afkcode - LLM-Powered Checklist Management & Autonomous Development Loop

A Rust port and enhancement of codex_loop.py that provides a Swiss Army knife for managing project checklists and running autonomous development loops with LLM CLI tools.

Features

• Autonomous Development Loop: Worker-only loop by default, with optional controller/worker alternation
• Parallel LLM Execution: Run multiple LLM instances simultaneously with staggered warmup delays to maximize throughput
• Gimme Mode: Automatic work item checkout from AGENTS.md files - LLMs get assigned specific tasks to prevent conflicts
• Checklist Generation: Create checklists from high-level prompts using LLM
• Checklist Management: Add, remove, and update checklist items with or without LLM assistance
• Standing Orders: Built-in project invariants ensure consistent LLM behavior
• Standing Orders Audit: One-time alignment keeps in-repo Standing Orders synced with the core rules
• Multi-LLM Support: Automatic fallback between Gemini, Codex CLI, Claude Code, and Warp Agent API on rate limits
• Model Selection: Specify custom models for each LLM tool (e.g., gemini-2.5-pro, o3, opus)
• Custom AGENTS.md: Separate file for LLM instructions and Standing Orders (see AGENTS_GUIDE.md)
• Smart Rate Limit Handling: Automatically switches to backup LLM tool when quota exhausted with temporary 5-minute squelching
• Completion Token Verification: LLM confirms intentional completion to prevent accidental loop exits
• Output Logging: Automatically streams all LLM output and console messages to a configurable log file during run mode
• Failed LLM Recovery: Automatically restores work items if an LLM subprocess crashes or hits rate limits
• Sleep Prevention: Prevents system sleep (S3) while LLM subprocesses are running - safe even if afkcode is forcibly killed

Installation

cd afkcode
cargo build --release

The binary will be in target/release/afkcode.

Quick Start

# Create a new checklist with examples
afkcode init my_project.md --title "My Project" --examples

# Generate a checklist from a description
afkcode generate api_project.md "Build a REST API for user management with PostgreSQL"

# Run the autonomous development loop
afkcode run my_project.md

# Add a single task
afkcode add my_project.md "Implement authentication" --section "Tasks"

# Remove completed items
afkcode remove my_project.md "DONE" --yes

Commands

run - Autonomous Development Loop

Runs the worker loop against a checklist until completion or all LLM tools are exhausted. Use --mode controller to opt into the legacy controller/worker alternation.

afkcode run <checklist> [OPTIONS]

Options:
  --controller-prompt <TEMPLATE>     Custom controller prompt
  --worker-prompt <TEMPLATE>         Custom worker prompt
  --completion-token <TOKEN>         Completion detection string
  --sleep-seconds <N>                Delay between iterations (default: 15)
  --mode <worker|controller>         Loop mode (default: worker)
  --run-audit                        Run the Standing Orders alignment audit (disabled by default)
  --audit-orders-path <PATH>         Override the Standing Orders audit target file (see AGENTS_GUIDE.md)
  --tools <TOOLS>                    Comma-separated list of LLM tools (default: gemini,codex,claude)
  --log-file <PATH>                  Log file path for streaming output (default: afkcode.log)
  --gemini-model <MODEL>             Model to use for Gemini CLI (e.g., gemini-2.5-pro)
  --claude-model <MODEL>             Model to use for Claude CLI (e.g., sonnet, opus)
  --codex-model <MODEL>              Model to use for Codex CLI (e.g., o3, o4-mini)

Parallel Execution Options:
  --num-instances <N>                Number of parallel LLM instances (default: 1)
  --warmup-delay <SECONDS>           Delay between launching instances (default: 30, 0 to disable)
  --no-gimme                         Disable gimme mode (work item checkout)
  --gimme-path <PATH>                Base path for AGENTS.md file search (default: current directory)
  --items-per-instance <N>           Number of work items each instance checks out (default: 1)

Warp Agent API requires WARP_API_KEY environment variable or warp_api_key in config.
See "Warp Agent API" section below for details.

