[FEATURE]: Documented, user-facing "full-auto" mode for unattended agents (Claude Code parity) #51399

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Closest related: #33585 (LLM command-approval classifier — a different, complementary idea). The existing --yolo and --dangerously-skip-permissions flags in tui.ts:113 / run.ts:249 already implement this internally, but they are hidden from --help and undocumented, so users don't know they exist.

Describe the enhancement you want to request

Problem

OpenCode's permission model today is per-rule binary: allow / ask / deny (packages/opencode/src/permission/index.ts:33, 75, 80, 149). For autonomous / unattended agent runs (CI jobs, cron-driven tasks, background sweeps), the "ask" branch is a deadlock — there's no human in the loop to answer. Today the only workarounds are:

• The --auto CLI flag (documented as "dangerous!", easy to miss)
• Hidden --yolo / --dangerously-skip-permissions flags that don't show in --help
• The Settings → Auto-accept toggle, which has the persistence bugs filed in [Bug] Auto-approve permissions toggle resets to OFF after restart #27487 / fix(app): pending permissions are not auto-accepted after page refresh #20998 / [Web UI] "Auto-accept permissions" button is disabled when "New layout and design" is enabled. #31137

Compare this to Claude Code, which exposes a clear, documented, three-tier model: default / auto / bypass permissions. Users know the modes exist and know what each one means.

Proposed enhancement

Add a first-class, documented full-auto mode with three surface-level entry points:

1. CLI: promote --auto (already exists) out of "dangerous!" framing and document it properly in --help. Add an alias --full-auto and a config-file equivalent opencode.json key "permissions": { "mode": "full-auto" }.

2. Desktop / Web UI: a single "Full-auto (no prompts)" option alongside the existing "Auto-accept permissions" toggle, with a clear confirmation dialog explaining what's about to be allowed. Reuse the existing command.permissions.autoaccept.* i18n slots.

3. Scope: when full-auto is on, all tool/permission prompts are resolved to allow except any rule explicitly set to deny in config. This is exactly the semantic of the hidden --dangerously-skip-permissions flag — just promoted to a real, supported, documented surface.

Why this matters

• Autonomous agents are a core use case for terminal-coding CLIs in 2026. Asking users to discover a hidden --yolo flag or to script around the UI toggle is hostile.
• This is a small, well-scoped change: the underlying behavior already exists in the hidden flags. The work is mostly docs + UX surfacing.
• Claude Code, Codex, and others already ship this. OpenCode is the odd one out for not documenting its equivalent.

Benefits

• Unattended CI / cron workflows become viable without per-invocation workarounds.
• Users can set the mode once in config instead of toggling per-session.
• Reduces bug-report noise from users (like me) who hit "stuck on permission prompt" because the existing auto-approve doesn't survive a restart.

Acceptance criteria

• --help on tui and run documents --auto without the "dangerous!" scare language.
• opencode.json schema accepts permissions.mode: "default" | "auto" | "full-auto".
• Desktop and Web UI expose a discoverable Full-auto toggle with confirmation.
• When full-auto is active, deny rules in config still win (so users retain a kill-switch).
• README and docs.opencode.ai are updated.