feat(polish): multi-provider UX improvements

[JeanBrasse]

Summary

• Fix duplicate React key warnings by filtering .worktrees/ paths from project lists
• Rename ZhipuAI → Zai across registry, settings, and UI
• Add SVG provider logos for OpenRouter, DeepSeek, Moonshot, MiMo, Qwen, Zai
• Block provider selection without configured API key (tooltip: "Add API key in Settings")
• Replace fixed 6-model grid with dynamic dropdown in NewChatModal
• Usage page: multi-provider pricing table, per-provider breakdown, updated header
• Settings: move AI Providers below Terminal, merge OpenCode/Pi into CLI Paths, remove standalone pages
• Update Git/Permissions/Skills/Terminal descriptions for multi-provider context

Test plan

• TypeScript: zero source errors (npx tsc --noEmit)
• QA-Dorothy: 9/9 checks pass
• No duplicate React key warnings in project pickers
• Provider selection disabled without API key
• Settings sidebar reordered correctly
• Usage page shows all-provider pricing and breakdown

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features

  • Added support for multiple AI providers (OpenRouter, DeepSeek, Moonshot, MiMo, Qwen, Zhipu)
  • Introduced Orchestrator mode for agents with restricted tool access
  • Added skills configuration for agents and scheduled tasks
  • Implemented usage tracking and cost breakdown by provider
  • Enhanced terminal with manual scroll-to-bottom functionality

• Bug Fixes

  • Improved PTY working directory tracking to prevent stale session issues
  • Better workspace trust handling for agent initialization

• UI/UX

  • New AI Providers settings section for API key management
  • Enhanced provider icons and badges throughout the interface
  • Skills selection UI for task creation

[vercel]

@JeanBrasse is attempting to deploy a commit to the Charlie Rabiller's projects Team on Vercel.

A member of the Team first needs to authorize it.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR introduces multi-provider LLM support with stale PTY lifecycle management, orchestrator mode for tool restrictions, skill injection into scheduled tasks and automations, and comprehensive frontend configuration UI for managing external AI provider credentials and defaults.

Changes

Cohort / File(s)	Summary
Agent PTY Lifecycle Management electron/core/agent-manager.ts, electron/services/api-routes/agent-routes.ts, electron/handlers/ipc-handlers.ts, electron/main.ts	Added killStalePty() and ensureProjectTrusted() functions to detect and terminate out-of-date PTYs based on working directory mismatch; integrated PTY working directory tracking via agent.ptyCwd across startup, message handling, and worktree updates.
New Provider Implementations electron/providers/deepseek-provider.ts, electron/providers/mimo-provider.ts, electron/providers/moonshot-provider.ts, electron/providers/openrouter-provider.ts, electron/providers/qwen-provider.ts, electron/providers/zhipu-provider.ts, electron/providers/index.ts	Added six new CLI-compatible provider classes implementing CLIProvider interface with MCP support, skill directory handling, scheduled script generation, and environment variable wiring for OpenRouter, DeepSeek, MiMo, Moonshot, Qwen, and Zai; registered in provider index.
Existing Provider Updates electron/providers/claude-provider.ts, electron/providers/codex-provider.ts, electron/providers/gemini-provider.ts, electron/providers/opencode-provider.ts, electron/providers/pi-provider.ts, electron/providers/cli-provider.ts	Extended all providers with CLAUDE_PROVIDER environment variable injection, optional skills parameter in scheduled script generation with prompt prefixing, and optional AppSettings parameter in PTY env var construction; added orchestratorMode field to interactive command params.
Orchestrator Mode & Tool Restriction electron/handlers/ipc-handlers.ts, electron/services/api-routes/agent-routes.ts, electron/services/slack-bot.ts, electron/services/telegram-bot.ts, electron/preload.ts	Added orchestratorMode flag to agent creation/update payloads; when enabled, restricts file-mutation tools (Edit/Write/MultiEdit/NotebookEdit) via --disallowed-tools CLI flag for delegation-only execution.
Scheduling & Automation Skills Integration electron/handlers/scheduler-handlers.ts, electron/handlers/automation-handlers.ts	Extended SchedulerTaskMetadata and Automation with optional skills field; updated task/automation creation to accept provider and skills; modified scheduled script generation to prepend skill instructions to prompts when skills are present.
Frontend Provider Registry & UI src/lib/providers.ts, src/components/ProviderBadge.tsx, src/components/AgentList/AgentCard.tsx, src/components/NewChatModal/index.tsx	Created centralized provider registry with model metadata, display properties, and icon definitions; introduced ProviderIcon component for dynamic icon rendering; updated agent/chat UI to render provider badges/icons from registry.
Settings UI - AI Providers src/components/Settings/AIProvidersSection.tsx, src/components/Settings/CLIPathsSection.tsx, src/components/Settings/constants.ts, src/components/Settings/index.ts, src/components/Settings/types.ts, src/components/Settings/GeneralSection.tsx	Added AIProvidersSection for configuring external provider API keys and enablement flags; extended CLIPathsSection with OpenCode/Pi toggles; updated GeneralSection to dynamically populate default provider selection from registry with availability detection for CLI and API-key-based providers.
Recurring Tasks & Automations UI src/app/recurring-tasks/components/CreateTaskModal.tsx, src/app/recurring-tasks/page.tsx, src/app/recurring-tasks/hooks/useTaskForm.ts, src/app/recurring-tasks/types.ts, src/app/automations/page.tsx	Extended task/automation creation forms with provider selection and skills multi-select; updated task listing to display provider and skills badges; integrated Claude-provided defaults for available skills and default provider.
Usage Analytics Per Provider src/app/usage/page.tsx, src/hooks/useClaude.ts	Extended token statistics tracking to aggregate usage by provider via providerTotals; added provider-specific usage table and breakdown cards; introduced fallback pricing for non-Claude providers.
Terminal Improvements src/components/AgentWorld/useAgentDialogTerminal.ts, src/components/AgentWorld/AgentTerminalDialog.tsx	Added bottom-tracking state to detect viewport position; conditionally auto-scroll on resize/fullscreen only when at bottom; exposed scrollToBottom callback and isAtBottom state; added manual scroll-to-bottom button when not at bottom.
Type System Extensions electron/types/index.ts, src/types/electron.d.ts, src/lib/claude-code.ts	Expanded AgentProvider union to include new providers; added ptyCwd and orchestratorMode fields to AgentStatus; extended AppSettings with provider enablement/API-key fields; added defaultProvider to ClaudeSettings.
Miscellaneous src/app/agents/page.tsx, src/components/CanvasView/hooks/useAgentActions.ts, src/components/Settings/GitSection.tsx, src/components/Settings/PermissionsSection.tsx, src/components/Settings/SkillsSection.tsx, src/components/Settings/TerminalSection.tsx, src/components/Settings/McpSection.tsx, src/components/Memory/page.tsx, src/app/skills/page.tsx, src/components/NewChatModal/StepModel.tsx, src/components/NewChatModal/StepTools.tsx, src/components/NewChatModal/types.tsx, src/hooks/useElectron.ts, src/components/KanbanBoard/components/NewTaskModal.tsx, electron/utils/statusline.ts, mcp-orchestrator/src/tools/agents.ts, __tests__/electron/services/api-routes/agent-routes.test.ts	Updated agent editing to persist orchestratorMode; refactored provider badge rendering from hardcoded provider lists to dynamic registry lookup; generalized UI copy from "Claude" to "AI" or "all providers"; extended test suite for PTY lifecycle and message reconnection; added provider field to token statistics logging; updated MCP tools model descriptions.

Sequence Diagrams

sequenceDiagram
    participant Client
    participant AgentRoutes as Agent Routes
    participant AgentManager as Agent Manager
    participant Provider
    participant PTY as Node-PTY
    participant FileSystem as File System

    Client->>AgentRoutes: POST /api/agents/:id/start
    AgentRoutes->>AgentManager: killStalePty(agent)
    AgentManager->>FileSystem: Compare agent.ptyCwd vs expected cwd
    alt PTY mismatch detected
        AgentManager->>PTY: Kill stale PTY process
        AgentManager->>AgentManager: Clear ptyCwd & ptyId
    end
    AgentRoutes->>AgentManager: ensureProjectTrusted(cwd)
    AgentManager->>FileSystem: Mark ~/.claude.json project trust
    AgentRoutes->>Provider: getProvider(agent.provider)
    Provider->>AgentRoutes: Return provider instance
    AgentRoutes->>Provider: resolveBinaryPath(appSettings)
    Provider->>AgentRoutes: Return CLI binary path
    AgentRoutes->>Provider: buildInteractiveCommand(params + orchestratorMode)
    Provider->>AgentRoutes: Return command with tool restrictions if orchestrator
    AgentRoutes->>Provider: getPtyEnvVars(agentId, projectPath, skills, appSettings)
    Provider->>AgentRoutes: Return env vars including CLAUDE_PROVIDER
    AgentRoutes->>PTY: spawn(cmd, cwd, env)
    PTY->>AgentRoutes: PTY process created
    AgentRoutes->>AgentManager: Store agent.ptyCwd = cwd
    AgentRoutes->>Client: Return { success, agent }

Loading

sequenceDiagram
    participant Client
    participant AgentRoutes
    participant AgentManager
    participant Provider
    participant Prompt as Prompt Builder

    Client->>AgentRoutes: POST /api/agents/:id/message (PTY missing)
    AgentRoutes->>AgentManager: killStalePty(agent)
    AgentManager->>AgentManager: Detect & clear stale PTY
    AgentRoutes->>Provider: getProvider(agent.provider)
    Provider->>AgentRoutes: Provider instance
    AgentRoutes->>Prompt: Build prompt with skills injection
    alt skills present
        Prompt->>Prompt: Prepend "IMPORTANT: Use these skills" + list
    else no skills
        Prompt->>Prompt: Use raw prompt
    end
    AgentRoutes->>Provider: buildInteractiveCommand(params)
    Provider->>Prompt: Return command
    AgentRoutes->>AgentRoutes: Spawn auto-reconnect one-shot PTY
    AgentRoutes->>AgentManager: Update agent.ptyCwd & ptyId
    AgentRoutes->>Client: Return { success }

