devin

Imported from masmuss/dotfiles (devin/AGENTS.md). Install upstream with npx skills add masmuss/dotfiles --skill devin. Copyright stays with the author.

Codebase Memory

Codebase Knowledge Graph (codebase-memory-mcp)

This project uses codebase-memory-mcp to maintain a knowledge graph of the codebase. ALWAYS prefer MCP graph tools over grep/glob/file-search for code discovery.

Priority Order

1. search_graph — find functions, classes, routes, variables by pattern
2. trace_path — trace who calls a function or what it calls
3. get_code_snippet — read specific function/class source code
4. check_index_coverage — validate candidate paths and missed ranges before claims
5. query_graph — run Cypher queries for complex patterns
6. get_architecture — high-level project summary

Evidence tiers

• Scout (Tier 1): quick positive lookup with few calls and targeted source checks. Mark it provisional; do not make negative or exhaustive claims.
• Verify (Tier 2, default): task-directed graph evidence, relevant trace directions, exact snippets for material claims, and relevant pagination.
• Auditor (Tier 3): bounded-scope full verification with current generation, complete relevant pagination, both call directions and broader relationships when material, and every limitation disclosed.
• After candidate paths are known in any tier, call check_index_coverage once with every evidence path. Add relevant scopes for negative or exhaustive claims. A clean result means no recorded gap, not proof of completeness. For partial, skipped, excluded, stale, pending, or unknown coverage, read/grep the reported ranges or scope before relying on graph results.

When to fall back to grep/glob

• Searching for string literals, error messages, config values
• Searching non-code files (Dockerfiles, shell scripts, configs)
• When MCP tools return insufficient results

Examples

• Find a handler: search_graph(name_pattern=".*OrderHandler.*")
• Who calls it: trace_path(function_name="OrderHandler", direction="inbound")
• Read source: get_code_snippet(qualified_name="pkg/orders.OrderHandler")

Session resets and subagents

• At session start or after compaction, confirm the nearest graph project and generation with list_projects or index_status, then choose Scout, Verify, or Auditor.
• The graph project is the list_projects entry whose root_path contains your working directory; pass its name (or the absolute directory itself) as project, never a repo or folder name.
• Before spawning a subagent, query the graph and coverage in the parent. Pass the tier, project, generation/freshness, bounded scope, queries and pagination state, qualified symbols, paths, call-chain findings, coverage evidence with ranges/reasons, source fallback already performed, and unresolved questions in the delegated task context.
• Do not assume subagents inherit MCP access or the parent conversation. If a child lacks MCP tools, it must not call or claim MCP access. It should use the supplied evidence and read/grep exact source, especially every reported missed-coverage range.