zhanlin-liu-EPOCH

EPOCH

Multi-round optimization framework that runs iterative optimize-evaluate loops: investigate failures, implement a fix, evaluate the result, accept or reject with evidence, repeat. Works with both Claude Code and OpenAI Codex.

Installation

Claude Code

claude plugin marketplace add zhanlin-liu/EPOCH
claude plugin install epoch@zhanlin-liu-EPOCH

Then use /epoch in any project.

OpenAI Codex

$skill-installer install https://github.com/zhanlin-liu/EPOCH/tree/main/skills/epoch

Restart Codex after installation, then use /epoch.

From Source

git clone https://github.com/zhanlin-liu/EPOCH.git
cd EPOCH
claude  # or codex

Quick Start

Option 1: Guided Setup (No Config)

/epoch

EPOCH will interview you (3–5 questions) and scaffold the project automatically.

Option 2: Existing Config

/epoch projects/my_project_run.yaml

EPOCH reads the config, sets up the environment, and starts the optimization loop.

Supported Task Types

Task Type	What It Optimizes	Example
prompt_tune	LLM prompt text	Sentiment classification with GPT-4.1-nano
finetune	ML hyperparameters	MobileNetV2 learning rate and optimizer
rule_based	Rule conditions and thresholds	Iris species classifier
code_improvement	Algorithm correctness and performance	Fibonacci calculator speed

How It Works

Round 1: Baseline
  └─ Scaffold project → Run evaluation → Save baseline metrics → Open PR

Round 2+: Optimization Loop
  └─ Investigate failures (TRAIN only)
     └─ Implement fix
        └─ Evaluate (EVAL only)
           ├─ ACCEPT → Merge PR → Next round
           └─ REJECT → Retry or next round

Each round produces:

• A git branch (epoch/<slug>/<run-id>/round-N)
• A pull request with metrics tables and evidence
• Structured JSON artifacts (baseline_metrics.json, delta_round_N.json)

Plugin Structure

.claude-plugin/
└── plugin.json
skills/
└── epoch/
    ├── SKILL.md                    # Entry point — dispatches by task_type
    ├── agents/                     # Role definitions
    │   ├── orchestrator.md
    │   ├── investigator.md
    │   ├── executor.md
    │   ├── baseline_executor.md
    │   ├── seed_planner.md
    │   └── reviewer.md
    └── references/                 # Task-type workflows
        ├── prompt_tune.md
        ├── finetune.md
        ├── rule_based.md
        ├── code_improvement.md
        ├── create_project.md
        └── create_skill.md

Writing a Config

Create projects/<slug>_run.yaml. See projects/ for example configs.

ML Tasks (prompt_tune, finetune, rule_based)

project:
  name: "My Classifier"
  slug: my_classifier

run:
  id: null                    # auto-generated at runtime
  goal: "Optimize classification accuracy"
  task_type: rule_based       # or prompt_tune, finetune
  max_rounds: 5
  max_retries_per_round: 1

env:
  manager: uv
  path: "projects/my_classifier"

evaluation:
  primary_metric: accuracy
  min_delta: 0.01             # minimum improvement to accept a round
  deterministic: true
  train_cmd: "python projects/my_classifier/evaluate.py train"
  eval_cmd: "python projects/my_classifier/evaluate.py eval"

git:
  push_to_remote: true
  create_prs: true
  target_branch: develop

Code Improvement

project:
  name: "My Algorithm"
  slug: my_algo

run:
  task_type: code_improvement
  max_rounds: 5
  max_retries_per_round: 1

env:
  manager: uv
  path: "projects/my_algo"

evaluation:
  primary_metric: execution_time
  min_delta: 0.05             # 5% minimum speedup per round
  deterministic: true
  cmd: "python projects/my_algo/main.py"
  test_cmd: "pytest projects/my_algo/tests/"

git:
  push_to_remote: true
  create_prs: true
  target_branch: develop

Hyperparameter Tuning (finetune)

Add ml and tune sections to control the search space:

ml:
  base_model: mobilenetv2
  framework: pytorch
  seed: 42
  max_train_epochs: 3

tune:
  optimizer: [adam, adamw, sgd]       # discrete options
  learning_rate: [0.0001, 0.01]      # continuous range [min, max]

Prompt Tuning (prompt_tune)

Add llm and tune sections:

llm:
  model: "gpt-4.1-nano"
  async: true
  concurrency: 12

tune:
  strategy: ["few shots", "chain of thought"]

Key Concepts

Train/Eval Separation

For ML tasks, EPOCH enforces strict separation:

• Investigation: Analyzes failures on TRAIN split only
• Evaluation: Accepts/rejects based on EVAL split only
• This prevents overfitting to the evaluation set

For code_improvement: all tests are visible (no split needed).

Acceptance Criteria