ham-tdd

name	ham-tdd
description	Orchestrate task-agnostic, multi-agent Test-Driven Development (TDD) loops between Antigravity and Hermes.
version	1.0.0
author	Antigravity & Hermes
license	MIT
platforms	["windows","macos","linux"]
metadata	{"hermes":{"tags":["tdd","multi-agent","collaboration","orchestrator","testing"],"related_skills":["hermes-agent","install-hermes-desktop"]}}

HAM-TDD: Multi-Agent TDD Collaboration Skill

Use this skill to orchestrate or participate in a collaborative Test-Driven Development (TDD) loop between Antigravity (Google DeepMind's orchestrating assistant) and Hermes (Nous Research's system-execution agent). It enables task-agnostic execution, automated error feedback, and continuous verification loops until a target task passes.

---

1. Skill Capabilities

This skill wraps the orchestration logic in harness.py to allow:

• Dynamic Agent-to-Agent Handshakes: Initiates, resumes, and keeps track of Hermes execution sessions (--resume <session_id>).
• Continuous Test & Repair Loops: Evaluates implementations against a user-specified command (e.g. unit tests, build commands, checkers) and feeds failing outputs back to the worker session.
• Platform-Agnostic Parsing: Strips TUI framing characters, ANSI formatting, and parses exit codes to accurately assess success/failure.
• Windows CMD/PowerShell Safety: Automatically repairs quoting artifacts, escaped characters, and trailing backslashes when command-line parameters get absorbed during multi-layered shell spawns.

---

2. Command Line API

Run the harness directly from the command line in the repository workspace:

python harness.py --task "<task_description>" --verify "<verification_command>" [options]

Parameters

• --task: (Required) Text description detailing the exact task, feature, or file(s) Hermes needs to create or modify.
• --verify: Optional shell command run by the harness to verify execution status (e.g. pytest tests/, npm run build, python verify.py).
• --workspace: Target workspace directory for execution (defaults to parent directory C:\Users\admin).
• --hermes: Path to the hermes.exe CLI binary.
• --rounds: Maximum feedback iterations (default: 5).

---

3. How the Agents Interact

1. Kickoff: Antigravity (via the harness) initiates the session by passing the task instructions to Hermes.
2. Implementation: Hermes proposes files or patches them directly in the workspace.
3. Verification: The harness executes the test script.
   • If success (exit code 0): The harness reports success to Hermes, logs the session, and exits.
   • If failure (exit code non-zero): The harness captures the exact terminal error traceback, feeds it back to Hermes as the next round's query, and prompts for code repair.

---

4. Best Practices & Windows Tips

• Quoting Arguments: When passing complex nested strings in Command Prompt or PowerShell, double quotes may be absorbed. The harness has built-in regex filters to strip absorbed trailing --rounds, --workspace, or --hermes arguments from the verification block.
• ComSpec Overrides: Because Windows child processes can inherit shell environments differently, the harness explicitly overrides the child process environment ComSpec variable to target cmd.exe directly when spawning Hermes CLI instances to avoid path resolution errors.
• Persistent Logs: Complete round histories and verification output logs are saved in collab_logs/collaboration_transcript.log for inspection.

---

5. Hierarchical Corporate Delegation Pattern

To solve complex engineering objectives, agents can delegate sub-tasks within a hierarchical "corporate structure" to isolate context and scale capability:

1. Executive Board (Antigravity Orchestrator): Manages high-level requirements and serves as the quality gatekeeper (runs tests).
2. Project Manager (Lead Hermes): Coordinates the repository workspace, runs build steps, and allocates specialized worker subagents for sub-tasks.
3. Specialized Worker Subagents:
   • Research Subagent: Handles documentation lookups and web searches (keeps lead developer's context clean).
   • Feature Coder Subagent: Implements target features and modules.
   • QA Subagent: Writes test harnesses and targets edge cases to break and strengthen the software.

Recommended Skill Prompting: When giving tasks to Hermes, explicitly instruct the agent:

"Evaluate if this task requires documentation research or unit testing. If so, spawn specialized subagents to implement or research, and integrate their outputs back in your primary session."