OctoCounts: Instructions file for Codex

OctoCounts agent instructions

Code discovery: use codebase-memory MCP first

This repository is indexed by codebase-memory-mcp. For symbols, implementations, callers, dependencies, architecture, and impact analysis, use the graph before filesystem search. Do not start with repository-wide rg, grep, find, globbing, or reading whole source files to discover code.

Project configuration

• MCP server: codebase-memory-mcp.
• Graph project: OctoCounts (case-sensitive; pass project: "OctoCounts").
• Local checkout: /Users/bytedance/traeProjects/OctoCounts.
• Index mode: full, including frontend, backend, extension, scripts, and docs that the indexer supports; existing .gitignore exclusions remain effective.
• Local MCP executable: /Users/bytedance/.local/bin/codebase-memory-mcp.
• The host already registers this server in ~/.codex/config.toml; do not add a duplicate server or overwrite unrelated host configuration.
• The index is local to this host. Another host/checkout must register the MCP server and index its own absolute repository path before using these examples.

Session startup and freshness

1. At session start or after compaction, call list_projects, then index_status({"project":"OctoCounts"}). Verify the returned root matches the active checkout; do not use an index of another worktree by accident.

2. Use Verify (Tier 2) by default. For a quick positive lookup, Scout is allowed with provisional conclusions; for an exhaustive bounded audit, use Auditor.

3. If the project is missing, initialize it with:

   {"repo_path":"/absolute/path/to/OctoCounts","name":"OctoCounts","mode":"full"}
   

   Pass this object to index_repository. Use the actual checkout path.

4. Before relying on a result, check file freshness with check_index_coverage. The host's auto_watch was false when configured on 2026-09-08. Do not assume background refresh is active, or enable global watching for unrelated repositories merely to work here. When relevant source has changed, refresh this repository with index_repository before subsequent graph analysis. Batch refreshes after coherent edits, not after every keystroke.

5. If watching is enabled later, allow normal background refresh; reindex only when missing/stale or an immediate refresh is needed. Query current status; do not treat historical node counts or generation timestamps as freshness.