.file_path).toMatch(/sample\.ts$/);
+ });
+
+ test("no missing-parameter warning is emitted for it", async () => {
+ const text = extractText(await replayEditCapture());
+ expect(text).not.toContain("missing required parameters");
+ expect(text).not.toContain('Tool call "Edit" failed');
+ });
+
+ test("the turn still ends as a tool call", async () => {
+ expect(extractStopReason(await replayEditCapture())).toBe("tool_use");
+ });
+
+ test("a genuinely absent required argument still fails visibly", async () => {
+ // The same capture, validated against a schema declaring a parameter the
+ // model never sent. The warning block is NOT weakened by item 1 — absence
+ // and emptiness are now different things, and this is the absence half.
+ const events = await replayEditCapture([
+ {
+ name: "Edit",
+ input_schema: {
+ type: "object",
+ properties: EDIT_SCHEMA[0].input_schema.properties,
+ required: [...EDIT_SCHEMA[0].input_schema.required, "never_sent"],
+ },
+ },
+ ]);
+
+ expect(extractText(events)).toContain(
+ 'Tool call "Edit" failed: missing required parameters: never_sent'
+ );
+ expect(extractToolNames(events)).not.toContain("Edit");
+ });
+});
diff --git a/packages/cli/src/handlers/shared/schema-validate.ts b/packages/cli/src/handlers/shared/schema-validate.ts
new file mode 100644
index 00000000..6d6b8b14
--- /dev/null
+++ b/packages/cli/src/handlers/shared/schema-validate.ts
@@ -0,0 +1,59 @@
+/**
+ * A minimal JSON Schema reader for TOOL schemas.
+ *
+ * Scope is deliberate, and recorded in this session's `decisions.md` (ruling 1):
+ * the keywords a Claude Code tool `input_schema` actually uses — `type`,
+ * `properties`, `required`, `enum`, `items`, `default` — and nothing else. It is
+ * not a general-purpose validator and must not grow into one. The trigger to
+ * replace it with a real library (`ajv`) is a tool schema in the wild that it
+ * mis-validates; until then a published CLI with a `--compile` binary target
+ * (`packages/cli/package.json` `build:binary`) carries no new runtime dependency
+ * for this.
+ *
+ * The one rule that matters more than any other lives in {@link missingRequired}:
+ * **presence is key presence, never truthiness.** The hand-written filter this
+ * replaces tested `parsedArgs[param] === ""` and so declared a legitimately empty
+ * string missing. A live `[email protected]` turn emitted
+ * `Edit{…,"new_string":""}` — a deletion, the whole point of the call — and
+ * claudish rejected it as "missing required parameters: new_string". The file was
+ * left unchanged and the model was told its own correct call was malformed. See
+ * `test-fixtures/sse-responses/grok-4.6-openai-edit-empty-new-string.sse`.
+ */
+
+/** The slice of JSON Schema this module reads. Anything else is ignored. */
+export interface JsonSchemaNode {
+ type?: string | string[];
+ properties?: Record<string, JsonSchemaNode | undefined>;
+ required?: string[];
+ enum?: unknown[];
+ items?: JsonSchemaNode;
+ default?: unknown;
+ [key: string]: unknown;
+}
+
+/**
+ * Which of the schema's `required` keys are genuinely absent from `args`?
+ *
+ * Present means **the key exists and its value is neither `undefined` nor
+ * `null`**. `""`, `0`, `false`, `[]` and `{}` are all present values a model may
+ * legitimately mean: an empty replacement string, a zero offset, a disabled flag,
+ * an empty list.
+ *
+ * `null` counts as absent rather than present because JSON's `null` is how a
+ * model spells "I have no value for this", and no Claude Code tool declares a
+ * nullable required parameter.
