Turnout for search_code returns "search failed: temp file" on Windows 11 (Issue #196)

[kevalrajpalknight]

Hello Community,

I was struggling with using search_code within claude code on my windows machine, so here is what alternative you can use to have turnout for your task

Tested and verified tool behaviour for mcp__codebase-memory__*. Use this reference to pick the right tool on the first try and avoid wasting tokens on tools that are broken or redundant.

---

Status at a Glance

Tool	Status	When to Use
search_graph	✅ Working	Primary search — replaces search_code entirely
get_code_snippet	✅ Working	Read full source once you have a qualified_name
trace_path	✅ Working	Call chains, impact analysis, data flow
detect_changes	✅ Working	What changed + all impacted symbols since a git ref
get_architecture	✅ Working	Orientation — node/edge overview of a project
get_graph_schema	✅ Working	See available node labels and edge types before querying
index_status	✅ Working	Confirm a project is indexed and ready
list_projects	✅ Working	List all indexed repos with node/edge counts
manage_adr	✅ Working	Persist and retrieve architectural decisions
query_graph	✅ Working	Raw Cypher — use only when search_graph is insufficient
search_code	❌ Broken (Windows)	Fails with temp file error — use search_graph instead
search_graph (semantic_query)	⚠️ Degraded	Returns all nodes unranked unless index was built in full mode

---

Recommended Workflow (minimum tokens)

Always follow this sequence — each step feeds the next, no wasted calls:

1. search_graph(query="...")          → get qualified_name(s)
2. get_code_snippet(qualified_name)   → read source
3. trace_path(function_name)         → understand impact
4. detect_changes(since="HEAD~N")    → see what's in flight

Never call get_architecture or get_graph_schema on every task — only on first orientation in a new project or after a reindex.

---

search_graph — Three Modes

Mode 1 — BM25 full-text (recommended)

search_graph(query="user profile update", project="...", limit=5)

• Splits camelCase: updateProfile → update, profile
• Boosts: Functions/Methods +10, Routes +8, Classes/Interfaces +5
• Filters noise: File, Folder, Module, Variable nodes excluded
• Use label param to restrict to one node type (e.g. label="Route")

Mode 2 — Regex name match

search_graph(name_pattern=".*Service.*", project="...", limit=10)

• Exact regex on symbol name — good for listing all classes of a pattern
• query overrides name_pattern if both are set — use one at a time

Mode 3 — Semantic vector (only if index built with full mode)

search_graph(semantic_query=["publish", "send", "queue"], project="...", limit=5)

• Bridges vocabulary gaps — finds dispatch when you search send
• Results appear in semantic_results field, not results
• Falls back to returning all nodes unranked if embeddings are missing — skip this mode unless you know the index was built with full mode

Useful filters (combinable with any mode):

• label="Route" — only routes
• label="Class" — only classes
• file_pattern="*.py" — restrict to file type
• include_connected=true — pull connected nodes too
• limit=5 — always set this; default is unlimited

---

trace_path — Three Modes

# Who calls / is called by this function
trace_path(function_name="admin_login", project="...", mode="calls", direction="outbound", depth=3)

# Full data flow with argument expressions at each hop
trace_path(function_name="admin_login", project="...", mode="data_flow", risk_labels=True)

# Cross-service through HTTP/async Route nodes
trace_path(function_name="sendEvent", project="...", mode="cross_service")

• direction: outbound (callees), inbound (callers), both
• risk_labels=true adds CRITICAL/HIGH/MEDIUM/LOW per hop — useful for security review
• include_tests=false by default — set true only when debugging test coverage

---

detect_changes — Impact Analysis

detect_changes(project="...", since="HEAD~5")
detect_changes(project="...", since="v1.2.0")
detect_changes(project="...", since="2026-01-01")

Returns: changed files + every symbol (function, class, route) defined in those files. Use before starting a task to know what's already in motion.

---

get_code_snippet — Reading Source

Always call search_graph first to get the exact qualified_name. Passing a short name works but returns suggestions if ambiguous.

get_code_snippet(qualified_name="repo-1.app.services.admin_auth_service.AdminAuthService.login", project="...")

Set include_neighbors=true to also get caller and callee names in the response — saves a trace_path call for shallow lookups.

---

manage_adr — Architectural Memory

Stores decisions that survive across sessions. No ADR exists yet in any project.

# Read
manage_adr(project="...", mode="get")

# Write / update
manage_adr(project="...", mode="update", content="## PURPOSE\n...\n\n## STACK\n...\n\n## ARCHITECTURE\n...\n\n## PATTERNS\n...\n\n## TRADEOFFS\n...")

Write an ADR after any significant architectural decision so future sessions don't re-derive it.

---

Node Labels by Project

repo-1 (richest — use for deep analysis)

• Node types: Method, Class, Variable, File, Module, Function, Route, Section, Folder
• Edge types: DEFINES, USAGE, CALLS, IMPORTS, DEFINES_METHOD, HANDLES, TESTS, THROWS, DECORATES, WRITES

repo-2 (frontend — functions + interfaces)

• Node types: Function, Section, File, Module, Interface, Variable, Folder, Type, Route
• Edge types: DEFINES, CALLS, CONTAINS_FILE, CONTAINS_FOLDER, IMPORTS, USAGE, HTTP_CALLS

repo-3 (frontend — functions + utilities)

• Node types: Function, Section, File, Module, Interface, Variable, Folder, Type
• Edge types: DEFINES, CALLS, CONTAINS_FILE, IMPORTS, USAGE

---

What NOT to Do

• Do not call search_code — it is broken on Windows (temp file error).
• Do not call get_architecture on every task — only once per project for orientation.
• Do not use semantic_query unless you confirmed the index was built with mode="full".
• Do not pass limit as unlimited — always set limit=5 to limit=10 to control token usage.
• Do not call query_graph with raw Cypher for simple lookups — search_graph is faster and token-cheaper.