Claude cli

First things first: your phase checklist at .wazir/runs/latest/phases/ lists exactly what needs to happen. Read it. Follow it. Don't skip ahead because the task looks simple. Which phase are you starting?

Claude Code CLI Integration

Command Routing

Follow the Canonical Command Matrix in hooks/routing-matrix.json.

• Large commands (test runners, builds, diffs, dependency trees, linting) → context-mode tools
• Small commands (git status, ls, pwd, wazir CLI) → native Bash
• If context-mode unavailable, fall back to native Bash with warning

Codebase Exploration

1. Query wazir index search-symbols <query> first
2. Use wazir recall file <path> --tier L1 for targeted reads
3. Fall back to direct file reads ONLY for files identified by index queries
4. Maximum 10 direct file reads without a justifying index query
5. If no index exists: wazir index build && wazir index summarize --tier all

Reference for using the Claude Code CLI (Anthropic's official CLI for Claude) in Wazir pipelines. Claude Code is an agentic coding tool that operates in your terminal with access to tools like file operations, search, and bash execution.

Commands

claude (interactive)

Launch the interactive TUI for ad-hoc work.

claude
claude "Fix the failing test in src/auth.ts"

Print Mode (non-interactive)

The -p (or --print) flag is the primary mode for Wazir automation. It runs Claude Code non-interactively, outputs the result to stdout, and exits.

# Basic non-interactive prompt
claude -p "Explain the architecture of this project"

# Pipe data from stdin (avoids command-line length limits)
git diff main | claude -p "Review this diff for bugs"

# Chain with other tools
claude -p "List all exported functions" --output-format json | jq '.result'

# Save output to file
claude -p "Summarize the test coverage" > summary.md

Key flags (all work with -p):

Flag	Description
-p, --print	Run non-interactively; print response to stdout and exit
--model <MODEL>	Select model: opus, sonnet, haiku, or full name (e.g., claude-opus-4-6)
--fallback-model <MODEL>	Fallback model when the primary is overloaded (print mode)
--output-format <FORMAT>	Output format: text (default), json, stream-json
--json-schema <SCHEMA>	Enforce structured output conforming to a JSON schema (requires --output-format json)
--allowedTools <TOOLS...>	Pre-approve specific tools without prompting (space-separated)
--disallowedTools <TOOLS...>	Block specific tools from being used
--max-turns <N>	Limit agentic turns in a session
--append-system-prompt <TEXT>	Add custom instructions while keeping default capabilities (safe choice)
--system-prompt <TEXT>	Replace the entire system prompt (removes all defaults; use with caution)
--dangerously-skip-permissions	Bypass all permission barriers (CI/CD and dev containers only)
--verbose	Enable verbose output for debugging
--input-format <FORMAT>	Input format: text (default)

Session Management

# Continue most recent session in current directory
claude -c

# Resume a specific session by ID
claude -r <SESSION_ID>

# Resume from a PR
claude --from-pr <NUMBER>

# Fork a session into a new thread
claude --fork-session <SESSION_ID>

# Set a custom session ID
claude --session-id <ID>

Subcommands

Subcommand	Description
claude mcp add	Add an MCP server (--transport http|stdio, -s for scope, -e for env vars)
claude mcp serve	Expose Claude Code itself as an MCP server
claude agents	List all configured agents
claude config	Manage configuration settings
claude remote-control	Serve your local environment for external builds

Model Selection

Model	Best For	Notes
opus / claude-opus-4-6	Complex reasoning, architecture, multi-step tasks	Most capable, highest cost
sonnet / claude-sonnet-4-6	Daily coding, balanced performance	Recommended default
haiku / claude-haiku-4-5	Quick tasks, fast responses, high volume	Lowest cost, fastest

Select via:

• CLI flag: --model opus or --model claude-opus-4-6
• Interactive: /model slash command
• Environment variable: CLAUDE_MODEL
• Settings: .claude/settings.json

Fallback: Use --fallback-model haiku with -p to auto-switch when the primary model is overloaded.

Permission Management

allowedTools

Pre-approve tools to avoid interactive permission prompts. Critical for non-interactive automation.

# Allow specific tools
claude -p --allowedTools "Read" "Grep" "Glob" "Bash(npm run test:*)" \
  "Review the test suite"

# Allow all tools (equivalent to dangerously-skip-permissions but scoped)
claude -p --allowedTools "Read" "Write" "Edit" "Bash" "Grep" "Glob" \
  "Fix the bug in auth.ts"

Tool name patterns:

• Exact: "Read", "Write", "Edit", "Bash", "Grep", "Glob"
• Scoped: "Bash(npm run test:*)" allows only matching bash commands
• MCP tools: "mcp__servername__toolname"

disallowedTools

Block specific tools:

claude -p --disallowedTools "Write" "Edit" "Bash" \
  "Analyze this codebase for security issues"

Project permissions

Store permanent permissions in .claude/settings.json:

{
  "permissions": {
    "allowedTools": ["Read", "Grep", "Glob"],
    "deny": ["Bash(rm *)"]
  }
}

Non-Interactive Usage

Piping data

# Pipe a diff for review
The middle of a task is where shortcuts creep in. You tell yourself you'll catch up on process later. You won't — that never happens. Check your checklist now. What are you behind on?
git diff main | claude -p "Review this diff for correctness"

# Pipe file content
cat src/auth.ts | claude -p "Find potential bugs"

# Pipe combined context
{ echo "## Spec"; cat spec.md; echo "## Code"; cat src/main.ts; } | \
  claude -p "Does the code match the spec?"

