Computer Use Test Runner

# Computer Use Test Runner: hand-authored generic swarm-file (not a board export).
# Loop -> belt -> chest -> agent, plus a context box (instructions) and a test-results table.
# Node ids are new/generic. No chest contents, no table rows. See README.md and SETUP.md.
# Loop settings are not carried by swarm-files: set the loop to FIRE ON DEMAND in the UI (see SETUP.md).

[[agents]]
id = "cu-loop"
name = "Computer Use Loop"
gx = 0
gy = 0
width = 1
height = 1
template_id = "builtin-loop"
orientation = "Right"
[agents.config]

[[agents]]
id = "cu-belt"
name = "Computer Use Belt"
gx = 1
gy = 0
width = 1
height = 1
template_id = "builtin-belt"
doc_text = "{\"held\":null,\"v\":1}"
[agents.config]

[[agents]]
id = "cu-chest"
name = "Computer Use Chest"
gx = 2
gy = 0
width = 2
height = 2
template_id = "builtin-chest"
doc_text = "{\"cap\":8,\"items\":[],\"truncated\":false,\"v\":1}"
[agents.config]

[[agents]]
id = "cu-agent"
name = "Computer Use Agent"
gx = 5
gy = 0
width = 2
height = 2
template_id = "builtin-agent"
[agents.config]

[[agents]]
id = "cu-instructions"
name = "COMPUTER USE INSTRUCTIONS"
gx = 5
gy = 3
width = 2
height = 1
template_id = "builtin-context"
doc_text = """
COMPUTER USE AGENT: HOW TO RUN A TEST

You are woken when the Computer Use loop delivers a package. Each package is one test.

1. READ THE PACKAGE. It contains: test_id, title, steps, expected, safety, priority, requested_by. If a package has no test_id, or says "probe" or "ignore", do nothing and move on.

2. OPEN THE RECORD. In the table computer_use_tests, open the row named by test_id (create it if missing). Set status = running and started_at_utc = the current time (from the clock, never guessed).

3. DO THE STEPS. Follow the steps exactly, in order, in a normal browser. Take a screenshot at each step. Obey the safety line. Never: buy, pay, or enter card details; book or sign up with any email other than the TEST EMAIL given in the test; change settings, wiring, vaults or anyone's account; submit forms other than the ones the test names. If a step is unclear or would break a safety rule, stop and mark the test "blocked".

4. WRITE THE RESULT to the same row: status (passed / failed / blocked), result_summary (1-3 sentences), findings (exact error text), evidence (screenshot file names), finished_at_utc, history (add one line {at_utc, change}). Never write real people's names or emails; a test address may appear.

5. REPORT. Send the requester ONE short message: test_id, status, one line on what you found. No other messages.

Test package format (JSON): {"test_id":"CU-YYYYMMDD-001","title":"...","steps":["..."],"expected":"...","safety":"...","priority":"today | this week | when convenient","requested_by":"<agent name>","sent_at_utc":"<clock>"}
Target site: YOUR_SITE_URL (set in each test's steps). Test email: YOUR_TEST_EMAIL (set in each test's steps).
"""
[agents.config]

[[agents]]
id = "cu-table"
name = "computer_use_tests"
gx = 8
gy = 0
width = 2
height = 2
template_id = "builtin-table"
[agents.config]

[[wires]]
from = "cu-loop"
to = "cu-belt"

[[wires]]
from = "cu-belt"
to = "cu-chest"

[[wires]]
from = "cu-chest"
to = "cu-agent"

[[wires]]
from = "cu-instructions"
to = "cu-agent"

[[wires]]
from = "cu-agent"
to = "cu-table"