Loading

sequenceDiagram
    participant User
    participant SchedulerUI as Scheduler UI
    participant SchedulerHandler as Scheduler Handler
    participant Provider
    participant FileSystem
    participant CronJob as Cron/Launchd

    User->>SchedulerUI: Create scheduled task + provider + skills
    SchedulerUI->>SchedulerHandler: scheduler:createTask(config)
    SchedulerHandler->>Provider: getProvider(config.provider)
    Provider->>SchedulerHandler: Provider instance
    SchedulerHandler->>Provider: buildScheduledScript({prompt, skills, ...})
    Provider->>Provider: Build prompt with skills prefix if present
    Provider->>SchedulerHandler: Return bash script with skill-annotated prompt
    SchedulerHandler->>FileSystem: Save skills + provider to SchedulerTaskMetadata
    SchedulerHandler->>CronJob: Register scheduled job with script
    CronJob->>User: Task scheduled with provider & skills

Loading

Estimated code review effort

🎯 5 (Critical) | ⏱️ ~120 minutes

Possibly related PRs

• feat: terminal fixes, agent permission modes, effort levels, usage page overhaul & UI polish #43: Overlapping changes to agent/PTY lifecycle management—both modify ptyId handling, add ptyCwd tracking, refactor provider binary resolution, and update PTY spawning with env var plumbing across agent-manager, ipc-handlers, and service routes.
• fix: CLI path detection, settings sidebar scroll, and UX improvements #46: Touches the same useAgentDialogTerminal hook for terminal auto-scroll behavior; both PRs add scroll-to-bottom logic but with different triggering conditions (one adds bottom-tracking and conditional scroll, the other adds fit/resize handlers).

Poem

🐰 Hoppy hops through provider lands so wide,
Six new companions now by our side—
DeepSeek, Moonshot, Qwen in the fray,
With skills injected and PTYs at bay,
Orchestrator mode keeps the toolbox tight,
Dorothy's expanding—what a delight! 🚀

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 40.48% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The PR title 'feat(polish): multi-provider UX improvements' directly and clearly summarizes the main change: UI/UX enhancements focused on supporting multiple AI providers.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 8

Note

Due to the large number of review comments, Critical severity comments were prioritized as inline comments.

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (10)

electron/services/telegram-bot.ts (1)

1181-1188: ⚠️ Potential issue | 🟠 Major

Don't reuse the active-session path after killing a stale PTY.

Lines 1181-1188 can tear down and recreate the Super Agent PTY, but the next branch still trusts superAgent.status. If that status is still running/waiting, Line 1210 writes raw input into a fresh shell instead of launching a new provider session. Reset the session state when killStalePty(...) swaps the PTY, or force the new-session branch in that case.

Also applies to: 1197-1210

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/services/telegram-bot.ts` around lines 1181 - 1188, The code may
kill and recreate a Super Agent PTY via killStalePty(superAgent) but then still
trusts superAgent.status and may write raw input into a fresh shell; after
killStalePty(superAgent) detect that the agent's pty was replaced (compare
previous superAgent.ptyId vs current or check ptyProcesses) and clear or reset
the session-related state on superAgent (e.g., set session id/state to null or a
sentinel and set superAgent.status to a non-running value) so the subsequent
logic (including the branch that writes to an existing session) will instead
take the "create new provider session" path; apply the same reset when
initAgentPty(superAgent) returns a new ptyId to ensure new PTY never receives
raw input intended for the old session.

electron/providers/index.ts (1)

24-35: ⚠️ Potential issue | 🟠 Major

Don't silently downgrade unknown providers to Claude.

With several external providers now registered, Line 46 turns any typo/rename/config drift into “run it with Claude anyway”. That is the wrong failure mode here: it can hit the wrong account, wrong model, or wrong billing path and is hard to diagnose. Keep the Claude fallback only for undefined/local and reject unsupported ids explicitly.

🛡️ Suggested fix

 export function getProvider(id?: AgentProvider): CLIProvider {
   if (!id || id === 'local') {
     return providers.claude;
   }
-  return providers[id] || providers.claude;
+  const provider = providers[id];
+  if (!provider) {
+    throw new Error(`Unsupported provider: ${id}`);
+  }
+  return provider;
 }

Also applies to: 42-46

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/index.ts` around lines 24 - 35, The providers map
currently lists available providers (providers: Record<string, CLIProvider>) and
the code falls back to new ClaudeProvider() for unknown ids; change the lookup
logic so that only the specific cases 'undefined' or 'local' still map to
Claude, but any other unrecognized provider id causes an explicit
rejection/throw (or returns an error) instead of silently using Claude; update
the resolution path that currently uses ClaudeProvider as a generic fallback to
validate the id against the providers map and throw a clear error mentioning the
unsupported provider id, referencing the providers constant and ClaudeProvider
to locate the implementation.

src/components/CanvasView/hooks/useAgentActions.ts (1)

4-14: ⚠️ Potential issue | 🟠 Major

Forward the selected provider/model into createAgent.

handleCreateAgent accepts model, _provider, _localModel, and _obsidianVaultPaths, but the payload sent to createAgent(...) drops all of them. Agents created from this flow will ignore the user's provider/local-model/vault choices and fall back to the default configuration.

💡 Suggested fix

-import type { AgentCharacter, AgentStatus } from '@/types/electron';
+import type { AgentCharacter, AgentProvider, AgentStatus } from '@/types/electron';

 interface CreateAgentConfig {
   projectPath: string;
   skills: string[];
   worktree?: { enabled: boolean; branchName: string };
   character?: AgentCharacter;
   name?: string;
   secondaryProjectPath?: string;
   permissionMode?: 'normal' | 'auto' | 'bypass';
   effort?: 'low' | 'medium' | 'high';
+  provider?: AgentProvider;
+  model?: string;
+  localModel?: string;
+  obsidianVaultPaths?: string[];
   orchestratorMode?: boolean;
 }

-    _provider?: string,
-    _localModel?: string,
-    _obsidianVaultPaths?: string[],
+    provider?: AgentProvider,
+    localModel?: string,
+    obsidianVaultPaths?: string[],
     effort?: 'low' | 'medium' | 'high',
     orchestratorMode?: boolean,
   ) => {
     try {
-      const agent = await createAgent({ projectPath, skills, worktree, character, name, secondaryProjectPath, permissionMode, effort, orchestratorMode });
+      const agent = await createAgent({
+        projectPath,
+        skills,
+        worktree,
+        character,
+        name,
+        secondaryProjectPath,
+        permissionMode,
+        effort,
+        provider,
+        model,
+        localModel,
+        obsidianVaultPaths,
+        orchestratorMode,
+      });

Also applies to: 19-20, 68-85

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/CanvasView/hooks/useAgentActions.ts` around lines 4 - 14, The
create flow currently drops the selected
model/provider/localModel/obsidianVaultPaths when calling createAgent from
handleCreateAgent, so forward the user's choices through the payload: ensure
handleCreateAgent passes model, _provider, _localModel and _obsidianVaultPaths
into the object sent to createAgent (and update createAgent's parameter shape if
needed to accept provider/localModel/obsidianVaultPaths), and propagate these
fields where agents are constructed so created agents respect the selected
provider, model and vault settings (check occurrences around handleCreateAgent
and createAgent to update all three call sites mentioned).

src/components/NewChatModal/index.tsx (1)

339-397: ⚠️ Potential issue | 🟠 Major

handleSubmit closes over a stale isOrchestrator value.

isOrchestrator is used in both the update payload and the onSubmit(...) call, but it is missing from this callback's dependency list. Toggling orchestrator off does not change any listed dependency, so Save/Start can still submit orchestratorMode: true.

💡 Suggested fix

-  }, [projectPath, prompt, selectedSkills, useWorktree, branchName, showSecondaryProject, selectedSecondaryProject, customSecondaryPath, model, permissionMode, effort, provider, localModel, selectedObsidianVaults, onSubmit, isEditMode, editAgent, onUpdate, onClose]);
+  }, [projectPath, prompt, selectedSkills, useWorktree, branchName, showSecondaryProject, selectedSecondaryProject, customSecondaryPath, model, permissionMode, effort, provider, localModel, selectedObsidianVaults, onSubmit, isEditMode, editAgent, onUpdate, onClose, isOrchestrator]);

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/NewChatModal/index.tsx` around lines 339 - 397, The
handleSubmit callback is closing over a stale isOrchestrator value (used in
onUpdate payload and the onSubmit(...) call); update the useCallback dependency
list for handleSubmit to include isOrchestrator so the latest value is captured,
ensuring orchestratorMode is correct when calling onUpdate and onSubmit
(references: handleSubmit, isOrchestrator, onSubmit, onUpdate).

electron/providers/gemini-provider.ts (1)

350-375: ⚠️ Potential issue | 🔴 Critical

Escape promptWithSkills before embedding it in the shell script.

Unlike buildScheduledCommand(), this path interpolates raw prompt text inside single quotes. A prompt or skill name containing ' will break the generated script, and crafted input can execute extra shell commands when a scheduled task runs.

💡 Suggested fix

   const promptWithSkills = (params.skills && params.skills.length > 0)
     ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
     : params.prompt;
+  const escapedPrompt = promptWithSkills.replace(/'/g, `'"'"'`);

   return `#!/bin/bash
@@
- "${params.binaryPath}" --output-format stream-json --debug --include-directories "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+ "${params.binaryPath}" --output-format stream-json --debug --include-directories "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
 echo "=== Task completed at $(date) ===" >> "${params.logPath}"
 `;

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/gemini-provider.ts` around lines 350 - 375, The generated
shell script embeds promptWithSkills directly inside single quotes, allowing
single quotes in the prompt to break the script and enable shell injection; fix
by producing an escapedPrompt (escape single quotes using the standard
shell-quote technique: replace ' with '\'' or use a safe sh-quoting helper)
before building the template and interpolate escapedPrompt instead of raw
promptWithSkills in the returned string in gemini-provider.ts (update the code
that constructs promptWithSkills and the template line that currently uses
"'${promptWithSkills}'" to use the escaped value).

electron/providers/pi-provider.ts (1)

209-235: ⚠️ Potential issue | 🟠 Major

Don't add unsupported, unescaped skill prompts to Pi scheduled jobs.