+ */
+export function missingRequired(
+ schema: JsonSchemaNode | undefined,
+ args: Record<string, unknown> | undefined
+): string[] {
+ const required = schema?.required;
+ if (!Array.isArray(required) || required.length === 0) return [];
+ const supplied = args ?? {};
+ return required.filter((key) => {
+ if (typeof key !== "string") return false;
+ if (!Object.hasOwn(supplied, key)) return true;
+ const value = supplied[key];
+ return value === undefined || value === null;
+ });
+}
diff --git a/packages/cli/src/handlers/shared/tool-call-recovery.ts b/packages/cli/src/handlers/shared/tool-call-recovery.ts
index 350bb536..af40df2f 100644
--- a/packages/cli/src/handlers/shared/tool-call-recovery.ts
+++ b/packages/cli/src/handlers/shared/tool-call-recovery.ts
@@ -1,15 +1,27 @@
/**
* Tool Call Recovery Module
*
- * Handles recovery from malformed tool calls generated by local models.
- * Implements multiple strategies:
- * 1. Text-based tool call extraction (parse JSON/XML from text)
- * 2. Parameter inference for missing required fields
- * 3. Retry prompt generation with error feedback
+ * Handles recovery from malformed tool calls generated by local models:
+ * text-based tool call extraction (parse JSON/XML from text), and validation of
+ * a call's arguments against the tool's own JSON Schema.
+ *
+ * It does NOT invent arguments. It used to: `inferMissingParameters` filled a
+ * missing `Glob.pattern` with `**\/*` (walks the whole tree), took `Task.prompt`
+ * from "the last meaningful sentence" of the model's prose, and rewrote
+ * `subagent_type` against a hardcoded roster of eight agent names that claudish
+ * cannot discover and that was already stale. A MISSING argument surfaces as a
+ * visible harness error; a FABRICATED one is a confident wrong action with no
+ * error anywhere. The codebase already prices that asymmetry the same way for
+ * tool `pattern` constraints (`ai-docs/architecture/adapters.md:471-474`), and
+ * the inference is gone for the same reason.
+ *
+ * What survives is only what RENAMES OR TYPES A VALUE THE MODEL SUPPLIED —
+ * see `schema-validate.ts`.
*/
import { TOOL_NAME_SHAPE, TOOL_NAME_SOURCE } from "../../adapters/tool-name-utils.js";
import { log } from "../../logger.js";
+import { type JsonSchemaNode, missingRequired } from "./schema-validate.js";
export interface ExtractedToolCall {
name: string;
@@ -293,7 +305,14 @@ export function extractToolCallsFromText(
knownTools.find((t) => t.toLowerCase() === toolName.toLowerCase()) || toolName;
const args: Record<string, any> = {};
- // Extract key=value pairs
+ // Extract key=value pairs. This is the ONLY argument source here: a value
+ // the model wrote down, under the name the model gave it.
+ //
+ // The per-tool synthesis that used to follow is deleted. It read arguments
+ // out of the surrounding prose — `Bash.command` from the words after "run",
+ // `Task.prompt` from the words after "to", `Grep.pattern` from the words
+ // after "search" — and, failing that, made them up. A sentence merely
+ // DESCRIBING a tool then dispatched that tool with invented arguments.
