Modes, Models & Usage

The ActionRow contains two configuration surfaces: the ModeSelector (an expanding modal with four axes) and the ContextRing (a live token-usage indicator). Together they decide which AI runs, how it behaves, and how much room is left in the context window.

ModeSelector

Click the mode chip on the right of the ActionRow. An expanding modal opens with three or four columns:

+---------------------------------------------------+
|  Agent       | Model       | Mode        | Effort |
|  (*) Default | (*) Opus    | (*) Normal  |  ===   |
|  ( ) Custom  | ( ) Sonnet  | ( ) Plan    |   |    |
|              | ( ) Haiku   | ( ) Auto    |        |
+---------------------------------------------------+

Selections take effect on the next send. The active selection is the chip’s label.

Agent

Choice	What it is
Claude	The default. Drives the chat panel through Claude with the bundled Blueprint AI skill.
Command Line	An always-available escape hatch. The panel skips chat and stays in Terminal mode — run claude directly, or any other CLI you need.

Claude is the supported assistant. Cursor, GitHub Copilot, Windsurf, and OpenAI Codex ship next, both in this Agent dropdown and as auto-installed skill targets. For now, the Command Line agent is the in-panel escape hatch for any non-Claude tool.

The selection is per-session and persisted in the JSONL so re-opening the chat restores the agent it ran with.

Model

Visible only if the selected agent exposes multiple models. For Claude:

Model	What it’s for
Opus	Most capable. Long, multi-step Blueprint workflows. Default.
Sonnet	Balanced. Fastest acceptable choice for routine ops.
Haiku	Cheapest / fastest. Quick queries, simple ops, fast feedback loops.

Switching models mid-session is fine — the AI keeps the same conversation history.

Mode

Mode	Behavior
Normal	The AI executes operations, asking permission per tool call as configured.
Plan	The AI proposes operations without executing. Useful for reviewing a complex change before approving it. No PermissionCards appear in this mode because no destructive ops are attempted.
Auto	The AI auto-approves non-destructive tool calls (creates, reads, layouts) but still asks for permission on destructive ops (deletes, overwrites). Saves clicks for routine work.

Mode also sticks to the session.

Effort

A slider that controls reasoning depth — low, medium, high. Higher effort lets the model spend more thinking tokens before each tool call. The marker on the slider shows your last-used value.

For most Blueprint ops, medium is plenty. High is useful for long, ambiguous prompts where the model needs to plan across many domains.

Recently used (MRU)

The ModeSelector remembers the last 5 distinct configurations (agent + model + mode + effort) and surfaces them as quick-pick chips at the bottom of the modal:

Recent
[ Sonnet · Plan · med ]
[ Opus · Normal · high ]

Click one to apply the entire combination in a single click.

ContextRing

To the left of the SelectionToggle, a small circular arc tracks context-window usage in real time — hover for a tooltip, click to expand the breakdown. As the conversation grows the ring fills and transitions through colors:

• Cream — under 50 %
• Cream→blue gradient — 50–80 %
• Pulse blue — 80–95 %
• Pulse warning — > 95 % (you’re close to a forced compaction)

Hover for a one-line summary: 12,400 / 200,000 tokens (6 %).

ContextUsageMenu

Click the ring to open the breakdown popup:

+----------------------------------+
|  Context usage                   |
|  System prompt        2,100 tok  |
|  Skill / tools        4,500      |
|  Conversation         5,800      |
|  Attachments          0          |
|  -------------------- ---------  |
|  Total used           12,400     |
|  Limit                200,000    |
|  Free                 187,600    |
|                                  |
|  [ /compact ]   [ /clear ]       |
+----------------------------------+

Each line is the count Anthropic reported for the last assistant turn. The two action buttons run the corresponding slash commands inline.

Live updates

The ring updates after every assistant turn — it reads the usage field from the streaming response, so it’s authoritative, not estimated. Attachments are counted only after send, since their size depends on the final serialization.

Thinking indicator

While the AI is generating, a small animated label appears at the end of the message list:

thinking...

Hovering shows the elapsed time. The SendButton turns into a stop square; clicking it cancels the in-flight run cleanly (no orphaned tool calls).

Stop and resume

• Stop — clicking the stop square cancels the active generation. The partial assistant message is preserved in the session JSONL with an interrupted flag.
• Resume — when you re-open an interrupted session, the panel offers a Resume button at the bottom of the last message. Clicking it sends a continuation cue to the AI.

Effort and price

Higher effort costs more per turn — both in tokens and in dollars — even when the output is the same length, because the model emits more thinking tokens before responding. For routine ops, leaving effort at medium is the right default — bump to high only when the AI is making mistakes on ambiguous prompts.