What the chatbox can do today
Every message you type is routed through the same provider pipeline the CLIs use. Claude Code, OpenCode, Copilot, and Local (vLLM) are all live. Switching the agent dropdown changes which binary/SDK handles the run — the chatbox is the unified mouth for all of them.
Claude Code flags — supported via UI
| UI Control | Equivalent CLI flag | Notes |
|---|
| Agent dropdown → Claude Code | claude | Launches with bypassPermissions by default |
| Model picker (sonnet/opus/haiku/fable) | --model claude-sonnet-5 | Aliases resolved on backend (sonnet→claude-sonnet-5, opus→claude-opus-4-8, haiku→claude-haiku-4-5, fable→claude-fable-5) |
| Mode selector (plan/auto/bypass) | --permission-mode | plan = read-only; auto = AI decides; bypass = no prompts |
| CWD input | --add-dir + working directory | Sets the root for all file tools |
| Session ID field | --session-id <uuid> | Resume exact prior conversation |
| Continue button | -c / --continue | Resumes most recent session in that CWD |
| Send message after task done | Multi-turn stdin loop | Keeps session alive, sends follow-up prompt |
| Stop button | Ctrl-C / SIGINT | Graceful provider.stop() |
| Advisor panel | --agents '{"reviewer": {...}}' | Second agent reviews primary output |
| Task cost display | (not in CLI, only JSON output) | Cost tracked per session in DB |
| Scheduler jobs | --print -p "…" | cron | UI native; CLI needs external cron |
OpenCode flags — supported via UI
| UI Control | Equivalent CLI flag | Notes |
|---|
| Agent dropdown → OpenCode | opencode run | Headless server mode under the hood |
| Model picker | -m provider/model | Format: anthropic/claude-sonnet-4-6 |
| Continue session | -c --continue | Session persisted in opencode DB |
| Session ID field | -s --session <id> | Resume specific session by ID |
| Fork checkbox (planned) | --fork | Branch a session without mutating original |
| Thinking toggle (planned) | --thinking | Show reasoning blocks in output |
| CWD input | --dir | Working directory for the agent |
Gemini CLI flags — supported via UI
| UI Control | Equivalent CLI flag | Notes |
|---|
| Agent dropdown → Gemini (planned) | gemini -p "…" | Provider not wired yet — see roadmap |
| Model picker | -m gemini-2.5-pro | Any Gemini model string |
| Mode → plan | --approval-mode plan | Read-only, no writes |
| Mode → auto | --approval-mode auto_edit | Auto-approve edits, prompt for rest |
| Mode → bypass | --approval-mode yolo | Auto-approve everything |
| Resume session | -r latest | Gemini uses index-based resume |
| CWD / extra dirs | --include-directories | Comma-separated extra workspace dirs |
| Sandbox toggle (planned) | -s --sandbox | Isolated execution environment |
Feature Roadmap — CLI parity gaps to close
| Feature | CLI Source | Priority | Status |
|---|
| Gemini provider wiring | Gemini CLI -p headless | High | Done |
| Session fork UI | OpenCode --fork, Claude --fork-session | High | Done |
| Thinking blocks toggle | OpenCode --thinking | Medium | Done |
| Effort level selector | Claude --effort low/medium/high/xhigh/max | Medium | Done |
| Max budget USD input | Claude --max-budget-usd | Medium | Done |
| Allowed/disallowed tools | Claude --allowedTools / --disallowedTools | Medium | Done |
| Custom system prompt | Claude --system-prompt / --append-system-prompt | High | Done |
| MCP server config panel | Claude/Gemini --mcp-config | High | TODO |
| File attachments on prompt | OpenCode -f --file, Claude --file | High | Done |
| Worktree per task | Claude/Gemini -w --worktree | Medium | TODO |
| Sandbox toggle | Gemini -s --sandbox | Low | TODO |
| Session export/import | OpenCode export / import | Medium | TODO |
| OpenCode stats panel | OpenCode stats (token usage/cost history) | Low | TODO |
| PR auto-run (opencode pr) | OpenCode pr <number> | Medium | TODO |
| Output format (JSON/stream) | Claude --output-format stream-json | Low | Done (internal) |
| Multi-agent (advisor) | Claude --agents | High | Done |
| Session resume / continue | Claude -c, OpenCode -c | High | Done |
| Background scheduler | (CLI needs cron; UI native) | High | Done |
| Git commit/branch/PR from UI | (CLI needs gh + shell) | High | Done |
UI-only superpowers (no CLI equivalent)
- Task history DB — every run is persisted, searchable, resumable. CLI sessions are ephemeral unless you pipe to a file.
- Multi-agent advisor loop — one agent generates, another reviews. No CLI does this natively.
- Scheduler — cron/delayed/interval jobs fire tasks automatically. No daemon needed.
- Analytics dashboard — cost per folder, per model, daily timeline. CLI gives nothing.
- Git panel — stage, commit, diff, push, PR create/merge all inside the same window.
- Parallel providers — switch between Claude Code, OpenCode, Copilot, Local without re-opening terminals.
- Soft-delete & trash — tasks never permanently lost unless you choose.
- Real-time streaming — tool calls, thinking blocks, agent groups rendered live.
Chatbox slash commands (planned)
| Command | What it does | Status |
|---|
/model sonnet | Switch model mid-session | Done |
/mode plan | Switch permission mode | Done |
/effort high | Set Claude effort level | Done |
/budget 2.00 | Cap spend for this run | Done |
/fork | Fork session before next message | Done |
/system <text> | Inject system prompt override | Done |
/file <path> | Attach a file to the next message | Done |
/mcp <server> | Toggle an MCP server on/off | Planned |
/tools Bash,Edit | Restrict available tools | Done |
/clear | Clear history (keep session ID) | Done |