fix: gate startup checks strictly on first submission, remove grace period (issue #363)

As gnanam1990 pointed out, the 3s grace period still allows the failure
mode: if a user pauses for a few seconds before typing, startup checks
fire and recommendation dialogs steal focus. A grace period is still a
timing mitigation, not a reliable fix.

New approach: startup checks only run after the user has submitted their
first message (submitCount > 0). No grace period, no timeout. This
guarantees the prompt gets first interaction — no dialog can steal focus
before the user has actually used the CLI.

If the user never submits a message, startup checks never run. That's
acceptable because with no user interaction there's no need for plugin
installations or marketplace seeding.

src/components/agents/generateAgent.ts

@@ -68,11 +68,11 @@ When a user describes what they want an agent to do, you will:
       assistant: "Now let me use the test-runner agent to run the tests"
     </example>
     - <example>
-      Context: User is creating an agent to respond to the word "hello" with a friendly jok.
-      user: "Hello"
-      assistant: "I'm going to use the ${AGENT_TOOL_NAME} tool to launch the greeting-responder agent to respond with a friendly joke"
+      Context: User is creating an agent for Claude Code product questions.
+      user: "How do I configure Claude Code hooks?"
+      assistant: "I'm going to use the ${AGENT_TOOL_NAME} tool to launch the claude-code-guide agent to answer the question"
       <commentary>
-      Since the user is greeting, use the greeting-responder agent to respond with a friendly joke. 
+      Since the user is asking how to use Claude Code, use the claude-code-guide agent.
       </commentary>
     </example>
   - If the user mentioned or implied that the agent should be used proactively, you should include examples of this.

src/screens/REPL.tsx

@@ -238,6 +238,7 @@ import { usePromptsFromClaudeInChrome } from 'src/hooks/usePromptsFromClaudeInCh
 import { getTipToShowOnSpinner, recordShownTip } from 'src/services/tips/tipScheduler.js';
 import type { Theme } from 'src/utils/theme.js';
 import { isPromptTypingSuppressionActive } from './replInputSuppression.js';
+import { shouldRunStartupChecks } from './replStartupGates.js';
 import { checkAndDisableBypassPermissionsIfNeeded, checkAndDisableAutoModeIfNeeded, useKickOffCheckAndDisableBypassPermissionsIfNeeded, useKickOffCheckAndDisableAutoModeIfNeeded } from 'src/utils/permissions/bypassPermissionsKillswitch.js';
 import { SandboxManager } from 'src/utils/sandbox/sandbox-adapter.js';
 import { SANDBOX_NETWORK_ACCESS_TOOL_NAME } from 'src/cli/structuredIO.js';
@@ -792,10 +793,8 @@ export function REPL({
   // accepts, and only then is the REPL component mounted and this effect runs.
   // This ensures that plugin installations from repository and user settings only
   // happen after explicit user consent to trust the current working directory.
-  useEffect(() => {
-    if (isRemoteSession) return;
-    void performStartupChecks(setAppState);
-  }, [setAppState, isRemoteSession]);
+  // Deferring startup checks is handled below (after promptTypingSuppressionActive
+  // is declared) to avoid temporal dead zone issues.
 
   // Allow Claude in Chrome MCP to send prompts through MCP notifications
   // and sync permission mode changes to the Chrome extension
@@ -1429,6 +1428,25 @@ export function REPL({
   const activeRemote = sshRemote.isRemoteMode ? sshRemote : directConnect.isRemoteMode ? directConnect : remoteSession;
   const [pastedContents, setPastedContents] = useState<Record<number, PastedContent>>({});
   const [submitCount, setSubmitCount] = useState(0);
+
+  // Defer startup checks until the user has submitted their first message.
+  // A timeout or grace period is insufficient (issue #363): if the user pauses
+  // before typing, startup checks can still fire and recommendation dialogs
+  // steal focus. Only the user's first submission guarantees the prompt was
+  // the first thing they interacted with.
+  const startupChecksStartedRef = React.useRef(false);
+  const hasHadFirstSubmission = (submitCount ?? 0) > 0;
+  useEffect(() => {
+    if (isRemoteSession) return;
+    if (startupChecksStartedRef.current) return;
+    if (!shouldRunStartupChecks({
+      isRemoteSession,
+      hasStarted: startupChecksStartedRef.current,
+      hasHadFirstSubmission,
+    })) return;
+    startupChecksStartedRef.current = true;
+    void performStartupChecks(setAppState);
+  }, [setAppState, isRemoteSession, hasHadFirstSubmission]);
   // Ref instead of state to avoid triggering React re-renders on every
   // streaming text_delta. The spinner reads this via its animation timer.
   const responseLengthRef = useRef(0);
@@ -2061,13 +2079,14 @@ export function REPL({
     if (allowDialogsWithAnimation && showRemoteCallout) return 'remote-callout';
 
