thinkflowlab/system1-agents/s1a

s1a

S1A runs openJiuwen agents with a System 1 decision model in the model slot: TypeSafe Jev over HTTP by default, Laya or Cua-S1 in process. About 400 ms per request, one selection over the options the page or the game enumerates, a probability per option. The command is s1a, run from a checkout of https://github.com/ThinkFlowLab/system1-agents: uv run --project <checkout> s1a .... Inside the Claude Code plugin the checkout is ${CLAUDE_PLUGIN_ROOT}. python -m s1a from the checkout is the same command.

When to delegate

• A web task that is a sequence of selections over visible controls: search forms, filters, date pickers, result

lists, quizzes. The values to type come from the task text.

• A game or environment that enumerates its legal moves each step and reports a score.
• One selection over options you already hold, where a probability per option helps: routing, ranking, gating.

Not for: a plain page fetch (use curl), arithmetic, constraint puzzles, search over move trees, free-text generation, or a task that needs a value the page never shows.

Commands

s1a list
s1a run flights --model jev --goal "<site URL first, the values to enter, the stop condition>"
s1a run <agent> --model jev --rethink off --episodes 3
s1a decide --state '<JSON object>' --option key="what it means" --option other="what it means" --rules "<facts>"
s1a decide --state '{"title": "Charged twice", "description": "I was charged twice for order 4411 and I want the second charge refunded.", "order_status": "delivered"}' \
  --option logistics="delivery tracking, delivery progress or delivery problems" --option payment="charges, failed payments or duplicate payments" \
  --option returns="requests for returns, exchanges or refunds" --option account="login or account access problems" \
  --option human="insufficient information, several independent requests, or an explicit request for a human" \
  --rules "Route the current explicit request to exactly one queue. A payment problem with an explicit request for a refund belongs to returns. If no unique queue fits, choose human."
s1a run ticket_router --model jev --rethink off --episodes 1      # the shipped router: 30 labelled tickets, five queues, scores the correct routes

Every run prints one JSON object on stdout and nothing else there. The harness logs go to files under the checkout's runs/logs. A browser agent's object has ok, final, error and usage with both models; final is the answer. screenshot is the path of final.png, a PNG of the page the task ended on, or null when no page was open; a failed screenshot leaves it null too and names the failure in screenshot_error. With --model jev it also has report and, when present (absent on a timeout), status and terminal: terminal.url and terminal.title are where the answer was read. With --model llm it has browser_result, the subagent's own verdict; status, report and terminal are absent there. A task takes seconds to a few minutes; --timeout sets the wall clock (the agent's own default, 180 s for flights) and --headed shows the browser. A tool agent's object is the series summary with job_dir, the job folder it wrote. decide prints choice, probabilities, confidence and ms. Exit codes: 0 for a finished run, including one whose JSON has ok: false and an error; 1 for a run, key or model error, one line on stderr; 2 for a usage error (an unknown agent, bad flags, a malformed --state, --option or @file). --model jev is TypeSafe Jev; --model laya is Laya, an open-weight System 1 decision model that runs in process after uv sync --extra laya, with the same outputs and no Jev key; --model cua is Cua-S1 Nano, a small option scorer in process after uv sync --extra cua, a baseline; tool agents also take llm, random and rule.

Keys

TYPESAFE_API_KEY for Jev, or OPENROUTER_API_KEY for the proxy; OPENAI_API_KEY or LLM_API_KEY, the base URL and MODEL_NAME for the chat model that types values and writes the answer. Export them in the host's environment. The command also reads the .env of its checkout. Exported variables win over the file.