const kvPattern = /(\w+)\s*=\s*["']?([^"',\s]+)["']?/g;
let kvMatch: RegExpExecArray | null;
// biome-ignore lint/suspicious/noAssignInExpressions: canonical RegExp.exec() iteration idiom
@@ -301,86 +320,6 @@ export function extractToolCallsFromText(
args[kvMatch[1]] = kvMatch[2];
}
- // Extract quoted strings as potential file paths or commands
- const quotedPattern = /["']([^"']+)["']/g;
- let quotedMatch: RegExpExecArray | null;
- const quotedValues: string[] = [];
- // biome-ignore lint/suspicious/noAssignInExpressions: canonical RegExp.exec() iteration idiom
- while ((quotedMatch = quotedPattern.exec(paramText)) !== null) {
- quotedValues.push(quotedMatch[1]);
- }
-
- // Tool-specific parameter extraction from natural language
- if (normalizedToolName === "Task") {
- // Look for subagent_type mentions
- if (!args.subagent_type) {
- const stMatch = paramText.match(/subagent_type\s*[=:]\s*["']?(\w+)["']?/i);
- if (stMatch) {
- args.subagent_type = stMatch[1];
- } else if (/explore|codebase|structure/i.test(paramText)) {
- args.subagent_type = "Explore";
- } else if (/plan|architect/i.test(paramText)) {
- args.subagent_type = "Plan";
- } else {
- args.subagent_type = "general-purpose";
- }
- }
- // Extract task intent as prompt
- if (!args.prompt) {
- // Use the text after "to" as the prompt
- const toMatch = paramText.match(/\bto\s+(.+)/i);
- if (toMatch) {
- args.prompt = toMatch[1].trim();
- } else {
- args.prompt = paramText.trim();
- }
- }
- if (!args.description) {
- args.description = (args.prompt || paramText).substring(0, 50).trim();
- }
- } else if (normalizedToolName === "Read") {
- // Extract file path
- if (!args.file_path) {
- if (quotedValues.length > 0) {
- args.file_path = quotedValues[0];
- } else {
- // Look for path-like strings
- const pathMatch = paramText.match(/(?:read|file)\s+([\/\w.-]+)/i);
- if (pathMatch) {
- args.file_path = pathMatch[1];
- }
- }
- }
- } else if (normalizedToolName === "Bash") {
- // Extract command
- if (!args.command) {
- if (quotedValues.length > 0) {
- args.command = quotedValues[0];
- } else {
- // Look for "run X" or "execute X"
- const cmdMatch = paramText.match(/(?:run|execute)\s+(.+)/i);
- if (cmdMatch) {
- args.command = cmdMatch[1].trim();
- }
- }
- }
- if (args.command && !args.description) {
- args.description = `Run ${args.command.split(" ")[0]} command`;
- }
- } else if (normalizedToolName === "Grep" || normalizedToolName === "Glob") {
- // Extract pattern
- if (!args.pattern) {
- if (quotedValues.length > 0) {
- args.pattern = quotedValues[0];
- } else {
- const searchMatch = paramText.match(/(?:search|find|look for)\s+(.+)/i);
- if (searchMatch) {
- args.pattern = searchMatch[1].trim();
- }
- }
- }
- }
-
// Only add if we extracted meaningful arguments
if (Object.keys(args).length > 0) {
extracted.push({
@@ -399,273 +338,17 @@ export function extractToolCallsFromText(
}
/**
- * Infer missing parameters for known tools
- */
-export function inferMissingParameters(
- toolName: string,
- args: Record<string, any>,
- missingParams: string[],
- context?: string
-): Record<string, any> {
- const inferred = { ...args };
-
- // Task tool inference
- if (toolName === "Task") {
- // Valid subagent types
- const validSubagentTypes = [
- "general-purpose",
- "Explore",
- "Plan",
- "claude-code-guide",
- "code-analysis:detective",
- "feature-dev:code-architect",
- "feature-dev:code-explorer",
- "feature-dev:code-reviewer",
- ];
-
- // Normalize subagent_type - models often use variations
- if (inferred.subagent_type) {
- const st = inferred.subagent_type.toLowerCase();
- // Map common variations to valid types
- if (st.includes("explore") || st.includes("codebase") || st.includes("file")) {
- inferred.subagent_type = "Explore";
- } else if (st.includes("plan") || st.includes("architect")) {
- inferred.subagent_type = "Plan";
- } else if (
- st.includes("analysis") ||
- st.includes("analyz") ||
- st.includes("config") ||
- st.includes("git") ||
- st.includes("test") ||
- st.includes("doc") ||
- st.includes("version")
- ) {
- inferred.subagent_type = "general-purpose";