This path now prepends /<skill-name> guidance even though PiProvider.supportsSkills() still reports false. It also injects the combined prompt into a single-quoted shell argument without escaping, so apostrophes in the prompt or skill names will break the script.

Suggested fix

-    const promptWithSkills = (params.skills && params.skills.length > 0)
+    const promptWithSkills = (this.supportsSkills() && params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPromptWithSkills = promptWithSkills.replace(/'/g, "'\\''");
@@
-"${params.binaryPath}" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" -p '${escapedPromptWithSkills}' >> "${params.logPath}" 2>&1
 echo "=== Task completed at $(date) ===" >> "${params.logPath}"
 `;
   }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/pi-provider.ts` around lines 209 - 235, The generated
scheduled-job script wrongly injects an unescaped skill hint even when
PiProvider.supportsSkills() is false and places the combined prompt into a
single-quoted shell argument (breaking on apostrophes); fix by only prepending
the skills hint when PiProvider.supportsSkills() returns true and by safely
escaping or quoting the prompt when passing it to the CLI (avoid single-quoted
literal with embedded apostrophes) — update the code that builds
promptWithSkills (referencing params and promptWithSkills) to check
PiProvider.supportsSkills() before adding the skills string and change the
invocation that uses "${params.binaryPath}" -p '${promptWithSkills}' to emit a
safely escaped argument (e.g., use a here-doc, printf %q equivalent, or properly
escape single quotes) so prompt and skill names with apostrophes do not break
the shell.

src/components/Settings/GeneralSection.tsx (1)

48-68: ⚠️ Potential issue | 🟠 Major

The provider availability list here never stays in sync with the current settings state.

It is populated once from appSettings.get() and then cached, while the dropdown only disables entries whose value is strictly false. That means new providers are briefly selectable before detection finishes, and API-key changes in the same settings session will not re-enable/disable options.

Also applies to: 344-354

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/Settings/GeneralSection.tsx` around lines 48 - 68, The
installedProviders state is only populated once from appSettings.get() on mount,
causing transient selectable providers and no updates when settings change; fix
by creating a reusable updateInstalledProviders function that calls
Promise.all([window.electronAPI?.cliPaths?.detect(),
window.electronAPI?.appSettings?.get()]) and calls setInstalledProviders(...)
with explicit booleans (defaulting to false) and then (1) call that function on
mount, (2) initialize installedProviders to all false to avoid transient enabled
UI, and (3) subscribe to settings changes (or add a useEffect that depends on
your settings state / use window.electronAPI?.appSettings?.onChange if
available) to re-run updateInstalledProviders whenever app settings change so
the dropdown stays in sync (refer to setInstalledProviders and the Promise.all
detection block in the current code).

electron/providers/codex-provider.ts (1)

279-306: ⚠️ Potential issue | 🟠 Major

Escape promptWithSkills before embedding it in the shell script.

This now splices raw prompt text and skill names into a single-quoted Bash argument. Any apostrophe in either value will terminate the string and make the scheduled run fail.

Suggested fix

   buildScheduledScript(params: {
     binaryPath: string;
     binaryDir: string;
     projectPath: string;
@@
   }): string {
     const flags = params.autonomous ? '--full-auto' : '';
     const promptWithSkills = (params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPromptWithSkills = promptWithSkills.replace(/'/g, "'\\''");
@@
-"${params.binaryPath}" ${flags} --json exec '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" ${flags} --json exec '${escapedPromptWithSkills}' >> "${params.logPath}" 2>&1
 echo "=== Task completed at $(date) ===" >> "${params.logPath}"
 `;
   }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/codex-provider.ts` around lines 279 - 306, The embedded
promptWithSkills is injected into a single-quoted shell argument and can break
the script if it contains apostrophes; update the code that builds
promptWithSkills (the variable derived from params.prompt and params.skills) to
escape single quotes before embedding (e.g. replace every ' with '\'' or
implement a small shellEscape helper used where promptWithSkills is
interpolated), then use the escaped value in the exec line
("${params.binaryPath}" ${flags} --json exec '<escapedPromptWithSkills>') so the
generated bash stays syntactically valid.

electron/providers/opencode-provider.ts (1)

210-237: ⚠️ Potential issue | 🔴 Critical

Shell-escape promptWithSkills before calling opencode run.

The new skills prefix is concatenated into a single-quoted shell argument without escaping the skill names first. If any selected skill contains ', the generated script becomes invalid and can escape the quoted prompt at runtime. Build the full prompt from raw text, then escape it once before writing the command line.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/opencode-provider.ts` around lines 210 - 237, The
generated script inserts promptWithSkills directly into a single-quoted shell
argument which breaks if a skill contains a single quote; fix by building the
full prompt (promptWithSkills) and then shell-escaping it before writing the
command line — implement an escaping helper (e.g., escapeShellArg or inline
replace to turn each ' into '\'' and wrap the result in single quotes) and use
that escaped value in the command invocation where "${params.binaryPath}" run
'...'; update the template to use escapedPrompt instead of raw promptWithSkills
so the produced script is safe for any skill text.

electron/providers/claude-provider.ts (1)

403-431: ⚠️ Potential issue | 🔴 Critical

Escape the skills directive before embedding it in the bash script.

params.prompt is already shell-quoted upstream, but params.skills.join(', ') is interpolated raw into the same single-quoted -p argument here. A skill name containing ' breaks the script and can escape the quoted prompt when the automation runs. Compose the full prompt from raw text and escape it once at this boundary.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/claude-provider.ts` around lines 403 - 431, The embedded
skills string can break the single-quoted -p argument if a skill contains a
single quote; update the prompt composition so you build the full raw prompt
(use params.prompt and params.skills to form fullPrompt instead of
promptWithSkills) and then shell-escape that fullPrompt before inserting it into
the bash script. Add or use a helper (e.g., shellEscape or
escapeForSingleQuotes) and apply it to fullPrompt prior to the -p substitution
(the location currently using -p '${promptWithSkills}'), ensuring single quotes
inside skills are transformed into the safe sequence for single-quoted shell
literals.

🟠 Major comments (14)

src/hooks/useClaude.ts-29-29 (1)

29-29: ⚠️ Potential issue | 🟠 Major

providerTotals can stay stale due to missing change detection in setData.

You added the new field to TokenStats, but the memo-like comparison in fetchData doesn’t consider tokenStats, so provider usage updates can be dropped when project/session counts are unchanged.

Proposed fix

             const rlChanged = JSON.stringify(prev.rateLimits) !== JSON.stringify(newData.rateLimits);
             if (rlChanged) return newData;
+            const tokenStatsChanged =
+              JSON.stringify(prev.tokenStats) !== JSON.stringify(newData.tokenStats);
+            if (tokenStatsChanged) return newData;
             // No significant changes
             return prev;

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/hooks/useClaude.ts` at line 29, The memo-like comparison in fetchData is
ignoring changes in TokenStats.providerTotals so updates get dropped; update the
comparison logic used before calling setData (in fetchData) to include
tokenStats (or specifically tokenStats.providerTotals) in the equality check or
perform a deep-compare of providerTotals, and then call setData when
providerTotals differ; alternatively remove the memo gate for tokenStats and
always merge/update tokenStats into the state in setData (referencing
TokenStats, providerTotals, fetchData, and setData) so provider usage changes
are never silently discarded.

src/components/KanbanBoard/components/NewTaskModal.tsx-86-88 (1)

86-88: ⚠️ Potential issue | 🟠 Major

Worktree filter is not cross-platform (/ only).

This misses Windows-style paths (\...\.worktrees\...), so duplicate project entries can still leak through there.

Proposed fix

-        const projectList = rawProjects.filter((p: Project) => !p.path.includes('/.worktrees/'));
+        const worktreePattern = /[\\/]\.worktrees[\\/]/;
+        const projectList = rawProjects.filter((p: Project) => !worktreePattern.test(p.path));

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/KanbanBoard/components/NewTaskModal.tsx` around lines 86 - 88,
The current filter in NewTaskModal.tsx uses p.path.includes('/.worktrees/')
which only matches POSIX separators and misses Windows backslashes; update the
filter for rawProjects -> projectList to detect ".worktrees" cross-platform by
normalizing the path or using a robust check (e.g. import Node's path and call
path.normalize(p.path) and then check for `${path.sep}.worktrees${path.sep}` or
split by path.sep and test for a ".worktrees" segment, or use a regex like
/[\/\\]\.worktrees[\/\\]/) so duplicate entries are reliably excluded; apply
this change to the projectList construction that filters Project objects.

src/app/recurring-tasks/hooks/useTaskForm.ts-21-22 (1)

21-22: 🛠️ Refactor suggestion | 🟠 Major

Make defaultProvider part of the hook contract.

useTaskForm is still the source of truth for what gets submitted, but it initializes and resets provider to 'claude', and Line 83 always serializes that state. That leaves the page-level defaultProvider working only if the modal imperatively syncs the hook on every open/reset. Seed and reset the hook from the configured default instead.

