On-Device Delegate

Send the bulky reading in a task to the on-device worker, then check its answer once. The main session keeps planning and answering. The worker reads the big inputs on the local model, so they never enter the main context.

Steps

1. Gate on what you would read. If one command or a short script can compute the answer, do it yourself now: no plan, no notes about this skill. Delegate only when you would otherwise READ more than about 25 KB of text into your context (prose, many files). Never delegate security review or design decisions.
2. Check the worker. The worker is the custom agent local-coding:on-device-worker (the plugin id). If the task tool lists it as plain on-device-worker instead (copied into an agents folder), use that. If neither is available, work normally.
3. Brief. Do not read the big input first. Make ONE task call:
   • name: "on-device-<short-kebab>"
   • description: 3-5 words
   • agent_type: the worker id from step 2
   • mode: "sync"
   • model: the on-device model id named in the session's routing line, exactly (if there is none: "winml-azureml-mai-code-1.1-flash-on-device/MAI-Code-1.1-Flash-On-Device"). The worker does not pin a model, so without this it runs on your own model.

   • prompt, exactly:

     GOAL: <one sentence>
     INPUTS: <exact paths>
     RULES: Read only INPUTS. At most 8 tool calls. Edit nothing but OUTPUT. No questions.
     OUTPUT: <file path and exact format, or: reply only>
     ANSWER: Reply ONCE with only <exact format>, then stop.
     If you cannot, reply only: UNABLE: <reason>

4. Check once. At most ONE cheap check (format, a count, or one spot check). Never recompute the answer or re-read the input. If the answer is wrong, UNABLE, or the task call fails: do it yourself. Then finish.

Example

User: "Use the on-device-delegate skill. From the logs under logs\, list every distinct error message with its count, most frequent first."

Task prompt:

GOAL: Count each distinct error message in the logs.
INPUTS: logs\
RULES: Read only INPUTS. At most 8 tool calls. Edit nothing but OUTPUT. No questions.
OUTPUT: reply only
ANSWER: Reply ONCE with only lines of the form "<count> <message>", most frequent first, then stop.
If you cannot, reply only: UNABLE: <reason>

Check: confirm the lines match the format and spot-check one count with a single search. Then answer the user.

Limits

• The worker can return a confident but wrong answer. The single check catches format errors, not every wrong fact. Tell the user what you checked.
• Expect minutes, not seconds, per delegated task, especially on the first call while the local model loads.

Setup scripts

scripts/ holds two setup scripts for the user, described in the local-coding plugin README. They are not part of a task: never run them while you work on the user's request.

• scripts/Set-LocalOutputCap.ps1 caps the local model's output tokens in the user's providers.json.
• scripts/Update-LocalCodingPlugin.ps1 installs or updates the plugin, then runs the cap script.