l123-harness

In AI-assisted development, agents routinely suffer from context bloat, uncontrolled reasoning drift, and an inability to accumulate knowledge across sessions. l123-harness packages a battle-tested methodology into a plugin so the host kernel stays lean while memory gating, event logging, and TDD execution are mounted as independent, swappable components. The host is the kernel; the plugin only provides capability components. Project knowledge is not baked in—it accumulates in the project itself after installation.

The core is a three-tier memory architecture: L1 (CLAUDE.md) holds stable principles, priorities, safety baselines, and a routing table as a small ever-present prefix; L2 (INDEX.md) provides fast symptom-to-rule-to-pointer indexing and is auto-injected at SessionStart; L3 (.agents/) stores knowledge, pitfalls, runbooks, current state, an immutable event log, and a weekly review inbox. Combined with gate approval for protected files (explain, obtain explicit approval, touch a gate token, then retry), zero-judgement immutable event records, periodic knowledge distillation, and TDD-driven stopping conditions, every agent action is traceable and bounded. Knowledge promotion follows a strict path—scratch into memory, weekly review into the inbox, and escalation to knowledge or L2 always requires explicit user approval, preventing silent scope creep.

The plugin ships for both Claude Code and DeepSeek Harness simultaneously, reusing the same hooks.json and four methodology skills (init-harness, event-review, tdd, write-a-skill) across both hosts. On mount it auto-registers skills into the context and bridges the official DSH hooks layer so the single hooks.json runs natively on either platform. It suits independent developers and small teams who maintain projects long-term and want their agents to behave in a traceable, progressively improving way.