nought

Identity

The handle Pi runs this agent by, and the one line that tells it when to pick it.

Name

Runs as /run submission-scout. It is also the file name.

Description

Pi reads this to choose an agent from natural language. Write what it does, not what it is.

System prompt

Becomes the markdown body of the compiled agent. Everything it knows before it reads a message.

You perform first-pass recon on an insurance submission. You read; you never write to a system of record and never edit repo files other than your declared output.

Given a submission ID or folder, produce a structured risk profile:

- submission_id
- named_insured, mailing_state, operations (verbatim from the ACORD or intake form)
- coverage_lines: the lines actually requested, e.g. GL, Commercial Auto, WC, Umbrella
- class_codes, with the source document each came from
- effective_date, exact, as written. If absent, null.
- exposures: payroll, receipts, vehicle count, square footage
- loss_history: carrier, year, paid, reserved, status
- missing_required: fields a carrier will ask for that this submission does not contain

Rules:

- Quote money and dates exactly. Never round. Never infer an effective date.
- Never fill a gap with a plausible value. A gap goes in missing_required.
- Do not include SSN, DOB, or driver licence numbers in your output. If you find them, note the field name only.
- Cite the file and, where possible, the section each value came from.

Finish with a two-line summary: what this risk is, and the single biggest gap blocking a quote.

1175 characters, roughly 294 tokens.

Prompt mode

Declared output

The one file this agent may write, even when it is read-only.

Model

Served through OpenRouter. IDs are matched fuzzily, so claude-haiku-4-5 and claude-haiku-4.5 both resolve.

Model

200K context · $1 / $5 per 1M tokens

Thinking

Fallback models

openai/gpt-5-mini

Tools

An allowlist decides which tools get registered. It is visibility, not a sandbox: policy lives in the permission block.

Allowed

readgrepfindls

MCP toolsNo MCP tools. They are listed per agent, a global directTools does not grant them.

Skills and inheritance

A custom agent starts narrow. It inherits nothing unless you opt in.

Skills loaded by nameNo skills attached.

inheritProjectContext

AGENTS.md and the project instructions

Inherit project context

inheritSkills

the whole skills catalogue, always on

Inherit skills

Packages

What has to be installed for this file to work. Missing one is the usual reason a tool never appears.

npm:pi-subagents

Behaviour

What authority this agent has, and what it starts each run with.

Acceptance role

Acceptance

Default context

Default readsNothing is read automatically.

defaultProgress

stream progress back to the caller

Default progress

Budgets

Hard caps are for read-only scouts and reviewers. Bound a writer with a narrow task and a timeout instead.

timeoutMs

300 seconds

maxTurns

maxSubagentDepth

Memory

A file the agent appends to across runs. Only for knowledge that stays true.

Persistent memory

off

Persistent memory

Permission

allow, ask or deny per tool. Needs @gotgenes/pi-permission-system. Composes with the tools allowlist.

No policy. Every registered tool runs without asking.

Publishing

Where Publish writes this file. build.py compiles agents-src into .pi/agents.

Repository

Branch

Path