nibzard/looper-go

nibzard 资料

最近记录的来源活动: 2026年3月9日 11:48

SkillsMP 收录数据更新: 2026年10月17日 00:09

已收集 skills: 7

GitHub 星标: 5
GitHub Forks: 0

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# Looper A deterministic, autonomous loop runner for AI agents (Codex, Claude). It processes exactly one task per iteration from a JSON backlog, with fresh context each run, and keeps a JSONL audit log for traceability. ## Quick Start ```bash make install looper ``` If `to-do.json` is missing, Looper bootstraps it from project docs, validates it against `to-do.schema.json`, and repairs it if needed. ## What It Does - Bootstraps `to-do.json` and `to-do.schema.json` if missing. - Validates `to-do.json` (using JSON Schema when available). - Repairs invalid task files via configured agent (any registered agent). - Runs one task per iteration (doing > todo > blocked). - When tasks are exhausted, runs a review pass; appends a final `project-done` marker task if no new work is found. - Supports multiple iteration schedules (agent name, odd-even, round-robin). - Enforces JSON output from the model and logs JSONL per run. - Optionally applies model summaries back into `to-do.json`. ## Install ### From source (Unix/Linux/macOS) ```bash make install ``` The binary is installed to `~/.local/bin/looper` by default. ### From source (Windows) On Windows, build the binary directly with Go: ```powershell go build -o looper.exe ./cmd/looper ``` Or use the provided build script: ```powershell go build -ldflags "-X main.Version=dev" -o bin/looper.exe ./cmd/looper ``` The Makefile and install/uninstall scripts are Unix-only; on Windows use `go build` or `go install` and place `looper.exe` on your PATH. The binary can be placed anywhere on your PATH. A common location is `%USERPROFILE%\bin` or a directory added to your PATH. To add to PATH (PowerShell): ```powershell $env:PATH += ";$env:USERPROFILE\bin" # To make permanent, add to System Environment Variables ``` ### Uninstall **Unix/Linux/macOS:** ```bash make uninstall ``` **Windows:** Simply delete the binary file. ### Homebrew (Unix/Linux/macOS only) ```bash brew install nibzard/tap/looper ``` ## Usage ```bash # Initialize a new project looper init # Run the loop (default command) looper # Run with specific todo file looper run path/to/tasks.json # Run with specific schedule looper run --schedule odd-even # Workflow commands looper workflow list # List available workflows looper workflow describe parallel # Describe a specific workflow # List tasks by status looper ls looper ls todo looper ls --status doing path/to/tasks.json # Tail the latest log looper tail looper tail --follow # Check dependencies and configuration looper doctor # Show version looper version # Run with a specific workflow looper run --workflow parallel looper run --workflow code-review # Run a release workflow via the agent looper push looper push --agent claude looper push -y # Validate task file against schema looper validate looper validate path/to/tasks.json looper --todo path/to/tasks.json validate --schema custom-schema.json # Format task file with stable ordering looper fmt looper fmt -w # Write formatted file back to disk looper fmt -check # Check if file is formatted looper fmt -d # Display diffs of formatting changes # Show effective configuration looper config looper config --json # Output in JSON format # Clean up old log runs looper clean --dry-run # Show what would be deleted looper clean --keep 10 # Keep 10 most recent runs looper clean --age 7d # Delete logs older than 7 days looper clean --keep 5 --age 30d # Keep 5 recent, delete any older than 30 days # Generate shell completion script looper completion bash # Output bash completion to stdout looper completion zsh # Output zsh completion to stdout looper completion fish # Output fish completion to stdout looper completion powershell # Output PowerShell completion to stdout ``` ## Shell Completion Looper provides shell auto-completion for bash, zsh, fish, and PowerShell. ### Installation **Bash:** ```bash # For current session looper completion bash | source # For persistent completion, add to ~/.bashrc or ~/.bash_profile echo 'source <(looper completion bash)' >> ~/.bashrc ``` **Zsh:** ```bash # For current session looper completion zsh | source # For persistent completion, add to ~/.zshrc echo 'source <(looper completion zsh)' >> ~/.zshrc # Or place the completion script in the zsh completion directory looper completion zsh > ~/.zsh/completion/_looper ``` **Fish:** ```bash # For persistent completion, add to fish config directory looper completion fish > ~/.config/fish/completions/looper.fish ``` **PowerShell:** ```powershell # For current session looper completion powershell | Invoke-Expression # For persistent completion, add to PowerShell profile looper completion powershell >> $PROFILE ``` ## Iteration Schedules Looper supports multiple iteration schedules for balancing agent usage: - `--schedule codex` (default) - All iterations use Codex - `--schedule claude` - All iterations use Claude - `--schedule <agent>` - All iterations use any registered agent type - `--schedule odd-even` - Odd iterations use one agent, even use another - `--schedule round-robin` - Rotate through a list of agents ### Schedule Options **Odd-even schedule:** ```bash looper run --schedule odd-even looper run --schedule odd-even --odd-agent claude --even-agent codex ``` **Round-robin schedule:** ```bash # Default: claude,codex pattern looper run --schedule round-robin # Custom pattern (2x Claude, then Codex, repeating) looper run --schedule round-robin --rr-agents claude,claude,codex ``` **Note:** The final review pass uses the configured review agent (defaults to Codex if not