AI execution runtime

Experimental

Prompt-pack execution and provider connectors are functional but evolving. They are separate from the stable MCP execution boundary.

run_prompt_pack() resolves an experimental provider connector, sends one system/user prompt pair, and returns a normalized dictionary.

from aksara.ai.limits import AgentRuntimeLimits
from aksara.ai.runtime import run_prompt_pack

result = await run_prompt_pack(
    {
        "system_prompt": "Answer concisely.",
        "user_prompt": "Explain this migration plan.",
        "provider": "ollama",
        "model": "llama3",
        "max_tokens": 500,
    },
    limits=AgentRuntimeLimits(
        run_timeout_seconds=30,
        provider_timeout_seconds=20,
        token_budget=2_000,
    ),
)

The result contains ok, provider, model, response, tokens, elapsed_ms, and error. Provider and model resolution uses explicit function overrides, prompt-pack values, AI Hub configuration, then provider defaults.

The runtime does not define an AgentRuntime class, persist sessions, resume calls, or make provider quality stable. See Agent runtime for the exact boundary and AI providers for configuration.