RE-PLAN Overview

RE-PLAN lets a chain react at runtime: after a step, one or more checkers evaluate progress and can decide to retry the step, roll back to an earlier checkpoint and re-run, restart, or fail — all bounded by a budget so it can’t loop forever.

Mark checkpoints

Roll-back targets are steps you mark as checkpoints:

LLMStepDescription(
    number=2,
    title="Risk assessment",
    aim="Assess the risks.",
    checkpoint=True,
    checkpoint_name="risk_phase",
)

Attach a policy

Configure RE-PLAN with a ReplanPolicy on the chain. Here a rule-based checker retries the current step whenever an error contains “rate limit”:

from mmar_carl import (
    ReasoningChain, ReplanPolicy, RuleBasedReplanCheckerConfig, ReplanAction,
)

policy = ReplanPolicy(
    checkers=[
        RuleBasedReplanCheckerConfig(
            error_substrings=["rate limit"],
            action_on_match=ReplanAction.RETRY_CURRENT_STEP,
        ),
    ],
)

chain = ReasoningChain(steps=steps, replan_policy=policy)

Toggle it per step with replan_enabled=True/False (defaults to the chain policy).

Actions

A checker’s verdict requests one of these ReplanAction values:

Action	Effect
`CONTINUE`	Proceed normally.
`RETRY_CURRENT_STEP`	Re-run the step that just finished.
`REPLAN_FROM_CHECKPOINT`	Roll back to a checkpoint and re-run from there.
`RESTART_CHAIN`	Start the whole chain over.
`FAIL`	Fail the chain immediately.

Policy structure

ReplanPolicy bundles:

Field	Type	Purpose
`enabled`	`bool`	Master switch (default `True`).
`checkers`	`list[...CheckerConfig]`	The checkers that vote.
`aggregation`	`ReplanAggregationConfig`	How votes combine (see checkers).
`trigger`	`ReplanTriggerConfig`	When to evaluate (after each step / only failures / only checkpoints / specific steps).
`budgets`	`ReplanBudgetConfig`	Guards against infinite replanning.

Budget guards

ReplanBudgetConfig prevents runaway loops:

Field	Default	Purpose
`max_replans_per_chain`	`3`	Total replan actions per run.
`max_replans_per_step`	`2`	Replans attributable to one step.
`max_visits_per_checkpoint`	—	Cap re-entries of a checkpoint.
`fail_on_budget_exhaustion`	—	Fail (vs continue) when the budget runs out.

Example

The repo’s deterministic RE-PLAN example (mock client, no API key) runs a 3-step memo chain. Step 2 emits NEEDS_REPLAN on its first attempt; a rule-based checker matches that substring, retries the step with feedback, and the second attempt succeeds.

from mmar_carl import ReplanPolicy, RuleBasedReplanCheckerConfig, ReplanAction

policy = ReplanPolicy(
    enabled=True,
    checkers=[
        RuleBasedReplanCheckerConfig(
            name="risk_quality_guard",
            result_substrings=["NEEDS_REPLAN"],
            action_on_match=ReplanAction.RETRY_CURRENT_STEP,
            feedback_on_match=["Add explicit mitigation ownership and concrete risks."],
        )
    ],
)

chain = ReasoningChain(steps=steps, max_workers=1, replan_policy=policy)
result = chain.execute(context)

for event in result.replan_events:
    print(event.step_number, event.final_action, event.rollback_target)

result.replan_events records each RE-PLAN decision — the step, the chosen action, the checker vote tally (event.aggregation), and any rollback target.