set). ### Role-Based Agent Configuration Looper supports role-based agent configuration that allows you to assign different agents to different phases of the loop. This is more powerful than simple iteration scheduling as it gives you fine-grained control over which agent handles each operation. #### Available Roles - **iter** - Agent used for main task iterations (overrides `schedule` when set) - **review** - Agent used for the review pass when no tasks are found - **repair** - Agent used for repairing invalid task files - **bootstrap** - Agent used for creating the initial task file #### Configuration via Config File The preferred way to configure roles is via the `[roles]` section in `looper.toml`: ```toml # Assign different agents to different loop phases [roles] iter = "claude" # Use Claude for task iterations review = "codex" # Use Codex for review passes repair = "claude" # Use Claude for repair operations bootstrap = "codex" # Use Codex for bootstrapping ``` #### Configuration via Environment Variables Roles can also be set via environment variables: - `LOOPER_ROLES_ITER` - Agent for iterations - `LOOPER_ROLES_REVIEW` - Agent for review passes - `LOOPER_ROLES_REPAIR` - Agent for repair operations - `LOOPER_ROLES_BOOTSTRAP` - Agent for bootstrap ```bash export LOOPER_ROLES_ITER="claude" export LOOPER_ROLES_REVIEW="codex" looper run ``` #### Configuration via CLI Flags For one-off overrides, use the step-specific CLI flags: ```bash looper run --repair-agent claude looper run --review-agent claude looper run --bootstrap-agent claude ``` #### Precedence Order When multiple configuration sources specify the same role, the precedence order is: 1. **CLI flags** (highest priority) - e.g., `--review-agent` 2. **Environment variables** - e.g., `LOOPER_ROLES_REVIEW` 3. **Project config file** - `looper.toml` [roles] section 4. **User config file** - `~/.looper/looper.toml` [roles] section 5. **Legacy config keys** - `repair_agent`, `review_agent`, `bootstrap_agent` 6. **Schedule defaults** - Uses `schedule` setting as fallback for iterations #### When to Use Role-Based Configuration **Use role-based configuration when:** - You want different agents for different phases (e.g., Claude for coding, Codex for review) - You want to override the schedule for specific operations - You want a consistent assignment regardless of iteration count **Use simple schedule when:** - You want the same agent for all operations - You want to alternate agents between iterations - You only need simple round-robin patterns #### Example Configurations **Claude for development, Codex for review:** ```toml [roles] iter = "claude" review = "codex" repair = "claude" bootstrap = "codex" ``` **Codex for everything except repair:** ```toml [roles] iter = "codex" review = "codex" bootstrap = "codex" repair = "claude" # Use Claude for complex repairs ``` **Use schedule for iterations, override specific roles:** ```toml schedule = "odd-even" # Alternate for iterations odd_agent = "claude" even_agent = "codex" [roles] review = "claude" # Always use Claude for review repair = "claude" # Always use Claude for repair ``` ### Step Agents Agents for specific loop steps can also be configured via CLI flags for one-off overrides: ```bash looper run --repair-agent claude looper run --review-agent claude looper run --bootstrap-agent claude ``` - **Repair agent** - Used for repairing invalid task files (default: `codex`) - **Review agent** - Used for the review pass when no tasks are found (configurable, defaults to `codex`) - **Bootstrap agent** - Used for creating the initial task file (default: `codex`) **Note:** For persistent configuration, prefer the `[roles]` section in `looper.toml` over CLI flags. ## Initializing a Project The `looper init` command scaffolds the initial project files: ```bash looper init ``` This creates: - `to-do.json` - Task file with one example task - `to-do.schema.json` - JSON Schema for validation - `looper.toml` - Configuration file with defaults ### Init Options ```bash # Overwrite existing files looper init --force # Skip creating looper.toml (use defaults or existing config) looper init --skip-config # Specify custom file paths looper init --todo my-tasks.json --schema my-schema.json --config my-config.toml ``` ### What Gets Created **to-do.json** - Contains a minimal task structure: ```json { "schema_version": 1, "project": { "root": "." }, "source_files": ["README.md"], "tasks": [ { "id": "T001", "title": "Example: Add project documentation", "description": "Create a README.md file documenting the project setup and usage.", "reference": "README.md", "priority": 1, "status": "todo" } ] } ``` **looper.toml** - Contains commented configuration with all available options. ## Configuration Looper uses a configuration hierarchy (later entries override earlier): 1. Built-in defaults 2. User config file (`~/.looper/looper.toml` or OS-specific config directory) 3. Project config file (`looper.toml` or `.looper.toml` in the current directory) 4. Environment variables 5. CLI flags ### Config File Locations **User-level config** (for global defaults): - `~/.looper/looper.toml` (preferred) - OS-specific config directories (if `~/.looper` doesn't exist): - **Windows**: `%APPDATA%\looper\looper.toml` - **macOS**: `~/Library/Application Support/looper/looper.toml` - **Linux/BSD**: `$XDG_CONFIG_HOME/looper/looper.toml` or `~/.config/looper/looper.toml` **Project-level config** (overrides user config): - `./looper.toml` (preferred) - `./.looper.toml` ### Config File Create `looper.toml` (or `.looper.toml`) in your project directory: ```toml # Paths todo_file = "to-do.json" schema_file = "to-do.schema.json" log_dir = "~/.looper" log_file = "" # Optional: write