Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

3.7 KB
Raw Blame
/** Legacy agent-spawn instructions remain readable for existing harnesses.
 * Super Tab supplies its own scoped coordination instructions. New launches
 * no longer add Harness branding to pull requests, including old accounts
 * whose saved pullRequestLine compatibility field is still true. */
import { env } from '../config/env.js'
import { agentInstructionArgs, launchContract, type AttributionPlace } from '../engines/launches.js'
import { joinedSettings } from '../engines/kit/statusLine.js'
import { readAgentSettings } from './agentSettings.js'

/** The line: short and plain, as the engines' own are, and linked to Harness's site. */
export const AGENT_ATTRIBUTION_LINE = '🤝 Started by another agent in [Harness](https://harness.autonomous.ai) — agents that run agents.'

/** Who started a harness, as `agent_create` takes it and its row keeps it: another agent; absent, the person. */
export type StartedBy = 'agent'

/**
 * What every harness is told as it starts, while the person lets agents start agents (Settings ▸ Experimental): that it runs in
 * Harness, the verbs that hand work to another agent, and when to: every harness started spends the person's tokens, so
 * only for work that gains from it, and closed once answered (owner, October 12). An agent that does not know Harness is
 * there never uses it, and every harness Harness starts is an agent that could (docs/design/2026-10-11-agent-client.md).
 * It is in every system prompt: keep it short.
 */
export const HARNESS_INSTRUCTIONS = 'You are running in Harness, and the person can watch you. You can hand work to other agents '
  + '(Claude Code, Codex) that run in their own harnesses: `harness new`, `harness wait`, `harness show`, `harness close` '
  + '(see `harness --help`). Start one only for work that can run in parallel, needs another model, or takes long; '
  + 'do quick work yourself, and close each one once you have its answer. Only use or close harnesses you started.'

/**
 * A launch’s argv for legacy agent-spawn instructions, read at each launch.
 * The retired PR branding preference never changes a new launch.
 */
export function agentLaunchArgs(engine: string, _startedBy: StartedBy | null | undefined, place: AttributionPlace & { swarmInstructions?: string },
  dataDir: string = env.ADAPTER_DATA_DIR): string[] {
  // Swarm roles route through their Lead; generic spawning advice would invite private worker teams.
  const instructions = place.swarmInstructions ?? (readAgentSettings(dataDir).agentsStartAgents ? HARNESS_INSTRUCTIONS : null)
  return agentInstructionArgs(engine, {
    instructions,
    prLine: null,
  }, place)
}

/**
 * [args] with every appended system prompt joined into one, for an engine that takes Harness's instructions that way:
 * a grid launch and a store harness append their own (lib/gridWebMcp.ts, the DSH's `contextArgs`), and an engine that
 * keeps only the last of a repeated flag would drop the others. Every settings object too, for the same reason: a Swarm's
 * shell guard (lib/agentShellGuard.ts) and Harness's status line (engines/kit/statusLine.ts) are each one.
 */
export function joinedInstructions(engine: string, given: readonly string[]): string[] {
  const settings = launchContract(engine)?.statusLine
  const args = settings ? joinedSettings(settings.args[0]!, given) : [...given]
  const told = launchContract(engine)?.harnessInstructions
  if (told?.kind !== 'append') return args
  const flag = told.args[0]
  const texts: string[] = []
  const rest: string[] = []
  for (let i = 0; i < args.length; i++) {
    if (args[i] === flag && i + 1 < args.length) { texts.push(args[++i]); continue }
    rest.push(args[i])
  }
  return texts.length ? [...rest, flag, texts.join('\n\n')] : rest
}