一键在 Manus 中运行任何 Skill

cocoindex-code-search

星标1

分支0

更新时间2026年7月15日 09:55

Use the pinned external CocoIndex Code MCP workflow for broad repository routing, then prove exact candidates with rg and direct file reads.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

strmt7

strmt7/simple_ai_trading

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

CocoIndex Code Search

Use this skill only when a broad question would otherwise require a large repository scan. Exact strings, symbols, filenames, and small result sets go directly to rg.

Required Flow

Check for an existing MCP server/tool named cocoindex-code.
If absent or stale, run python tools/cocoindex_agent_search.py mcp-install.
Prove registration with python tools/cocoindex_agent_search.py mcp-smoke. Registration alone is not evidence; the JSON-RPC initialize and tool-list probes must pass.
Create or refresh an index only when the disk write is intentional: index --allow-dirty-index for an intentional worktree snapshot, or search --refresh "<query>" on a clean tree.
Route broad queries through MCP or python tools/cocoindex_agent_search.py search --limit 5 "<query>".
Confirm every candidate with exact rg and direct reads in the live repo.
Run the ten-case benchmark after changing this workflow or upgrading the pinned package.

Safety Contract

Keep cocoindex-code[full]==0.2.37 pinned.
Never run ccc init in the live checkout and never commit .cocoindex_code/.
Store the shared install and per-repository mirrors/databases under AGENT_COCOINDEX_HOME or the platform-specific external default.
Do not mirror real .env files, credentials, ignored files, or private artifacts. Binary files are not semantic-search evidence.
MCP search is read-only and never refreshes an index implicitly. Treat stale results as routing hints, not current truth.
Keep routing output bounded: five results is the default and ten is the hard maximum. Refine the query or add path/language filters instead of loading a larger result set.
Benchmark UTF-8 output bytes as a context-volume proxy, but never report those bytes as measured model tokens.
The wrapper may reuse an existing daemon and must stop only daemons it owns.
Do not build a cold index during high-load model training or without enough disk headroom.

The portable contract, package hashes, and benchmark inputs are documented in docs/AGENT_WORKFLOWS.md and docs/reference/cocoindex-code-agent-benchmark-2026-07-11.md.

同仓库更多 Skills

同仓库

context-budget

strmt7/simple_ai_trading

Bound repository context reads and semantic-search output without weakening source, test, or evidence verification.

2026-07-151

ai-regression-testing

strmt7/simple_ai_trading

Choose staged contract tests that catch partial AI-generated fixes across models, data, risk, execution ownership, and CLI/Windows parity.

2026-07-111

docs-knowledge-maintainer

strmt7/simple_ai_trading

Keep concise user docs and machine-readable model evidence synchronized with real behavior without losing provenance or caveats.

2026-07-111

karpathy-guidelines

strmt7/simple_ai_trading

Apply four anti-error coding principles plus this repository's stricter financial-evidence, safety, and single-session rules.

2026-07-111

search-first

strmt7/simple_ai_trading

Research the existing implementation, tests, primary technical sources, and financial rationale before adding trading code or dependencies.

2026-07-111

source-audit

strmt7/simple_ai_trading

Audit financial, data, model, and dependency claims for primary-source quality, dates, provenance, and reproducibility before implementation or publication.

2026-07-111

name	cocoindex-code-search
description	Use the pinned external CocoIndex Code MCP workflow for broad repository routing, then prove exact candidates with rg and direct file reads.
metadata	{"origin":"adapted from ZMB-UZH/omero-docker-extended at b27dbe990703d64d13e540c40cf4e122954c664d"}

CocoIndex Code Search

Use this skill only when a broad question would otherwise require a large repository scan. Exact strings, symbols, filenames, and small result sets go directly to rg.

Required Flow

Check for an existing MCP server/tool named cocoindex-code.
If absent or stale, run python tools/cocoindex_agent_search.py mcp-install.
Prove registration with python tools/cocoindex_agent_search.py mcp-smoke. Registration alone is not evidence; the JSON-RPC initialize and tool-list probes must pass.
Create or refresh an index only when the disk write is intentional: index --allow-dirty-index for an intentional worktree snapshot, or search --refresh "<query>" on a clean tree.
Route broad queries through MCP or python tools/cocoindex_agent_search.py search --limit 5 "<query>".
Confirm every candidate with exact rg and direct reads in the live repo.
Run the ten-case benchmark after changing this workflow or upgrading the pinned package.

Safety Contract

Keep cocoindex-code[full]==0.2.37 pinned.
Never run ccc init in the live checkout and never commit .cocoindex_code/.
Store the shared install and per-repository mirrors/databases under AGENT_COCOINDEX_HOME or the platform-specific external default.
Do not mirror real .env files, credentials, ignored files, or private artifacts. Binary files are not semantic-search evidence.
MCP search is read-only and never refreshes an index implicitly. Treat stale results as routing hints, not current truth.
Keep routing output bounded: five results is the default and ten is the hard maximum. Refine the query or add path/language filters instead of loading a larger result set.
Benchmark UTF-8 output bytes as a context-volume proxy, but never report those bytes as measured model tokens.
The wrapper may reuse an existing daemon and must stop only daemons it owns.
Do not build a cold index during high-load model training or without enough disk headroom.

The portable contract, package hashes, and benchmark inputs are documented in docs/AGENT_WORKFLOWS.md and docs/reference/cocoindex-code-agent-benchmark-2026-07-11.md.