verification-slots

name	verification-slots
description	This skill should be used when the user asks about "verification slots", "integrating a linter", "adding a deterministic tool", "harness-enforcer", "constraint enforcement interface", "wrapping a tool", or needs the technical reference for how deterministic and agent-based checks work in the harness framework.

Verification Slots

The verification slot is the harness framework's core technical abstraction. Every constraint — whether backed by a linter, a formatter, a structural test, or an LLM agent — is checked through the same interface. This uniformity means the rest of the system (hooks, CI, commands) does not care how a constraint is verified.

The Contract

Input:

• Constraint definition (rule, enforcement type, tool)
• Scope (commit, pr, weekly, manual)
• File set (changed files, or all files)

Output:

• Result: pass or fail
• Findings: list of {file, line, message} (empty on pass)

How the Enforcer Decides

The harness-enforcer agent reads each constraint from HARNESS.md and dispatches verification based on the enforcement field:

Enforcement	What happens
deterministic	Execute the tool command, interpret exit code
agent	Read the rule text, review files, produce findings
deterministic + agent	Run both, merge findings
unverified	Skip — log as unchecked

Deterministic Tool Integration

To fill a verification slot with a deterministic tool:

1. Identify a tool that checks the constraint (linter, formatter, scanner, custom script)
2. Verify it runs locally and produces usable output
3. Update the constraint in HARNESS.md: set enforcement to deterministic and tool to the exact command
4. Run /harness-audit to confirm the tool works in the harness context

For detailed integration patterns for common tools (ESLint, Prettier, gitleaks, ArchUnit, custom scripts), consult references/tool-integration.md.

Agent-Based Verification

When no deterministic tool exists, the enforcer reads the constraint's prose rule and reviews code against it. The output format is identical to deterministic verification — pass/fail with file:line findings.

Agent-based checks are:

• Flexible — work with any constraint that can be described in prose
• Non-deterministic — may give different results on identical input
• Token-intensive — each check costs LLM inference
• Nuance-aware — can evaluate intent and context, not just syntax

Mixed Enforcement

Some constraints benefit from both levels. For example, a linter checks that doc comments exist (deterministic presence check) while the agent reviews that comments explain reasoning rather than restating signatures (quality check).

Use deterministic + agent in the enforcement field and list both tools.

The Fan-out Slot

The fan-out slot is a first-class agent-backed verification slot for large harnesses (Claude Code runtime only). Instead of one context checking every constraint — where the enforcer tends to tire and report "all constraints checked" after checking only some — the fan-out slot spawns one verifier subagent per rule, each with a single constraint and a clean context window, then passes every candidate violation through a skeptic persona that tries to refute it before it reaches the report.

A synthesis barrier waits for all N verifiers to return and reconciles their results into the same uniform contract every slot produces: pass/fail with {file, line, message} findings. Because the report cannot form until every verifier has returned, the count of results equals the count of enforceable constraints — no rule is silently dropped. The slot is read-only and propose-only (INV-1): it reports findings and never writes a durable artefact.

The harness-enforcer agent elects this slot automatically when the enforceable-constraint count exceeds its threshold (default 8); below it, the single-context path is used unchanged. The slot adapts the shipped enforcer-fanout.workflow.js template from the dynamic-workflows skill.

Additional Resources

Reference Files

• references/tool-integration.md — Step-by-step integration patterns for linters, formatters, type checkers, secret scanners, structural test frameworks, and custom scripts