     // LSP plugin recommendation (lowest priority - non-blocking suggestion)
-    if (allowDialogsWithAnimation && lspRecommendation) return 'lsp-recommendation';
+    // Suppress during startup window to prevent stealing focus from the prompt (issue #363)
+    if (allowDialogsWithAnimation && lspRecommendation && startupChecksStartedRef.current) return 'lsp-recommendation';
 
     // Plugin hint from CLI/SDK stderr (same priority band as LSP rec)
-    if (allowDialogsWithAnimation && hintRecommendation) return 'plugin-hint';
+    if (allowDialogsWithAnimation && hintRecommendation && startupChecksStartedRef.current) return 'plugin-hint';
 
     // Desktop app upsell (max 3 launches, lowest priority)
-    if (allowDialogsWithAnimation && showDesktopUpsellStartup) return 'desktop-upsell';
+    if (allowDialogsWithAnimation && showDesktopUpsellStartup && startupChecksStartedRef.current) return 'desktop-upsell';
     return undefined;
   }
   const focusedInputDialog = getFocusedInputDialog();

src/screens/replStartupGates.test.ts

@@ -0,0 +1,53 @@
+import { describe, expect, test } from 'bun:test'
+
+import { shouldRunStartupChecks } from './replStartupGates.js'
+
+describe('shouldRunStartupChecks', () => {
+  test('runs checks after first message submission', () => {
+    expect(shouldRunStartupChecks({
+      isRemoteSession: false,
+      hasStarted: false,
+      hasHadFirstSubmission: true,
+    })).toBe(true)
+  })
+
+  test('skips checks in remote sessions even after submission', () => {
+    expect(shouldRunStartupChecks({
+      isRemoteSession: true,
+      hasStarted: false,
+      hasHadFirstSubmission: true,
+    })).toBe(false)
+  })
+
+  test('skips checks if already started', () => {
+    expect(shouldRunStartupChecks({
+      isRemoteSession: false,
+      hasStarted: true,
+      hasHadFirstSubmission: true,
+    })).toBe(false)
+  })
+
+  test('does not run checks before first submission', () => {
+    expect(shouldRunStartupChecks({
+      isRemoteSession: false,
+      hasStarted: false,
+      hasHadFirstSubmission: false,
+    })).toBe(false)
+  })
+
+  test('does not run checks when idle before first submission', () => {
+    expect(shouldRunStartupChecks({
+      isRemoteSession: false,
+      hasStarted: false,
+      hasHadFirstSubmission: false,
+    })).toBe(false)
+  })
+
+  test('skips checks in remote session regardless of other conditions', () => {
+    expect(shouldRunStartupChecks({
+      isRemoteSession: true,
+      hasStarted: false,
+      hasHadFirstSubmission: false,
+    })).toBe(false)
+  })
+})

src/screens/replStartupGates.ts

@@ -0,0 +1,35 @@
+/**
+ * Startup gates for the REPL.
+ *
+ * Prevents startup plugin checks and recommendation dialogs from stealing
+ * focus before the user has interacted with the prompt.
+ *
+ * This addresses the root cause of issue #363: on mount, performStartupChecks
+ * triggers plugin loading, which populates trackedFiles, which triggers
+ * useLspPluginRecommendation to surface an LSP recommendation dialog. Since
+ * promptTypingSuppressionActive is false before the user has typed anything,
+ * getFocusedInputDialog() returns the dialog, unmounting PromptInput entirely.
+ *
+ * The fix gates startup checks on actual prompt interaction. A pure timeout
+ * or grace period is insufficient because pausing before typing would still
+ * allow dialogs to steal focus. Only the user's first submission guarantees
+ * the prompt is no longer in the vulnerable pre-interaction window.
+ */
+
+/**
+ * Determines whether startup checks should run.
+ *
+ * Startup checks are deferred until the user has submitted their first
+ * message. This guarantees the prompt was the first thing the user interacted
+ * with, so no recommendation dialog can steal focus before the first keystroke.
+ */
+export function shouldRunStartupChecks(options: {
+  isRemoteSession: boolean;
+  hasStarted: boolean;
+  hasHadFirstSubmission: boolean;
+}): boolean {
+  if (options.isRemoteSession) return false;
+  if (options.hasStarted) return false;
+  if (!options.hasHadFirstSubmission) return false;
+  return true;
+}

