Active context

Active context is an optional sidecar pattern for Vivary-backed codebases. The simple mental model is:

Vivary routes the work. CocoIndex-code helps find fuzzy source-code candidates.

The baseline remains small and deterministic: tropo owns the typed graph, strato owns the agent loop, and no index, daemon, embedding model, or MCP server is enabled by default.

For coding workspaces, create-vivary can scaffold a CocoIndex-code profile:

create-vivary init my-workspace --preset coding --active-context cocoindex-code

This writes:

• an active-context skill for Claude/Codex-style agents;
• docs/active-context.md with the local policy;
• graph nodes under modules/active-context/index.md, decisions/, and verification/;
• .cocoindex_code/ in .gitignore.

It does not automatically install CocoIndex-code, initialize an index, run embeddings, enable MCP, or send source text anywhere. Those are explicit gates after the user approves active context. For copy/paste agent instructions, use LLM-ACTIVE-CONTEXT.md.

The Rule

Agents ask before using active context unless the workspace already has explicit approval. Installing, initializing, indexing, enabling MCP, and using external embedding providers are Vivary gates.

When CocoIndex-code Helps

Use it when semantic retrieval is likely to beat raw text search:

• a large or unfamiliar codebase;
• implementation questions where exact names are unknown;
• cross-module flows, feature tracing, refactors, or review;
• rg returns too much, too little, or only naming coincidences;
• the user explicitly asks for CocoIndex, ccc, semantic code search, or active context.

Skip it when the exact file/string is known, the codebase is tiny, the source sensitivity is unknown, or the task is about Vivary graph truth rather than source-code semantics.

Retrieval Order

1. Ask Vivary what to open first: tropo find "<task>" --budget 1200 --json.
2. Use graph truth for ids, types, edges, and blast radius: tropo graph, tropo blast <id>, ozone impact <id>.
3. Use CocoIndex-code only for fuzzy code candidates: ccc search --refresh "<query>".
4. Read matched files directly before acting.
5. Verify with project tests and Vivary checks before a gate.

Install And Prove It

Useful commands after approval:

uv tool install --python 3.11 --upgrade "cocoindex-code[full]"
ccc init -f
ccc doctor
ccc index
ccc status
ccc search --refresh "authentication flow"
ccc search --path "src/db.py" "database connection pool"
codex mcp add cocoindex-code -- ccc mcp

When filtering with ccc search --path, prefer an exact indexed path such as src/db.py. Broad folder globs like src/* or docs/* can return no results even when matching files are indexed; run the query without --path first if you need to discover the exact path.

On non-interactive Windows agent runs, drive init through stdin so it chooses the local sentence-transformers default instead of opening an interactive prompt:

cmd /c "echo. | ccc init -f"

The sidecar should report the query, refresh status, useful file paths/line ranges, files read, and how the semantic hits changed or confirmed the graph-based context.