Planner

Experimental

Aksara 0.7.0 does not define or export a Planner class. The plan schemas and deterministic handlers described here are functional but evolving.

The real planning surface is AiPlan plus AiPlanStep. External code creates the plan. Aksara validates its shape and can execute known step handlers.

from aksara.ai.planner import AiPlan, AiPlanStep, validate_plan

plan = AiPlan(
    intent="Inspect application health",
    steps=[
        AiPlanStep(
            id="health",
            type="run_health_check",
            description="Check the running application",
        )
    ],
)

assert validate_plan(plan) == []

To execute this experimental plan against an application:

from aksara.ai.limits import AgentRuntimeLimits
from aksara.ai.planner import execute_plan

result = await execute_plan(
    app,
    plan,
    dry_run=True,
    limits=AgentRuntimeLimits(max_steps=5, run_timeout_seconds=30),
)
print(result.success)

Import execute_plan from aksara.ai.planner for this API. The top-level aksara.ai.execute_plan name belongs to the separate experimental investigation orchestrator and accepts a different plan type.

What the planner does

• validates plan step IDs, dependencies, and known step types;
• rejects cycles;
• sorts dependencies deterministically;
• runs registered handlers with step and overall time limits;
• stops after the first failed step; and
• supports preview behavior where the underlying handler implements it.

Some handlers can inspect an application or produce deterministic artifacts. Code generation and patch application are experimental, and generated test steps still contain incomplete behavior. Review every preview before applying changes.

The planner does not call an LLM, choose goals, persist a session, resume after restart, or provide a stable autonomous workflow contract.