console logs to a file instead of stdout # Loop settings max_iterations = 50 schedule = "codex" # agent name|odd-even|round-robin # Role-based agent configuration (preferred over repair_agent/review_agent/bootstrap_agent) [roles] iter = "codex" # Agent for task iterations (overrides schedule) review = "codex" # Agent for review passes repair = "codex" # Agent for repair operations bootstrap = "codex" # Agent for bootstrap # Legacy single-step configuration (still supported, use [roles] for clarity) # repair_agent = "codex" # any registered agent type # review_agent = "codex" # any registered agent type (default: codex) # bootstrap_agent = "codex" # any registered agent type (default: codex) # Odd-even schedule options odd_agent = "codex" # any registered agent type even_agent = "claude" # any registered agent type # Round-robin schedule options rr_agents = ["claude", "codex"] # any registered agent types # Agents [agents.codex] binary = "codex" model = "" # reasoning = "medium" # Optional: low, medium, or high reasoning effort # args = ["--flag", "value"] # Optional extra args to pass to codex # parser = "codex_parser.py" # Optional: parser for agent output [agents.claude] binary = "claude" model = "" # args = ["--flag", "value"] # Optional extra args to pass to claude # parser = "builtin:claude" # Optional: parser for agent output # Custom agents can be registered and configured # To add a custom agent, register it via init() in your code # and configure it under the agents map: # [agents.opencode] # binary = "opencode" # model = "custom-model" # parser = "opencode_parser.py" # Output apply_summary = true # Git # git_init = true # Run 'git init' before bootstrap when enabled # Hooks hook_command = "/path/to/hook.sh" # Delay between iterations loop_delay_seconds = 0 # Logging configuration log_level = "info" # Log level: debug, info, warn, error, fatal log_format = "text" # Log format: text, json, logfmt log_timestamps = true # Show timestamps in log output log_caller = false # Show caller location (file:line) in logs # Parallel task execution (disabled by default for backward compatibility) [parallel] enabled = false # Enable parallel task execution max_tasks = 4 # Maximum concurrent tasks (0 = unlimited) max_agents_per_task = 1 # Agents per task for consensus (1 = single agent) strategy = "priority" # Task selection: priority|dependency|mixed fail_fast = false # Stop all on first failure output_mode = "multiplexed" # Output: multiplexed|buffered|summary # Workflow selection (default: "traditional") workflow = "traditional" # Workflow-specific configuration [workflows.parallel] max_concurrent = 3 # Max concurrent tasks (default: 3) fail_fast = false # Stop on first error (default: false) [workflows.code-review] diff_path = "." # Path to review (default: ".") review_stages = ["analyze", "security", "style"] require_approval = true # Require manual approval approval_file = ".looper/approval.txt" [workflows.incident-triage] severity_levels = ["critical", "high", "medium", "low"] auto_assign = true notify_slack = false # slack_webhook = "https://hooks.slack.com/services/..." # Plugin configuration (settings for installed plugins) [plugins.claude] timeout = "30m" # Execution timeout (e.g., "30s", "5m", "1h") # work_dir = "." # Working directory for plugin execution # binary = "claude" # Override plugin binary path # model = "" # Model for agent plugins # reasoning = "medium" # Reasoning effort for agent plugins # enabled = true # Enable/disable plugin without uninstalling ``` Looper reads the config file from the current working directory (not the todo file directory). ### Plugin Configuration Plugins can be configured via `[plugins.<name>]` sections in `looper.toml`. This allows you to customize plugin behavior without modifying the plugin manifest or reinstalling the plugin. #### Configuration Pattern ```toml [plugins.<plugin-name>] timeout = "30m" # Maximum execution time work_dir = "." # Working directory for plugin execution binary = "claude" # Override binary path from manifest model = "" # Model for agent plugins reasoning = "" # Reasoning effort for agent plugins enabled = true # Enable or disable plugin ``` #### Plugin Settings | Setting | Type | Description | |---------|------|-------------| | `timeout` | string | Maximum execution duration (e.g., `"30s"`, `"5m"`, `"1h"`) | | `work_dir` | string | Working directory for plugin execution (relative to project root) | | `binary` | string | Override the plugin binary path from the manifest | | `model` | string | Model to use for agent plugins (e.g., `"claude-3-5-sonnet-20241022"`) | | `reasoning` | string | Reasoning effort for agent plugins (e.g., `"low"`, `"medium"`, `"high"`) | | `enabled` | bool | Enable or disable the plugin without uninstalling it | #### Configuration Examples **Set a longer timeout for a slow plugin:** ```toml [plugins.my-slow-plugin] timeout = "1h" ``` **Override the binary path for local development:** ```toml [plugins.claude] binary = "/path/to/dev/claude" ``` **Configure model and reasoning for an agent plugin:** ```toml [plugins.codex] model = "gpt-4" reasoning = "high" ``` **Disable a plugin temporarily:** ```toml [plugins.experimental-plugin] enabled = false ``` **Set a specific working directory:** ```toml [plugins.build-tool] work_dir = "./build" ``` #### Configuration Precedence Plugin configuration settings override the corresponding values in the plugin manifest (`looper-plugin.toml`). This allows you to: 1. Override timeouts for specific projects 2. Use different binary paths for development vs production 3. Configure agent plugins with project-specific models 4. Temporarily disable plugins without uninstalling them #### Relationship to Plugin Manifests The `[plugins.