- } else if (!validSubagentTypes.includes(inferred.subagent_type)) {
- log(
- `[ToolRecovery] Unknown subagent_type "${inferred.subagent_type}", mapping to general-purpose`
- );
- inferred.subagent_type = "general-purpose";
- }
- }
-
- if (missingParams.includes("subagent_type") && !inferred.subagent_type) {
- // Default to general-purpose if not specified
- inferred.subagent_type = "general-purpose";
- log("[ToolRecovery] Inferred subagent_type: general-purpose");
- }
-
- // Try to extract meaningful task description from context
- let extractedTask = "";
- if (context) {
- // Look for common patterns that indicate the model's intent
- const patterns = [
- /(?:I(?:'ll| will| need to| want to| am going to)|Let me|Going to)\s+([^.!?\n]+)/i,
- /(?:help you|assist with)\s+([^.!?\n]+)/i,
- /(?:explore|search|find|look for|investigate)\s+([^.!?\n]+)/i,
- /(?:implement|create|build|add|fix|update)\s+([^.!?\n]+)/i,
- ];
- for (const pattern of patterns) {
- const match = context.match(pattern);
- if (match?.[1] && match[1].length > 10) {
- extractedTask = match[1].trim();
- log(`[ToolRecovery] Extracted task from context: "${extractedTask.substring(0, 50)}..."`);
- break;
- }
- }
- // Fallback: use the last meaningful sentence as context
- if (!extractedTask && context.length > 20) {
- const sentences = context.split(/[.!?\n]+/).filter((s) => s.trim().length > 15);
- if (sentences.length > 0) {
- extractedTask = sentences[sentences.length - 1].trim();
- }
- }
- }
-
- if (missingParams.includes("prompt") && !inferred.prompt) {
- // Try to use description, task content, query, or extracted context
- // Some models use "query" instead of "prompt"
- if (inferred.query) {
- inferred.prompt = inferred.query;
- log(`[ToolRecovery] Mapped query -> prompt: "${inferred.query.substring(0, 50)}..."`);
- } else if (inferred.description && inferred.description !== "Execute task") {
- inferred.prompt = inferred.description;
- } else if (inferred.task) {
- inferred.prompt = inferred.task;
- } else if (extractedTask) {
- inferred.prompt = extractedTask;
- } else if (context && context.length > 20) {
- // Use the full context if nothing else works
- inferred.prompt = context.substring(0, 500).trim();
- }
- if (inferred.prompt) {
- log(`[ToolRecovery] Inferred prompt: "${inferred.prompt.substring(0, 50)}..."`);
- }
- }
-
- if (missingParams.includes("description") && !inferred.description) {
- // Generate description from prompt or extracted task
- if (inferred.prompt) {
- // Take first 50 chars of prompt as description
- inferred.description = inferred.prompt.substring(0, 50).replace(/\s+/g, " ").trim();
- if (inferred.description.length < inferred.prompt.length) {
- inferred.description += "...";
- }
- } else if (extractedTask) {
- inferred.description = extractedTask.substring(0, 50).trim();
- } else {
- inferred.description = "Execute task";
- }
- log(`[ToolRecovery] Inferred description: ${inferred.description}`);
- }
- }
-
- // Bash tool inference
- if (toolName === "Bash") {
- if (missingParams.includes("command") && !inferred.command) {
- // Check for common alternative parameter names
- inferred.command = inferred.cmd || inferred.shell || inferred.script || "";
- }
- if (missingParams.includes("description") && !inferred.description) {
- if (inferred.command) {
- // Generate description from command
- const cmd = inferred.command.split(" ")[0];
- inferred.description = `Run ${cmd} command`;
- }
- }
- }
-
- // Read tool inference
- if (toolName === "Read") {
- if (missingParams.includes("file_path") && !inferred.file_path) {
- inferred.file_path = inferred.path || inferred.file || inferred.filename || "";
- }
- }
-
- // Write tool inference
- if (toolName === "Write") {
- if (missingParams.includes("file_path") && !inferred.file_path) {
- inferred.file_path = inferred.path || inferred.file || inferred.filename || "";
- }
- if (missingParams.includes("content") && !inferred.content) {
- inferred.content = inferred.text || inferred.data || inferred.body || "";
- }
- }
-
- // Grep tool inference
- if (toolName === "Grep") {
- if (missingParams.includes("pattern") && !inferred.pattern) {