♻️ Suggested refactor

-const INITIAL_FORM = {
+const createInitialForm = (defaultProvider: string) => ({
   agentId: '',
   projectPath: '',
   title: '',
   prompt: '',
   schedulePreset: 'daily',
   customCron: '',
   time: '09:00',
   days: ['1', '2', '3', '4', '5'],
   intervalDays: 2,
   selectedDays: ['1'] as string[],
   autonomous: true,
   useWorktree: false,
   notifyTelegram: false,
   notifySlack: false,
-  provider: 'claude',
+  provider: defaultProvider,
   skills: [] as string[],
-};
+});

 export function useTaskForm(
   agents: Agent[],
   loadTasks: () => Promise<void>,
   showToast: (msg: string, type: 'success' | 'error' | 'info') => void,
+  defaultProvider = 'claude',
 ) {
   const [showCreateForm, setShowCreateForm] = useState(false);
-  const [formData, setFormData] = useState(INITIAL_FORM);
+  const [formData, setFormData] = useState(() => createInitialForm(defaultProvider));
@@
-        setFormData(INITIAL_FORM);
+        setFormData(createInitialForm(defaultProvider));

-  const taskForm = useTaskForm(scheduled.agents, scheduled.loadTasks, showToast);
+  const taskForm = useTaskForm(scheduled.agents, scheduled.loadTasks, showToast, defaultProvider);

Also applies to: 30-31, 77-84, 93-96

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/app/recurring-tasks/hooks/useTaskForm.ts` around lines 21 - 22, The hook
useTaskForm currently hardcodes provider to 'claude' and always serializes that
value; change it so defaultProvider is accepted and used as part of the hook's
contract: add defaultProvider to the hook parameters signature, initialize
provider state from that default (instead of the fixed 'claude'), ensure
reset/initialize logic in the hook (and any serialize/submit function referenced
in useTaskForm) resets provider back to the passed defaultProvider, and replace
any hardcoded 'claude' occurrences (including places around provider
initialization, reset, and serialization) with the hook's provider state so the
page-level defaultProvider drives the form value.

src/hooks/useElectron.ts-295-297 (1)

295-297: ⚠️ Potential issue | 🟠 Major

Make the worktree filter path-separator agnostic.

p.path.includes('/.worktrees/') only catches POSIX-style paths. On Windows, worktree paths are typically ...\ .worktrees\..., so those entries will still leak into the list and the duplicate-key issue this PR is fixing will persist there.

💡 Suggested fix

-      // Filter out worktree paths to avoid duplicate React keys
-      setProjects(list.filter((p: { path: string }) => !p.path.includes('/.worktrees/')));
+      // Filter out worktree paths to avoid duplicate React keys
+      setProjects(list.filter((p: { path: string }) => !/[\\/]\.worktrees[\\/]/.test(p.path)));

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/hooks/useElectron.ts` around lines 295 - 297, The filter currently only
detects POSIX separators (p.path.includes('/.worktrees/')), so Windows paths
with backslashes slip through; update the predicate used in setProjects to
detect worktree segments with either separator (e.g. test p.path against a
regexp like /[\/\\]\.worktrees[\/\\]/ or normalize separators before checking)
so entries containing ".worktrees" with either '/' or '\' are excluded; change
the filtering logic near window.electronAPI!.fs.listProjects() and the
setProjects call to use the new check on p.path.

src/app/automations/page.tsx-242-243 (1)

242-243: ⚠️ Potential issue | 🟠 Major

Block unavailable providers in the automation form.

This form exposes every registry entry and forwards the chosen/default provider straight into the create payload without checking whether its CLI or API key is actually available. That lets users save automations that only fail later when the job starts.

Also applies to: 1018-1028

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/app/automations/page.tsx` around lines 242 - 243, The form currently
passes formData.agentProvider (or defaultProvider) and agentSkills into the
create payload without validating availability, which allows saving automations
that will later fail; update the form submit/validation logic (where
agentProvider/defaultProvider and agentSkills are used) to verify the chosen
provider has required credentials (CLI or API key) available in the registry
before allowing selection or submission, block/disable unavailable providers in
the provider selection UI, and return a validation error if an unavailable
provider is chosen so the create payload never includes a provider lacking
credentials.

electron/services/slack-bot.ts-460-481 (1)

460-481: ⚠️ Potential issue | 🟠 Major

Use the provider command builders here instead of a Claude-shaped command template.

These branches resolve a provider-specific binary, but they still append Claude-only flags and argument ordering. The Codex and Pi providers in this PR use different command shapes, so Slack-triggered launches will break as soon as the agent is not Claude.

Also applies to: 595-621

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/services/slack-bot.ts` around lines 460 - 481, Replace the manual
Claude-shaped command string assembly around
slackAgentProvider/resolveBinaryPath with the provider-specific command builder
on slackAgentProvider (e.g., a buildCommand/buildLaunchCommand method); gather
the same inputs you currently compute (binaryPath, skipPermissions flag derived
from agent.permissionMode or agent.skipPermissions, mcpConfigPath when
slackAgentProvider.getMcpConfigStrategy() === 'flag' and file exists,
agent.secondaryProjectPath and the default homedir /.dorothy add-dir, disallowed
tools when isSuperAgent(agent) || agent.orchestratorMode, and the session prompt
which should include agent.skills when present and not isSuperAgent); pass these
as structured options into the provider builder so quoting/flag ordering is
handled by the provider implementation instead of concatenating Claude-only
flags in this file.

src/app/recurring-tasks/components/CreateTaskModal.tsx-35-36 (1)

35-36: ⚠️ Potential issue | 🟠 Major

This modal bypasses the new provider-availability guardrails.

It renders every PROVIDER_REGISTRY entry with no disabled state, so recurring tasks can still be created for providers whose CLI or API key is missing. Pass the same availability map used by StepModel and disable or filter unsupported options here.

Also applies to: 166-177

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/app/recurring-tasks/components/CreateTaskModal.tsx` around lines 35 - 36,
The modal currently renders every PROVIDER_REGISTRY entry with no disabled
state; update CreateTaskModal (and the rendering block around the later lines
noted) to accept and use the same provider availability map that StepModel uses
(e.g., pass in the availability map prop from the parent or reuse the existing
availability prop) and either filter out unsupported providers or render them
disabled based on that map; specifically, when iterating PROVIDER_REGISTRY to
build options in CreateTaskModal, consult the availability lookup (the same
shape StepModel uses) and set disabled=true or omit the option for providers
marked unavailable to prevent creating recurring tasks for providers missing
CLI/API keys.

src/components/Settings/constants.ts-77-91 (1)

77-91: ⚠️ Potential issue | 🟠 Major

Add piEnabled to the shared settings defaults.

CLIPathsSection now reads and saves this flag, but DEFAULT_APP_SETTINGS never initializes it. Fresh settings objects therefore start with an implicit undefined state, and the UI has to cast around the missing field instead of reading a typed default. Mirror the field in the shared settings type as well.

Suggested fix

   opencodeEnabled: false,
+  piEnabled: false,
   opencodeDefaultModel: '',

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/Settings/constants.ts` around lines 77 - 91,
DEFAULT_APP_SETTINGS is missing the piEnabled boolean default which causes fresh
settings to have undefined for that flag; update DEFAULT_APP_SETTINGS to include
piEnabled: false and ensure the shared settings type (e.g., the
AppSettings/shared settings interface used by CLIPathsSection) includes
piEnabled: boolean so the initializer and type definitions match; locate
DEFAULT_APP_SETTINGS in src/components/Settings/constants.ts and add the
piEnabled entry alongside other provider flags and update the shared settings
type declaration to include piEnabled with type boolean.

src/components/Settings/CLIPathsSection.tsx-198-207 (1)

198-207: ⚠️ Potential issue | 🟠 Major

Test the path the user is editing, not a different shell command.

The OpenCode button always runs opencode --version, and the Pi button reads appSettings.cliPaths?.pi instead of localPaths.pi while interpolating that path directly into a shell string. Users cannot validate unsaved edits, and paths with spaces or quotes will fail unexpectedly. Use the current input value and invoke the binary with args instead of shell concatenation.

Also applies to: 242-247

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/Settings/CLIPathsSection.tsx` around lines 198 - 207, The test
buttons are running a hardcoded shell invocation and/or reading saved
appSettings instead of the current input state and are building shell strings
that break on spaces/quotes; update the handlers (the OpenCode handler which
uses setTestingOpencode and setOpencodeResult and the Pi handler that reads
appSettings.cliPaths?.pi) to read the current input state (localPaths.opencode
and localPaths.pi or the component state that backs the inputs) and call
window.electronAPI.shell.exec with a command+args payload (e.g. { command:
localPaths.opencode, args: ['--version'] }) rather than interpolating into one
shell string; also keep the same success/error handling but ensure you pass the
raw path value (fallback-guard it) and avoid using appSettings.cliPaths?.pi so
unsaved edits can be tested.

electron/services/api-routes/agent-routes.ts-234-244 (1)

234-244: ⚠️ Potential issue | 🟠 Major

Ignore exit callbacks from superseded PTYs.

/start now kills the previous PTY before spawning a new one, and /message can also replace a PTY after reconnect. Both delayed onExit handlers still mutate agent.status unconditionally, so the old PTY can flip the new run to error/completed ~1.5s later.

Suggested fix

     ptyProcess.onExit(({ exitCode }) => {
       // Delay status change to let hooks (on-stop.sh, task-completed.sh) finish
       // capturing output before wait_for_agent resolves.
       setTimeout(() => {
+        if (agent.ptyId !== ptyId) return;
         if (agent.status === 'running') {
           agent.status = exitCode === 0 ? 'completed' : 'error';
         } else if (agent.status === 'waiting') {
           // PTY exited while agent was waiting for input — the claude process
           // crashed. Mark as error so /wait is unblocked and the orchestrator
           // can retry rather than hanging until timeout.
           agent.status = 'error';
         }
         if (exitCode !== 0) {
           agent.error = `Process exited with code ${exitCode}`;
         }
         agent.lastActivity = new Date().toISOString();
+        agent.ptyId = undefined;
+        agent.ptyCwd = undefined;
         ptyProcesses.delete(ptyId);
         saveAgents();
         ctx.agentStatusEmitter.emit(`status:${agent.id}`);
       }, 1500);
     });

Apply the same guard/cleanup to the reconnectPtyId exit handler below as well.

Also applies to: 297-301, 432-445

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/services/api-routes/agent-routes.ts` around lines 234 - 244, Delayed
onExit handlers for superseded PTYs (the block that checks agent.ptyId and
ptyProcesses) can still run and mutate agent.status for a newer PTY; update the
exit handlers that reference reconnectPtyId (and any other PTY onExit callbacks
around lines with reconnectPtyId and the 432-445 block) to early-return unless
the PTY id that triggered the callback matches the current agent.ptyId (or
reconnectPtyId as intended), and remove/cleanup the listener when
killing/replacing a PTY; in practice, add a guard at the start of the onExit
callbacks (compare the captured ptyId to agent.ptyId/reconnectPtyId) and only
update agent.status or perform cleanup when they match, and ensure you delete
the old pty from ptyProcesses when you kill it so delayed handlers can detect it
was superseded.

electron/handlers/scheduler-handlers.ts-881-891 (1)

881-891: ⚠️ Potential issue | 🟠 Major

Don't let task edits strip the saved skills list.

createTask() now stores skills, but the update/regenerate path never reads metadata[taskId].skills back into buildScheduledScript() or the recreated launchd/cron job. Any later edit will silently drop the skill preamble from future runs.

