Large Language Models and agentic systems are increasingly deployed as
collaborative components in software development workflows. In practice,
however, system failures are often not caused by model limitations but by
the absence of explicit governance and shared operational memory.
Project state becomes fragmented across version control, chat logs,
documentation, and ad-hoc trackers. As a result, LLM-based agents operate
with partial or stale information, or worse, overwrite human-owned semantics
such as priorities, intent, and roadmap decisions.
This paper introduces DevTracker, a lightweight governance layer that
formalizes the separation between human-owned semantics and machine-written
evidence. DevTracker treats a tracker as a governance contract: humans retain
control over meaning (purpose, priority, intent), while automation is limited
to auditable evidence (timestamps, lifecycle signals, quality metrics).
DevTracker audits a Git repository, executes a quality suite, proposes
reviewable updates, and applies changes only under explicit policy constraints.
All updates are attributable, reversible, and recorded in an append-only
journal. Outputs are designed to be both human-readable and machine-consumable,
enabling safe integration with dashboards and LLM tool-calling interfaces.
We argue that explicit governance and externalized operational memory are
necessary conditions for scalable human–LLM collaboration. DevTracker provides
a minimal, auditable reference implementation of this principle.
Project state becomes fragmented across version control, chat logs, documentation, and ad-hoc trackers. As a result, LLM-based agents operate with partial or stale information, or worse, overwrite human-owned semantics such as priorities, intent, and roadmap decisions.
This paper introduces DevTracker, a lightweight governance layer that formalizes the separation between human-owned semantics and machine-written evidence. DevTracker treats a tracker as a governance contract: humans retain control over meaning (purpose, priority, intent), while automation is limited to auditable evidence (timestamps, lifecycle signals, quality metrics).
DevTracker audits a Git repository, executes a quality suite, proposes reviewable updates, and applies changes only under explicit policy constraints. All updates are attributable, reversible, and recorded in an append-only journal. Outputs are designed to be both human-readable and machine-consumable, enabling safe integration with dashboards and LLM tool-calling interfaces.
We argue that explicit governance and externalized operational memory are necessary conditions for scalable human–LLM collaboration. DevTracker provides a minimal, auditable reference implementation of this principle.