# Analyze Harness Model Protocols

Agent framework audits often miss the translation layer between internal state and provider APIs. This skill guides Claude, Codex, and Claude Code through message formats, tool calls, streaming events, and provider abstractions.

## Install

```bash
npx skillstore add dowwie/harness-model-protocol
```

## Metadata

- Status: approved
- Slug: dowwie-harness-model-protocol
- Skillstore revision: r1
- Version status: missing
- Tree hash: dee3ae522c4625409d81ed272fd0a0c91ccaddd7b92250a55897c5a815559cf1
- Author: Dowwie
- GitHub username: Dowwie
- License: MIT
- Repository: https://github.com/Dowwie/agent\_framework\_study/tree/main/.claude/skills/harness-model-protocol
- Ref: 1ffa7643651792ccb4bd3b15d924d2c97edff755
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network
- Quality score: 77
- Quality tier: bronze
- Public page: https://skillstore.pages.dev/skills/dowwie-harness-model-protocol
- Manifest: https://skillstore.pages.dev/api/skills/dowwie-harness-model-protocol/manifest

## Capabilities

- Compares OpenAI, Anthropic, Gemini, and custom message wire formats.
- Maps how tool schemas are sent to LLM APIs and parsed from responses.
- Examines streaming event handling, including partial tool call accumulation.
- Catalogs agentic primitives such as system prompts, scratchpads, interrupts, and confirmations.
- Produces a structured markdown analysis with code references and architecture implications.
- Identifies provider abstraction patterns such as thin adapters, unified clients, and gateways.

## Use Cases

- Review an agent framework protocol layer: Map how a framework converts internal messages and tool calls into provider-specific API requests.
- Compare provider adapter designs: Evaluate whether OpenAI, Anthropic, Gemini, and local model adapters use a thin or unified abstraction.
- Document streaming and tool-call behavior: Create a clear report on streaming events, partial JSON handling, tool result attribution, and human approval points.

## Prompt Templates

### Basic protocol overview

```
Use harness-model-protocol to summarize how this framework sends messages to its LLM provider and receives responses.
```

### Tool call encoding review

```
Use harness-model-protocol to trace how tool definitions are encoded for the model and how tool calls are parsed back into framework actions.
```

### Streaming behavior analysis

```
Use harness-model-protocol to analyze streaming event types, partial argument accumulation, and final tool call reconstruction in this codebase.
```

### Multi-provider architecture comparison

```
Use harness-model-protocol to compare provider adapters, message translation, tool choice support, graceful degradation, and implications for a new framework design.
```

## Limitations

- Requires access to the target framework source code to produce useful references.
- Does not execute or benchmark LLM provider calls.
- Does not replace a full security review of tool execution or credential handling.
- May need a separate codebase mapping step for large or unfamiliar repositories.

## Best Practices

- Run codebase mapping first when provider adapter files are not obvious.
- Ask for file and line references for each protocol claim in the final report.
- Review tool execution security separately from wire-format translation.

## Anti Patterns

- Do not assume OpenAI-compatible APIs behave the same across all providers.
- Do not treat streaming text deltas and tool argument deltas as the same event type.
- Do not document prompt assembly without checking how dynamic context is inserted.

## Security Audit

- Audited at: 2026-07-09T04:16:04.803\+00:00
- Summary: The static findings are false positives caused by Markdown code fences, inline protocol examples, and report-template prompts. I found no executable shell commands, no real network calls, no prompt injection attempt, and no intent to extract hidden system instructions.

## Stats

- Views: 310
- Downloads: 9
- Favorites: 0
- Popularity score: 0