Build

Recipes

Tested procedures an agent can find, adapt and rehearse. A recipe is an ordinary script with a header and a PARAMS block; the agent searches them by describing the experiment, overrides the parameters, plans the run, and only then executes.

How an agent uses one

Find. mhp_lab op="recipes" query="pcr with plate transfer" returns the name, which device classes it needs, what it will ask the human, and its parameters.
Adapt. The agent fills PARAMS from the conversation: which plate slot, how many cycles, which device if there are several.
Plan. mhp_run recipe="pcr-with-plate-transfer" params={...} plan=true rehearses it. Reads are real; every write and action is a dry run through all the safety gates; nothing moves. The result lists each step, the ones that will ask a person, the long-running ones, and the first one the driver would refuse. The agent shows this to you.
Run. The same call without plan, or with background=true for anything long. Output and saved files land in the run folder; mhp_data reads them back.

Bundled recipes

Recipe	Needs	Does
`pcr-with-plate-transfer`	robot arm, thermocycler	moves a plate from the deck to the thermocycler, pre-heats the lid, runs the program, holds at 4 °C, returns the plate
`timed-hold`	any device with a setpoint	reaches a setpoint, holds it for a duration while logging readings to CSV, returns to a safe setpoint
`setpoint-sweep`	any device with a setting and a signal	steps through a list of setpoints, settles, records the measured value at each, saves a CSV and prints the table; calibration and characterisation
`periodic-readings`	any device with readable signals	logs chosen signals every N seconds for hours, with optional alert bands; run it in the background overnight

Writing your own

Put a file in ~/.openmhp/recipes/ (or a package's scripts/ folder for a single-instrument procedure). The header lines are what the agent searches; PARAMS is what it overrides.

# recipe: bradford-assay
# needs: liquid_handler, plate_reader
# asks: which plate holds the samples; the standard curve wells
# summary: Dispense Bradford reagent into a sample plate, incubate 5 min, read absorbance at 595 nm, save the plate as CSV.
PARAMS = {"handler": None, "reader": None, "sample_plate": "deck_A1", "reagent": "trough_1", "volume_ul": 200}

import time, os
lh = lab[PARAMS["handler"] or lab.find("pipette reagent into plate", cls="liquid_handler", state="idle")[0]["id"]]
rd = lab[PARAMS["reader"] or lab.find("absorbance plate reader", cls="plate_reader", state="idle")[0]["id"]]
with lh, rd:
    lh.wait(lh.invoke("dispense", source=PARAMS["reagent"], target=PARAMS["sample_plate"], volume_ul=PARAMS["volume_ul"]))
    time.sleep(300)
    job = rd.wait(rd.invoke("read_absorbance", wavelength_nm=595))
with open(os.path.join(run_dir, "absorbance.csv"), "w") as f:
    f.write(job["result"]["csv"])
print("bradford read; absorbance.csv saved")

Rules that make a recipe safe to hand to an agent: choose devices by lab.find with a class and state="idle" unless a PARAMS id is given; hold leases with with; save data to run_dir; print one summary line; never catch and hide a refusal.

What the plan looks like

{"verdict": "ok; 1 step(s) need a human's confirmation",
 "steps": [
  {"step": 1, "device": "arm-01", "kind": "write",  "target": "speed", "value": 30, "gate": "ok"},
  {"step": 2, "device": "arm-01", "kind": "invoke", "target": "home", "gate": "ok"},
  {"step": 3, "device": "thermocycler-01", "kind": "read", "target": "lid_closed", "value": true},
  {"step": 4, "device": "arm-01", "kind": "invoke", "target": "pick_plate", "params": {"location": "deck_A1"}, "gate": "ok"},
  {"step": 7, "device": "thermocycler-01", "kind": "invoke", "target": "run_protocol", "gate": "ok", "long_running": true},
  {"step": 9, "device": "thermocycler-01", "kind": "invoke", "target": "open_lid", "gate": "ok", "asks_human": true}],
 "refused": [], "asks_human": [{"step": 9, "target": "open_lid"}]}