Show a real context-usage breakdown (/context) and reflect the always-sent baseline in the indicator

[umidjon-2231]

What feature would you like to see?

Two related gaps in how Kimi Code reports context usage:

1. The status-bar indicator under-reports. On a fresh session it shows context: 0.0% (0/262.0k) even though every request already carries a large always-sent baseline — the system prompt, tool schemas, MCP tools, skills listing, and memory (AGENTS.md). The number only becomes meaningful after the first turn, and there is no way to see what fills the window.

2. There is no /context command. Claude Code has a /context command that breaks usage down by category (system prompt, system tools, MCP tools, custom agents, memory files, skills, messages, free space). Kimi Code has /usage (totals + quotas) but nothing that shows the per-category composition.

Proposal

• Add a /context slash command that prints a category breakdown with token counts, percentages, and a progress bar, reusing the existing /usage panel style.
• Add a [context_window] baseline_mode config (off / include / subtract) so users can choose whether the status-bar indicator (and /usage / /status) folds in the baseline. Default off preserves current behavior.

I'm happy to open a PR — I already have an implementation with tests and bilingual docs ready, pending maintainer direction on the approach (especially the default baseline_mode and the three-mode design).

Additional information

• The breakdown is computed in agent-core and surfaced through a new getContextBreakdown() SDK session method (mirrors the existing getUsage / getTools RPC path).
• To avoid double-counting, the /context total is anchored to the model's real reported input-token count after the first turn, so it matches the footer rather than drifting on the character-based estimate.

[umidjon-2231]

cc @liruifengv — would appreciate a maintainer's take on whether this fits the roadmap, and on the proposed baseline_mode default, before I open the PR. A full implementation (with tests + bilingual docs) is already ready on my fork. Thanks!

[ktwu01]

Related UI angle from the legacy kimi-cli feedback: I opened #563 for making context usage quieter by default or threshold-based.

This issue asks for a more accurate /context breakdown. The complementary UX question is whether the detailed usage should be primarily on-demand, with footer/status warnings only when usage crosses an actionable threshold.

Original kimi-cli request: MoonshotAI/kimi-cli#2291.

[umidjon-2231]

This issue asks for a more accurate /context breakdown. The complementary UX question is whether the detailed usage should be primarily on-demand, with footer/status warnings only when usage crosses an actionable threshold.

The per-category breakdown is what makes optimization actionable. Today "context filling up" gives no lever to pull; with category-level numbers you can see which MCP server's schemas cost the most tokens, decide whether a rarely-used MCP is worth keeping connected, prune a bloated skills listing, or trim AGENTS.md. The threshold footer tells you when to look /context tells you what to do about it.