K3 (kimi-k3) excessive input token consumption in auto mode — 18.3M input for a single session

What version of Kimi Code is running?

kimi code cli 0.27.0

Which open platform/subscription were you using?

ark-coding plan

Which model were you using?

K3

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Describe the bug
In --auto mode with model ark/kimi-k3, a single session consumed 18.3 million input tokens within a few hours. The context window was only ~20% full (~200k tokens/round), so the issue is not context overflow but aggressive history accumulation and sub-agent forking, where each sub-agent independently burns tokens without the user being aware.

Observed data

Metric	Value
Total input tokens	18,300,000
Total output tokens	104,000
Input/output ratio	176:1
Context usage per round	~20% (204k out of 1M)
Estimated rounds	~90+

---

Expected behavior
For a similar task, Claude Code (also with 200k+ context) consumes 5–10x fewer input tokens in my experience. I expect Kimi Code CLI's context management — tool result dedup/summarization, sub-agent result compression, and automatic compaction — to be on par.

---

Environment

• OS: Windows
• Kimi Code CLI version:0.27.0
• Model: ark/kimi-k3 (1M context)
• Mode: --auto
• No custom proxy or config overrides

---

Additional context

• /compact was not called during the session — but with only 20% context usage, I did not expect it to be necessary
• Sub-agents were forked automatically by the main agent to explore the codebase; each sub-agent independently consumes tokens from the same quota
• Claude Code handles the same scenario with far less consumption, suggesting this is a CLI-level optimization issue rather than a model-level one

What steps can reproduce the bug?

1. cd into a mid-size codebase (few thousand files)
2. Run kimi --auto -m ark/kimi-k3
3. Give a moderately complex task (e.g., "explain the architecture and fix any lint errors")
4. Let the agent run for 30–60 minutes
5. Check usage panel

What is the expected behavior?

╭ Usage ───────────────────────────────────────────────╮
│ Session usage │
│ ark/kimi-k3 input 18.3M output 104k total 18.4M │
│ │
│ Context window │
│ ████░░░░░░░░░░░░░░░░ 20% (204k / 1M) │

╭ Usage ────────────────────────────────────────────╮
│ Session usage │
│ ark/kimi-k3 input 5.9M output 81.9k total 6M │
│ │
│ Context window │
│ ███░░░░░░░░░░░░░░░░░ 14% (139k / 1M) │
╰───────────────────────────────────────────────────╯

Additional information

No response

Reproduce the report with kimi --auto -m ark/kimi-k3 in a mid-size codebase and monitor the usage panel, comparing input-token and context-window totals. Trace the auto-mode history accumulation and automatically forked sub-agent activity; done means the same scenario no longer produces excessive input consumption while maintaining expected context usage.