---

name: simulation-workflows description: "Run OASIS Reddit, Twitter, custom-platform, manual-action, and LLM-action simulations with safe environment lifecycle and model-backend guidance." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Simulation Workflows

Use this sub-skill when a task asks a future agent to run or adapt an OASIS simulation: Reddit-style or Twitter-style runs, custom Platform setup, manual interventions, selected LLMAction steps, concurrency controls, or a small no-credential smoke test.

Start Here

• For the current async lifecycle and env.step action contract, read environment-lifecycle.md.
• For Reddit, Twitter, hybrid manual/LLM, and custom-platform recipes, read workflows.md.
• For CAMEL/OpenAI/VLLM/local-server/DeepSeek model setup, credentials, and cost controls, read model-backends.md.
• For common failures, stale-doc traps, database locks, and provider recovery, read troubleshooting.md.
• For a tiny no-LLM check, run or adapt scripts/oasis_manual_smoke.py.

Core Runtime Contract

Use public imports from oasis where possible:

from oasis import (
    make, DefaultPlatformType, ActionType, ManualAction, LLMAction,
    AgentGraph, SocialAgent, UserInfo,
)

The active environment lifecycle is:

env = make(agent_graph=agent_graph,
           platform=DefaultPlatformType.REDDIT,
           database_path=db_path,
           semaphore=8)
await env.reset()
await env.step(actions)
await env.close()

actions must be keyed by SocialAgent objects, not integer IDs:

actions: dict[SocialAgent, ManualAction | LLMAction | list[ManualAction | LLMAction]]

For manual actions, use the source-backed dataclass fields ManualAction(action_type=..., action_args={...}). Some public docs still show older EnvAction/SingleAction or action/args spellings; those names are not the current package contract.

Boundary Routing

• Agent profile schema validation, UserInfo templates, and AgentGraph construction details belong to the agent-profiles sibling sub-skill.
• Detailed action argument tables, platform/recsys internals, SQLite schema, and DB trace analysis belong to platform-actions.
• Legacy paper experiment YAMLs, large-scale runs, and post-simulation analysis belong to experiments-analysis.

Do not run real LLMAction steps unless provider credentials, a model with function/tool-calling support, concurrency limits, and an explicit token/cost budget are available.