<name>]` configuration complements the plugin manifest: - **Manifest** (`looper-plugin.toml`): Defines the plugin's default settings and metadata - **Config** (`looper.toml`): Overrides manifest settings for a specific project For example, if a plugin manifest declares: ```toml [plugin] binary = "./bin/my-plugin" ``` You can override it in your project config: ```toml [plugins.my-plugin] binary = "/usr/local/bin/my-plugin" timeout = "5m" ``` #### Environment Variables Plugin settings can also be configured via environment variables using the pattern `LOOPER_PLUGINS_<NAME>_<SETTING>`: ```bash # Set timeout for claude plugin export LOOPER_PLUGINS_CLAUDE_TIMEOUT="1h" # Disable a plugin export LOOPER_PLUGINS_EXPERIMENTAL_ENABLED="0" # Set model for codex plugin export LOOPER_PLUGINS_CODEX_MODEL="gpt-4" ``` ### Advanced Logging Configuration Looper provides fine-grained control over log output through four configuration options: #### Log Level (`log_level`) Controls the verbosity of log messages. Available levels (from most to least verbose): - **`debug`** - Detailed information for debugging, including all internal operations - **`info`** (default) - General informational messages about loop progress - **`warn`** - Warning messages for potential issues - **`error`** - Error messages only - **`fatal`** - Critical errors that cause termination **When to use each level:** - Use `debug` when troubleshooting issues or developing features - Use `info` for normal operation (default) - Use `warn` or `error` to reduce noise in production logs - Use `fatal` for minimal logging (rarely needed) ```bash # Via config file log_level = "debug" # Via environment variable export LOOPER_LOG_LEVEL=debug # Via CLI flag looper run --log-level debug ``` #### Log Format (`log_format`) Controls how log messages are formatted: - **`text`** (default) - Human-readable text format with colors and styling - **`json`** - Structured JSON output, one log per line (for log aggregation tools) - **`logfmt`** - Key-value pair format (common in cloud-native systems) **When to use each format:** - Use `text` for interactive terminal sessions and development (default) - Use `json` for log aggregation systems (e.g., ELK, Splunk, CloudWatch) - Use `logfmt` for systems that expect key-value logging ```bash # JSON logging for production log_format = "json" # Via environment export LOOPER_LOG_FORMAT=json ``` #### Log Timestamps (`log_timestamps`) Controls whether timestamps are included in log output: - **`true`** (default) - Include timestamps in each log message - **`false` - Omit timestamps for cleaner output **When to disable timestamps:** - When logging to a system that adds its own timestamps - For cleaner local development output - When timing information is not needed ```bash # Disable timestamps log_timestamps = false # Via environment export LOOPER_LOG_TIMESTAMPS=0 ``` #### Log Caller (`log_caller`) Controls whether the source location (file and line number) is included: - **`false`** (default) - Do not show source location - **`true`** - Show `file:line` for each log message **When to enable caller information:** - When debugging to trace where log messages originate - During development to identify code paths - Not recommended for production (adds verbosity) ```bash # Enable caller information for debugging log_caller = true # Via environment export LOOPER_LOG_CALLER=1 ``` #### Example Configurations **Development/Debugging:** ```toml # Maximum verbosity for troubleshooting log_level = "debug" log_format = "text" log_timestamps = true log_caller = true ``` **Production with Log Aggregation:** ```toml # Structured logs for ELK/Splunk log_level = "info" log_format = "json" log_timestamps = false # Log system adds timestamps log_caller = false ``` **Minimal Local Output:** ```toml # Clean output for local development log_level = "warn" log_format = "text" log_timestamps = false log_caller = false ``` ### Environment Variables - `LOOPER_TODO` - Task file path (default: `to-do.json`) - `LOOPER_SCHEMA` - Schema file path - `LOOPER_BASE_DIR` / `LOOPER_LOG_DIR` - Log directory (supports `~`, `$HOME`, or `%USERPROFILE%` on Windows) - `LOOPER_MAX_ITERATIONS` - Maximum iterations - `LOOPER_ITER_SCHEDULE` / `LOOPER_SCHEDULE` - Iteration schedule (agent name or odd-even/round-robin) **Role-based environment variables (preferred):** - `LOOPER_ROLES_ITER` - Agent for task iterations (overrides schedule when set) - `LOOPER_ROLES_REVIEW` - Agent for review passes - `LOOPER_ROLES_REPAIR` - Agent for repair operations - `LOOPER_ROLES_BOOTSTRAP` - Agent for bootstrap **Legacy single-step environment variables (still supported):** - `LOOPER_REPAIR_AGENT` - Agent for repair operations - `LOOPER_REVIEW_AGENT` - Agent for review pass (default: codex) - `LOOPER_BOOTSTRAP_AGENT` - Agent for bootstrap (default: codex) **Schedule options:** - `LOOPER_ITER_ODD_AGENT` - Agent for odd iterations - `LOOPER_ITER_EVEN_AGENT` - Agent for even iterations - `LOOPER_ITER_RR_AGENTS` - Comma-separated agent list for round-robin **Other settings:** - `LOOPER_APPLY_SUMMARY` - Apply summaries to task file (1/0) - `LOOPER_GIT_INIT` - Run 'git init' before bootstrap when enabled (1/0) - `LOOPER_HOOK` - Hook command to run after each iteration - `LOOPER_LOOP_DELAY` - Delay between iterations (seconds) - `LOOPER_PROMPT_DIR` - Prompt directory override (dev only, requires `LOOPER_PROMPT_MODE=dev`) - `LOOPER_PROMPT` - User prompt for bootstrap (brainstorms tasks from your idea instead of scanning docs) - `LOOPER_PRINT_PROMPT` - Print rendered prompts (1/0, dev only) **Logging settings:** - `LOOPER_LOG_LEVEL` - Log level: `debug`, `info`, `warn`, `error`, `fatal` (default: `info`) - `LOOPER_LOG_FORMAT` - Log format: `text`, `json`, `logfmt` (default: `text`) - `LOOPER_LOG_TIMESTAMPS` - Show timestamps in logs (1/0, default: 1) - `LOOPER_LOG_CALLER` - Show caller location (file:line) in logs (1/0, default: 0) - `LOOPER_LOG_FILE` - Write console logs to a file instead of stdout (optional) **Agent settings:** - `CODEX_BIN` / `CLAUDE_BIN` - Agent binary paths (on Windows, use `codex.exe` / `claude.exe`) - `CODEX_MODEL` / `CLAUDE_MODEL` - Model selection - `CODEX_REASONING` / `CODEX_REASONING_EFFORT` - Codex reasoning effort (e.g., "low", "medium", "high") - `CODEX_ARGS` - Comma-separated extra args for Codex (e.g., `--flag,value`) - `CLAUDE_ARGS` - Comma-separated extra args for Claude (e.g., `--flag,value`) **Plugin settings:** - `LOOPER_PLUGINS_<NAME>_TIMEOUT` - Plugin execution timeout (e.g., `LOOPER_PLUGINS_CLAUDE_TIMEOUT="1h"`) - `LOOPER_PLUGINS_<NAME>_WORK_DIR` - Plugin working directory (e.g., `LOOPER_PLUGINS_BUILD_WORK_DIR="./build"`) - `LOOPER_PLUGINS_<NAME>_BINARY` - Override plugin binary path (e.g., `LOOPER_PLUGINS_CLAUDE_BINARY="/path/to/claude"`) - `LOOPER_PLUGINS_<NAME>_MODEL` - Model for agent plugins (e.g., `LOOPER_PLUGINS_CODEX_MODEL="gpt-4"`) - `LOOPER_PLUGINS_<NAME>_REASONING` - Reasoning effort for agent plugins (e.g., `LOOPER_PLUGINS_CODEX_REASONING="high"`) - `LOOPER_PLUGINS_<NAME>_ENABLED` - Enable/disable plugin (1/0, e.g., `LOOPER_PLUGINS_EXPERIMENTAL_ENABLED="0"`) ### CLI Flags Global flags (place before the subcommand): - `--todo`, `--schema`, `--log-dir`, `--log-file` - `--log-level`, `--log-format`, `--log-timestamps`, `--log-caller` - `--codex-bin`, `--claude-bin`, `--codex-model`, `--claude-model`, `--codex-reasoning`, `--codex-args`, `--claude-args` Run flags (use with `run`): - `--max-iterations`, `--schedule`, `--odd-agent`, `--even-agent`, `--rr-agents` - `--repair-agent`, `--review-agent`, `--bootstrap-agent` - `--apply-summary`, `--git-init`, `--hook`, `--loop-delay` Run `looper help` for the full list. Dev-only flags appear when `LOOPER_PROMPT_MODE=dev` is set. ## Task File (to-do.json) `to-do.json` is the source of truth for the loop and must match `to-do.schema.json`. The loop chooses a single task each iteration: 1) Any task in `doing` (lowest id wins) 2) Otherwise, highest priority `todo` (priority 1 is highest) 3) Otherwise, highest priority `blocked` Minimal example: ```json { "schema_version": 1, "source_files": ["README.md"], "tasks": [ { "id": "T1", "title": "Add README", "description": "Create a comprehensive README documenting the project setup and usage.", "reference": "README.md", "priority": 1, "status": "todo" } ] } ``` ### Task Fields - `id` (required) - Unique task identifier (e.g., "T1", "T002") - `title` (required) - Concise, actionable summary - `description` (optional) - Detailed explanation of what the task involves - `reference` (optional) - Relevant file paths, URLs, or documentation links - `priority` (required) - 1-5, where 1 is highest priority - `status` (required) - One of: "todo", "doing", "blocked", "done" - `details` (optional) - Additional implementation notes or constraints - `steps` (optional) - Array of specific sub-steps for complex tasks - `blockers` (optional) - Array of blocking reasons when status is "blocked" - `tags` (optional) - Array of category labels (e.g., "cli", "agents") - `files` (optional) - Array of files this task will modify - `depends_on` (optional) - Array of task IDs this task depends on - `created_at` (optional) - ISO 8601 timestamp - `updated_at` (optional) - ISO 8601 timestamp ## Logs and Output Logs are stored per project under: **Unix/Linux/macOS:** ``` ~/.looper/<project>-<hash>/ ``` **Windows:** ``` %USERPROFILE%\.looper\<project>-<hash>\ ``` Each run produces: - `<timestamp>-<pid>.jsonl` - Full JSONL trace - `<timestamp>-<pid>-<label>.last.json` - The last assistant message `looper tail --follow` watches the latest run in real time. ## Hooks Set the `hook_command` in config or use `--hook` / `LOOPER_HOOK` to run a script after each iteration. The hook receives these arguments: ``` <task_id> <status> <last_message_json_path> <label> ``` ## Parallelism Looper supports explicit parallel task execution while maintaining backward compatibility (sequential mode as default). The parallelism feature provides task-level parallelism (multiple tasks concurrently) and agent-level parallelism (multi-agent consensus per task). ### Enabling Parallelism Parallelism is **disabled by default** for backward compatibility. Enable it via the `[parallel]` config section or CLI flag: ```toml [parallel] enabled = true max_tasks = 4 ``` Or via CLI: ```bash looper run --parallel --max-tasks 4 ``` ### Configuration Options | Option | Type | Default | Description | |--------|------|---------|-------------| | `enabled` | bool | `false` | Enable parallel task execution | | `max_tasks` | int | `0` | Maximum concurrent tasks (0 = unlimited) | | `max_agents_per_task` | int | `1` | Agents per task for consensus (>1 = multi-agent) | | `strategy` | string | `"priority"` | Task selection strategy | | `fail_fast` | bool | `false` | Stop all on first failure | | `output_mode` | string | `"multiplexed"` | Output handling mode | ### Task Selection Strategies The `strategy` option determines how tasks are selected for parallel execution: - **`priority`** - Select highest priority tasks first (default) - **`dependency`** - Prioritize tasks that unblock other tasks - **`mixed`** - Balance priority and dependency awareness ### Output Modes The `output_mode` option controls how concurrent output is displayed: - **`multiplexed`** - Interleave output with task/agent ID prefixes (default) - **`buffered`** - Buffer output per task and display