Harness Engineering Lifecycle

Evolve a proven repo-local contract into a generalized, high-performance, declarative harness system, then safely dogfood those changes across a producer-consumer repository boundary. Start here only after the target repo has a steward.yaml contract and at least an H2 smoke proof; use mcp-harness-repo-maintainer before that.

When to use

• Evolving custom validation scripts into a central linter engine.
• Extending steward CLI features for large polyrepos.
• Testing a local CLI/harness build against a sibling repository to catch path-resolution crashes or integration friction.
• Turning repeated local benchmark/probe findings into a reusable harness feature.

When not to use

• Initial steward.yaml adoption or first quick probe in one repo — use mcp-harness-repo-maintainer.
• Skill creation, registry updates, or marketplace packaging — use skill-authoring-lifecycle or plugin-marketplace-setup.
• Product-specific diagnostics before a cold-start contract exists — first add a bounded action, probe, and scenario in the target repo.

Part 1: The Cascading Agent Surface (Architecture & Generalization)

When engineering a harness beyond one local contract, follow the Cascading Agent Surface guidelines. A strict separation of tools vs skills can fail once repo workflows span multiple packages, adapters, or consumers. Instead, link them using domain-agnostic abstractions:

1. Layer 0 (The Embedded Agent Surface): The target application or engine must natively expose its internal state via explicit hooks (e.g., RPC or memory probes). Do not rely on brittle UI scraping or black-box testing.
2. Layer 1 (The Protocol Adapter): Build generalized MCP servers or protocol adapters to connect to Layer 0. These tools provide raw visibility and actuation (e.g., taking screenshots, reading memory) but must contain NO business logic.
3. Layer 2 (The Orchestrator): Build specialized harness CLIs that use their own automation/scripting to chain multiple Layer 1 actions together. The Orchestrator's primary job is the Fast Feedback Loop: it must emit structured, diagnostic JSON to pinpoint exactly what broke across boundaries.
4. Layer 3 (The AI Wrapper): Following the agentskills.io spec, Skills can and should contain thin-wrapper tools (scripts). The Skill acts as the AI's brain: it teaches the AI how to trigger Layer 2, interpret its complex JSON heuristics, and safely execute domain-specific recovery tools.

Generalization Principles

1. Decouple transport from engine: Command interfaces (CLI) and JSON-RPC (MCP) are thin wrappers. All logic lives in a reusable core package.
2. Declarative configuration: Use a configuration file (like steward.yaml) to specify:
   • Branding: Name and description of the harness.
   • Actions: Typed, bounded repo-local commands with effects, limits, outputs, and evidence policy.
   • Documentation lattice: Maps of labels to specific doc files.
3. Pruned traversals: Never perform recursive file-walking (list(recursive: true)) without pruning standard build/VCS folders early.
4. Generalized checks: Do not hardcode linters. Use parameterized declarative engines (e.g. disallowed-substrings).

Part 2: Cross-Repo Remediation (Testing)

When you make changes to the harness (Producer), you must validate it against dependent downstream projects (Consumer).

1. Establish Sibling Baseline: Locate the consumer repository (e.g. ../<consumer-repo-name>). Run its tests to ensure it starts green. Do not change it yet.
2. Deploy Local Producer: Build and install the local development version of your harness.
3. Capture Adoption Friction: Run the consumer's validation suite using the local producer build. Document path crashes or fragile rules in a scratchpad.
4. Core Remediation: Fix the root causes in the Producer codebase. Generalize the fix; do not write consumer-specific hacks.
5. Consumer Configuration Remediation: Update the Consumer repository to leverage the new generalized features (e.g. updating its steward.yaml).
6. Dual Verification: Verify that the consumer suite passes, and the producer suite passes.
7. Durable Knowledge: Extract any learnings to ADRs or FAQs using the repository-governance-lifecycle skill.

Detour stop rule: If installing, activating, or wiring the local producer build fails twice, stop restoration, record the friction, use the consumer's native gate or portable fallback when possible, and return to the original acceptance check. Do not promote a generalized harness capability from that same detour; capture an observation or unknown case first.

Producer/consumer proof artifacts

Before promoting a generalized harness change, capture the proof from both sides.

Side	Required artifact	Purpose
Producer	Changed schema/action/rule with tests or validator output	Shows the generalized feature works in the source harness.
Consumer	Updated steward.yaml or scenario manifest	Shows adoption does not require local-path magic.
Consumer	Benchmark summary or truthful durability_blocked output	Shows a fresh agent can see what is proven and what is blocked.
Both	Validation commands and versions	Makes failures reproducible across repos.
Docs	ADR/FAQ update when the boundary changed	Keeps future agents from rediscovering the same split.

Do not promote a local fix into the generalized harness when only one consumer has a private path workaround. Keep it local until the contract, schema, or rule has a reusable shape and at least one held-out or future-agent repeat proves it transfers.

Install

npx skills add arenukvern/skill_steward --skill harness-engineering-lifecycle

Sources

See references/sources.md.