Examples:

# With default fallback (Gemini → Codex → Claude)
afkcode run project.md

# Specific tool with custom model
afkcode run project.md --tools gemini --gemini-model gemini-2.5-pro

# Claude with opus model
afkcode run project.md --tools claude --claude-model opus

# Controller/worker alternation
afkcode run project.md --mode controller

# Run the Standing Orders audit (see AGENTS_GUIDE.md)
afkcode run project.md --run-audit

# Use custom AGENTS.md location
afkcode run project.md --run-audit --audit-orders-path docs/AGENTS.md

# Custom sleep time with multiple tools
afkcode run project.md --tools codex,claude --sleep-seconds 30

# Run 3 parallel LLM instances with 30-second warmup delay between each
afkcode run project.md --num-instances 3 --warmup-delay 30

# Parallel execution with gimme mode (each instance gets its own work item)
afkcode run project.md --num-instances 4 --items-per-instance 2

# Disable gimme mode (all instances work on same checklist)
afkcode run project.md --num-instances 2 --no-gimme

How Fallback Works:

1. Starts with first tool in list (default: gemini)
2. If rate limit detected, automatically switches to next tool (e.g., codex, then claude)
3. Rate-limited tools are temporarily squelched for 5 minutes
4. After 5 minutes, the system automatically tries the most preferred tool again
5. Continues development loop seamlessly
6. Only exits if all tools exhausted or completion detected

Parallel Execution:

Run multiple LLM instances simultaneously to maximize throughput:

1. Staggered Launch: Instances launch with configurable warmup delay (default 30s) to prevent API rate limit spikes
2. Independent Fallback: Each instance has its own LlmToolChain with separate rate limit tracking
3. Coordinated Shutdown: When any instance confirms completion (stop token twice), all instances finish their current iteration and exit
4. Gimme Mode: By default, each instance checks out work items from AGENTS.md files, preventing multiple LLMs from working on the same task

Gimme Mode (Work Item Checkout):

Gimme mode automatically assigns work items to each LLM instance:

1. Parses all AGENTS.md files under the search path
2. Selects items marked [ ] (incomplete) - configurable to include [x] or [BLOCKED]
3. Marks selected items as [ip:XXXX] with unique checkout IDs
4. Injects work items into each instance's prompt
5. Failed LLM Recovery: If an LLM crashes or hits rate limits, its work items are automatically restored to [ ] using the checkout ID

Example AGENTS.md workflow:

# Before checkout
- [ ] Implement user authentication
- [ ] Add rate limiting
- [ ] Write unit tests

# After checkout (instance 0 gets first item)
- [ip:a3f7] Implement user authentication
- [ ] Add rate limiting
- [ ] Write unit tests

Completion Token Verification:

• Worker mode: The worker's stdout is scanned (case-insensitive) for the configured completion_token. If detected, afkcode runs a confirmation turn using a dedicated prompt. The loop exits only when the token appears again in that confirmation response; otherwise work resumes normally.
• Controller mode: Unchanged. When the controller emits the completion token, afkcode asks the LLM to re-confirm intent before exiting.

Exit Conditions:

• Worker mode: Token appears in consecutive worker and confirmation turns
• Controller mode: Controller emits completion token (default: __ALL_TASKS_COMPLETE__) and the verification prompt confirms intent
• All LLM tools exhausted due to rate limits
• User presses Ctrl+C

Output Logging: All console output during run mode (LLM responses, status messages, errors) is automatically streamed to a log file (default: afkcode.log). This can be customized via the --log-file CLI argument or the log_file config option. The log file uses buffered writing to maintain responsiveness while capturing all output for later review.

Checklist hygiene (short bullets, removing completed items, using sub-items for partials) is enforced by the Standing Orders that live in your repository; afkcode does not rewrite checklist content during worker turns.

init - Create New Checklist

Creates a bare checklist with standing orders.

afkcode init <checklist> [OPTIONS]

Options:
  -t, --title <TITLE>     Project title (default: "Project Checklist")
  -e, --examples          Include example sections

Example:

afkcode init webapp.md --title "Web Application" --examples

generate - Generate Checklist from Prompt

Uses LLM to generate a complete checklist from a high-level description.

afkcode generate <checklist> <prompt> [OPTIONS]