on completion - **`summary`** - Show only summaries without detailed output ### Environment Variables - `LOOPER_PARALLEL` - Enable parallel execution (1/0) - `LOOPER_MAX_TASKS` - Maximum concurrent tasks - `LOOPER_MAX_AGENTS_PER_TASK` - Agents per task - `LOOPER_PARALLEL_STRATEGY` - Task selection strategy - `LOOPER_FAIL_FAST` - Stop on first error (1/0) - `LOOPER_OUTPUT_MODE` - Output handling mode ### Multi-Agent Consensus Set `max_agents_per_task > 1` to run multiple agents per task for consensus/voting: ```toml [parallel] enabled = true max_tasks = 2 max_agents_per_task = 3 # Run 3 agents per task for consensus output_mode = "summary" # Recommended for multi-agent ``` When using multi-agent consensus: - Each agent runs independently with the same task - Results are collected and the first successful result is used - Use `output_mode = "summary"` to avoid verbose output ### Example Configurations **Basic parallel execution:** ```toml [parallel] enabled = true max_tasks = 4 # Run up to 4 tasks at once max_agents_per_task = 1 # Single agent per task strategy = "priority" # Select highest priority first fail_fast = false # Continue on errors output_mode = "multiplexed" # Interleave output with prefixes ``` **Multi-agent consensus:** ```toml [parallel] enabled = true max_tasks = 2 max_agents_per_task = 3 # Run 3 agents per task strategy = "mixed" output_mode = "summary" # Show only summaries ``` **Dependency-aware execution:** ```toml [parallel] enabled = true max_tasks = 0 # Unlimited (bounded by dependencies) strategy = "dependency" # Prioritize unblocking tasks fail_fast = false ``` ### Parallelism vs Workflows Looper offers two parallel execution mechanisms: 1. **`[parallel]` config** - Task-level parallelism within the traditional loop - Works with the traditional loop - Executes multiple independent tasks concurrently - Respects task dependencies - Configurable via `[parallel]` section 2. **`parallel` workflow** - A complete workflow implementation - Selected via `workflow = "parallel"` - Different execution model - Configured via `[workflows.parallel]` Both can be used together or independently. For most use cases, the `[parallel]` config is sufficient. ### When to Use Parallelism **Use parallelism when:** - You have many independent tasks - Tasks have minimal file conflicts - You want faster completion of large backlogs - You need multi-agent consensus for critical tasks **Avoid parallelism when:** - Tasks frequently modify the same files - Sequential task ordering is important - Debugging intermittent issues - Resource constraints (API rate limits, memory) ## Plugins Looper supports a plugin system for extending functionality with external agents and workflows. Plugins are external binaries that communicate via JSON-RPC over stdin/stdout. ### Plugin Commands ```bash # List installed plugins looper plugin list # Show plugin details looper plugin info claude # Create a new plugin skeleton looper plugin create my-agent --type agent # Validate a plugin looper plugin validate ./my-plugin # Install a plugin from git URL or local path looper plugin install https://github.com/user/looper-plugin looper plugin install ./my-plugin # Uninstall a plugin looper plugin uninstall my-agent ``` ### Plugin Validation The `looper plugin validate` command checks a plugin for correctness and completeness: ```bash # Basic validation looper plugin validate ./my-plugin # Strict validation mode (enforces all rules strictly) looper plugin validate ./my-plugin --strict # Quick validation (manifest only, skip binary checks) looper plugin validate ./my-plugin --quick # Verbose output (show all validation details) looper plugin validate ./my-plugin --verbose ``` **Validation Rules:** 1. **Manifest Checks** - `looper-plugin.toml` must exist in the plugin directory - Required fields: `name`, `version`, `category`, `plugin.binary` - Category must be `agent` or `workflow` - Plugin name must be alphanumeric with hyphens/underscores (cannot start with `-` or `_`) - Version should follow semver format (e.g., `1.0.0`) 2. **Category-Specific Configuration** - **Agent plugins** must have `[agent]` section with `type` field - **Workflow plugins** must have `[workflow]` section with `type` field 3. **Binary Validation** - Binary path must exist and be executable - Binary should respond to `--version` or `--help` flags - Warnings for script-based plugins (`.sh`, `.py`, etc.) 4. **Dependency Checks** - Required binaries in `PATH` are verified - API key dependencies are documented (warnings only) 5. **Capability Warnings** - Warns about dangerous capabilities (`can_execute_commands`, `can_access_network`) **Validation Output:** ``` Plugin: my-agent Status: VALID Warnings: - binary does not respond to --version or --help - plugin can execute commands (ensure you trust this plugin) ``` ### Plugin Creation The `looper plugin create` command scaffolds a new plugin with all required files: ```bash # Create an agent plugin (default) looper plugin create my-agent # Create a workflow plugin looper plugin create my-workflow --type workflow # Specify output directory looper plugin create my-agent --output ./plugins # Set metadata looper plugin create my-agent --author "Your Name" --license MIT --description "My custom agent" ``` **What Gets Created:** 1. **Directory Structure** ``` my-agent/ ├── looper-plugin.toml # Plugin manifest ├── README.md # Plugin documentation └── bin/ └── my-agent # Executable binary stub ``` 2. **Manifest File** (`looper-plugin.toml`) - Pre-populated with name, version, category - Category-specific configuration sections - Capability declarations - Default metadata fields (author, license, etc.) 3. **Binary Stub** - Executable shell script