Classifies mismatches between an agent's hypothesis and observations to provide structured feedback for autonomous error correction.
// Source: https://github.com/vinhnx/vtcode | License: Apache-2.0 | Copyright 2025 Vinh Nguyen // Provided via justclaw.it.com; keep this notice with the code. //! Hypothesis-loop mismatch classification for the agent runloop. //! //! desirable loop: hypothesis → observation → mismatch → inspect evidence → //! revise hypothesis. This module is the pure, testable core of that loop: //! it names the mismatch, but never blocks, retries, or performs I/O. The //! binary's failure diagnosis and loop notices call into it to frame //! model-facing feedback; the evaluator provides a count-based heuristic. //! //! Dimension key for [`MismatchEvidence`] (index → field → meaning): //! //! | field | meaning when set | //! | --------------------- ... (truncated -- full source via MCP)
See the full source, get the GitHub permalink, and search for more like it.
Get a free API key