Suggested fix

       // Resolve provider from metadata
       const meta = loadSchedulerMetadata();
       const taskProvider: AgentProvider = meta[taskId]?.provider || resolveDefaultProvider(deps);
+      const taskSkills = meta[taskId]?.skills;

       // Always regenerate the shell script using provider
       const claudePath = await getCLIPath(taskProvider);
@@
       const scriptContent = cliProvider.buildScheduledScript({
         binaryPath: claudePath,
         binaryDir: claudeDir,
         projectPath,
         prompt: escapedPrompt,
         autonomous,
         mcpConfigPath,
         logPath,
         homeDir,
+        skills: taskSkills,
       });
@@
-          await createLaunchdJob(taskId, schedule, projectPath, prompt, autonomous, taskProvider);
+          await createLaunchdJob(taskId, schedule, projectPath, prompt, autonomous, taskProvider, taskSkills);
@@
-          await createCronJob(taskId, schedule, projectPath, prompt, autonomous, taskProvider);
+          await createCronJob(taskId, schedule, projectPath, prompt, autonomous, taskProvider, taskSkills);

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/handlers/scheduler-handlers.ts` around lines 881 - 891, The
update/regenerate path is not preserving the saved skills because the code that
recreates jobs doesn't read metadata[taskId].skills; update the regenerate flow
to load the scheduler metadata (from where saveSchedulerMetadata writes), read
metadata[taskId].skills (falling back to config.skills or empty array), and pass
that skills array into buildScheduledScript and the job-creation calls
(createLaunchdJob and createCronJob) so the recreated job and script include the
original skill preamble.

electron/handlers/scheduler-handlers.ts-869-872 (1)

869-872: ⚠️ Potential issue | 🟠 Major

Keep legacy scheduled tasks on Claude when metadata is missing.

Now that new tasks persist provider, the update flow below still falls back to resolveDefaultProvider(deps) for older tasks with no stored provider. Editing a pre-provider task after changing Settings can silently switch its execution/auth provider.

Suggested fix

-      const taskProvider: AgentProvider = meta[taskId]?.provider || resolveDefaultProvider(deps);
+      const taskProvider: AgentProvider = meta[taskId]?.provider || 'claude';

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/handlers/scheduler-handlers.ts` around lines 869 - 872, The current
fallback uses resolveDefaultProvider(deps) which causes legacy tasks lacking a
stored provider to unintentionally pick up the current app default; update the
provider resolution in the task creation/update path so it checks
config.provider, then deps.agents.get(config.agentId)?.provider, then any stored
provider on the task (e.g., config.metadata?.provider or the legacy field that
held provider), and only if none of those exist fall back to the literal
'claude' (keep references to AgentProvider, taskProvider, config.provider,
deps.agents.get, and resolveDefaultProvider when locating the code to change).

electron/services/api-routes/agent-routes.ts-417-422 (1)

417-422: ⚠️ Potential issue | 🟠 Major

Reset stale transcript/error state before auto-respawn.

This branch starts a fresh one-shot session, but it keeps agent.output, lastCleanOutput, and error from the previous crashed run. A successful retry can therefore still surface the old error and interleave transcripts.

Suggested fix

       ptyProcesses.set(reconnectPtyId, reconnectPty);
       agent.ptyId = reconnectPtyId;
       agent.ptyCwd = rawWorkingDir;
       agent.status = 'running';
       agent.currentTask = message.slice(0, 100);
+      agent.output = [];
+      agent.lastCleanOutput = undefined;
+      agent.error = undefined;
       agent.lastActivity = new Date().toISOString();

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/services/api-routes/agent-routes.ts` around lines 417 - 422, Reset
the agent's previous run state before starting the one-shot respawn: clear
agent.output and agent.lastCleanOutput (set to empty strings) and clear
agent.error (set to null/undefined) in the same branch where you set
agent.ptyId, agent.ptyCwd, agent.status, and agent.currentTask so stale
transcripts or errors from the previous crashed run cannot surface or interleave
with the new session.

src/types/electron.d.ts-352-365 (1)

352-365: ⚠️ Potential issue | 🟠 Major

Mirror the new provider fields in appSettings.save().

appSettings.get() now exposes opencode* / openRouter* / deepSeek* / mimo* / moonshot* / qwen* / zhipu*, but the save() signature below still rejects them. The new Settings UI will otherwise need type escapes to persist its own fields.

Suggested fix

     save: (settings: {
       notificationsEnabled?: boolean;
       notifyOnWaiting?: boolean;
       notifyOnComplete?: boolean;
       notifyOnStop?: boolean;
       notifyOnError?: boolean;
       telegramEnabled?: boolean;
       telegramBotToken?: string;
       telegramChatId?: string;
       telegramAuthToken?: string;
       telegramAuthorizedChatIds?: string[];
       telegramRequireMention?: boolean;
       slackEnabled?: boolean;
       slackBotToken?: string;
       slackAppToken?: string;
       slackSigningSecret?: string;
       slackChannelId?: string;
       jiraEnabled?: boolean;
       jiraDomain?: string;
       jiraEmail?: string;
       jiraApiToken?: string;
       socialDataEnabled?: boolean;
       socialDataApiKey?: string;
       tasmaniaEnabled?: boolean;
       tasmaniaServerPath?: string;
       defaultProvider?: string;
+      opencodeEnabled?: boolean;
+      opencodeDefaultModel?: string;
+      openRouterEnabled?: boolean;
+      openRouterApiKey?: string;
+      deepSeekEnabled?: boolean;
+      deepSeekApiKey?: string;
+      mimoEnabled?: boolean;
+      mimoApiKey?: string;
+      moonshotEnabled?: boolean;
+      moonshotApiKey?: string;
+      qwenEnabled?: boolean;
+      qwenApiKey?: string;
+      zhipuEnabled?: boolean;
+      zhipuApiKey?: string;
       notificationSounds?: {
         waiting?: string;
         complete?: string;
         stop?: string;
         error?: string;
       };

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/types/electron.d.ts` around lines 352 - 365, The appSettings.save()
signature does not include the new provider fields added to the Settings type
(opencodeEnabled, opencodeDefaultModel, openRouterEnabled, openRouterApiKey,
deepSeekEnabled, deepSeekApiKey, mimoEnabled, mimoApiKey, moonshotEnabled,
moonshotApiKey, qwenEnabled, qwenApiKey, zhipuEnabled, zhipuApiKey), causing
saves from the new Settings UI to be rejected; update the parameter type (or
SaveSettings/interface used by the appSettings.save method) to include these
optional properties with the same types as declared in electron.d.ts so
appSettings.save accepts and persists those fields, then ensure any runtime
serialization logic inside the save implementation handles/passes through these
new fields unchanged.

🟡 Minor comments (12)

mcp-orchestrator/src/tools/agents.ts-181-181 (1)

181-181: ⚠️ Potential issue | 🟡 Minor

Extract the duplicated model description to a shared constant.

The identical model description at lines 181 and 472 should be stored as a constant to prevent drift when Claude models change. However, verify that the aliases 'opusplan' and 'sonnet[1m]' are current—the April 2026 API docs confirm 'sonnet', 'opus', 'haiku', and the listed full IDs (claude-sonnet-4-6, claude-opus-4-6, claude-haiku-4-5-20251001) are valid, but 'opusplan' and the '[1m]' notation do not appear in official Anthropic documentation. Also note that other files in this codebase define model enums as 'sonnet' | 'opus' | 'haiku' without 'opusplan', which suggests possible inconsistency across the tool definitions.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@mcp-orchestrator/src/tools/agents.ts` at line 181, The model description
string used in the zod schema (the value passed to model:
z.string().optional().describe(...)) is duplicated (at the current occurrences
around the model field in agents.ts and again later) and should be extracted to
a single exported constant (e.g., CLAUDE_MODEL_DESCRIPTION) so both schema calls
reference the same value; update the z.string().optional().describe(...)
invocations in the functions/objects that reference that field to use the
constant. While extracting, reconcile the alias list to match the codebase
canonical enum ('sonnet' | 'opus' | 'haiku') by removing or flagging nonstandard
entries ('opusplan' and 'sonnet[1m]') unless you confirm they’re intentionally
supported, and ensure the constant’s text matches the authoritative API docs so
future changes only need one edit.

src/app/recurring-tasks/components/TaskCard.tsx-90-100 (1)

90-100: ⚠️ Potential issue | 🟡 Minor

Keep a visible fallback when provider lookup misses.

if (!def) return null hides the provider entirely. With new providers and rename churn in this PR, that makes registry drift much harder to spot in the task list. Falling back to task.provider as plain text keeps the UI debuggable.

💡 Suggested fix

             {task.provider && task.provider !== 'claude' && (() => {
               const def = getProviderDef(task.provider);
-              if (!def) return null;
+              if (!def) {
+                return <div className="flex items-center gap-1">{task.provider}</div>;
+              }
               return (

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/app/recurring-tasks/components/TaskCard.tsx` around lines 90 - 100, The
provider lookup in TaskCard (getProviderDef(task.provider)) returns null and
currently causes the UI to hide the provider (`if (!def) return null`); change
this to render a visible fallback that displays task.provider as plain text
(e.g., a small label or span) so missing/renamed providers remain visible; keep
the existing rendering logic for when def exists (icon branches: def.icon.type
=== 'image' | 'svg-gemini' | 'cpu' | 'text') and only use the fallback when def
is falsy, referencing getProviderDef, task.provider, and the local def variable
to locate the code to modify.

src/components/AgentList/AgentCard.tsx-104-106 (1)

104-106: ⚠️ Potential issue | 🟡 Minor

Don't badge local agents as Claude.

