Start - Pre-Flight

Purpose

Prime the session for tool-first navigation. Load CodeGraph, check project state, establish what's changed. This replaces "read everything and hope" with "load tools and discover."

Principle

Do not create more nodes to avoid thinking.

Create nodes only when they reduce confusion.

Instructions

Step 0 — Read Permissions

Read the project settings files to know what tools are pre-approved and denied:

Read: .claude/settings.json
Read: .claude/settings.local.json  (optional — skip without error if missing)

.claude/settings.json is committed and always present. .claude/settings.local.json is per-developer and may not exist.

Internalize the allow and deny lists. Never ask permission for tools in the allow list. Never attempt tools in the deny list. Do not print the contents — just absorb them silently.

Step 1 - Load CodeGraph

Run ToolSearch to load the codegraph tools:

ToolSearch: "+codegraph status"

Then call codegraph_status to verify the graph is available and check when it was last updated.

If .codegraph/ does not exist or the graph is stale, tell the user:

"CodeGraph is missing/stale. Run codegraph init -i to rebuild."

Step 2 - Check Branch and Changes

Run:

git branch --show-current
git log --oneline -5
git status --short

Summarize: what branch, what recent work, anything uncommitted.

Step 3 - Load Core Tools

Run ToolSearch to pre-load the tools you will most need:

ToolSearch: "select:mcp__codegraph__codegraph_search,mcp__codegraph__codegraph_callers,mcp__codegraph__codegraph_callees,mcp__codegraph__codegraph_context,mcp__codegraph__codegraph_node,mcp__codegraph__codegraph_impact"

Step 4 - Establish Workflow Rules

State these out loud as a commitment for the session:

Session workflow:

1. Find symbols — codegraph_search over Grep
2. Trace callers/callees — codegraph_callers/codegraph_callees before editing
3. Check impact — codegraph_impact before changing a symbol
4. Read whole — Read (not partial) before edit
5. Grep as fallback — only for string literals, comments, or things outside the code graph

Step 5 - Output

Produce a short status block:

## Session Ready

**Branch:** <branch name>
**Recent:** <last 2-3 commits, one line each>
**Uncommitted:** <list of changed files>
**CodeGraph:** <available + last updated / missing / stale>
**Tools loaded:** codegraph_search, codegraph_callers, codegraph_callees, codegraph_context, codegraph_node, codegraph_impact

Ready. What are we working on?

Constraints

• Do NOT read MEMORY.md, CLAUDE.md, or large docs during pre-flight. Those are reference — read them only when a specific question needs them.
• Do NOT summarize the project or explain the architecture. The user knows.
• Keep output under 15 lines. This is a boot sequence, not a briefing.
• If CodeGraph is unavailable, still complete the other steps. Note it and move on.