feat(cli): add a read-only /context diagnostics view #1423

Problem

Maka’s TUI status line shows a compact context-usage indicator, but users cannot inspect what is consuming the current context window or understand whether the displayed values are measured or estimated.

This becomes increasingly important in long-running agent sessions involving tool calls, large results, and context compaction. Diagnostics should also remain consistent after switching, restarting, or resuming a session.

Comparable tools expose this as a read-only command:

• Claude Code /context
• Grok Build /context

Desired outcome

Add a read-only /context command to the Maka TUI that describes the latest completed model request.

It should:

• Show context used, total, free space, and percentage when reliable data is available.
• Clearly distinguish provider-reported totals from locally estimated categories.
• Provide a high-level breakdown such as system instructions, tool definitions, and messages.
• Show compaction information relevant to the displayed request when available.
• Produce the same result after session switch, restart, or resume.
• Degrade clearly when historical or trace data is unavailable.

The command must not send a model message, modify context, or trigger compaction.

Implementation direction

Prefer existing Runtime and provider-request trace data rather than introducing a second request-snapshot system.

The latest completed provider request should be the source for request usage and breakdown. Historical values must not be interpreted using the session’s currently selected model.

Existing promptSegments should not be treated as a complete latest-request snapshot because its inputs may represent different phases of a turn.

The exact rendering and internal projection API are intentionally left open.

Acceptance criteria

• /context is available as a local, read-only TUI command.
• The view represents the latest completed provider request.
• Actual and estimated values are visibly distinguished.
• Missing data results in an explicit unavailable or empty state rather than inferred values.
• Session switch, restart, and resume preserve the same available snapshot.
• Output remains readable in narrow terminals.
• The feature does not introduce context-management actions.

Non-goals

• Exact tokenizer-level category accounting.
• Per-file or per-message drill-down.
• Separate accounting for Skills, MCP servers, memory, or workspace instructions.
• Optimization recommendations.
• /compact integration or context editing.
• Session-wide cumulative tool-call or compaction statistics.

Related work

• fix(runtime): preserve token usage fields across replay #1422 preserves token-usage fields across persistence and replay.
• [MM-13] Migrate desktop and CLI context assembly to ContextEngine #792 tracks shared ContextEngine ownership and cross-surface diagnostics.