When agent.provider === 'local', this forces the header badge to claude. The card then shows a Claude icon for a local-model agent, which is misleading. Pass local through if the registry supports it, or omit the provider badge for local agents.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/AgentList/AgentCard.tsx` around lines 104 - 106, The
ProviderIcon call in AgentCard currently maps agent.provider === 'local' to
'claude', causing local agents to show a Claude badge; update the ProviderIcon
invocation to pass agent.provider through unchanged (i.e., pass 'local' when
agent.provider === 'local') or pass undefined/null to omit the badge for local
agents instead of forcing 'claude'—change the ProviderIcon prop usage in
AgentCard (the ProviderIcon component invocation) so it receives the real
agent.provider value or nothing for local agents.

src/components/ProviderBadge.tsx-19-28 (1)

19-28: ⚠️ Potential issue | 🟡 Minor

ProviderBadge is already out of sync with the multi-provider registry.

This local map still only covers claude, codex, gemini, opencode, and pi, so any openrouter / deepseek / moonshot / mimo / qwen / zai badge routed through this component will still hit if (!config) return null at Line 37. Reusing the shared provider registry here would keep badge coverage in sync with the rest of the UI.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/ProviderBadge.tsx` around lines 19 - 28, The component defines
a local PROVIDER_CONFIG map that is missing many providers and causes
ProviderBadge to return null for providers like
openrouter/deepseek/moonshot/mimo/qwen/zai; remove or stop using the local
PROVIDER_CONFIG inside ProviderBadge and instead import and use the shared
multi-provider registry (the central provider registry used by the app) to look
up provider metadata inside ProviderBadge (use the same lookup call the rest of
the UI uses), ensuring the lookup covers icon and label types expected by
ProviderBadge and fall back gracefully if a provider entry lacks an icon; update
references to PROVIDER_CONFIG in ProviderBadge to use the shared registry lookup
so badge coverage stays in sync.

src/components/AgentList/AgentCard.tsx-39-44 (1)

39-44: ⚠️ Potential issue | 🟡 Minor

This fallback drops the actual SVG assets.

For any provider icon that isn't image, svg-gemini, cpu, or text, the card now renders label.slice(0, 2) instead of the real icon. That means the new SVG provider logos added in this PR won't actually appear on agent cards.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/AgentList/AgentCard.tsx` around lines 39 - 44, The current
fallback unconditionally renders label.slice(0, 2) which drops any other SVG
assets; update the AgentCard icon rendering to first attempt to render the
provider's actual icon/SVG (e.g., check for a provider.icon / provider.svg /
iconType not equal to 'image'|'svg-gemini'|'cpu'|'text' and render that asset
using the existing SVG/image component or <img> wrapper), and only use the
compact badge (the label.slice(0, 2> span with className="shrink-0 inline-flex
font-bold text-[10px] text-text-muted") when no icon asset is available; modify
the return in AgentCard to conditionally render the real SVG/image asset before
falling back to the two-letter label.

src/components/Settings/AIProvidersSection.tsx-213-227 (1)

213-227: ⚠️ Potential issue | 🟡 Minor

The routing explainer contradicts the cards above it.

This footer says every provider routes through OpenRouter and that direct routing is future work, but the individual cards describe several providers as direct endpoints today. Users will get the wrong setup/debugging guidance from this.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/Settings/AIProvidersSection.tsx` around lines 213 - 227, The
"How routing works" footer text is incorrect and contradicts the provider cards;
update the copy in the <div> labeled "How routing works" (the JSX block
containing the ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY <code> elements) so it
accurately reflects current behavior: state which providers use direct API
endpoints today versus which route via OpenRouter, explain that
provider-specific keys are used when present and OpenRouter key is the fallback,
and remove the blanket statement that "All providers use Claude CLI... and
direct API routing is future work." Cross-check the provider cards above in
AIProvidersSection to ensure the footer matches those descriptions.

electron/providers/moonshot-provider.ts-67-73 (1)

67-73: ⚠️ Potential issue | 🟡 Minor

Model argument should be quoted for shell safety.

Same issue as zhipu-provider.ts — the --model argument is unquoted.

Proposed fix

-    if (params.model && params.model !== 'default') command += ` --model ${params.model}`;
+    if (params.model && params.model !== 'default') command += ` --model '${params.model}'`;

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/moonshot-provider.ts` around lines 67 - 73, The --model
argument in buildOneShotCommand is currently appended unquoted; update
buildOneShotCommand so when params.model && params.model !== 'default' you
escape any single quotes in params.model (like other fields use .replace(/'/g,
"'\\''")) and wrap it in single quotes before adding ` --model ...`, ensuring
params.binaryPath, params.prompt handling remains unchanged; locate the
buildOneShotCommand function and change the model concatenation to use the same
quoting/escaping pattern as params.binaryPath and params.prompt.

electron/providers/openrouter-provider.ts-124-137 (1)

124-137: ⚠️ Potential issue | 🟡 Minor

Model argument should be quoted for shell safety.

Proposed fix

-    if (params.model && params.model !== 'default') {
-      command += ` --model ${params.model}`;
-    }
+    if (params.model && params.model !== 'default') {
+      command += ` --model '${params.model}'`;
+    }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/openrouter-provider.ts` around lines 124 - 137, In
buildOneShotCommand the --model value is appended unquoted which can break
shells or allow injection; change the construction so that when params.model is
used you escape any single quotes in params.model (like
params.model.replace(/'/g, "'\\''")) and wrap it in single quotes when appending
(i.e. ` --model 'escapedModel'`), similar to how params.binaryPath and
params.prompt are escaped, ensuring params.model is safely quoted before
insertion.

electron/providers/deepseek-provider.ts-91-97 (1)

91-97: ⚠️ Potential issue | 🟡 Minor

Model argument should be quoted for shell safety.

Proposed fix

-    if (params.model && params.model !== 'default') command += ` --model ${params.model}`;
+    if (params.model && params.model !== 'default') command += ` --model '${params.model}'`;

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/deepseek-provider.ts` around lines 91 - 97, The model
value in buildOneShotCommand is interpolated unquoted into the shell
command—escape and quote it like binaryPath and prompt to avoid shell injection.
In buildOneShotCommand (OneShotCommandParams), when adding `--model
${params.model}` replace it with a safely escaped-and-quoted version (e.g. build
a quotedModel = `'${params.model.replace(/'/g, "'\\''")}'` and use `--model
${quotedModel}`) so the model argument is always shell-safe.

electron/providers/qwen-provider.ts-69-75 (1)

69-75: ⚠️ Potential issue | 🟡 Minor

Model argument should be quoted for shell safety.

Proposed fix

-    if (params.model && params.model !== 'default') command += ` --model ${params.model}`;
+    if (params.model && params.model !== 'default') command += ` --model '${params.model}'`;

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/qwen-provider.ts` around lines 69 - 75,
buildOneShotCommand builds a shell command but inserts params.model without
quoting, which is unsafe; update buildOneShotCommand so when adding `--model
${params.model}` it instead quotes and escapes the model value the same way as
binaryPath and prompt (e.g., replace any single quotes in params.model with
"'\\''" and wrap the result in single quotes) and only append the quoted
`--model 'escapedModel'` when params.model exists and is not 'default'.

electron/providers/mimo-provider.ts-67-73 (1)

67-73: ⚠️ Potential issue | 🟡 Minor

Model argument should be quoted for shell safety.

Proposed fix

-    if (params.model && params.model !== 'default') command += ` --model ${params.model}`;
+    if (params.model && params.model !== 'default') command += ` --model '${params.model}'`;

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/mimo-provider.ts` around lines 67 - 73, In
buildOneShotCommand, the --model value is added unquoted which is unsafe for
shell injection; change the branch that appends ` --model ${params.model}` to
quote and escape the model exactly like the binaryPath and prompt (e.g., use
params.model.replace(/'/g, "'\\''") and wrap in single quotes) so the command
becomes ` --model 'escapedModel'`; keep the existing check for params.model &&
params.model !== 'default' and otherwise leave behavior unchanged.

electron/providers/zhipu-provider.ts-69-75 (1)

69-75: ⚠️ Potential issue | 🟡 Minor

Model argument should be quoted for shell safety.

The --model argument is not quoted, unlike in buildInteractiveCommand (Line 43). While model IDs from getModels() appear safe, user-selected or custom models could contain characters that break the command.

Proposed fix

   buildOneShotCommand(params: OneShotCommandParams): string {
     let command = `'${params.binaryPath.replace(/'/g, "'\\''")}'`;
     command += ' -p';
-    if (params.model && params.model !== 'default') command += ` --model ${params.model}`;
+    if (params.model && params.model !== 'default') command += ` --model '${params.model}'`;
     command += ` '${params.prompt.replace(/'/g, "'\\''")}'`;
     return command;
   }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/zhipu-provider.ts` around lines 69 - 75, The
buildOneShotCommand function currently appends --model without quoting/escaping
the model value, which can break the shell command; update buildOneShotCommand
to quote and escape params.model the same way buildInteractiveCommand does
(escape single quotes and wrap the model value in single quotes when adding
`--model`), so the constructed command uses `--model 'escaped-model'` whenever
params.model is set and not 'default'.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

Shell injection risk: promptWithSkills is not escaped in the generated bash script.

Same unescaped prompt issue as other providers.

🐛 Proposed fix

   buildScheduledScript(params: {
     binaryPath: string; binaryDir: string; projectPath: string; prompt: string;
     autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string;
     skills?: string[];
   }): string {
     const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
     const promptWithSkills = (params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
     return `#!/bin/bash
 export HOME="${params.homeDir}"
 ...
 unset CLAUDECODE
-"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
 echo "=== Task completed at $(date) ===" >> "${params.logPath}"
 `;
   }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/deepseek-provider.ts` around lines 185 - 205, The
generated script in buildScheduledScript embeds promptWithSkills directly into a
single-quoted -p argument, allowing shell injection; instead write the prompt
safely to a temp file and pass it to the binary (or pass via a decoded base64
file) rather than inlining the raw string. Update buildScheduledScript to:
create a uniquely-named temp file (e.g., using mktemp) inside the returned
script, write a safely-escaped or base64-encoded promptWithSkills into that
file, pass the file contents to the binary (or use a --prompt-file option if
supported), and remove the temp file afterwards; reference buildScheduledScript
and the promptWithSkills variable when making the change.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

🧩 Analysis chain

🌐 Web query:

What is the official API endpoint URL for Xiaomi MiMo AI models?