placeholder - JSON-RPC communication example - Response format for plugin type 4. **README.md** - Installation instructions - Usage examples - Plugin protocol documentation **Example: Creating a Custom Agent** ```bash # Create the plugin skeleton looper plugin create my-agent --type agent --author "Jane Doe" # The command outputs: # Plugin "my-agent" created successfully at ./my-agent # # Next steps: # 1. cd my-agent # 2. Edit the manifest (looper-plugin.toml) if needed # 3. Implement the plugin binary in bin/my-agent # 4. Test: looper plugin validate . # 5. Install: looper plugin install . cd my-agent # Edit the binary to implement your agent # The binary must accept JSON-RPC requests via stdin and respond via stdout # Validate before installing looper plugin validate . # Install to project plugins looper plugin install . # Now use your agent looper run --schedule my-agent ``` **Plugin Binary Protocol** Plugins communicate via JSON-RPC over stdin/stdout: **Agent Request:** ```json { "jsonrpc": "2.0", "id": 1, "method": "run", "params": { "prompt": "...", "context": {...} } } ``` **Agent Response:** ```json { "jsonrpc": "2.0", "id": 1, "result": { "task_id": "T001", "status": "done", "summary": "...", "files": ["..."], "blockers": [] } } ``` **Workflow Request:** ```json { "jsonrpc": "2.0", "id": 1, "method": "run", "params": { "config": {...}, "work_dir": "...", "todo_file": "..." } } ``` **Workflow Response:** ```json { "jsonrpc": "2.0", "id": 1, "result": { "success": true, "message": "..." } } ``` **Adapter Best Practices** See `AGENT_PLUGIN_ADAPTER.md` for a minimal adapter contract and reference approach that keeps JSON-RPC stdout clean and maps Looper prompts to a single vendor CLI/API invocation. ### Plugin Locations Plugins are searched in priority order: 1. **Project plugins** - `.looper/plugins/` (highest priority) 2. **User plugins** - `~/.looper/plugins/` 3. **Built-in plugins** - claude, codex, traditional (lowest priority) ### Plugin Manifest Plugins require a `looper-plugin.toml` manifest file: ```toml name = "my-agent" version = "1.0.0" category = "agent" description = "My custom AI agent" [plugin] binary = "./bin/my-agent" author = "Your Name" homepage = "https://github.com/user/my-plugin" license = "MIT" min_looper_version = "0.1.0" [agent] type = "my-agent" supports_streaming = false supports_tools = true default_prompt_format = "stdin" [capabilities] can_modify_files = true can_execute_commands = true can_access_network = false can_access_env = true ``` ### Plugin Categories - **agent** - AI agent plugins (claude, codex, or custom) - **workflow** - Workflow plugins (traditional, parallel, or custom) ### Plugin Capabilities Plugins declare their security permissions via a `[capabilities]` section in their manifest. This system allows looper to enforce security boundaries and provide transparency about what plugins can do. #### Capability Flags The `[capabilities]` section in `looper-plugin.toml` declares which operations the plugin can perform: ```toml [capabilities] can_modify_files = true # Can create, read, and modify files can_execute_commands = true # Can run shell commands can_access_network = false # Can make network requests can_access_env = true # Can read environment variables ``` #### Available Capabilities | Capability | Type | Description | Security Consideration | |------------|------|-------------|------------------------| | `can_modify_files` | bool | Plugin can create, modify, and delete files | **High risk** - allows plugin to change any file in the project | | `can_execute_commands` | bool | Plugin can run shell commands via `exec.Command` | **High risk** - allows arbitrary command execution | | `can_access_network` | bool | Plugin can make HTTP/network requests | **Medium risk** - could exfiltrate data or download malicious content | | `can_access_env` | bool | Plugin can read environment variables | **Medium risk** - could access API keys and secrets | **Note:** File reading is implied by `can_modify_files = true`. If a plugin only needs to read files, set `can_modify_files = true` and document that it only reads. #### How Looper Enforces Capabilities Looper uses a capability manager that enforces these security boundaries: 1. **Manifest Declaration** - Plugins must declare capabilities in their manifest 2. **Runtime Checks** - Looper checks capabilities before each sensitive operation 3. **Permission Levels** - Three permission levels are available: - `granted` - Capability is allowed - `denied` - Capability is explicitly blocked - `prompt` - User is prompted before allowing (future feature) 4. **Default Deny** - If a plugin doesn't declare a capability, it cannot use that operation #### Restricted Execution Helpers Looper provides helper types that automatically enforce capability checks: **`RestrictedCommandBuilder`** - For shell command execution with capability checks: ```go builder := plugin.NewRestrictedCommandBuilder(capManager, plugin) cmd, err := builder.Command(ctx, "git", "status") // Returns error if !can_execute_commands ``` **`RestrictedFileWriter`** - For file operations with capability checks: ```go writer := plugin.NewRestrictedFileWriter(capManager, plugin, baseDir) err := writer.WriteFile(ctx, "output.txt", data, 0644) // Returns error if !can_modify_files ``` #### Example Manifests **Minimal Agent (Read-Only):** ```toml [capabilities] can_modify_files = false can_execute_commands = false can_access_network = false can_access_env = true ``` **Full-Featured Agent:** ```toml [capabilities] can_modify_files = true can_execute_commands = true can_access_network = true can_access_env = true ``` **Network-Only Plugin:** ```toml [capabilities] can_modify_files = false