feat: token-efficiency improvements

[jaissebastian]

What is this feature?

When an AI agent uses code-review-graph MCP tools, some tools returned large dumps of data even when the agent only needed a quick answer. This caused:

• High token usage per tool call
• Agents making 3–4 separate tool calls just to review a change
• Flow results being walls of raw node data, hard to read

This feature fixes all three problems: it adds a compact output mode to every tool, introduces a single-call review tool, and makes flow output readable.

---

What changed

1. detail_level="minimal" extended to all 29 tools

Previously, only 7 tools (like detect_changes, get_impact_radius) supported detail_level. Now every tool does.

When an agent passes detail_level="minimal", the tool strips verbose fields and returns only what's needed to make a decision. The agent can then call again with detail_level="standard" only for the specific item it wants to drill into.

Tools that gained detail_level support:

Tool	What minimal returns
get_hub_nodes_tool	name, kind, in_degree, out_degree, total_degree
get_bridge_nodes_tool	name, kind, betweenness
get_knowledge_gaps_tool	counts per category only
get_surprising_connections_tool	source, target, surprise_score, reasons
get_suggested_questions_tool	top 5 question texts + priority counts
get_architecture_overview_tool	community names + counts + warnings only
get_community_tool	name, size, cohesion, dominant_language
list_graph_stats_tool	total_nodes, total_edges, files_count, last_updated
find_large_functions_tool	name, kind, line_count, relative_path
refactor_tool (dead_code/suggest modes)	top 10 items, name/kind only
cross_repo_search_tool	name, kind, repo, file_path
list_repos_tool	alias, path
get_affected_flows_tool	narrative summary + compact bullets

---

2. New tool: get_review_bundle_tool

A single tool call that replaces the typical 3–4 call review workflow.

Before (old way):

detect_changes_tool → get_affected_flows_tool → query_graph_tool (callers_of) → get_review_context_tool

After (new way):

get_review_bundle_tool

What it returns in one response:

• risk_score + risk — high / medium / low
• changed_functions — list of changed functions, with optional source snippets
• subgraph — one-hop callers and callees for each changed symbol (capped at max_nodes=20)
• affected_flows — compact summaries of impacted execution flows
• test_gaps — changed functions that have no test coverage
• review_checklist — structured [ ] action items generated from the analysis
• review_priorities — top 5 items to review first

Parameters:

get_review_bundle_tool(
  repo_root="<absolute path>",
  base="HEAD~1",              # git ref to diff against
  max_nodes=20,               # how many caller/callee neighbours to include
  include_snippets=false      # set true to get source code in the response
)

---

3. get_affected_flows_tool — readable summaries

Previously returned raw node dumps per flow. Now detail_level="minimal" returns:

• A narrative paragraph: e.g. "3 flows affected by changes in 2 files: 1 high-criticality, 1 medium, 1 low."
• Compact bullets per flow: flow_id, name, entry_point, criticality, depth, node_count, file_count, summary
• Raw step data is still available under the detail key if needed

A new helper summarize_flow() was added to flows.py — it is reused by get_review_bundle to generate the same compact format.