Options:
  --tools <TOOLS>        Comma-separated list of LLM tools (default: gemini,codex,claude)
  --gemini-model <MODEL> Model to use for Gemini CLI (e.g., gemini-2.5-pro)
  --claude-model <MODEL> Model to use for Claude CLI (e.g., sonnet, opus)
  --codex-model <MODEL>  Model to use for Codex CLI (e.g., o3, o4-mini)
  

Example:

afkcode generate microservice.md \
  "Create a microservice in Rust using Actix-web and PostgreSQL for product inventory tracking"

add - Add Single Item

Adds one checklist item without LLM assistance.

afkcode add <checklist> <item> [OPTIONS]

Options:
  --sub                  Add as sub-item (indented)
  -s, --section <NAME>   Add to specific section

Examples:

# Add to end of file
afkcode add project.md "Write unit tests"

# Add as sub-item
afkcode add project.md "Test error handling" --sub

# Add to specific section
afkcode add project.md "Deploy to staging" --section "Deployment"

add-batch - Add Multiple Items via LLM

Uses LLM to expand a high-level description into multiple checklist items.

afkcode add-batch <checklist> <description> [OPTIONS]

Options:
  --tools <TOOLS>        Comma-separated list of LLM tools (default: gemini,codex,claude)
  --gemini-model <MODEL> Model to use for Gemini CLI
  --claude-model <MODEL> Model to use for Claude CLI
  --codex-model <MODEL>  Model to use for Codex CLI

Example:

afkcode add-batch project.md \
  "Add comprehensive error handling to the authentication module"

remove - Remove Items

Removes items matching a pattern (substring search).

afkcode remove <checklist> <pattern> [OPTIONS]

Options:
  -y, --yes    Skip confirmation prompt

Examples:

# Interactive removal
afkcode remove project.md "COMPLETED"

# Non-interactive
afkcode remove project.md "old feature" --yes

update - Update Checklist via LLM

Uses LLM to update the checklist according to instructions.

afkcode update <checklist> <instruction> [OPTIONS]

Options:
  --tools <TOOLS>        Comma-separated list of LLM tools (default: gemini,codex,claude)
  --gemini-model <MODEL> Model to use for Gemini CLI
  --claude-model <MODEL> Model to use for Claude CLI
  --codex-model <MODEL>  Model to use for Codex CLI

Example:

afkcode update project.md \
  "Reorganize tasks by priority, moving critical items to the top"

Note: Creates a .md.bak backup before updating.

Standing Orders and Custom AGENTS.md

Afkcode uses Standing Orders - a set of 9 immutable rules that govern LLM behavior during autonomous development. These ensure consistent, predictable behavior across sessions.

Core Standing Orders Summary

1. Minimal Information: Checklist items contain only the minimum needed for LLM action
2. Completion Handling: Delete complete items; use [~] with sub-items for partials
3. Discovery: Add newly discovered work succinctly
4. Git Commit: Always commit before finishing a turn
5. Immutability: Standing Orders can't be altered except during audit
6. No Manual Work: Prefer automation over manual steps
7. "Do the thing": Review, pick, implement, update, build, commit
8. "Fix shit": Identify and fix broken code/design, update, commit
9. Stop Token Etiquette: Only emit completion token when truly done

Custom AGENTS.md

You can store Standing Orders and other LLM instructions in a separate AGENTS.md file. This provides:

• Cleaner checklists: Separate task lists from meta-instructions
• Reusability: Share one AGENTS.md across multiple checklists
• Better version control: Track behavioral changes separately
• Warp integration: Reference via @AGENTS.md in Warp's Agent Mode

Quick Start with AGENTS.md:

# Create AGENTS.md
touch AGENTS.md

# Populate it with Standing Orders
afkcode run project.md --run-audit

# Now AGENTS.md contains core orders + space for project-specific rules

For complete documentation on Standing Orders, audit process, and AGENTS.md usage, see AGENTS_GUIDE.md.

