pyocloop 0.2.0

TUI loop harness that orchestrates OpenCode to execute tasks from a PLAN.md file iteratively

pyocloop

A Python TUI that orchestrates OpenCode to execute tasks from a PLAN.md file iteratively, one session at a time.

pyocloop is inspired by and compatible with ocloop by @d3vr. It reimplements the same concept in pure Python using Textual, fixing TUI display issues and a path resolution bug present in the original.

How it works

1. pyocloop starts an opencode serve subprocess
2. On each iteration it creates a session, sends your loop prompt (with the plan file path injected), and waits for the session to go idle
3. OpenCode reads the plan, executes the next task, marks it [x], and optionally appends <plan-complete> when done
4. The TUI shows live progress: task counter, progress bar, current task, token count, and elapsed/average time per iteration
5. The loop stops when OpenCode writes <plan-complete> to the plan file

Requirements

• Python 3.11+
• OpenCode installed and configured (opencode on your PATH)

Installation

git clone https://github.com/rbarzic/pyocloop
cd pyocloop
pip install .

For development (editable install):

pip install -e .

Usage

ocloop COMMAND [OPTIONS]

Commands:
  run        Run the OCLoop orchestration loop
  bootstrap  Create a starter PLAN.md and .loop-prompt.md in a directory

ocloop run

ocloop run [OPTIONS]

Options:
  -m, --model TEXT   Model to use (format: providerID/modelID).
                     List available models with: opencode models
  -a, --agent TEXT   Agent to use
  --prompt PATH      Path to loop prompt file  [default: .loop-prompt.md]
  --plan PATH        Path to plan file         [default: PLAN.md]
  -p, --port INT     OpenCode server port      [default: 4096]
  -r, --run          Start iterations immediately (no keypress needed)
  -d, --debug        Debug mode (skip plan file validation)
  --verbose          Log every SSE event in the activity panel
  --log PATH         Write all log entries to a file
  --help             Show this message and exit

ocloop bootstrap

ocloop bootstrap [DIRECTORY] [OPTIONS]

Arguments:
  DIRECTORY          Directory to initialise  [default: current directory]

Options:
  -f, --force        Overwrite existing files
  --help             Show this message and exit

Key bindings

Key	Action
S	Start the loop
Space	Pause / Resume
R	Retry after an error
Q	Quit (aborts current session)

Example

The examples/ directory contains a self-contained demo: answering EU capitals quiz questions.

File layout:

your-project/
├── PLAN.md              # task list (or use --plan to point elsewhere)
├── .loop-prompt.md      # instructions for OpenCode (or use --prompt)
└── ...

examples/PLAN.md — task list:

# EU Capitals Quiz Plan

## Backlog

### Phase 1: Western Europe

- [ ] **1** What is the capital of Austria?
- [ ] **2** What is the capital of Belgium?
- [ ] **3** What is the capital of France?
...

examples/loop-prompt.md — loop prompt (note the {{PLAN_FILE}} placeholder):

Execute the next task from {{PLAN_FILE}}.

Before starting:
1. Read {{PLAN_FILE}} fully

Task selection:
- Pick the FIRST uncompleted task
- Mark it [x] after completion

Completion check:
- If all tasks are [x] or [BLOCKED], append:
  <plan-complete>SUMMARY</plan-complete>

Bootstrap and run:

# Create starter files in a new directory
ocloop bootstrap ./myproject

# Edit the generated files, then run from inside the directory
# (--plan defaults to PLAN.md and --prompt defaults to .loop-prompt.md)
cd myproject
ocloop run --model openai/gpt-4.5 --run

Or run from outside the directory by specifying paths explicitly:

ocloop run \
  --model openai/gpt-4.5 \
  --plan ./myproject/PLAN.md \
  --prompt ./myproject/.loop-prompt.md \
  --run

Plan file format

- [ ] Pending task
- [x] Completed task
- [MANUAL] Task requiring human intervention (skipped by the loop)
- [BLOCKED: reason] Task that could not be completed

The loop ends when OpenCode appends this tag to the plan file:

<plan-complete>Summary of what was done</plan-complete>

Loop prompt tips

• Use {{PLAN_FILE}} as the placeholder — pyocloop replaces it with the absolute path at runtime
• Instruct OpenCode to mark tasks [x] after completion and append <plan-complete> when all done
• Keep prompts focused: one task per session works best for reliable progress tracking

Available models

Run opencode models [provider] to list models for a specific provider. Below are the models available at time of writing (your installation may differ):

Provider	--model value	Notes
OpenAI	openai/gpt-5.5	Latest flagship
OpenAI	openai/gpt-5.5-pro	Pro tier
OpenAI	openai/gpt-5.5-fast	Faster/cheaper
OpenAI	openai/gpt-5.4-mini	Lightweight
z.ai (coding)	zai-coding-plan/glm-5.1	Coding-optimised
z.ai (coding)	zai-coding-plan/glm-5-turbo	Fast coding model
z.ai (coding)	zai-coding-plan/glm-4.7	Previous gen
z.ai (limited ctx)	zai-limited-context/glm-5.1	Smaller context window
DeepSeek	opencode/deepseek-v4-flash-free	Free via opencode
GitLab Duo	gitlab/duo-chat-sonnet-4-6	Anthropic Sonnet via GitLab
GitLab Duo	gitlab/duo-chat-opus-4-7	Anthropic Opus via GitLab
GitLab Duo	gitlab/duo-chat-gpt-5-4	GPT-5.4 via GitLab

Anthropic models can be used directly if you configure the anthropic provider in opencode.

Architecture

pyocloop does not call any LLM directly. It delegates all AI work to the opencode binary:

ocloop (Textual TUI)
  └─ opencode serve  (subprocess)
       ├─ POST /session          create session
       ├─ POST /session/{id}/prompt_async  send prompt
       ├─ GET  /event?directory=…          SSE stream (progress events)
       └─ GET  /config                     detect active model