Files
platform/todo/sub-agents/an-api-endpoint-for-spawning-an-agent-with-a.md
pastilhasandClaude Opus 5 9b0376ec39 todo: close the spawning endpoint
Both modes verified against the live server, not reasoned about.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-23 05:02:45 +00:00

70 lines
3.9 KiB
Markdown

---
id: sub-agents/an-api-endpoint-for-spawning-an-agent-with-a
title: 'An API endpoint for spawning an agent with a prompt'
theme: sub-agents
status: done
created: 2026-08-21
date_source: stated
order: 1
---
**An API endpoint for spawning an agent with a prompt.** Wanted 2026-08-21. An agent calls it, names a
prompt, and gets another agent running. A sub-agent system — but **defined by the owner**, which is the
whole point of building one rather than leaning on the harness's own: the shapes, the limits and the
vocabulary are Officer's, so every harness gets the same door and the definition does not change when a
CLI does.
It has more of a starting point than it looks. `startClaudeTurn` already runs a turn with no browser
attached (`nudger.ts:102`, `ws: null`), so the mechanism for "run an agent from the backend" is proven —
what is missing is spawning a NEW one on request rather than feeding a known session, plus the surface
an agent can actually call, which is MCP or an authenticated route.
Open: whether a spawned agent is a Command HQ commander (named, listed, costed — `command_hq_sessions`
already does all three) or something lighter that disappears when it finishes; how its output returns to
the caller, streamed or awaited; and what the caller may set — model, effort, cwd, permission mode — versus
what Officer fixes.
**Bounded by the budget rule in `sub-agents/spawning-must-respect-the-account-s-own`, which is not
optional.** An endpoint that spawns freely is exactly how a weekly limit gets spent by something nobody
was watching.
---
**2026-08-21, narrowing this considerably.** The endpoint largely exists:
`POST /api/command-hq/commanders/new` takes a prompt, starts a turn with `ws: null`, waits only for
`session:claude` rather than the whole turn, and returns Claude's own transcript id. A commander is named,
listed and costed by `command_hq_sessions` already.
So the open question "commander or something lighter" answers itself from the code: making a spawned agent
a commander costs nothing and gets all three for free, where a lighter thing needs its own registry to
become visible at all. **The budget condition is now on that route**, so anything spawned through it
already respects the window.
What is genuinely left is the SURFACE an agent can call — MCP tool or authenticated route — and how output
returns to the caller. That is the remaining work on this item; the mechanism underneath it is done.
---
**Closed 2026-08-23.** `POST /api/sub-agents/spawn` — prompt, optional cwd and model, and either the session
key immediately or the result if you ask it to wait. `GET /api/sub-agents/budget` answers whether there is
room without consuming any.
Verified both modes against the live server rather than reasoned about: the synchronous call returned the
four words the agent was told to say, and the async call returned in milliseconds with the agent's file on
disk five seconds later.
Three things settled while building it, each worth not re-deriving:
- **The budget check is a precondition, not a refinement.** It runs before the spawn and refuses 429 naming
the window. `checkBudget()` already existed and was already tested, which is the only reason this was a
short job.
- **`confined`, and true on the day it shipped.** The registry test's own comment records that `chat` sat on
that list for a day before its turns actually ran as the member. This one inherits the mechanism rather
than claiming it — `chatIdentity` off the token, `member: { osUser, home }` to the sidecar — and no path in
the endpoint chooses a directory.
- **`member` is passed explicitly, never omitted.** The field is required-but-may-be-undefined on purpose:
optional let a call site forget it, and forgetting it means a member's agent running as the owner in the
owner's home.
What it deliberately is not: a queue, a supervisor, or a retry. One call starts one agent.