Workflow Examples

Example 1: Starting a New Project

# Generate initial checklist
afkcode generate api.md "Build RESTful API for blog platform with Rust and PostgreSQL"

# Review the generated checklist
cat api.md

# Run autonomous development
afkcode run api.md --sleep-seconds 20

Example 2: Adding Features

# Add high-level requirement
afkcode add api.md "Add rate limiting to API endpoints" --section "Requirements"

# Expand into tasks
afkcode add-batch api.md \
  "Implement rate limiting using token bucket algorithm with Redis backend"

# Continue development
afkcode run api.md

Example 3: Cleanup and Maintenance

# Remove completed items
afkcode remove api.md "DONE" --yes

# Reorganize
afkcode update api.md \
  "Group related tasks together and add section headers for API, Database, and Testing"

Custom Prompts

Controller Prompt

The controller reviews progress and assigns tasks. Default:

You are the controller in an autonomous development loop.
Study the shared checklist in @{checklist}, summarize current progress, and update it as needed.
Assign the next actionable task to the worker so momentum continues.
If—and only if—all high-level requirements and every checklist item are fully satisfied,
output {completion_token} on a line by itself at the very end of your reply;
otherwise, do not print that string.

Worker Prompt

The worker implements tasks. Default:

@{checklist} Do the thing.

This assumes your checklist defines what "Do the thing" means (see Standing Orders #7).

Custom Example

afkcode run project.md \
  --controller-prompt "Review @{checklist} and decide what's next. Output {completion_token} when done." \
  --worker-prompt "@{checklist} Implement the highest priority incomplete task."

Configuration

Configuration File Support

afkcode supports TOML configuration files to avoid repeating common CLI arguments. Configuration precedence (highest to lowest):

1. CLI arguments (always win)
2. Config file (if present)
3. Built-in defaults (fallback)

Default Configuration File

afkcode automatically loads afkcode.toml from the current directory if it exists. You can also specify a custom config file:

afkcode --config /path/to/custom.toml run checklist.md

Configuration File Format

Create an afkcode.toml file in your project directory. See afkcode.toml.example for a complete example with all available options.

# LLM tools to use (comma-separated: gemini, codex, claude)
# Default: "gemini,codex,claude"
tools = "gemini,codex,claude"

# Sleep duration between LLM calls in seconds
# Default: 15
sleep_seconds = 20

# Log file path for streaming output during run mode
# Default: "afkcode.log"
log_file = "afkcode.log"

# Controller prompt template
# Default: built-in template (see below)
controller_prompt = """
You are the controller in an autonomous development loop.
Study the shared checklist in @{checklist}, summarize current progress, and update it as needed.
Assign the next actionable task to the worker so momentum continues.
If—and only if—all high-level requirements and every checklist item are fully satisfied, output {completion_token} on a line by itself at the very end of your reply; otherwise, do not print that string.
"""

# Worker prompt template
# Default: "@{checklist} Do the thing."
worker_prompt = "@{checklist} Do the thing."

# Completion detection token
# Default: "__ALL_TASKS_COMPLETE__"
completion_token = "__ALL_TASKS_COMPLETE__"

# Run mode (worker or controller)
# Default: "worker"
# mode = "worker"

# Standing Orders audit configuration (see AGENTS_GUIDE.md)
# skip_audit = true
# orders_path = "AGENTS.md"
# commit_audit = true

# Model selection for LLM tools
# gemini_model = "gemini-2.5-pro"
# claude_model = "opus"
# codex_model = "o3"

# Parallel execution settings
# num_instances = 3           # Number of parallel LLM instances
# warmup_delay = 30           # Seconds between launching instances
# gimme_mode = true           # Enable work item checkout (default: true)
# gimme_base_path = "."       # Base path for AGENTS.md search
# gimme_items_per_instance = 1  # Work items each instance checks out

Configuration Examples

Example 1: Use Claude by default with longer sleep time

# afkcode.toml
tools = "claude"
sleep_seconds = 30

Then simply run:

afkcode run checklist.md  # Uses Claude with 30s sleep automatically

Example 2: Override config with CLI arguments

# afkcode.toml
tools = "claude"
sleep_seconds = 30

# CLI args override config file
afkcode run checklist.md --tools codex --sleep-seconds 10
# Uses Codex with 10s sleep, not Claude with 30s

Example 3: Project-specific configurations

# Different config for each project
cd project-a
echo 'tools = "codex"' > afkcode.toml
afkcode run checklist.md  # Uses Codex

cd ../project-b
echo 'tools = "claude"' > afkcode.toml
afkcode run checklist.md  # Uses Claude

