feat: Update opinionated CLI config to include Opus, Sonnet, and Haiku models in order

[sestinj]

Summary

Updates the CLI's opinionated default configuration to include all three Claude 4-6 models in the preferred order when using the ANTHROPIC_API_KEY fallback.

Changes

• Modified updateAnthropicModelInYaml() to create all three Claude models instead of just Sonnet
• Model order: Opus (first/default), Sonnet (second), Haiku (third)
• Updated filtering logic to handle all three Claude model types
• Updated all test cases to verify the new three-model configuration

When This Gets Used

This opinionated config is created when:

• Headless mode with ANTHROPIC_API_KEY environment variable set (no config file)
• Interactive onboarding when user chooses Enter your Anthropic API key
• Normal flow fallback when user has completed onboarding but no config exists and ANTHROPIC_API_KEY is available

Testing

• All existing tests pass
• Updated test expectations for three-model configuration
• E2E tests with ANTHROPIC_API_KEY flow still work

---

Continue Tasks: ✅ 7 no changes — View all

---

Summary by cubic

Updates the CLI’s default Anthropic config to include Opus, Sonnet, and Haiku in order, adds skill slash commands with direct invocation, enables image-aware file reads, and introduces a structured plan mode with an ExitPlanMode tool and UI polish. Also fixes assistant prefill errors after compaction by appending a user “continue” message.

• New Features

  • Fallback config (when ANTHROPIC_API_KEY is set) now creates Opus, Sonnet, and Haiku in that order.
  • Skills: invoke via /skills, /skills , or /; autocomplete shows skills loaded from .continue/skills/, .claude/skills/, or ~/.continue/skills/. Optional when_to_use frontmatter; structured tool prompts/output with <available_skills>, , and <skill_files>.
  • Plan mode: new ExitPlanMode tool (plan-only), clearer system message/workflow, and permission policy updated to allow ExitPlanMode.
  • Read tool supports images (png/jpg/gif/webp) and returns multipart results; added an AI SDK adapter so image parts are sent to models. Tests cover text and image cases.
  • UI: user messages render as full-width white blocks, leading newlines are trimmed from assistant output, Bash output lines capped at 200 chars, and Esc dismisses Quiz prompts. Permission selector replaces the “dangerous commands” warning with an Auto mode tip. AGENTS.md adds local build/link steps.

• Bug Fixes

  • Compaction now appends a user “continue” message to avoid assistant prefill errors with Anthropic models.

^{Written for commit d546f88. Summary will update on new commits.}

[continue]

Docs Review: No documentation updates needed.

This PR changes the internal default model configuration for the Anthropic API key flow—adding Opus, Sonnet, and Haiku instead of just Sonnet. The CLI documentation intentionally stays at a high level ("you can use an Anthropic API key directly") without specifying which exact models get auto-configured.

This is appropriate because:

• Users see the configured models in their config.yaml after setup
• Documenting specific model defaults would create maintenance burden
• The change doesn't affect how users interact with the CLI or configure it

[cubic-dev-ai]

2 issues found across 10 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="extensions/cli/src/ui/SlashCommandUI.tsx">

<violation number="1" location="extensions/cli/src/ui/SlashCommandUI.tsx:43">
P2: Async command loading can race with assistant changes or unmounts; the later-resolving promise will overwrite newer commands or trigger setState on an unmounted component. Add a cancellation/active flag in the effect so only the latest request updates state.</violation>
</file>

<file name="extensions/cli/src/ui/UserInput.tsx">

<violation number="1" location="extensions/cli/src/ui/UserInput.tsx:190">
P2: When `assistant`/`isRemoteMode` becomes false, the cached slash commands are never reset, so the UI can keep showing stale assistant/remote commands after switching modes. Reset the ref to the fallback list in the `else` branch.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P2: Async command loading can race with assistant changes or unmounts; the later-resolving promise will overwrite newer commands or trigger setState on an unmounted component. Add a cancellation/active flag in the effect so only the latest request updates state.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At extensions/cli/src/ui/SlashCommandUI.tsx, line 43:

