Liteflow: A C program for running YAML-defined DAGs with LLM-driven runtime edits.

Liteflow occupies an interesting niche at the intersection of DevOps automation and generative AI. It is fundamentally a specialized workflow runtime, rather than an LLM orchestration agent. This distinction is critical: it doesn't aim to generate the full workflow itself, but rather to provide a highly constrained, auditable mechanism for an external model to correct or augment a workflow while it is in motion. The implementation as a single, straightforward C application enhances its appeal to developers who prioritize minimal overhead and maximum control. The technical brilliance of Liteflow rests in its mutation grammar. By limiting the LLM's actionable verbs to `RETRY`, `PATCH`, `INSERT_BEFORE`, and `ABORT`, the developers have successfully mitigated the 'hallucination risk' inherent in passing free-form natural language output to critical infrastructure logic. These verbs correspond to tangible changes—re-running a step, modifying a field, injecting a remediation task, or failing gracefully. This rigorous control means the LLM's power is channeled through a strictly defined, machine-readable interface, vastly improving reliability for critical paths like CI/CD pipelines. Furthermore, the system's commitment to auditability is exceptional. Every change—whether an original task failure, a planner invocation, or a successful mutation application—is logged into a structured, replayable event log (`events.jsonl`). This provides the necessary forensic trail for complex, semi-autonomous systems. Developers can not only see *what* failed but *why* the LLM decided to intervene, and precisely *how* the graph was altered to attempt recovery. This shifts the focus from 'did it run?' to 'can we prove the path it took?' For the intended audience—developers and sysadmins building critical pipelines—Liteflow’s architecture is a strength. It avoids the over-engineering trap of abstract, high-level workflow engines, instead offering a lean, fast, and deterministic CLI experience. While the limitation to YAML subsets and file-based state passing keeps the scope intentionally small (avoiding the pitfalls of general graph synthesis), it achieves powerful results for the defined use cases: error handling and iterative refinement within a known workflow structure. It is a pragmatic tool for integrating the variability of LLMs into the rigidity demanded by enterprise pipelines.