LLM Backend Configuration

Supported Tools

afkcode supports these LLM CLI tools and APIs with automatic fallback:

1. Gemini CLI (gemini) - Default first choice

   • Command: gemini --yolo
   • Flags automatically added by afkcode for unattended operation
   • Rate limit detection: "rate limit", "429", "quota"
   • Model selection: --gemini-model gemini-2.5-pro

2. Codex CLI (codex)

   • Command: codex exec
   • Configuration: Requires ~/.codex/config.toml with approval_policy = "never"
   • Rate limit detection: "rate limit reached", "429", "rate_limit_error"
   • Model selection: --codex-model o3 or --codex-model o4-mini

3. Claude Code (claude)

   • Command: claude --print --dangerously-skip-permissions
   • Flags automatically added by afkcode for unattended operation
   • Rate limit detection: "usage limit reached", "limit will reset", "429"
   • Model selection: --claude-model opus or --claude-model sonnet

4. Warp Agent API (warp) - HTTP-based, multi-model

   • Type: REST API (not CLI-based)
   • API endpoint: https://app.warp.dev/api/v1
   • Authentication: Bearer token via WARP_API_KEY environment variable or config
   • Rate limit detection: "rate limit", "429", "quota exceeded"
   • Supports many models via warp_model config:
     • Claude: claude-sonnet-4-5, claude-opus-4-1, claude-haiku-4-5, claude-sonnet-4
     • OpenAI: gpt-5, gpt-5-1 (with low/medium/high reasoning modes)
     • Google: gemini-3-pro, gemini-2-5-pro
     • z.ai: glm-4-6
     • Auto modes: auto-cost-efficient, auto-responsiveness

Using Default Fallback

# Tries Gemini first, then Codex, then Claude on rate limits
afkcode run checklist.md
afkcode run checklist.md --tools gemini,codex,claude

Using Single Tool

# Gemini only (no fallback)
afkcode run checklist.md --tools gemini

# Gemini with specific model
afkcode run checklist.md --tools gemini --gemini-model gemini-2.5-pro

# Codex only with specific model
afkcode run checklist.md --tools codex --codex-model o3

# Claude only with specific model
afkcode run checklist.md --tools claude --claude-model opus

Custom Tool Order

# Try Claude first, fall back to Codex, then Gemini
afkcode run checklist.md --tools claude,codex,gemini

# Try Codex only, no Gemini
afkcode run checklist.md --tools codex

Model Selection Examples

# Use specific models for each tool
afkcode run checklist.md \
  --tools gemini,claude \
  --gemini-model gemini-2.5-pro \
  --claude-model opus

# Configure via config file
echo 'gemini_model = "gemini-2.5-pro"' >> afkcode.toml
echo 'claude_model = "sonnet"' >> afkcode.toml
afkcode run checklist.md

Note: CLI tools (gemini, codex, claude) must be installed and available on your PATH. Warp Agent API only requires an API key.

Warp Agent API

Warp Agent API is a powerful option that provides access to multiple LLM models through a single HTTP API.

Key Benefits:

• Multi-model access: Switch between Claude, GPT, Gemini, and more via configuration
• No CLI installation: Just needs an API key
• Unified billing: One account for all models
• Latest models: Access to Claude 4.5 Sonnet, GPT-5, Gemini 3 Pro, etc.
• Auto modes: Let Warp choose the best model automatically

Setup:

1. Get your API key from Warp
2. Set environment variable:

   export WARP_API_KEY="your-key-here"

3. Or add to afkcode.toml:

   warp_api_key = "your-key-here"
   warp_model = "claude-sonnet-4-5"  # or any supported model

Usage:

# Use Warp Agent API
afkcode run checklist.md --tools warp

# With specific model
echo 'warp_model = "gpt-5"' >> afkcode.toml
afkcode run checklist.md --tools warp

# As fallback after Gemini
afkcode run checklist.md --tools gemini,warp

Available Models:

Warp Agent API supports a curated set of top LLMs:

• Claude (Anthropic):

  • claude-sonnet-4-5 - Latest Sonnet model
  • claude-opus-4-1 - Most capable Claude model
  • claude-haiku-4-5 - Fast and efficient
  • claude-sonnet-4 - Previous generation

• GPT (OpenAI):

  • gpt-5 - Latest GPT model
  • gpt-5-1 - GPT-5.1 with configurable reasoning

• Gemini (Google):

  • gemini-3-pro - Latest Gemini
  • gemini-2-5-pro - Previous generation

• GLM (z.ai):

  • glm-4-6 - Hosted by Fireworks AI

• Auto Modes:

  • auto-cost-efficient - Optimizes for lower credit consumption
  • auto-responsiveness - Prioritizes highest quality and speed

See Warp's model documentation for the latest model list.

