Events

litellm.agent(..., stream=True) yields eight kinds of event. Each adapter translates its runtime's native events into these.

from litellm.harness import (
    Text, Reasoning, ToolCall, ToolResult, FileChange, Compaction, Approval, Done,
)

for event in litellm.agent(Harness.CLAUDE_CODE, "Add type hints to utils.py", sandbox=box, model="litellm_proxy/coder", stream=True):
    match event:
        case Text(delta=d):
            print(d, end="")
        case Reasoning():
            pass
        case ToolCall(name=name, input=args):
            log.info("tool %s %s", name, args)
        case ToolResult(is_error=True, output=out):
            log.warning("tool failed: %s", out)
        case ToolResult():
            pass
        case FileChange(kind=kind, path=path):
            changed.append(path)
        case Compaction():
            pass
        case Approval() as a:
            a.allow() if a.tool == "read" else a.deny("read-only review")
        case Done(cost=cost):
            print(f"\n${cost:.4f}")

That match has no case _. Event is a closed union, so mypy and pyright both treat it as exhaustive.

Event	Fields	Notes
Text	delta	Assistant text, streamed.
Reasoning	delta	Only when the model and the harness emit it.
ToolCall	id, name, native_name, input, builtin	builtin is False for your own tools=.
ToolResult	id, output, is_error	Paired with ToolCall.id.
FileChange	path, kind, diff	kind is "created", "modified" or "deleted". diff is a unified diff for text files up to 256 KB, else None.
Compaction	tokens_before, tokens_after	The runtime summarized its own history.
Approval	tool, input, allow(), deny(reason)	Only from harnesses with tool approval.
Done	result, usage, cost, stop_reason	Always the last event, exactly once.

FileChange events come from diffing a sandbox snapshot taken before and after each turn, so edits appear even when the runtime doesn't report them as a tool call.

Tool names

Built-in tools that do the same thing get the same name on every harness, so UI code doesn't need a branch per runtime. native_name always has the runtime's original name.

name	Claude Code	Codex	OpenCode	Deep Agents
read	Read		read	read_file
write	Write		write	write_file
edit	Edit		edit	edit_file
bash	Bash	command execution	bash	execute
glob	Glob		glob	glob
grep	Grep		grep	grep
ls	LS		list	ls
web_search	WebSearch	web search	webfetch

Tools that aren't in this table keep their native name in name.

Stream result

stream = litellm.agent(Harness.OPENCODE, prompt, sandbox=box, model="litellm_proxy/coder", stream=True)
for event in stream:
    render(event)
result = stream.result  # the same Result the Done event carries