Structured output

# JSON output with full metadata (tool calls, token usage)
claude -p --output-format json "List all API endpoints"

# Streaming JSONL (real-time events)
claude -p --output-format stream-json "Analyze the codebase"

# Schema-enforced structured output
claude -p --output-format json --json-schema '{"type":"object","properties":{"findings":{"type":"array"},"summary":{"type":"string"}}}' \
  "Review this code and return findings"

System prompt customization

# Append instructions (keeps default Claude Code capabilities)
claude -p --append-system-prompt "You are a security auditor. Focus only on vulnerabilities." \
  "Review src/auth.ts"

# Full override (removes all defaults; use when you need a clean slate)
claude -p --system-prompt "You are a JSON-only responder. Return only valid JSON." \
  --output-format json "List all functions in this file"

MCP Server Integration

Claude Code supports MCP (Model Context Protocol) servers for extended capabilities.

Adding MCP servers

# stdio transport
claude mcp add github -- npx -y @modelcontextprotocol/server-github

# HTTP transport
claude mcp add api --transport http https://api.example.com

# With scope (project vs user)
claude mcp add myserver -s project -- node server.js

# With environment variables
claude mcp add myserver -e API_KEY=xxx -- node server.js

Using MCP tools in automation

# Allow specific MCP tools in print mode
claude -p --allowedTools "mcp__github__create_pull_request" "mcp__github__list_issues" \
  "Create a PR for the current changes"

Claude Code as MCP server

# Expose Claude Code as an MCP server for other tools
claude mcp serve

Tool Search (lazy loading): Since early 2026, Claude Code uses Tool Search for MCP tools by default, loading tool schemas on demand rather than all at once. This reduces context usage by ~95%.

Built-in Slash Commands

Command	Description
/model	Switch model
/cost	Show token usage and cost
/clear	Clear conversation context
/compact	Compress conversation to save tokens
/help	Display available commands
/review	Review current changes
/debug	Debug a failing test or error
/effort	Set reasoning effort level

Wazir Integration Patterns

Secondary Review (used by wz:reviewer)

CLAUDE_MODEL=$(jq -r '.multi_tool.claude.model // empty' .wazir/state/config.json 2>/dev/null)
CLAUDE_MODEL=${CLAUDE_MODEL:-sonnet}

# Review uncommitted changes
git diff | claude -p --model "$CLAUDE_MODEL" \
  --allowedTools "Read" "Grep" "Glob" \
  "Review this diff against these acceptance criteria: <criteria>" \
  2>&1 | tee .wazir/runs/latest/reviews/claude-review.md

# Review a spec or design artifact
cat artifact.md | claude -p --model "$CLAUDE_MODEL" \
  "Review this spec against these criteria: <criteria>" \
  2>&1 | tee .wazir/runs/latest/reviews/claude-review.md

Structured Review Output

claude -p --model "$CLAUDE_MODEL" --output-format json \
  --json-schema '{"type":"object","properties":{"findings":{"type":"array","items":{"type":"object","properties":{"severity":{"type":"string"},"description":{"type":"string"},"location":{"type":"string"}}}},"summary":{"type":"string"}}}' \
  "Review the changes in src/auth/" \
  > .wazir/runs/latest/reviews/claude-review.json

Parallel Execution as External Validator

# Run Claude review in background
git diff main | claude -p --model haiku \
  --allowedTools "Read" "Grep" \
  "Quick security scan of this diff" \
  > .wazir/runs/latest/reviews/claude-security.md 2>&1 &

Multi-Turn Programmatic Sessions

For complex automation requiring multiple turns:

# Limit turns to prevent runaway sessions
claude -p --max-turns 5 --allowedTools "Read" "Grep" "Bash(npm test)" \
  "Run the tests, analyze any failures, and suggest fixes"

Error Handling

Error	Handling
Non-zero exit (auth/rate-limit/transport)	Log full stderr, mark pass as claude-unavailable, use self-review only. Next pass re-attempts.
Timeout	Wrap with timeout 120 claude -p .... Treat timeout as claude-unavailable.
Model overloaded	--fallback-model haiku auto-switches. Without it, retry after backoff.
Permission denied	Add required tools to --allowedTools or use --dangerously-skip-permissions in CI.
Max turns reached	Increase --max-turns or break the task into smaller prompts.

Configuration

Claude Code reads configuration from (highest to lowest precedence):

1. CLI flags (--model, --allowedTools, etc.)
2. Environment variables (CLAUDE_MODEL, ANTHROPIC_API_KEY)
3. .claude/settings.json (project-level)
4. ~/.claude/settings.json (user-level)
5. .claude/rules/ directory (modular rule files)
6. CLAUDE.md (project instructions)
7. Auto Memory (persisted learnings)

Key config fields in settings.json: model, maxTokens, permissions.allowedTools, permissions.deny, env.

Completion claims without verification evidence are worthless. For each phase checklist item — what specific output proves it was actually done? Can you point to something concrete for every single one?