Unattended Configuration

For autonomous operation, afkcode requires all LLM tools to be configured for unattended/automatic execution without prompts.

Gemini CLI Configuration

Gemini CLI is configured with the --yolo flag for unattended operation. afkcode automatically passes this flag:

gemini --yolo

• --yolo: Enables fully unattended execution without approval prompts

Note: Ensure Gemini CLI is installed and authenticated. No additional configuration file is needed.

Codex CLI Configuration

Codex CLI is configured via ~/.codex/config.toml. Required settings for unattended operation:

approval_policy = "never"
sandbox_mode    = "danger-full-access"

This disables all approval prompts and grants full system access. The approval_policy = "never" setting ensures Codex executes all commands automatically without asking for confirmation.

Claude Code Configuration

Claude Code uses CLI flags for unattended operation. afkcode automatically passes the following flags:

claude --print --dangerously-skip-permissions

• --print: Non-interactive output mode
• --dangerously-skip-permissions: Enables "Safe YOLO mode" for fully unattended execution without approval prompts

Security Note: These settings grant the LLM unrestricted access to your system. Only use in trusted development environments. Consider running in an isolated container or VM for additional safety.

Tips & Best Practices

1. Be Specific: Write detailed, actionable tasks in your checklist
2. Small Increments: Break large features into small, testable chunks
3. Regular Commits: Standing orders ensure commits happen automatically
4. Monitor Progress: Check git log periodically to review autonomous changes
5. Rate Limits: Increase --sleep-seconds if hitting rate limits frequently
6. Backup: Run git status before starting to ensure clean state
7. Review: Always review LLM-generated code before deploying

Safety Features

• Backup Creation: update command creates .md.bak backups
• Confirmation Prompts: remove asks for confirmation (unless --yes)
• Rate Limit Detection: Automatically stops on rate limit
• Ctrl+C Handling: Graceful shutdown on interrupt
• Sleep Prevention: System sleep is inhibited during LLM execution. Uses OS-native process-bound facilities that automatically release when afkcode exits (including crashes or SIGKILL), so your laptop won't stay awake indefinitely if something goes wrong

Troubleshooting

Problem: Tool not found error

# Make sure your LLM tools are installed and on PATH
which gemini
which codex
which claude

# Or use only the tool you have installed
afkcode run checklist.md --tools claude

Problem: Loop exits immediately

# Check LLM command works independently
codex exec   # Should start Codex
claude --print "test"  # Should output response

# Try with single tool to debug
afkcode run checklist.md --tools codex

Problem: All tools rate limited

# Increase sleep time to reduce API usage
afkcode run checklist.md --sleep-seconds 60

# Use only one tool and wait for reset
afkcode run checklist.md --tools claude

Problem: Generated checklist has wrong structure

# Try more specific prompt
afkcode generate project.md "Detailed description with technology stack, architecture, and specific features..."

# Or try a different tool
afkcode generate project.md "..." --tools claude

Problem: Fallback not working

# Verify tools are installed
which gemini && which codex && which claude

# Check rate limit detection messages in output
# Should see "Rate limit detected for gemini"
# Followed by "Switching to fallback tool: codex"

Problem: Standing Orders being modified unexpectedly

# Audit is disabled by default to prevent unwanted changes
# Only run audit when you explicitly want to sync Standing Orders
afkcode run project.md --run-audit

# To completely disable audit in config:
echo 'skip_audit = true' >> afkcode.toml

Problem: Want to use custom AGENTS.md location

# Specify custom path via CLI
afkcode run project.md --run-audit --audit-orders-path docs/AGENTS.md

# Or via config file
echo 'orders_path = "docs/AGENTS.md"' >> afkcode.toml

# See AGENTS_GUIDE.md for complete documentation

Development

Building

cargo build

Running Tests

cargo test

Running from Source

cargo run -- run checklist.md
cargo run -- init new.md --examples
cargo run -- --help