- inferred.pattern = inferred.query || inferred.search || inferred.regex || "";
- }
- }
-
- // Glob tool inference
- if (toolName === "Glob") {
- if (missingParams.includes("pattern") && !inferred.pattern) {
- inferred.pattern = inferred.glob || inferred.path || inferred.search || "**/*";
- }
- }
-
- // ToolSearch inference
- // max_results has a schema default of 5; query must be extracted from context
- if (toolName === "ToolSearch") {
- if (missingParams.includes("max_results") && inferred.max_results === undefined) {
- inferred.max_results = 5;
- log("[ToolRecovery] Inferred max_results: 5 (default)");
- }
- if (missingParams.includes("query") && !inferred.query) {
- inferred.query = inferred.search || inferred.keyword || inferred.tool || "";
- if (inferred.query) {
- log(`[ToolRecovery] Inferred ToolSearch query: "${inferred.query}"`);
- }
- }
- }
-
- return inferred;
-}
-
-/**
- * Generate a retry prompt with error feedback
- */
-export function generateRetryPrompt(
- toolName: string,
- missingParams: string[],
- providedArgs: Record<string, any>,
- toolSchema?: ToolSchema
-): string {
- let prompt = `Your previous tool call to "${toolName}" was incomplete. `;
- prompt += `Missing required parameters: ${missingParams.join(", ")}.\n\n`;
-
- if (toolSchema?.input_schema?.properties) {
- prompt += `The ${toolName} tool requires:\n`;
- for (const param of missingParams) {
- const propSchema = toolSchema.input_schema.properties[param];
- if (propSchema) {
- prompt += `- ${param}: ${propSchema.description || propSchema.type || "required"}\n`;
- } else {
- prompt += `- ${param}: required\n`;
- }
- }
- prompt += "\n";
- }
-
- prompt += `You provided: ${JSON.stringify(providedArgs, null, 2)}\n\n`;
- prompt += "Please try again with ALL required parameters included.";
-
- return prompt;
-}
-
-/**
- * Check if a tool call can be repaired
- */
-export function canRepairToolCall(
- toolName: string,
- args: Record<string, any>,
- missingParams: string[]
-): boolean {
- // Check if we have enough context to infer the missing params
- const inferred = inferMissingParameters(toolName, args, missingParams);
-
- // Verify all missing params are now present
- for (const param of missingParams) {
- if (!inferred[param] || inferred[param] === "") {
- return false;
- }
- }
-
- return true;
-}
-
-/**
- * Get tool calling guidance to add to system prompt for local models
- */
-export function getToolCallingGuidance(): string {
- return `
-IMPORTANT TOOL CALLING INSTRUCTIONS:
-When calling tools/functions, you MUST include ALL required parameters. Incomplete tool calls will fail.
-
-For the Task tool, you MUST always provide:
-- description: A short (3-5 word) description of the task
-- prompt: The detailed task instructions
-- subagent_type: The type of agent (e.g., "general-purpose", "Explore", "Plan")
-
-For the Bash tool, you MUST always provide:
-- command: The shell command to execute
-- description: A brief description of what the command does
-
-For file tools (Read, Write, Edit), always provide the full file_path.
-
-For the ToolSearch tool, you MUST always provide:
-- query: The search query string (keywords or "select:tool_name")
-- max_results: Maximum number of results (default: 5)
-
-Format your tool calls as valid JSON with all required fields populated.
-`;
-}
-
-/**
- * Validate and potentially repair a tool call
- * Returns the repaired arguments if successful, null if repair failed
+ * Validate a tool call's arguments against the tool's own JSON Schema.
+ *
+ * `repaired` is reserved for a NON-DESTRUCTIVE transform of what the model
+ * supplied — schema `default`s applied, a value moved to the key the schema
+ * declares, a string coerced to its declared type. It must never again mean
+ * "invented": the inference that used to set it is deleted (see the file header).
+ * Nothing in this commit sets it; item 9 re-establishes it with that meaning.
+ *
+ * `valid: false` therefore means the model genuinely omitted a required
+ * argument, which the caller surfaces to the user as a visible failure rather
+ * than guessing.