💡 Result:

The official API endpoint URL for Xiaomi MiMo AI models is https://api.xiaomimimo.com/v1. This is the OpenAI-compatible base URL for accessing models like MiMo-V2-Pro, MiMo-V2-Omni, MiMo-V2-Flash, and MiMo-V2-TTS. The official platform is at https://platform.xiaomimimo.com/.

Citations:

• 1: https://platform.xiaomimimo.com/
• 2: http://mimo-v2.org/
• 3: [Feature]: Add support for Xiaomi MiMo API and MiMo-V2-Flash model CherryHQ/cherry-studio#11959
• 4: https://www.mimo-v2.com/docs
• 5: https://docs.litellm.ai/docs/providers/xiaomi_mimo
• 6: https://mimo-v2.org/

---

Correct the MiMo API base URL to the official endpoint.

The URL https://api.mimo.com/v1 is incorrect. The official API endpoint for Xiaomi's MiMo AI models is https://api.xiaomimimo.com/v1. Update the constant accordingly.

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/mimo-provider.ts` at line 14, Update the MIMO_BASE_URL
constant to use the official Xiaomi MiMo API endpoint: replace the current value
of MIMO_BASE_URL (const MIMO_BASE_URL = 'https://api.mimo.com/v1') with
'https://api.xiaomimimo.com/v1' so all requests from the mimo-provider.ts module
use the correct base URL.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

Shell injection risk: promptWithSkills is not escaped in the generated bash script.

Same unescaped prompt issue as in other providers.

🐛 Proposed fix

   buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
     const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
     const promptWithSkills = (params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
     return `#!/bin/bash
 export HOME="${params.homeDir}"
-"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
 ...
 `;
   }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
	const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
	const promptWithSkills = (params.skills && params.skills.length > 0)
	? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
	: params.prompt;
	return `#!/bin/bash
	export HOME="${params.homeDir}"
	if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true; fi
	if [ -f "${params.homeDir}/.zshrc" ]; then source "${params.homeDir}/.zshrc" 2>/dev/null || true; fi
	export PATH="${params.binaryDir}:$PATH"
	cd "${params.projectPath}"
	echo "=== Task started at $(date) ===" >> "${params.logPath}"
	unset CLAUDECODE
	"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
	echo "=== Task completed at $(date) ===" >> "${params.logPath}"
	`;
	}
	buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
	const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
	const promptWithSkills = (params.skills && params.skills.length > 0)
	? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
	: params.prompt;
	const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
	return `#!/bin/bash
	export HOME="${params.homeDir}"
	if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true; fi
	if [ -f "${params.homeDir}/.zshrc" ]; then source "${params.homeDir}/.zshrc" 2>/dev/null || true; fi
	export PATH="${params.binaryDir}:$PATH"
	cd "${params.projectPath}"
	echo "=== Task started at $(date) ===" >> "${params.logPath}"
	unset CLAUDECODE
	"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
	echo "=== Task completed at $(date) ===" >> "${params.logPath}"
	`;
	}

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/mimo-provider.ts` around lines 132 - 148, The
buildScheduledScript function embeds promptWithSkills directly into a
single-quoted shell argument, which allows shell injection; instead, write
promptWithSkills to a temporary file inside the generated script (e.g., create a
TMP prompt file under ${params.homeDir} or use mktemp), use a safe write (printf
'%s' to the temp file) and pass the prompt to the binary by reading the file
(e.g., -p "$(cat "$TMPFILE")"), then remove the temp file; update
buildScheduledScript to perform those steps so promptWithSkills is never
interpolated into the command line.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

Shell injection risk: promptWithSkills is not escaped in the generated bash script.

Same issue as in zhipu-provider.ts. The prompt is embedded without escaping single quotes, which can break the script or allow injection.

🐛 Proposed fix

   buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
     const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
     const promptWithSkills = (params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
     return `#!/bin/bash
 export HOME="${params.homeDir}"
 if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true; fi
 if [ -f "${params.homeDir}/.zshrc" ]; then source "${params.homeDir}/.zshrc" 2>/dev/null || true; fi
 export PATH="${params.binaryDir}:$PATH"
 cd "${params.projectPath}"
 echo "=== Task started at $(date) ===" >> "${params.logPath}"
 unset CLAUDECODE
-"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
 echo "=== Task completed at $(date) ===" >> "${params.logPath}"
 `;
   }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
	const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
	const promptWithSkills = (params.skills && params.skills.length > 0)
	? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
	: params.prompt;
	return `#!/bin/bash
	export HOME="${params.homeDir}"
	if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true; fi
	if [ -f "${params.homeDir}/.zshrc" ]; then source "${params.homeDir}/.zshrc" 2>/dev/null || true; fi
	export PATH="${params.binaryDir}:$PATH"
	cd "${params.projectPath}"
	echo "=== Task started at $(date) ===" >> "${params.logPath}"
	unset CLAUDECODE
	"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
	echo "=== Task completed at $(date) ===" >> "${params.logPath}"
	`;
	}
	buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
	const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
	const promptWithSkills = (params.skills && params.skills.length > 0)
	? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
	: params.prompt;
	const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
	return `#!/bin/bash
	export HOME="${params.homeDir}"
	if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true; fi
	if [ -f "${params.homeDir}/.zshrc" ]; then source "${params.homeDir}/.zshrc" 2>/dev/null || true; fi
	export PATH="${params.binaryDir}:$PATH"
	cd "${params.projectPath}"
	echo "=== Task started at $(date) ===" >> "${params.logPath}"
	unset CLAUDECODE
	"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
	echo "=== Task completed at $(date) ===" >> "${params.logPath}"
	`;
	}

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/moonshot-provider.ts` around lines 148 - 164, The
buildScheduledScript function embeds promptWithSkills directly into a
single-quoted bash argument, creating a shell injection/breakage risk; update
buildScheduledScript to safely quote/escape the prompt before embedding (e.g.,
transform single quotes in promptWithSkills to the safe bash sequence '"'"' or
switch to a quoted heredoc token such as <<'EOF' to pass the raw prompt) so that
any single quotes or shell metacharacters in params.prompt or params.skills
cannot terminate the surrounding quotes or inject commands; ensure the
escaped/quoted value is used where '-p '${promptWithSkills}'' appears.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

Shell injection risk: promptWithSkills is not escaped in the generated bash script.

Same issue as other providers — the prompt needs single-quote escaping before embedding in the bash script.

🐛 Proposed fix

   buildScheduledScript(params: {
     ...
   }): string {
     const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
     const promptWithSkills = (params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");

     return `#!/bin/bash
 ...
 unset CLAUDECODE
-"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
 ...
 `;
   }

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	buildScheduledScript(params: {
	binaryPath: string;
	binaryDir: string;
	projectPath: string;
	prompt: string;
	autonomous: boolean;
	mcpConfigPath: string;
	logPath: string;
	homeDir: string;
	skills?: string[];
	}): string {
	const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
	const promptWithSkills = (params.skills && params.skills.length > 0)
	? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
	: params.prompt;
	return `#!/bin/bash
	export HOME="${params.homeDir}"
	if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then
	source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true
	fi
	if [ -f "${params.homeDir}/.zshrc" ]; then
	source "${params.homeDir}/.zshrc" 2>/dev/null || true
	elif [ -f "${params.homeDir}/.bash_profile" ]; then
	source "${params.homeDir}/.bash_profile" 2>/dev/null || true
	fi
	export PATH="${params.binaryDir}:$PATH"
	cd "${params.projectPath}"
	echo "=== Task started at $(date) ===" >> "${params.logPath}"
	unset CLAUDECODE
	"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
	echo "=== Task completed at $(date) ===" >> "${params.logPath}"
	`;
	}
	buildScheduledScript(params: {
	binaryPath: string;
	binaryDir: string;
	projectPath: string;
	prompt: string;
	autonomous: boolean;
	mcpConfigPath: string;
	logPath: string;
	homeDir: string;
	skills?: string[];
	}): string {
	const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
	const promptWithSkills = (params.skills && params.skills.length > 0)
	? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
	: params.prompt;
	const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
	return `#!/bin/bash
	export HOME="${params.homeDir}"
	if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then
	source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true
	fi
	if [ -f "${params.homeDir}/.zshrc" ]; then
	source "${params.homeDir}/.zshrc" 2>/dev/null || true
	elif [ -f "${params.homeDir}/.bash_profile" ]; then
	source "${params.homeDir}/.bash_profile" 2>/dev/null || true
	fi
	export PATH="${params.binaryDir}:$PATH"
	cd "${params.projectPath}"
	echo "=== Task started at $(date) ===" >> "${params.logPath}"
	unset CLAUDECODE
	"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
	echo "=== Task completed at $(date) ===" >> "${params.logPath}"
	`;
	}

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/openrouter-provider.ts` around lines 261 - 298, The
generated bash script embeds promptWithSkills directly into a single-quoted
argument, creating a shell injection risk; update buildScheduledScript to escape
the prompt before embedding (e.g., transform single quotes in promptWithSkills
into a safe shell-escaped sequence and also handle backslashes/newlines as
needed) and use that escaped string in the returned script instead of
promptWithSkills so the produced -p '...' argument cannot break out of the
single quotes.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

Shell injection risk: promptWithSkills is not escaped in the generated bash script.

Same unescaped prompt issue as other providers.

