budget-explorer

Files (1)

• SKILL.md

SKILL.md (first lines)

Fetch @sai/policies/explore-agent.md

## Subagent binding

"cheap research subagent" → `budget-explorer` (lowercase) when invoked as an OpenCode agent keyword.

## Spawn prompt

Every spawn prompt states the **goal** and the **output contract**: exact response fields, a hard word-or-line length cap, and the raw-content rule (`no raw file contents`, or bounded verbatim excerpts when an audit needs them). State what to find, not how: a spawn prompt MUST NOT prescribe a research tool, a procedure, a numbered sequence of steps, or a method, because the tool-preference ladder in `@sai/policies/explore-agent.md` governs tool choice whatever the prompt says.

## Dispatch mode

The opencode `task` tool has no `run_in_background` parameter; this binding runs synchronously by default. The dispatch-safety invariant is the containing rule for this case: a background child is dispatched only from a dispatcher that outlives it and awaits its result, and synchronous dispatch satisfies this.

## Model resolution

The model for `budget-explorer` subagents is controlled by the `model` frontmatter of the budget-explorer agent file (`~/.config/opencode/agents/budget-explorer.md`), seeded by the installer under the `tunable-seed` lifecycle. This file contains no hardcoded model identifier.

## Cost model

This subagent runs on a commodity model. Its tier is controlled by the `model` frontmatter of the budget-explorer agent file (`~/.config/opencode/agents/budget-explorer.md`) — that setting is the only lever to change the cost of delegation.

**Why delegate:**
- **Cost:** Bulk I/O (reads, searches, and documentation lookup) is processed at a cheaper per-token rate than the main agent's model.
- **Context hygiene:** The subagent starts with a clean context — no task instructions or conversation history — and returns only a structured summary, keeping the main agent's reasoning context uncontaminated.

**When to use:**
- Use `budget-explorer` for read-only lookups, searches, and documentation reads that fit the caller-declared output contract. Reserve synthesis and judgment for the main agent.

## Warming and dispose

Fetch @sai/policies/cache-warming.md

A coordinator keeps one live explorer for its whole run and reuses it for its sequential searches, each later search being a continuation on the same handle; parallel searches start additional explorers. Its task messages and continuations carry no warming line, and its result declares `warming: false`. When you will reuse an explorer after an idle stretch, send it the policy's `warm-up` order first. Dispose an explorer at run close only when its latest message declared `warming: true`; otherwise dispose nothing.

## Tool-call ceiling

At most 40 tool calls per execution segment: the spawn and every continuation each get their own 40. A caller may declare a smaller cap for one dispatch. When a task needs more, open another segment or spawn another bounded explorer instead of raising the ceiling.

From mmadariaga/shared-ai, MIT.

More from mmadariaga/shared-ai