feat: add exec command and inline pipeline execution

[numman-ali]

Summary

• agentflow exec — run a single agent with a prompt, get raw output back. No files needed.
• agentflow run -e — pass inline JSON or Python expression directly
• agentflow run - — read pipeline from stdin

Usage

# Single agent call — just prompt in, output out
agentflow exec shell "echo hello"
agentflow exec claude "Explain this codebase" --tools read_only
agentflow exec gemini "Search GitHub trends" --model gemini-3-pro-preview --output text

# Inline JSON pipeline
agentflow run -e '{"name":"q","nodes":[{"id":"q","agent":"shell","prompt":"echo works"}]}'

# Inline Python pipeline
agentflow run -e 'from agentflow import Graph, shell
with Graph("test") as g:
    shell(task_id="q", script="echo works")
print(g.to_json())'

# Pipeline from stdin
echo '{"name":"q","nodes":[...]}' | agentflow run -
python3 my_pipeline.py | agentflow run -

Design

• exec builds a single-node PipelineSpec in memory, runs through the standard orchestrator
• exec --output auto defaults to raw text on TTY, JSON when piped
• run -e auto-detects JSON (starts with {) vs Python (executed as subprocess, matching existing .py loader)
• run - reads stdin, same JSON/Python detection
• Backwards compatible: agentflow run <path> works unchanged

Test plan

• agentflow exec shell "echo hello" → hello
• agentflow exec shell "echo hello" --output text → hello
• agentflow run -e '{"name":"q","nodes":[...]}' → pipeline runs
• agentflow run -e 'from agentflow import ...' → Python inline works
• echo '...' | agentflow run - → stdin works
• agentflow run pipeline.py → backwards compatible
• Error cases: bad agent, empty prompt, both path and -e, no input
• All 156 non-CLI tests pass; CLI test failures are pre-existing

🤖 Generated with Claude Code