Checkers & Aggregation

A RE-PLAN policy runs one or more checkers after a step. Each returns a verdict (an action + reason + confidence); the policy’s aggregation combines them into one decision.

Rule-based checker

Deterministic rules on the step’s error/result text — fast and free.

from mmar_carl import RuleBasedReplanCheckerConfig, ReplanAction

RuleBasedReplanCheckerConfig(
    name="rule_based",
    error_substrings=["rate limit", "timeout"],   # match against the error
    result_substrings=["I don't know"],           # match against the output
    action_on_failure=ReplanAction.RETRY_CURRENT_STEP,        # step failed
    action_on_match=ReplanAction.REPLAN_FROM_CHECKPOINT,      # a substring matched
    confidence=0.9,
)

Key fields: trigger_on_failed_step, error_substrings, result_substrings, action_on_failure, action_on_match, rollback_target_on_failure / rollback_target_on_match, feedback_on_failure / feedback_on_match.

LLM checker

Asks an LLM whether (and how) to replan — flexible, costs a call.

from mmar_carl import LLMReplanCheckerConfig

LLMReplanCheckerConfig(
    name="llm_replan",
    model="anthropic/claude-3.5-sonnet",  # optional per-checker override
    history_entries=6,                    # how much history to show the judge
    error_entries=4,
    retries=1,
)

It also takes a custom prompt_template, temperature, max_tokens, and parse_error_action (what to do if the LLM’s reply can’t be parsed).

Registered checker

Reference a checker you registered in code with context.register_replan_checker(name, checker):

from mmar_carl import RegisteredReplanCheckerConfig

RegisteredReplanCheckerConfig(name="my_custom_checker")

Aggregation

When a policy has several checkers, ReplanAggregationConfig decides the outcome:

from mmar_carl import ReplanAggregationConfig, ReplanAggregationStrategy

ReplanAggregationConfig(
    strategy=ReplanAggregationStrategy.K_OF_N,
    k=2,                              # need 2 checkers to agree
    action_selection="priority",     # or "highest_confidence"
)

Strategy	Replans when…
ANY	any checker votes to replan.
ALL	every checker votes to replan.
K_OF_N	at least k checkers vote to replan.
MANDATORY_PLUS_K_OF_REST	all mandatory_checkers agree and k of the rest do.

action_selection decides which action wins when checkers disagree — "priority" (by action severity) or "highest_confidence".

See also

• RE-PLAN overview — policy, triggers, budgets, checkpoints.