Recent Changes

New Features (Latest)

• Parallel LLM Execution: Run multiple LLM instances simultaneously with staggered warmup delays (--num-instances, --warmup-delay)
• Gimme Mode: Automatic work item checkout from AGENTS.md files - each LLM instance gets assigned specific tasks to prevent conflicts
• Failed LLM Recovery: Automatically restores work items to [ ] if an LLM subprocess crashes or hits rate limits, using unique checkout IDs ([ip:XXXX])
• Coordinated Shutdown: When any LLM confirms completion, all instances finish their current iteration and exit gracefully
• Independent Fallback Tracking: Each parallel instance has its own LlmToolChain for separate rate limit tracking
• Sleep Prevention: Prevents system from sleeping (S3) while LLM subprocesses are running - uses OS-native facilities (macOS IOPMAssertion, Windows SetThreadExecutionState, Linux D-Bus inhibitor) that auto-release on process exit, so your laptop won't stay hot in your bag even if afkcode crashes
• Warp Agent API Support: HTTP-based access to multiple LLM models (Claude 4.5, GPT-5, Gemini 3 Pro, etc.) through Warp's unified API
• Multi-Model Selection via Warp: Single API key provides access to Claude, OpenAI, Google, and z.ai models
• Gemini Support: Gemini CLI is now the default first tool in the fallback chain
• Model Selection: Specify custom models for each LLM tool (--gemini-model, --claude-model, --codex-model, warp_model)
• Custom AGENTS.md: Comprehensive support for separate Standing Orders files (see AGENTS_GUIDE.md)
• Audit Disabled by Default: Standing Orders audit now requires --run-audit flag to prevent unwanted changes
• Enhanced Configuration: Added skip_audit, orders_path, commit_audit, warp_api_key, warp_model, and model selection config options
• Temporary Tool Squelching: Rate-limited tools are temporarily disabled for 5 minutes, then automatically retry
• Completion Token Verification: LLM confirms intentional completion to prevent false positives

Bug Fixes

• Fixed sub-item formatting: Sub-items now properly include the dash prefix ( - [ ] instead of [ ])
• Fixed section-specific add: Items added with --section now correctly appear in the specified section
• Fixed standing orders preservation: The update command now automatically restores standing orders if the LLM removes them

Previous Features

• Configuration file support: Added afkcode.toml support for persistent configuration
• Global --config flag: Specify custom configuration file paths
• Configuration precedence: CLI args override config file values, which override defaults

Differences from Python Version

1. Subcommands: Not just a loop - full checklist management toolkit
2. Standing Orders: Built-in invariants with audit support and custom AGENTS.md
3. Automatic Fallback: Seamless switching between LLM tools on rate limits
4. Supported Tools: Gemini, Codex CLI, Claude Code, and Warp Agent API with tool-specific rate limit detection
5. Warp Agent API: HTTP-based multi-model access (Claude 4.5, GPT-5, Gemini 3 Pro, GLM 4.6, Auto modes)
6. Model Selection: Specify custom models per tool (e.g., gemini-2.5-pro, o3, opus, claude-sonnet-4-5)
7. Better Error Handling: Comprehensive error messages with anyhow
8. Type Safety: Compile-time guarantees for correctness
9. Cross-Platform: Works on Linux, macOS, Windows
10. Configuration Files: TOML-based configuration for project-specific settings

Contributing

Contributions are welcome! Please open an issue or pull request at: https://github.com/allquixotic/afkcode

License

Licensed under the Apache License, Version 2.0. See LICENSE.txt for details.

Copyright (c) 2025 Sean McNamara smcnam@gmail.com

See Also

• AGENTS_GUIDE.md - Complete guide to Custom AGENTS.md and Standing Orders audit
• AGENTS.md - Template for LLM-focused project documentation (created via --run-audit)
• example_checklist.md - Sample checklist structure
• afkcode.toml.example - Complete configuration file example
• TESTING.md - Testing documentation and smoke test suite
• GitHub: https://github.com/allquixotic/afkcode