Using the Web App

The web app is a mobile-first PWA: a chat on a clean canvas, with every tool call, todo, plan, and approval rendered inline as the turn streams.

The chat

Type a request and send it (Shift+Enter for a new line). The agent works through the task autonomously — you watch:

• Text streams in as the model thinks and writes.
• Thinking blocks are collapsible — the model’s reasoning, shown when the provider reports it.
• Tool cards appear for every call: the tool name, its arguments, and the result. Command output renders in a terminal-style viewer with copy + expand; diff-shaped output renders in the unified diff viewer.
• Todos track the agent’s live task list (from the todo_write tool).

Approvals

By default (ACH_APPROVAL=require) mutating tools — bash, write, edit, apply_patch, code, repl — queue for human approval instead of executing:

• The turn pauses and an approval card shows the tool, its arguments, and a description.
• Approve runs the tool and the turn continues; Reject tells the agent and the turn continues without it.
• Multiple pending approvals offer Approve all.
• The Stop button in the header aborts the current turn.

Nothing mutates without your say-so. The approval queue exists but never blocks when ACH_APPROVAL=auto; off removes the queue entirely.

Plan mode

Set ACH_PLAN_MODE=1 to make the agent research first:

1. The agent investigates with read-only tools (read, glob, grep, web search) and proposes a plan.
2. A plan card renders the proposal with Approve plan / Reject plan.
3. Approving unlocks execution — the agent works through the plan.

Workspaces

The harness is universal — it works with any number of projects, not one directory:

• The header switcher (top-left) lists every known workspace with a check on the active one.
• Open workspace… opens a dialog to point the harness at a new directory; the agent gets full access to it, the same trust model as running a coding agent in your terminal.
• The sidebar shows only the active workspace’s conversations, grouped under a collapsible header with a filter search box.
• Every conversation runs inside its workspace: tool execution uses that directory, and the agent gets a workspace-specific system prompt (see below).

Conversations

• The sidebar lists the active workspace’s conversations, with New conversation, rename, archive, and delete, plus a filter box.
• Conversations persist to SQLite (ACH_DB_PATH) and resume with full history on reload or from another device pointed at the same server.
• The header shows the workspace name and current git branch.

Mobile & PWA

The app is installable: on Android via the browser menu, on iOS via Share → Add to Home Screen. It respects safe areas, works with the on-screen keyboard, and the shell (UI + assets) loads offline via the service worker — the agent itself needs the server.

Declarative agents

The harness speaks the ask-agent agents/ convention, per workspace:

my-project/
├── agents/
│   ├── refactorer/
│   │   ├── agent.rb          # Ask::Agent::Definition subclass
│   │   └── instructions.md   # becomes the system prompt base
│   └── shared/
│       └── tools/            # tools shared by all agents

• The agent picker in the composer lists every agent in the active workspace (agents/<name>/ or app/agents/<name>/), plus the default.
• Selecting one applies to the conversation: its instructions.md becomes the system prompt base and the session is built via Ask::Agent.new, so the definition’s tools, skills (agent_dir), and model apply.
• Conversations remember their agent; switching the picker updates it for the next turn.

System prompts

Each workspace gets a pi-style system prompt, composed from:

1. A base prompt — a default, or a custom one via configuration.
2. Guidelines — sensible defaults plus anything you add.
3. Project context — AGENTS.md / CLAUDE.md files discovered by walking from the workspace up to the filesystem root (nearest first), injected as <project_context>.
4. An append section (ACH_SYSTEM_PROMPT).
5. A Current working directory: footer.

See Configuration for the knobs.

Model selection

The gear icon opens Settings: workspace info (name, root, branch, adapter, features) and the model list. The chosen model is saved locally and used for new messages.