<comment>Async command loading can race with assistant changes or unmounts; the later-resolving promise will overwrite newer commands or trigger setState on an unmounted component. Add a cancellation/active flag in the effect so only the latest request updates state.</comment>

<file context>
@@ -29,20 +29,19 @@ const SlashCommandUI: React.FC<SlashCommandUIProps> = ({
+      getAllSlashCommands(assistant || ({} as AssistantConfig), {
         isRemoteMode,
-      });
+      }).then(setAllCommands);
     }
-
</file context>

[cubic-dev-ai]

P2: When assistant/isRemoteMode becomes false, the cached slash commands are never reset, so the UI can keep showing stale assistant/remote commands after switching modes. Reset the ref to the fallback list in the else branch.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At extensions/cli/src/ui/UserInput.tsx, line 190:

<comment>When `assistant`/`isRemoteMode` becomes false, the cached slash commands are never reset, so the UI can keep showing stale assistant/remote commands after switching modes. Reset the ref to the fallback list in the `else` branch.</comment>

<file context>
@@ -180,20 +180,24 @@ const UserInput: React.FC<UserInputProps> = ({
+    { name: "exit", description: "Exit the chat", category: "system" },
+  ]);
+
+  useEffect(() => {
     if (assistant || isRemoteMode) {
-      return getAllSlashCommands(assistant || ({} as AssistantConfig), {
</file context>

Suggested change

	useEffect(() => {
	if (assistant || isRemoteMode) {
	return getAllSlashCommands(assistant || ({} as AssistantConfig), {
	getAllSlashCommands(assistant || ({} as AssistantConfig), {
	isRemoteMode,
	}).then((commands) => {
	slashCommandsRef.current = commands;
	});
	}
	}, [assistant?.prompts, assistant?.rules, isRemoteMode]);
	useEffect(() => {
	if (assistant || isRemoteMode) {
	getAllSlashCommands(assistant || ({} as AssistantConfig), {
	isRemoteMode,
	}).then((commands) => {
	slashCommandsRef.current = commands;
	});
	return;
	}
	slashCommandsRef.current = [
	{ name: "help", description: "Show help message", category: "system" },
	{ name: "clear", description: "Clear the chat history", category: "system" },
	{ name: "exit", description: "Exit the chat", category: "system" },
	];
	}, [assistant?.prompts, assistant?.rules, isRemoteMode]);

[cubic-dev-ai]

4 issues found across 33 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="extensions/cli/src/tools/skills.ts">

<violation number="1" location="extensions/cli/src/tools/skills.ts:127">
P2: Derive the skill directory using both path separators (or a path utility). The current substring("/") logic fails on Windows paths and produces an empty skill directory.</violation>
</file>

<file name="extensions/cli/src/ui/components/MemoizedMessage.tsx">

<violation number="1" location="extensions/cli/src/ui/components/MemoizedMessage.tsx:171">
P2: Guard against non-positive terminal widths; `contentWidth` can be <= 0 (e.g., non‑TTY), which makes the wrap loop non-terminating and can hang rendering. Clamp `contentWidth` to at least 1 (or return early) before the wrapping loop.</violation>
</file>

<file name="extensions/cli/src/tools/readFile.ts">

<violation number="1" location="extensions/cli/src/tools/readFile.ts:140">
P2: Reading an image returns before markFileAsRead, so the Edit tool will still reject the same file as “not read.” If the Read tool now supports images, it should still record the file in readFilesSet to keep the “read before edit” guard consistent.</violation>
</file>

<file name="extensions/cli/src/tools/types.ts">

<violation number="1" location="extensions/cli/src/tools/types.ts:72">
P2: `typeof null === "object"` is `true` in JavaScript. If a `null` value reaches this type guard at runtime (e.g., from an untyped boundary or unexpected tool result), accessing `result.type` will throw a `TypeError`. Add a null check for defensive safety.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[cubic-dev-ai]

