Record what a subagent started with

Holds what a subagent was given when a delegation started: the task, evidence, system prompt, first message, model, tool names, and its permission and context settings. A LeadAgent keeps one in memory for each delegation; read them with $get_subagent_contexts(). A manifest covers the starting point only, not what happened during the run.

Usage

DelegationManifest(
  input,
  definition,
  sources,
  scope,
  system_prompt,
  message,
  model,
  tools,
  policies,
  size,
  schema_version = 1L
)

Arguments

input

The task, as a list of DelegationInput fields.

definition

The definition's name, prompt, memory, initial prompt and skill names, plus an id hash.

sources

The evidence given to the subagent, each with a SHA-256 digest of its text.

scope

The lead's delegation_scope.

system_prompt

The subagent's system prompt, including memory and skills.

message

The first message sent to the subagent, including the definition's initial_prompt and the evidence.

model

The model name. Credentials are never recorded.

tools

Names of the subagent's tools.

policies

A summary of the permission, context and DelegationPolicy settings, with hashes that identify them. Callbacks are not recorded.

size

text_bytes, manifest_bytes and estimated_tokens (NULL when unknown). The token count is an estimate, not a billed count.

schema_version

Format version; must be 1L.

Value

A DelegationManifest object.

Details

To store a manifest, save S7::props(x) (plain lists) and rebuild it with do.call(DelegationManifest, record). A manifest includes the task and evidence text, so check who may see it before showing it to users; $get_subagent_contexts(redact = TRUE) leaves that text out.