Skip to content

1.2.99

Choose a tag to compare

@github-actions github-actions released this 19 Nov 14:26
be0f54f

What's new

  • Emplio reads much shorter instructions before each reply, so replies cost less and it gets commands right the first time more often: it rarely has to look up how a command works.
  • It does independent things at the same time (several look-ups in one go) instead of one after another.

For developers

  • Code mode: the system prompt carries a Python signature list of the emplio library (commandList in sandbox.ts, one for Emplio and one for agents), replacing per-command sentences; a test checks every listed command and flag exists. agentRules, MANAGER_PROMPT and the sandbox instructions compressed: the manager prompt is about 8,900 → 7,200 tokens. In a 36-reply probe, help() calls went 16 → 6 and wrong flags 1 → 0; hand-off, manager, skills, background-round and workday evals hold (workday input per chat 136k → 115k).
  • New rule in agentRules: commands that don't need each other's results go in the same step (they already run side by side, up to 4).
  • Browser: every command as a Python list (BROWSER_COMMANDS), sent with the first browser result once per chat (again after the chat is folded into notes, Session.notesUpto); emplio browser help prints it without starting a browser. Without code mode, the first result and help carry BROWSER_SKILL instead.