Working with thronglets

A thronglet is a nested session of an agent harness that throng starts for you. It works in a directory you name, and its final message comes back as the result. It sees nothing of your conversation: no files you read, no decisions you made, no user messages. Everything it needs travels in the prompt.

Choosing the agent

agent is one string: <harness>/<model>[:<effort>], e.g. claude/opus:high, codex/gpt-6-sol:xhigh, opencode/openrouter/z-ai/glm-5.3-flash, gemini/gemini-2.5-pro, cursor/gpt-5.4-nano. The model is a value the harness itself offers: list_harnesses returns them with the effort levels (it probes the adapters: seconds, no tokens), and a model_rejected message lists them too. A value with bracketed parameters (Cursor's gpt-5.4-nano[reasoning=medium]) can be named by the part before the [ when only one value has it. Effort is low | medium | high | xhigh | max; omitted means the harness default.

Pick another harness or model when the task gains from a different model's view (a second review, a design critique) or when the user names one. Same-vendor subagents of your own harness stay the cheaper default for plain parallel work.

The prompt

Self-contained, written for a reader with no context: the task, the repository and the files by path, the constraints, what the result must contain, and what to leave alone. The final message is the result, so ask for the shape you need ("end with a list of findings: file, line, problem"). A harness that finds an AGENTS.md follows it, commits and reviews included: when you want only the work, say where the thronglet's job ends ("leave the changes uncommitted").

cwd is the live tree the agent edits: no sandbox, no copy. Thronglets that write in parallel each get their own tree, e.g. a worktree apiece.

One turn or a conversation

run_thronglet creates the session and runs the first turn; send_message runs the next one in the same session, with the harness's own memory of the earlier turns. One message is one turn. Use a follow-up instead of a new run whenever the context already lives in the session: "now fix what you found", "the tests fail with this output, continue", a corrected structured result.