Agent Mode

Agent Mode outputs structured NDJSON instead of the interactive terminal UI. It's how AI coding agents (Claude Code, Codex CLI, Gemini CLI) consume Kane CLI results: parse events programmatically, extract the final result, and present it to the user.

Enable Agent Mode

Add --agent to any run command:

kane-cli run "Verify the checkout flow completes successfully" \
  --url https://myapp.com \
  --agent \
  --headless

When --agent is active:

• The interactive TUI is fully suppressed (no boot screen, no menus)
• Stdin is not used (no interactive prompts)

tip

Always combine --agent with --headless in non-interactive environments (CI/CD, agent contexts) to avoid display server errors.

---

Output Format

Kane CLI outputs one JSON object per line (NDJSON) to stdout:

{"type":"run_start","objective":"Verify checkout","timestamp":"2026-04-14T10:30:45Z"}
{"type":"step_start","index":0,"objective":"Navigate to cart page"}
{"type":"step_event","index":0,"event":"action","detail":"Navigated to /cart","success":true}
{"type":"step_event","index":0,"event":"screenshot","detail":"Screenshot captured"}
{"type":"step_end","index":0,"status":"passed","duration":2.3,"summary":"Navigated to cart"}
{"type":"step_start","index":1,"objective":"Click checkout button"}
...
{"type":"run_end","status":"passed","summary":"...","duration":45.2,...}

---

Event Schema

Step Events

Type	Key Fields	Description
step_start	index, objective, child_id*	Step began
step_event	index, event**, detail, success*	Activity within a step
step_end	index, status (passed|failed), duration, summary	Step completed

*optional
**event values: screenshot, reasoning, action, vision, assertion, evaluation

Flow Events

Type	Key Fields	Description
run_start	objective, timestamp	Run started
bifurcation	flows[], count	Agent branched into sub-flows
child_agent_start	child_id, objective, parent_step	Child agent spawned
child_agent_end	child_id, success, steps_taken, summary	Child agent finished
ask_user	question, step_index, options*	Agent needs input (auto-disabled when stdin is not TTY)
error	message	Error occurred
run_end	(see below)	Terminal event: always the last line

---

The run_end Event

run_end is always the last line. It contains the complete test result:

{
  "type": "run_end",
  "status": "passed",
  "summary": "Searched for laptop and added first result to cart",
  "one_liner": "Searched for laptop on Amazon and added to cart",
  "reason": "Objective completed",
  "duration": 45.2,
  "final_state": {
    "price": "$29.99",
    "product_name": "Wireless Headphones"
  },
  "context": {
    "memory": {},
    "variables": {},
    "pointer": "(passed) Searched for laptop on Amazon"
  },
  "token_usage": {
    "reasoning_input": 12000,
    "reasoning_output": 800,
    "vision_input": 5000,
    "vision_output": 200
  },
  "session_dir": "~/.testmuai/kaneai/sessions/2026-04-14_10-30-45_a1b2c3",
  "run_dir": "~/.testmuai/kaneai/sessions/2026-04-14_10-30-45_a1b2c3/runs/0",
  "test_url": "https://test-manager.lambdatest.com/projects/123/test-cases/456"
}

Key Fields

Field	Description
status	"passed" or "failed"
summary	Full description of what the agent did
one_liner	Short single-sentence summary
reason	Why the run stopped (present on failure)
final_state	Values extracted via "store as" objectives
test_url	Link to the TestMu AI dashboard for this run
session_dir	Path to session-level logs
run_dir	Path to run-level logs, step JSON, screenshots
token_usage	Token consumption breakdown

---

The .evidence proof pack

Every Kane CLI run produces an .evidence pack stored natively in the results folder, so failure context travels with the run instead of living in a separate dashboard. Each pack contains the full execution proof for the run: step-by-step traces, screenshots, video, and command, network and console logs, plus the NDJSON event stream and result codes. Coding agents and humans read the same evidence to see exactly what happened, where it failed, and why.

---

Parsing Output

Get the run_end event:

# Get just the final result
kane-cli run "..." --agent 2>/dev/null | tail -1 | jq .

# Extract status
kane-cli run "..." --agent 2>/dev/null | tail -1 | jq -r '.status'

# Extract a stored value
kane-cli run "go to example.com, store the price as 'price'" --agent 2>/dev/null \
  | tail -1 | jq -r '.final_state.price'

---

Handling ask_user Events

If an objective requires user input mid-run, Kane CLI fires ask_user:

{"type": "ask_user", "question": "Which item should I select?", "options": ["Small", "Medium", "Large"]}

note

ask_user is auto-disabled when stdin is not a TTY. In CI/CD and AI agent contexts, stdin is never a TTY, so ask_user never fires. Write objectives that don't require interactive prompts.

If stdin IS a TTY, respond by writing JSON to stdin:

{"type": "user_response", "answer": "Medium"}

To cancel:

{"type": "cancel"}

---

Agent Mode vs Interactive Mode

	Agent Mode (--agent)	Interactive Mode
TUI	Suppressed	Full terminal UI
Output	NDJSON to stdout	Formatted text
Stderr	Logs only	Logs + progress UI
Use case	AI agents, programmatic tools	Human development

---

Next Steps

• Error Codes: Complete reference of result_code values and recommended actions
• Parallel Execution: Run multiple tests concurrently
• Skills: Install the Kane CLI skill for Claude, Codex, or Gemini
• CLI Reference: Full flag and command reference