Kimi-CLI-X

中文文档

Install from Source

python install.py

pip Install

pip install kimix
python -m kimix.cli
# or
kimix
python -m kimix

Note: This repo supports not only KIMI LLM but also various API keys! Like OpenAI, Anthropic, etc. Default config templates are in docs/; use kimix --config=xx.json after setup.

Why Kimi-CLI-X?

Kimi-CLI-X is a deep optimization of the original Kimi-CLI, focusing on prompt efficiency, tool reliability, and extensibility, plus new tools for real-world development.

Optimizations

1. Lean system prompts — Compressed initial prompts and tool descriptions down to ~2000 tokens while covering nearly all built-in tools.
2. Hardened permissions & validation — Properly handles Shell, Glob, and other tool validations to reduce retry loops from failures.
3. Better subprocess output — Redirects large outputs to temp files and filters redundant logs for easier backend retrieval.
4. Simpler concurrency — Streamlined design for subprocesses, sub-agents, and background tasks.
5. Programmable prompts — Allows custom system prompt injection at the upper layer for flexible scenarios.
6. Explicit conversation management — Clearer multi-task orchestration and state tracking.
7. Write-and-validate — Auto format checks and warnings on strict config files to prevent model-hallucinated errors.
8. Multi-API support — Import custom configs compatible with OpenAI, Anthropic, and more.
9. Verified backends — Tested against kimi, xai, anthropic, openai_legacy, openai_responses, google_genai/gemini, vertexai/vertex, plus 25+ Hermes-ported providers (deepseek, openrouter, xiaomi, zai, minimax, bedrock, …). See kimi-cli/tests/core/test_create_llm.py and the Supported Providers list.

New Capabilities

Scriptable Workflows (Core Advantage)

Unlike traditional CLI interaction where you type commands one by one, Kimi-CLI-X lets you write Python scripts to orchestrate entire workflows. You can combine prompts, loops, conditionals, and tool calls into fully automated, reproducible task pipelines:

from kimix import *
from pathlib import Path

clear_default_context()

for i in Path('docs').glob('*.md'):
    prompt(f'''According to the new git commits, update document `{i}`''')

Benefits:

• Batch automation: Use native Python syntax (for loops, file globbing) to fire tasks at multiple files at once.
• Complex orchestration: Freely compose mode switches, tool calls, and logic into multi-stage, multi-branch workflows.
• Reproducible & maintainable: Workflows live as version-controlled scripts, not ephemeral chat history.

---

Context Memory Architecture

Kimi-CLI-X embeds an automatic context memory system inside the KimiSoul core loop, keeping long conversations coherent without manual intervention. Three layers work together:

1. Conversation History Index (HistoryIndex)

Every user/assistant message is automatically indexed by BM25 inverted index (N-gram, n=2) on append, persisted to /history_index/.json, and survives process restarts. Cap at 500 rounds; oldest evicted automatically.

2. Automatic Context Compaction (SimpleCompaction)

Triggered when context token ratio hits compaction_trigger_ratio or free space falls below reserved_context_size:

• Retention policy: Recent N rounds kept verbatim (depth adapted by adaptive_preserve_depth — deepened on errors, thinking, multi-file edits, etc.); first message always kept (primacy effect).
• LLM summarization: Old messages compressed into structured summaries via a lightweight LLM call; thinking blocks discarded.
• Cascade handling: When already-compacted content is compressed again (depth ≥3), switches to COMPACT_CASCADE prompt to prevent information degradation.
• Post-compaction, all rounds marked is_compacted in HistoryIndex for future retrieval.

3. Auto History Retrieval + On-Demand Recall

• Auto retrieval (_maybe_auto_retrieve_history): Each round, if user input ≥10 chars, BM25-searches HistoryIndex for matching compacted rounds; injects matches above auto_retrieve_history_threshold as [Auto-retrieved from past conversation].
• retrieve tool: the agent can actively search all archived history (including compacted rounds) by natural-language query, returning verbatim excerpts with relevance scores (or fetch a turn by id).

┌──────────────┐    append     ┌──────────────┐    overflow    ┌──────────────────┐
│   Context    │ ───────────► │ HistoryIndex │ ────────────► │ SimpleCompaction │
│ (live window)│              │ (BM25 index) │               │ (LLM summary)    │
└──────────────┘              └──────────────┘               └──────────────────┘
       ▲                            │                               │
       │       auto-retrieve        │                               │
       └────────────────────────────┘                               │
       │              Retrieve (agent主动recall)                       │
       └────────────────────────────────────────────────────────────┘

---

Agent Harness: Self-Regulation & Reminders

The KimiSoul core loop actively keeps long runs on track — no manual babysitting. It works in CLI, server, and sub-agent sessions alike.

• Verification Gate — a turn can’t end while todos are unfinished, or while files were edited without running any check. Failing checks are fed back to the agent to fix.
• Anti-loop detection — catches repeated edits to the same file across different tools, and the same error recurring without a root-cause fix; nudges the agent to change strategy.
• Todo reminders — unfinished todos are periodically re-surfaced at the end of the context, so goals never drift out of attention.
• Compact reminders — when context fills up (~70%), the agent is prompted to compact on its own terms before forced auto-compaction kicks in.
• Budget reminders (opt-in) — wrap-up warnings as the per-turn step/time budget runs out, so the agent finishes gracefully instead of being cut off.
• Context meter — when context usage shifts materially, the agent is reminded to recall past history with the Retrieve tool.
• Decision-aware compaction — compaction summaries preserve a Decisions & Conclusions and a Verification Status section, so early decisions and verified work survive.
• Context pruning — stale tool outputs, thinking blocks, and near-duplicate content are automatically elided to reclaim context space.

todo_write

The todo_write tool tracks multi-step plans:

• Incremental updates with append/overwrite modes, fuzzy title matching, and per-todo notes.
• Nested sub-todos via todo_write (send the full tree) or todo_update(parent=...); todo_update(complete=True) finishes a subtree in one call.

Best-of-N Sampling

AgentSwarm’s parallel_sample mode runs the same task N times in isolated workspaces (git worktree / temp copy), picks the winner by model self-evaluation or majority vote, then applies and verifies the winning diff. Failures are explicit errors — never silently accepted.

---

Documentation Index

Tutorials

Config Reference