src/services/tools/toolExecution.test.ts

@@ -0,0 +1,33 @@
+import { describe, expect, test } from 'bun:test'
+
+import { SkillTool } from '../../tools/SkillTool/SkillTool.js'
+import {
+  getSchemaValidationErrorOverride,
+  getSchemaValidationToolUseResult,
+} from './toolExecution.js'
+
+describe('getSchemaValidationErrorOverride', () => {
+  test('returns actionable missing-skill error for SkillTool', () => {
+    expect(getSchemaValidationErrorOverride(SkillTool, {})).toBe(
+      'Missing skill name. Pass the slash command name as the skill parameter (e.g., skill: "commit" for /commit, skill: "review-pr" for /review-pr).',
+    )
+  })
+
+  test('does not override unrelated tool schema failures', () => {
+    expect(getSchemaValidationErrorOverride({ name: 'Read' } as never, {})).toBe(
+      null,
+    )
+  })
+
+  test('does not override SkillTool when skill is present', () => {
+    expect(
+      getSchemaValidationErrorOverride(SkillTool, { skill: 'commit' }),
+    ).toBe(null)
+  })
+
+  test('uses the actionable override for structured toolUseResult too', () => {
+    expect(getSchemaValidationToolUseResult(SkillTool, {} as never)).toBe(
+      'InputValidationError: Missing skill name. Pass the slash command name as the skill parameter (e.g., skill: "commit" for /commit, skill: "review-pr" for /review-pr).',
+    )
+  })
+})

src/services/tools/toolExecution.ts

@@ -43,6 +43,7 @@ import { FILE_READ_TOOL_NAME } from '../../tools/FileReadTool/prompt.js'
 import { FILE_WRITE_TOOL_NAME } from '../../tools/FileWriteTool/prompt.js'
 import { NOTEBOOK_EDIT_TOOL_NAME } from '../../tools/NotebookEditTool/constants.js'
 import { POWERSHELL_TOOL_NAME } from '../../tools/PowerShellTool/toolName.js'
+import { SKILL_TOOL_NAME } from '../../tools/SkillTool/constants.js'
 import { parseGitCommitId } from '../../tools/shared/gitOperationTracking.js'
 import {
   isDeferredTool,
@@ -596,6 +597,31 @@ export function buildSchemaNotSentHint(
   )
 }
 
+export function getSchemaValidationErrorOverride(
+  tool: Tool,
+  input: unknown,
+): string | null {
+  if (tool.name !== SKILL_TOOL_NAME || !input || typeof input !== 'object') {
+    return null
+  }
+
+  const skill = (input as { skill?: unknown }).skill
+  if (skill === undefined || skill === null) {
+    return 'Missing skill name. Pass the slash command name as the skill parameter (e.g., skill: "commit" for /commit, skill: "review-pr" for /review-pr).'
+  }
+
+  return null
+}
+
+export function getSchemaValidationToolUseResult(
+  tool: Tool,
+  input: unknown,
+  fallbackMessage?: string,
+): string {
+  const override = getSchemaValidationErrorOverride(tool, input)
+  return `InputValidationError: ${override ?? fallbackMessage ?? ''}`
+}
+
 async function checkPermissionsAndCallTool(
   tool: Tool,
   toolUseID: string,
@@ -614,7 +640,9 @@ async function checkPermissionsAndCallTool(
   // Validate input types with zod (surprisingly, the model is not great at generating valid input)
   const parsedInput = tool.inputSchema.safeParse(input)
   if (!parsedInput.success) {
-    let errorContent = formatZodValidationError(tool.name, parsedInput.error)
+    const fallbackErrorContent = formatZodValidationError(tool.name, parsedInput.error)
+    let errorContent =
+      getSchemaValidationErrorOverride(tool, input) ?? fallbackErrorContent
 
     const schemaHint = buildSchemaNotSentHint(
       tool,
@@ -672,7 +700,11 @@ async function checkPermissionsAndCallTool(
               tool_use_id: toolUseID,
             },
           ],
-          toolUseResult: `InputValidationError: ${parsedInput.error.message}`,
+          toolUseResult: getSchemaValidationToolUseResult(
+            tool,
+            input,
+            parsedInput.error.message,
+          ),
           sourceToolAssistantUUID: assistantMessage.uuid,
         }),
       },

src/tools/AgentTool/prompt.ts

@@ -156,34 +156,24 @@ ${AGENT_TOOL_NAME}({
   const currentExamples = `Example usage:
 
