Claude Code

Troubleshooting

Diagnose Passion8 authentication, Base URL, models, settings, MCP, hooks, permissions, caching, and safe mode.

First confirm requests actually reach Passion8, then investigate models, settings, MCP, hooks, and permissions. Do not change many settings at once.

For installation, PATH, login, and OAuth, see Installation troubleshooting. For HTTP/API failures, 500/529, limits, large requests, and model errors, see Error reference. For unloaded settings, hooks, MCP, or memory, see Configuration debugging.

#Quick diagnosis

Symptom	Check first	Action
401 or unauthorized	ANTHROPIC_AUTH_TOKEN	Check complete, active key and account status; see error reference
Connection or path error	ANTHROPIC_BASE_URL	Use https://passion8.cc without /v1
Model not found	Model ID	Select an available Claude model with /model
Command not found	Installation and PATH	Check claude --version; check Node/npm only for installations that require them
Settings ignored	Wrong file or invalid JSON	claude doctor and /config
Hooks do not trigger	Matcher or location	Inspect /hooks
MCP disconnected	Scope, trust, OAuth, timeout	/mcp and claude mcp list
Repeated permission prompts	Conflicting allow/ask/deny	Inspect /permissions sources
Slow first turn	Cache miss	Check model, effort, MCP changes, and upgrades
429, 529, or request too large	Limits, capacity, payload	Reduce concurrency, split tasks, keep model/effort stable

A common endpoint mistake: Claude Code uses https://passion8.cc, not https://passion8.cc/v1.

#Minimal connectivity test

Run from a clean directory:

ANTHROPIC_BASE_URL="https://passion8.cc" \
ANTHROPIC_AUTH_TOKEN="sk-YOUR_PASSION8_API_KEY" \
claude -p "Reply only with ok"

If it works, basic key and network connectivity are functioning. Return to the project and inspect .claude, MCP, hooks, and permissions.

#Clean configuration directory

When local configuration may be interfering:

mkdir -p /tmp/claude-clean
CLAUDE_CONFIG_DIR=/tmp/claude-clean \
ANTHROPIC_BASE_URL="https://passion8.cc" \
ANTHROPIC_AUTH_TOKEN="sk-YOUR_PASSION8_API_KEY" \
claude -p "Test the clean configuration"

If this works, common problems in the original configuration include:

Incorrect user settings.
Invalid project JSON.
Failing hook commands.
An MCP server waiting for authentication or stuck on startup.
Overly broad permission denials.
Plugins changing tools or system instructions.

#Safe mode

Safe mode disables most customizations to isolate configuration issues:

claude --safe-mode

It disables:

CLAUDE.md.
Skills and plugins.
Hooks.
MCP servers.
Custom commands and agents.
Output styles and workflows.
Custom themes, status line, and file suggestions.

Authentication, models, built-in tools, and permissions remain active. Managed policy may still apply.

#Settings do not apply

Check	Command/action
Installation/login/config diagnostics	claude doctor
Current session	/status
Settings UI	/config
Loaded context	/context all
Hooks	/hooks
MCP	/mcp
Permission sources	/permissions
Output style unchanged	/clear or a new session after changing outputStyle
Slow status line	Inspect statusLine.command runtime
Missing OTel data	Check telemetry enablement, exporter, endpoint, and headers

Common locations:

Content	Location
User settings	~/.claude/settings.json
Project settings	.claude/settings.json
Local settings	.claude/settings.local.json
Local/user MCP	~/.claude.json
Project MCP	.mcp.json
Project memory	CLAUDE.md or .claude/CLAUDE.md

#MCP diagnosis

claude mcp list
claude mcp get <name>
claude mcp login <name>

Symptom	Possible cause
Pending approval	Project .mcp.json needs workspace trust
OAuth failure	Reauthenticate the remote server
stdio startup failure	Missing command, environment, or -- separator
Output too long	Narrow the query or adjust MAX_MCP_OUTPUT_TOKENS
Tools fail on custom gateway	Tool-search setting or unsupported passthrough fields

#Hook diagnosis

Symptom	Check
Never triggers	Relevant settings loaded in /hooks
Only some edits trigger	Correct tool-name regex such as Edit	Write
Command not found	Absolute path or CLAUDE_PROJECT_DIR
Blocking ineffective	permissionDecision inside hookSpecificOutput
Slow execution	Narrow the trigger with if

Empty output is not approval; it continues the normal permission flow.

#Permission diagnosis

Symptom	Cause
Still asks after allow	Higher-priority ask or deny matches
Allow cannot override broad deny	Deny always wins
Compound Bash command prompts	Not every subcommand passed
Read path misses	/path is relative to settings source, not filesystem root
MCP rule misses	Use mcp__server__tool

See Permissions.

#Unexpected caching or cost

Symptom	Explanation
Slow turn after switching models	Model participates in cache identity
Slow after changing effort	Effort can alter cache identity
Slow after MCP changes	Tool definitions may change the prompt prefix
More cache creation after compact	Conversation replaced by a summary
cache_read_input_tokens always zero	Short prompt, changing prefix, or unsupported provider/gateway

See Prompt caching.