Harness Engineering

In this plugin, constraints live in HARNESS.md's Constraints section and are enforced at three timescales: advisory at edit time (hooks), strict at merge time (CI), and investigative on schedule (audit).

Garbage Collection

Periodic checks that fight entropy — the slow drift that neither real-time hooks nor PR gates catch. Documentation goes stale, conventions erode, dead code accumulates, dependencies fall behind. Garbage collection agents run on a schedule to find and fix (or flag) these issues.

In this plugin, GC rules live in HARNESS.md's Garbage Collection section and are run by the harness-gc agent.

The Living Harness

The central design principle of this plugin is that the harness is a living document — HARNESS.md — that generates its own enforcement. The document declares what should be true; the plugin's agents, hooks, and CI check whether it is true; the auditor updates the document's Status section to reflect reality.

This creates a self-referential feedback loop: the harness is harnessed by its own document. When the Status section shows drift between declared and actual enforcement, the team knows where to invest next.

Progressive Hardening

Constraints follow a promotion ladder:

1. Unverified — declared intent, no automation yet
2. Agent — LLM-based review against the constraint's prose rule
3. Deterministic — tool-backed enforcement (linter, formatter, test)

Start by declaring what should be true. Automate when ready. The harness improves over time without restructuring.

Testing the Harness Itself

The harness teaches TDD for code. But the harness's own artifacts — skills, conventions, CLAUDE.md directives — are specifications that produce agent behaviour. A skill without behavioural tests is an unverified claim.

Test-Driven Agentic Behaviours (TDAB, after Antony Marcano, 2026) applies TDD to guidance files: write a test describing desired agent behaviour, run the agent, observe the gap, modify the guidance, verify the behaviour. Red-green-refactor for skills.

On the promotion ladder, a skill without behavioural tests is unverified. A skill with passing tests is agent-verified. If you would not ship code without tests, do not ship skills without them either.

Implementation Patterns

Beneath the three components lies a layer of mechanical patterns that make the harness work in practice. Six patterns from production agent systems:

• Tiered Memory — always-loaded index, on-demand topic files, searchable archive
• Dream Consolidation — periodic pruning and reorganisation of compound learning memory
• Progressive Context Compaction — compression intensity increases with conversation age
• Progressive Tool Expansion — start agents with minimal tools, expand on demand
• Command Risk Classification — allow/ask/deny per tool, gated by reversibility and blast radius
• Single-Purpose Tool Design — typed inputs, constrained scope, individual permission surfaces

These are documented in detail in the framework (Theme 10, Appendices H, I, J).

Enforcement Timing

Three concentric feedback loops:

Plugin Components

For detailed guidance on each component, consult the relevant skill: context-engineering, constraint-design, garbage-collection, verification-slots.

Additional Resources

Reference Files

• references/boeckeler-summary.md — Full summary of the article, the three components, the four hypotheses, and related work