 <example_agent_descriptions>
-"test-runner": use this agent after you are done writing code to run tests
-"greeting-responder": use this agent to respond to user greetings with a friendly joke
+"claude-code-guide": use this agent when the user asks how Claude Code works or how to use its features
+"statusline-setup": use this agent to configure the user's Claude Code status line setting
 </example_agent_descriptions>
 
 <example>
-user: "Please write a function that checks if a number is prime"
-assistant: I'm going to use the ${FILE_WRITE_TOOL_NAME} tool to write the following code:
-<code>
-function isPrime(n) {
-  if (n <= 1) return false
-  for (let i = 2; i * i <= n; i++) {
-    if (n % i === 0) return false
-  }
-  return true
-}
-</code>
+user: "How do I configure Claude Code hooks?"
 <commentary>
-Since a significant piece of code was written and the task was completed, now use the test-runner agent to run the tests
+This is a Claude Code usage question, so use the claude-code-guide agent
 </commentary>
-assistant: Uses the ${AGENT_TOOL_NAME} tool to launch the test-runner agent
+assistant: Uses the ${AGENT_TOOL_NAME} tool to launch the claude-code-guide agent
 </example>
 
 <example>
-user: "Hello"
+user: "Set up my Claude Code status line"
 <commentary>
-Since the user is greeting, use the greeting-responder agent to respond with a friendly joke
+This matches the statusline-setup agent, so use it to configure the setting
 </commentary>
-assistant: "I'm going to use the ${AGENT_TOOL_NAME} tool to launch the greeting-responder agent"
+assistant: "I'm going to use the ${AGENT_TOOL_NAME} tool to launch the statusline-setup agent"
 </example>
 `
 

src/tools/SkillTool/SkillTool.test.ts

@@ -0,0 +1,31 @@
+import { describe, expect, test } from 'bun:test'
+
+import { SkillTool } from './SkillTool.js'
+
+describe('SkillTool missing parameter handling', () => {
+  test('missing skill stays required at the schema level', async () => {
+    const parsed = SkillTool.inputSchema.safeParse({})
+
+    expect(parsed.success).toBe(false)
+  })
+
+  test('validateInput still returns an actionable error when called with missing skill', async () => {
+    const result = await SkillTool.validateInput?.({} as never, {
+      options: { tools: [] },
+      messages: [],
+    } as never)
+
+    expect(result).toEqual({
+      result: false,
+      message:
+        'Missing skill name. Pass the slash command name as the skill parameter (e.g., skill: "commit" for /commit, skill: "review-pr" for /review-pr).',
+      errorCode: 1,
+    })
+  })
+
+  test('valid skill input still parses and validates', async () => {
+    const parsed = SkillTool.inputSchema.safeParse({ skill: 'commit' })
+
+    expect(parsed.success).toBe(true)
+  })
+})

src/tools/SkillTool/SkillTool.ts

@@ -352,6 +352,16 @@ export const SkillTool: Tool<InputSchema, Output, Progress> = buildTool({
   toAutoClassifierInput: ({ skill }) => skill ?? '',
 
   async validateInput({ skill }, context): Promise<ValidationResult> {
+    if (!skill || typeof skill !== 'string') {
+      return {
+        result: false,
+        message:
+          'Missing skill name. Pass the slash command name as the skill parameter ' +
+          '(e.g., skill: "commit" for /commit, skill: "review-pr" for /review-pr).',
+        errorCode: 1,
+      }
+    }
+
     // Skills are just skill names, no arguments
     const trimmed = skill.trim()
     if (!trimmed) {
@@ -434,7 +444,7 @@ export const SkillTool: Tool<InputSchema, Output, Progress> = buildTool({
     context,
   ): Promise<PermissionDecision> {
     // Skills are just skill names, no arguments
-    const trimmed = skill.trim()
+    const trimmed = skill ?? ''
 
