Skip to content
Keelson Rib · Swarm

Task context

A swarm investigating a GitHub issue can read the checkout. It cannot read the issue. Its agents have no shell, no token, and no forge access, and that is correct for read-only agents. The gap is on the intake side.

What goes wrong without it

Section titled “What goes wrong without it”

The task text is capped at 8,000 characters, so a caller summarizes. The summary drops acceptance criteria, and agents fill the gap from what a name suggests. In one real swarm, an issue defined an input named author as the model ID that wrote an artifact. No agent had read the issue body, and the conclusion treated author as a possible GitHub login.

Snapshot, then delegate

Section titled “Snapshot, then delegate”

Task context is evidence the caller retrieves before starting the swarm and passes in whole: issue bodies, PR descriptions, diffs, review comments, check results, or a plain note. Each item has an id the agents cite, a kind, a title, and the full text. It can carry where the text came from and when.

The swarm stores each body verbatim, and nothing can change it for the life of the swarm. Every agent, spawned workers included, sees the list of items in its system prompt and reads a body with chat_context. The bodies stay out of the prompt, so a large diff costs nothing until an agent asks for it.

The rib fetches nothing. There is no GitHub adapter in the rib or the harness, and an agent never holds a credential.

Attribution travels with the text

Section titled “Attribution travels with the text”

Each read is headed by the item’s source URL, retrieval time, and SHAs, so a quote stays attributable wherever an agent repeats it. Evidence about a moving target is only meaningful against the commit it was taken from, so a diff, review, or checks item is refused without a head_sha.

Missing beats invented

Section titled “Missing beats invented”

Every agent’s prompt carries three rules:

  • The task context is the only external evidence it has.
  • Requirements live in the context items, not in the task text’s summary of them. Read the item before relying on it, and cite its id.
  • If something is absent, has no retrieval time, or is bound to a different head SHA than the one under discussion, write MISSING EVIDENCE or STALE EVIDENCE and name what is needed.

A swarm started with no context says so in every agent’s prompt. Asking for an item that was not supplied returns an error that tells the agent to report it as missing, along with the list of items that do exist.

The list of items, without the bodies, is kept in the swarm’s summary, which chat_swarm_status returns and the run records. The record shows what evidence a conclusion rested on.

Section titled “Related”