P2: Derive the skill directory using both path separators (or a path utility). The current substring("/") logic fails on Windows paths and produces an empty skill directory.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At extensions/cli/src/tools/skills.ts, line 127:

<comment>Derive the skill directory using both path separators (or a path utility). The current substring("/") logic fails on Windows paths and produces an empty skill directory.</comment>

<file context>
@@ -63,20 +116,40 @@ ${skills.map((skill) => `\nname: ${skill.name}\ndescription: ${skill.description
-        content.push(
-          `<skill_files>${skill.files.join(",")}</skill_files>`,
-          `<other_instructions>Use the read file tool to access skill files as needed.</other_instructions>`,
+        const skillDir = skill.path.substring(
+          0,
+          skill.path.lastIndexOf("/"),
</file context>

[cubic-dev-ai]

P2: Guard against non-positive terminal widths; contentWidth can be <= 0 (e.g., non‑TTY), which makes the wrap loop non-terminating and can hang rendering. Clamp contentWidth to at least 1 (or return early) before the wrapping loop.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At extensions/cli/src/ui/components/MemoizedMessage.tsx, line 171:

<comment>Guard against non-positive terminal widths; `contentWidth` can be <= 0 (e.g., non‑TTY), which makes the wrap loop non-terminating and can hang rendering. Clamp `contentWidth` to at least 1 (or return early) before the wrapping loop.</comment>

<file context>
@@ -162,19 +164,49 @@ export const MemoizedMessage = memo<MemoizedMessageProps>(
+      const text = formatMessageContentForDisplay(message.content);
+      const termWidth = process.stdout.columns || 80;
+      // 1 char padding on each side
+      const contentWidth = termWidth - 2;
+      // Word-wrap then pad each line to fill the terminal width
+      const wrappedLines: string[] = [];
</file context>

Suggested change

	const contentWidth = termWidth - 2;
	const contentWidth = Math.max(termWidth - 2, 1);

[cubic-dev-ai]

P2: Reading an image returns before markFileAsRead, so the Edit tool will still reject the same file as “not read.” If the Read tool now supports images, it should still record the file in readFilesSet to keep the “read before edit” guard consistent.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At extensions/cli/src/tools/readFile.ts, line 140:

<comment>Reading an image returns before markFileAsRead, so the Edit tool will still reject the same file as “not read.” If the Read tool now supports images, it should still record the file in readFilesSet to keep the “read before edit” guard consistent.</comment>

<file context>
@@ -81,6 +135,12 @@ export const readFileTool: Tool = {
       const realPath = fs.realpathSync(filepath);
+
+      // Handle image files
+      if (isImageFile(filepath)) {
+        return readImageFile(realPath, filepath);
+      }
</file context>

Suggested change

	if (isImageFile(filepath)) {
	return readImageFile(realPath, filepath);
	}
	if (isImageFile(filepath)) {
	markFileAsRead(realPath);
	return readImageFile(realPath, filepath);
	}

[cubic-dev-ai]

P2: typeof null === "object" is true in JavaScript. If a null value reaches this type guard at runtime (e.g., from an untyped boundary or unexpected tool result), accessing result.type will throw a TypeError. Add a null check for defensive safety.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At extensions/cli/src/tools/types.ts, line 72:

<comment>`typeof null === "object"` is `true` in JavaScript. If a `null` value reaches this type guard at runtime (e.g., from an untyped boundary or unexpected tool result), accessing `result.type` will throw a `TypeError`. Add a null check for defensive safety.</comment>

<file context>
@@ -41,13 +41,57 @@ export interface ToolRunContext {
+export function isMultipartToolResult(
+  result: ToolResult,
+): result is MultipartToolResult {
+  return typeof result === "object" && result.type === "multipart";
+}
+
</file context>

Suggested change

	return typeof result === "object" && result.type === "multipart";
	return typeof result === "object" && result !== null && result.type === "multipart";