Agent Skills
› AjayIrkal23/agentic-mercy-10x
› codebase-intel-first
codebase-intel-first
GitHub作为代码库理解的强制编排步骤，在规划、审计或编码前优先使用索引和依赖图构建结构模型。解决发现与文件读取的优先级，通过一次调用回答结构性问题，避免低效的多轮搜索。
Trigger Scenarios
涉及代码库的任何任务（规划、审计、审查、实现、重构、调试）
需要理解代码结构或模块关系时
Install
npx skills add AjayIrkal23/agentic-mercy-10x --skill codebase-intel-first -g -y
SKILL.md
Frontmatter
{
"name": "codebase-intel-first",
"schema": 1,
"category": "intel",
"surfaces": [
"codebase"
],
"triggers": {
"paths": [],
"intents": [
"intel"
],
"keywords": [
"agent",
"agents",
"alike",
"already",
"analysis",
"answer",
"ast",
"audit",
"auditing",
"bigger",
"blast",
"body",
"boundaries",
"broader",
"build",
"call",
"callees",
"callers",
"change",
"changes",
"code",
"code-intelligence",
"codebase",
"coding",
"complex",
"config",
"confirm",
"context",
"contracts",
"create",
"cross-file",
"cross-module",
"ctx_read\/ctx_search",
"dead",
"dead-code",
"debugging",
"dependency",
"deterministic",
"discovery\/understanding",
"docs",
"dozens",
"dramatically",
"easy",
"edges",
"engineers",
"env",
"every",
"exact",
"execution",
"explore",
"fast",
"file",
"files",
"first",
"fits",
"flow",
"function\/class\/method",
"give",
"graph",
"graph\/index",
"graphify",
"graphs",
"grepping",
"help",
"higher-level",
"i\/o",
"identify",
"impacted",
"implementing",
"index",
"instead",
"intel",
"jcodemunch",
"jcodemunch-token-saver",
"jcodemunch\/graphify",
"json",
"know",
"known",
"layer",
"layers",
"lean-ctx",
"likely",
"locate",
"lookup",
"making",
"mandatory",
"map",
"markdown",
"mcp",
"mental",
"miss",
"model",
"need",
"never",
"non-code",
"orchestration",
"owns",
"path",
"pattern",
"perspective",
"picture",
"plan",
"planning",
"precedence",
"problem",
"progressively",
"questions",
"radius",
"read",
"read\/grep\/find",
"reading",
"reads",
"reduces",
"refactoring",
"refining",
"repository",
"resolves",
"retrieval",
"review",
"right",
"round-trips",
"scope",
"search",
"section",
"server",
"single-file",
"size",
"skip",
"small",
"solve",
"spawning",
"start",
"startup",
"step",
"structural",
"structure",
"subagent",
"symbol",
"task",
"tell",
"text",
"time",
"together",
"token",
"touches",
"trace",
"trigger",
"trivial",
"understand",
"unfamiliar",
"usage",
"whole",
"wire",
"work",
"zoom"
]
},
"platforms": [
"linux",
"darwin",
"windows"
],
"token-cost": 935,
"description": "MANDATORY orchestration step for ANY task that touches a codebase — planning, auditing, code review, implementing, refactoring, debugging, or \"help me understand X\". Run the jcodemunch symbol index + graphify dependency graph FIRST to build a structural mental model, BEFORE reading files, grepping, or spawning Explore agents. Resolves the precedence between lean-ctx (ctx_read\/ctx_search, which owns file I\/O) and code-intelligence retrieval (jcodemunch\/graphify, which own discovery\/understanding). Trigger at the START of every plan, audit, or coding task on easy and complex work alike — the graph\/index answer structural questions in one call instead of dozens of round-trips and never miss cross-module edges. Use together with jcodemunch-token-saver and graphify.",
"disable-model-invocation": false
}
Codebase Intel First
The orchestration layer over jcodemunch-token-saver (symbol index) and
graphify (dependency graph). Those two skills tell you HOW to use each tool.
This one tells you WHEN and IN WHAT ORDER — so code intelligence is the FIRST
move in every phase, not an afterthought.
The precedence rule (memorize this)
Discovering or understanding code → jcodemunch / graphify FIRST. Reading a file you already located, or any non-code file → lean-ctx
ctx_read.
lean-ctx's "ALWAYS use ctx_read instead of Read" mandate is about file I/O
efficiency. It does NOT mean "read files to discover structure." Discovery is
the graph's job. Never grep -r / find / ls -R across source to learn how
the code is shaped — query the graph.
Phase playbook — run these BEFORE anything else
Planning / "understand X" / architecture
mcp__graphify__graph_stats— size + shapemcp__graphify__god_nodes— entry points & most-connected modulesmcp__graphify__query_graph "<your question>"— NL structural querymcp__jcodemunch__get_repo_map/get_repo_outline— layoutmcp__jcodemunch__get_context_bundle/search_symbols "<name>"— the relevant code → Now write the plan. The graph IS your codebase map; don't rebuild it by reading dirs.
Auditing / review / dead-code / impact
mcp__jcodemunch__find_dead_code/get_dead_code_v2mcp__jcodemunch__get_blast_radius "<symbol>"mcp__jcodemunch__get_coupling_metrics/get_hotspotsmcp__jcodemunch__find_references/find_importers "<symbol>"mcp__graphify__god_nodes+get_neighbors "<node>"→ Complete + ranked. A grep-based audit silently misses cross-module call sites.
Implementing / changing code
mcp__jcodemunch__search_symbols "<name>"+get_symbol_source— locate preciselymcp__jcodemunch__find_references/find_importers— EVERY caller before a signature changemcp__jcodemunch__get_blast_radius "<symbol>"— before touching anything sharedmcp__graphify__get_neighbors "<file>"— downstream dependents → Then edit. Native Read/Grep on source is blocked during coding until jcodemunch is used.
Debugging
mcp__jcodemunch__get_call_hierarchy "<fn>"— callers + calleesmcp__jcodemunch__find_implementations/find_referencesmcp__jcodemunch__get_signal_chains— data/control flowmcp__graphify__shortest_path "<A>" "<B>"— connect two points
After the graph points you somewhere
Use lean-ctx to read the EXACT files/symbols surfaced:
ctx_read "<path>" mode=signatures|map|lines:N-Mfor the pinpointed filectx_readfor any doc / config /.env/ markdownctx_shellfor git / build / lint / test Read narrowly — jcodemunch already gave you byte-precise locations.
Freshness — do NOT rebuild unless told
systemd watch daemons + SessionStart guards keep the index/graph fresh. Rebuild ONLY when a guard prints STALE/MISSING:
- jcodemunch:
mcp__jcodemunch__index_folder({"path": "<root>", "incremental": true}) - graphify:
graphify update <root>(Bash, cheap, no LLM)
Red flags (you are doing it wrong)
| Thought | Correct move |
|---|---|
| "Let me grep the repo to find where X is" | mcp__jcodemunch__search_symbols "X" |
| "Let me read this whole file to understand it" | get_symbol_source / ctx_read mode=signatures |
| "Let me Explore-agent the architecture" | mcp__graphify__query_graph / god_nodes |
| "I'll just change this function" | find_references + get_blast_radius first |
| "ctx_read is mandatory so I'll read to discover" | ctx_read is for files you LOCATED; discover via the graph |
See also: jcodemunch-token-saver, graphify, rule ~/.claude/rules/codebase-intel-first.md.
Version History
- 581d130 Current 2026-07-19 09:05