🐛 Proposed fix

   buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
     const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
     const promptWithSkills = (params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
     return `#!/bin/bash
 ...
-"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
 ...
 `;
   }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/qwen-provider.ts` around lines 134 - 150, The generated
script embeds promptWithSkills directly in a single-quoted -p argument, which
allows shell injection; in buildScheduledScript, stop interpolating the raw
prompt into the command and instead either (A) write promptWithSkills to a
temporary file (via mktemp) and pass that file to the binary (e.g.,
--prompt-file or cat it into stdin) or (B) fully shell-escape single quotes in
promptWithSkills before embedding (replace every ' with '\''), or use a
single-quoted here-doc (<<'EOF') to safely supply the prompt; update the script
generation to create the temp file or here-doc, reference params.binaryPath and
params.logPath appropriately, and ensure the temp file is removed afterwards.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

Shell injection risk: promptWithSkills is not escaped in the generated bash script.

The promptWithSkills variable is embedded directly into the bash script using single quotes, but single quotes within the prompt are not escaped. If params.prompt or skill names contain ', the generated script will either break or potentially allow shell injection.

Compare with buildScheduledCommand (Line 65) which correctly escapes:

command += ` -p '${params.prompt.replace(/'/g, "'\\''")}'`;

🐛 Proposed fix: escape single quotes in promptWithSkills

   buildScheduledScript(params: { binaryPath: string; binaryDir: string; projectPath: string; prompt: string; autonomous: boolean; mcpConfigPath: string; logPath: string; homeDir: string; skills?: string[]; }): string {
     const flags = params.autonomous ? '--dangerously-skip-permissions' : '';
     const promptWithSkills = (params.skills && params.skills.length > 0)
       ? `[IMPORTANT: Use these skills for this session: ${params.skills.join(', ')}. Invoke them with /<skill-name> when relevant to the task.] ${params.prompt}`
       : params.prompt;
+    const escapedPrompt = promptWithSkills.replace(/'/g, "'\\''");
     return `#!/bin/bash
 export HOME="${params.homeDir}"
 if [ -s "${params.homeDir}/.nvm/nvm.sh" ]; then source "${params.homeDir}/.nvm/nvm.sh" 2>/dev/null || true; fi
 if [ -f "${params.homeDir}/.zshrc" ]; then source "${params.homeDir}/.zshrc" 2>/dev/null || true; fi
 export PATH="${params.binaryDir}:$PATH"
 cd "${params.projectPath}"
 echo "=== Task started at $(date) ===" >> "${params.logPath}"
 unset CLAUDECODE
-"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${promptWithSkills}' >> "${params.logPath}" 2>&1
+"${params.binaryPath}" ${flags} --output-format stream-json --verbose --mcp-config "${params.mcpConfigPath}" --add-dir "${params.homeDir}/.dorothy" -p '${escapedPrompt}' >> "${params.logPath}" 2>&1
 echo "=== Task completed at $(date) ===" >> "${params.logPath}"
 `;
   }

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@electron/providers/zhipu-provider.ts` around lines 138 - 154, The generated
bash script in buildScheduledScript inserts promptWithSkills directly into a
single-quoted -p argument which breaks or allows injection if the prompt or
skill names contain single quotes; fix this by escaping single quotes in
promptWithSkills (or when embedding it) the same way buildScheduledCommand does
(replace all ' with '\'' ) before constructing the script so the resulting -p
'...escaped...' is safe and robust.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
tailwind_files="$(fd -HI --glob 'tailwind.config.*' . || true)"

if [ -n "$tailwind_files" ]; then
  while IFS= read -r file; do
    echo "== $file =="
    sed -n '1,220p' "$file"
  done <<< "$tailwind_files"
fi

echo "== Dynamic class call sites =="
rg -n --fixed-strings 'text-${accent}' src/components/NewChatModal/StepModel.tsx
rg -n --fixed-strings 'border-${accent}' src/components/NewChatModal/StepModel.tsx
rg -n --fixed-strings 'bg-${accent}/10' src/components/NewChatModal/StepModel.tsx

echo "== Existing safelist / explicit class references =="
rg -n 'safelist|accent-blue|accent-green|accent-purple|accent-cyan|amber-500|sky-500|violet-500|orange-500|blue-500|indigo-500' .

Repository: Charlie85270/Dorothy

Length of output: 40495

---

🏁 Script executed:

# Check common tailwind config locations and next.js config
find . -maxdepth 2 -type f \( -name "next.config.*" -o -name "tailwind.config.*" -o -name "postcss.config.*" \) | head -20

# Check package.json for tailwind config
grep -A 20 '"tailwind"' package.json 2>/dev/null || echo "No tailwind config in package.json"

# Check if postcss.config has tailwind setup
cat postcss.config.js 2>/dev/null || echo "No postcss.config.js found"

Repository: Charlie85270/Dorothy

Length of output: 217

---

🏁 Script executed:

# Read StepModel.tsx to see the actual implementation
wc -l src/components/NewChatModal/StepModel.tsx
head -50 src/components/NewChatModal/StepModel.tsx
sed -n '35,55p' src/components/NewChatModal/StepModel.tsx
sed -n '190,210p' src/components/NewChatModal/StepModel.tsx

Repository: Charlie85270/Dorothy

Length of output: 3707

---

🏁 Script executed:

# Check providers.ts to understand accent value structure
sed -n '30,120p' src/lib/providers.ts

Repository: Charlie85270/Dorothy

Length of output: 3554

---

🏁 Script executed:

# Check the next.config.ts for tailwind safelist
cat next.config.ts

# Check postcss.config.mjs
cat postcss.config.mjs

Repository: Charlie85270/Dorothy

Length of output: 824

---

🏁 Script executed:

# Search for safelist configuration in the codebase
rg -i 'safelist|content.*safelist' --type ts --type js --type mjs

Repository: Charlie85270/Dorothy

Length of output: 91

---

🏁 Script executed:

# Verify the exact lines in StepModel.tsx where dynamic classes are used
sed -n '40,50p' src/components/NewChatModal/StepModel.tsx
sed -n '193,205p' src/components/NewChatModal/StepModel.tsx

Repository: Charlie85270/Dorothy

Length of output: 1261

---

🏁 Script executed:

# Check if there's a tailwind.config in a different location
find . -name "tailwind.config.*" -o -name "tailwind.js"

# Check app/globals.css for any `@apply` or other class definitions
head -100 src/app/globals.css

# Search for where badgeClass is used (this stores full class names)
rg -n 'badgeClass' src/components/NewChatModal/ | head -20

Repository: Charlie85270/Dorothy

Length of output: 2818

---

🏁 Script executed:

# Look for tailwind.css (v4 config file)
find . -name "tailwind.css" -type f

# Check the full postcss.config.mjs
cat postcss.config.mjs 2>/dev/null || cat postcss.config.js 2>/dev/null || cat postcss.config.ts 2>/dev/null

# Find where badgeClass is actually used in the template
rg -A 3 'badgeClass' src/components/NewChatModal/StepModel.tsx

Repository: Charlie85270/Dorothy

Length of output: 157

---

🏁 Script executed:

# Check if there's a content/safelist config anywhere in the codebase
rg -i '"content".*\[|content.*glob|safelist' --type ts --type js -B 2 -A 2

Repository: Charlie85270/Dorothy

Length of output: 46

---

🏁 Script executed:

# Check for tailwind.css in the src directory (Tailwind v4 config location)
find src -name "tailwind.css" -o -name "tailwind.config.*"

# Check the main tailwind.css file if it exists in src/app
ls -la src/app/ | grep -i tailwind

# Look at how Tailwind v4 is configured - check if there's a `@source` directive in globals.css
grep -n '@source\|@config\|content' src/app/globals.css

Repository: Charlie85270/Dorothy

Length of output: 46

---

🏁 Script executed:

# Verify the actual accent values being used and if they're ever fully specified in class names
rg 'accent-blue|accent-green|accent-purple|accent-cyan|amber-500|sky-500|violet-500|orange-500|blue-500|indigo-500' src/lib/providers.ts | head -20

Repository: Charlie85270/Dorothy

Length of output: 904

---

🏁 Script executed:

# Check if badgeClass (which stores full class names) is actually used or if it's dead code
rg -n 'badgeClass' src/components/NewChatModal/StepModel.tsx -A 2 -B 2

Repository: Charlie85270/Dorothy

Length of output: 46

---

🏁 Script executed:

# Check if globals.css has a complete theme extension for accent colors
sed -n '250,350p' src/app/globals.css

# Check how Tailwind v4 is processing content - look for any extend or theme config
rg -A 10 'theme|extend|colors' src/app/globals.css | head -30

Repository: Charlie85270/Dorothy

Length of output: 4437

---

🏁 Script executed:

# Verify if accent-* classes are actually rendered as utilities or just CSS vars
grep -n 'accent-blue\|accent-green\|accent-purple' src/app/globals.css | head -20

Repository: Charlie85270/Dorothy

Length of output: 777

---

🏁 Script executed:

# Check build/dist output to see if these classes actually make it through
find . -name "*.css" -path "*/dist/*" -o -path "*/.next/*" 2>/dev/null | head -5

Repository: Charlie85270/Dorothy

Length of output: 46

---

Dynamic accent classes will not be compiled by Tailwind and will break styling in production.

text-${accent}, border-${accent}, and bg-${accent}/10 compose class names at runtime. Tailwind v4 uses static content analysis and cannot extract these patterns—they won't appear in the compiled CSS. Only static class names like 'text-text-muted' (hardcoded fallback) are detected and compiled.

The codebase already defines badgeClass in the provider registry with full, pre-composed class names (e.g., 'bg-blue-500/15 text-blue-600 dark:text-blue-400'), but this field is not used in StepModel.tsx. Use the pre-composed class names from the registry instead of composing fragments at runtime.

Also applies to: 193-199

🤖 Prompt for AI Agents

Verify each finding against the current code and only fix it if needed.

In `@src/components/NewChatModal/StepModel.tsx` around lines 42 - 43, The dynamic
Tailwind classes in ProviderIcon (e.g., `text-${accent}`) and elsewhere are not
compiled; switch to using the provider registry's pre-composed `badgeClass`
instead of composing `text-${accent}`, `border-${accent}`, or `bg-${accent}/10`
at runtime. Update the `ProviderIcon` component to accept/use the provider's
`badgeClass` (from the provider registry entry) for its className logic and
remove runtime template strings; likewise replace similar dynamic class
composition in the other block mentioned (lines ~193-199) to use the registry
`badgeClass` so Tailwind can statically include the styles.