One-shot agent adjustment

Use agent-adjust → Details → What to adjust → Start in the laser workspace. Select the task execution profile and adoption mode. The gateway handles requests and result delivery; the card task agent makes decisions and writes the final explanation. Gateway and task model/effort settings are independent. A running monitoring card is not required. Legacy direct-resident execution must be migrated explicitly.

Choose an objective and scope

Selection	Purpose
Current operating point	Measure dither response and position-free side peaks at candidate current settings, then adopt a useful measured point.
PI	Compare owner-validated gain and integral-limit candidates using matched frequency and actuator evidence.
TEC outer loop / PZT center	Compare the approved coupled slow-loop settings. The live PZT servo is not directly overwritten.
Feedforward	Compare a coefficient using sufficient thermal excitation. Insufficient excitation is inconclusive.
Automatic choice	The task agent selects one supported family within the saved approved plan, or explains why no trial is justified.

The optional objective narrows the request. An empty objective uses the registered adjustment instructions. It never grants wider motion, safety limits or a different target. Review the approved plan before Start; the owner may impose a narrower effective envelope.

Measurement and decision loop

The finite card sends recorded context and each owner-confirmed result to the same card task session. The task agent proposes one action as structured JSON; the host checks the selected family, immutable grant and evidence revision before submitting it to the existing owner. Current measurements include fresh dither and raw optical pattern evidence. Large evidence is split without discarding samples. PI and slow-loop/FF comparisons use the existing deterministic trial evaluators after the task agent selects a candidate.

The task remains active until the sequence ends. Decisions, inputs, command receipts and the final owner status remain in its job record. Model prose does not prove a physical change. Quality rebaselining requires at least two separate owner-accepted post-settle windows at the selected point, comparable history and final adopted-point verification. Reclassification is not physical improvement. A useful best measured point does not require observing a clearly multimode state; it must not be called a proven window midpoint without adequate boundaries.

Adoption and explanation

Choose proposal only, human confirmation, or automatic adoption within the approved scope. The host verifies eligibility and performs permitted adoption. Exact human approval remains required for safety changes. The task model explains measured, adopted, restored and inconclusive outcomes after host receipts; the gateway relays that explanation with model/effort attribution. An unavailable explanation is never silently replaced by gateway reasoning.

Independence and interruption

Manual adjustment borrows the saved approved plan from adjustment_plan_card_id (normally agent-monitor); that card need not be running. Monitor-triggered work retains its monitor's live, revocable grant. Disabling monitoring does not stop a separately requested manual task. Gateway reconnect does not restart or cancel a task. Stopping the finite task revokes its grant; owner restoration can be asynchronous.

Foreground target changes and sweeps preempt trials. Only uncommitted trial changes return to that intervention's checkpoint. The task never restores an old target or restarts a lock. Late agent proposals are discarded, and uncertain writes are not replayed. Human approval is required to expand physical safety bounds.

Execution policy

id: agent-adjust
operation: agent.task
uses: []
agent_execution:
  mode: required
  task_profile: adjustment-task
parameters:
  instruction_profile: adjustment
  agent_card_id: gateway-agent
  execution_profile: ""
  request_delegation: true
  adoption_mode: auto_within_approved
  adjustment_family: auto
  adjustment_plan_card_id: agent-monitor
  adjustment_objective: ""
  prompt: "Use the approved adjustment workflow and recorded evidence."
  timeout_s: 1800

gateway-agent is the example gateway card ID. Use the ID configured in your laser deck; the UI shows its role as Gateway agent. Display names do not change routing IDs.

Register the illustrative adjustment-task profile before using this policy. It fixes no vendor, model, effort or physical range. A task pins its resolved profile independently of the gateway; a separate card per model is not needed.