*/
export function validateAndRepairToolCall(
toolName: string,
@@ -698,31 +381,11 @@ export function validateAndRepairToolCall(
}
}
- const required = schema.input_schema.required || [];
- const missingParams = required.filter(
- (param) =>
- parsedArgs[param] === undefined || parsedArgs[param] === null || parsedArgs[param] === ""
- );
+ const missingParams = missingRequired(schema.input_schema as JsonSchemaNode, parsedArgs);
if (missingParams.length === 0) {
return { valid: true, args: parsedArgs, repaired: false, missingParams: [] };
}
- // Try to infer missing parameters
- const repairedArgs = inferMissingParameters(toolName, parsedArgs, missingParams, textContent);
-
- // Check if repair was successful
- const stillMissing = required.filter(
- (param) =>
- repairedArgs[param] === undefined ||
- repairedArgs[param] === null ||
- repairedArgs[param] === ""
- );
-
- if (stillMissing.length === 0) {
- log(`[ToolRecovery] Successfully repaired tool call ${toolName}`);
- return { valid: true, args: repairedArgs, repaired: true, missingParams: [] };
- }
-
- return { valid: false, args: repairedArgs, repaired: false, missingParams: stillMissing };
+ return { valid: false, args: parsedArgs, repaired: false, missingParams };
}
diff --git a/packages/cli/src/test-fixtures/sse-responses/grok-4.6-openai-edit-empty-new-string.sse b/packages/cli/src/test-fixtures/sse-responses/grok-4.6-openai-edit-empty-new-string.sse
new file mode 100644
index 00000000..61826dfc
--- /dev/null
+++ b/packages/cli/src/test-fixtures/sse-responses/grok-4.6-openai-edit-empty-new-string.sse
@@ -0,0 +1,101 @@
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":"The","role":"assistant"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" user"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" wants"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" me"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" to"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" delete"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" the"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" line"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" `"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":"const"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" dead"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" ="}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" "}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":"1"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":";`"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" entirely"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576863,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" using"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" the"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" Edit"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" tool"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":"."}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" Change"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" nothing"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" else"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576864,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":".\n"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576869,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":"I understand you want me to delete the line `const dead = 1;` entirely using the Edit tool, changing nothing else."}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576870,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":" However, looking at the current code structure, removing just that line would leave two consecutive blank lines between the alpha and beta functions, which might not be ideal."}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576874,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"reasoning_content":"\n\nLet me check what the exact current code looks like to make sure I'm deleting it correctly."}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":"Removing"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" the"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" `"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":"const"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" dead"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" ="}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" "}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":"1"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":";`"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" line"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" and"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" leaving"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576885,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" the"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576886,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" rest"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576886,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" of"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576886,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" the"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576886,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" file"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576886,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":" unchanged"}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576886,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"content":"."}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576887,"model":"grok-4.6-build","choices":[{"index":0,"delta":{"tool_calls":[{"id":"call-5aff48b6-257c-41f4-bac8-0b177cc3b2c2-3","function":{"name":"Edit","arguments":"{\"file_path\":\"/private/tmp/claude-501/-Users-jack-mag-claudish--claude-worktrees-comparte/441af1f4-d089-45e9-bf08-3261eee149c8/scratchpad/live-capture/sample.ts\",\"old_string\":\"const dead = 1;\\n\",\"new_string\":\"\"}"},"index":0,"type":"function"}]}}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576887,"model":"grok-4.6-build","choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}],"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: {"id":"3c04e832-460c-9258-8d61-f1e5f3c5d93e","object":"chat.completion.chunk","created":1789576887,"model":"grok-4.6-build","choices":[],"usage":{"prompt_tokens":18800,"completion_tokens":109,"total_tokens":20240,"prompt_tokens_details":{"text_tokens":18800,"audio_tokens":0,"image_tokens":0,"cached_tokens":18560},"completion_tokens_details":{"reasoning_tokens":1331,"audio_tokens":0,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0},"num_sources_used":0,"cost_in_usd_ticks":62560000},"system_fingerprint":"fp_08d0bc26c22b024e","service_tier":"default"}
+
+data: [DONE]