     // Remove leading slash if present (for compatibility)
     const commandName = trimmed.startsWith('/') ? trimmed.substring(1) : trimmed
@@ -592,7 +602,7 @@ export const SkillTool: Tool<InputSchema, Output, Progress> = buildTool({
     // - Skill is a prompt-based skill
 
     // Skills are just names, with optional arguments
-    const trimmed = skill.trim()
+    const trimmed = skill ?? ''
 
     // Remove leading slash if present (for compatibility)
     const commandName = trimmed.startsWith('/') ? trimmed.substring(1) : trimmed

src/utils/api.test.ts

@@ -1,6 +1,7 @@
 import { expect, test } from 'bun:test'
 import { z } from 'zod/v4'
 import { getEmptyToolPermissionContext, type Tool, type Tools } from '../Tool.js'
+import { SkillTool } from '../tools/SkillTool/SkillTool.js'
 import { toolToAPISchema } from './api.js'
 
 test('toolToAPISchema preserves provider-specific schema keywords in input_schema', async () => {
@@ -64,3 +65,16 @@ test('toolToAPISchema preserves provider-specific schema keywords in input_schem
     },
   })
 })
+
+test('toolToAPISchema keeps skill required for SkillTool', async () => {
+  const schema = await toolToAPISchema(SkillTool, {
+    getToolPermissionContext: async () => getEmptyToolPermissionContext(),
+    tools: [] as unknown as Tools,
+    agents: [],
+  })
+
+  expect((schema as { input_schema: unknown }).input_schema).toMatchObject({
+    type: 'object',
+    required: ['skill'],
+  })
+})

src/utils/context.test.ts

@@ -94,3 +94,22 @@ test('gpt-5.4 family keeps large max output overrides within provider limits', (
   expect(getMaxOutputTokensForModel('gpt-5.4-mini')).toBe(128_000)
   expect(getMaxOutputTokensForModel('gpt-5.4-nano')).toBe(128_000)
 })
+
+test('MiniMax-M2.7 uses explicit provider-specific context and output caps', () => {
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  delete process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
+
+  expect(getContextWindowForModel('MiniMax-M2.7')).toBe(204_800)
+  expect(getModelMaxOutputTokens('MiniMax-M2.7')).toEqual({
+    default: 131_072,
+    upperLimit: 131_072,
+  })
+  expect(getMaxOutputTokensForModel('MiniMax-M2.7')).toBe(131_072)
+})
+
+test('unknown openai-compatible models still use the conservative fallback window', () => {
+  process.env.CLAUDE_CODE_USE_OPENAI = '1'
+  delete process.env.CLAUDE_CODE_MAX_OUTPUT_TOKENS
+
+  expect(getContextWindowForModel('some-unknown-3p-model')).toBe(8_000)
+})

src/utils/context.ts

@@ -74,7 +74,7 @@ export function getContextWindowForModel(
 
   // OpenAI-compatible provider — use known context windows for the model.
   // Unknown models get a conservative 8k default so auto-compact triggers
-  // before hitting a hard context_window_exceeded error (issue #248 finding 3).
+  // before hitting a hard context_window_exceeded error.
   const isOpenAIProvider =
     isEnvTruthy(process.env.CLAUDE_CODE_USE_OPENAI) ||
     isEnvTruthy(process.env.CLAUDE_CODE_USE_GEMINI) ||

src/utils/model/openaiContextWindows.ts

@@ -44,6 +44,10 @@ const OPENAI_CONTEXT_WINDOWS: Record<string, number> = {
   'mistral-large-latest':     131_072,
   'mistral-small-latest':     131_072,
 
+  // MiniMax
+  'MiniMax-M2.7':             204_800,
+  'minimax-m2.7':             204_800,
+
   // Google (via OpenRouter)
   'google/gemini-2.0-flash':1_048_576,
   'google/gemini-2.5-pro':  1_048_576,
@@ -110,6 +114,10 @@ const OPENAI_MAX_OUTPUT_TOKENS: Record<string, number> = {
   'mistral-large-latest':     32_768,
   'mistral-small-latest':     32_768,
 
+  // MiniMax
+  'MiniMax-M2.7':            131_072,
+  'minimax-m2.7':            131_072,
+
   // Google (via OpenRouter)
   'google/gemini-2.0-flash':   8_192,
   'google/gemini-2.5-pro':    65_536,