cocoindex-code-search

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

1. Check for an existing MCP server/tool named cocoindex-code.
2. If absent or stale, run python tools/cocoindex_agent_search.py mcp-install.
3. 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.
4. 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.
5. Route broad queries through MCP or python tools/cocoindex_agent_search.py search --limit 5 "<query>".
6. Confirm every candidate with exact rg and direct reads in the live repo.
7. 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.