본문으로 건너뛰기

Built-in subagents

This page is the Grida binding of the specialized subagent pattern (Master / delegation and verification). The master names the shape: an internal agent the host invokes with a constrained prompt and bounded cost. This page records what Grida runs and the product decisions for each.

Every built-in operation stays on the chat's provider authority. Titling and compaction never move a chat to a different capacity source; if that provider cannot serve the request, the operation fails without changing provider.

Catalog​

OperationTriggers whenOutput
Chat titleThe first text request of an untitled chat is acceptedThe chat's catalogue label
CompactionContext nears the verified window (automatic) or the user asksA bounded summary that replaces earlier context

Chat title​

Opt-in. A client enables titles at launch, because a title costs one model request that one-shot runs and scripted providers do not expect. The temporary titler never titles.

A temporary side chat is titled like a chat, from its own first request. Its title is presentation only: it is never listed, a person cannot rename it, and it ends with the side chat.

Trigger. When a chat's first text request is accepted, the title runs in the background alongside the first turn. Subagent chats and chats that already have a turn are not titled.

Shape. A temporary, in-memory, tool-free conversation under the same launch's provider execution. Input is the bounded first request text only.

Bounds. The request text is capped near 1 KB, the run has a fixed timeout, and the output is reduced to the first line without quotes, reasoning markup, or trailing punctuation, then bounded to a short label. Empty output leaves the chat untitled.

Race safety. The title is written only if the chat is still untitled at the same catalogue revision. A person's rename always wins. Failure is silent; the chat stays untitled.

Compaction​

Contract. Realizes Master / context and tool design as one bounded, tool-free summary of portable earlier context.

Triggers. Automatic when the assembled context approaches the model's verified context window, and manual on request.

Bounds. The summarized source and the summary have fixed byte ceilings and the summary a fixed output-token ceiling. A summary that is invalid or does not reduce context is refused.

Authority. Compaction runs on the chat's own model and captured provider authority. If the live principal, scope, or provider differs during the run or its recovery, it fails as an authority change instead of continuing elsewhere.

Outcome. An installed summary replaces the covered records; every other outcome (stale, stopped, uncertain, failed, no change) leaves history intact.

Adding a built-in subagent​

Bar: would every Grida agent want this, regardless of host environment? If yes, it is a built-in. Titling and compaction pass; canvas-specific extraction or form-builder helpers do not. Planning is an operating regime of the chat itself (Plan mode), not a built-in subagent.

  1. Confirm the master carries the pattern shape — if not, propose it to Master / delegation and verification first.
  2. Add a row to the catalog and a section here.
  3. State the cost bounds (input, output, timeout, tools) explicitly. Specialized subagents are where runaway cost compounds silently.

See also​