can_execute_commands = false can_access_network = true # Only network access can_access_env = false ``` #### Security Best Practices 1. **Principle of Least Privilege** - Only enable capabilities your plugin actually needs 2. **Document Intent** - Clearly explain why each capability is needed in your plugin's README 3. **Warn Users** - Use validation warnings for dangerous capabilities (`can_execute_commands`, `can_modify_files`) 4. **Sandbox When Possible** - Consider running untrusted plugins in isolated environments 5. **Audit Regularly** - Review plugin manifests for excessive capability requests > **Important Security Notice**: Looper's capability system provides transparency but does not enforce OS-level restrictions. Plugins run with the same permissions as the looper process itself. For detailed information about the security model, threat model, and sandboxing limitations, see [SECURITY.md](SECURITY.md#plugin-security-model). **Security model summary:** - Capabilities are enforced only when plugins use looper-provided helper APIs. - There is no OS-level sandbox; plugins run with the same user permissions as looper. - Use a container, VM, or dedicated user account for untrusted plugins. #### Validation Warnings When validating plugins, looper warns about dangerous capabilities: ```bash $ looper plugin validate ./my-plugin Plugin: my-plugin Status: VALID Warnings: - plugin can execute commands (ensure you trust this plugin) - plugin can modify files (ensure you trust this plugin) ``` #### Capability Discovery To see what capabilities a plugin has declared: ```bash # View plugin info including capabilities looper plugin info my-plugin # The output shows the [capabilities] section from the manifest ``` ### Core Plugins Bundle Looper ships with a bundle of core plugins that are automatically extracted on first run. These core plugins provide the essential functionality out of the box without requiring manual installation. #### Core Plugin Extraction On the first run, looper automatically extracts core plugins to `~/.looper/plugins/`: - **claude** - Claude AI agent integration (built-in) - **codex** - Codex AI agent integration (built-in) - **traditional** - Traditional looper workflow (built-in) The extraction process: 1. Creates `~/.looper/plugins/` if it doesn't exist 2. Extracts each core plugin to its own subdirectory 3. Creates a README.md in each plugin directory with usage instructions #### Core Plugin vs User Plugin Priority When plugins with the same name exist in multiple locations, looper uses this priority order: 1. **Project plugins** (`.looper/plugins/`) - highest priority 2. **User plugins** (`~/.looper/plugins/`) 3. **Core plugins** (bundled) - lowest priority This means you can override a core plugin by installing your own version to either the project or user plugins directory. For example, to use a custom version of the claude plugin: ```bash # Install to project directory (overrides core plugin for this project only) looper plugin install ./my-claude-plugin # Or install to user directory (overrides core plugin globally) mkdir -p ~/.looper/plugins cp -r my-claude-plugin ~/.looper/plugins/claude ``` #### Core Plugin Manifests Core plugins include embedded manifests that define their capabilities. The manifests are used to register plugins with the plugin registry and are available via `GetCoreManifests()`. Core plugin capabilities: **Claude Agent:** - Supports streaming, tools, and MCP (Model Context Protocol) - Can modify files and execute commands - Default prompt format: stdin **Codex Agent:** - Supports streaming and tools - Can modify files and execute commands - Default prompt format: stdin **Traditional Workflow:** - Sequential task execution with review passes - Supports repair operations - Maximum 50 iterations (configurable) #### Checking Core Plugin Status To verify core plugins are installed: ```bash # List all plugins (includes core plugins) looper plugin list # Show info about a specific core plugin looper plugin info claude looper plugin info traditional ``` #### Core Plugins in the Architecture The core plugins bundle is implemented in `internal/coreplugins/`: - `bundle.go` - Extraction logic and manifest registry - `claude.go` - Claude agent manifest - `codex.go` - Codex agent manifest - `traditional.go` - Traditional workflow manifest Core plugins are extracted automatically via `EnsureExtracted()` which is called during looper initialization. The extraction is thread-safe and only happens once per session. ### See Also - [ARCHITECTURE.md](ARCHITECTURE.md) - Detailed architecture documentation - [CONTRIBUTING.md](CONTRIBUTING.md) - Contributing guidelines ## Workflows Looper supports a pluggable workflow system that allows you to customize how tasks are executed. Workflows define the execution strategy for processing tasks, from simple iterative loops to complex multi-stage processes. ### Available Workflows ```bash # List all available workflows looper workflow list # Describe a specific workflow looper workflow describe traditional ``` **Built-in workflows:** - **traditional** - The default looper loop with sequential task execution, review passes, and repair - **parallel** - Concurrent task execution with configurable concurrency limits and fail-fast option - **code-review** - Multi-stage code review workflow with analyze, security, and style stages - **incident-triage** - Incident classification, assignment, and notification workflow ### Running with a Workflow Select a workflow via the `--workflow` flag or config file: ```bash # Run with a specific workflow looper run --workflow parallel looper run --workflow code-review looper run --workflow incident-triage # Use workflow via config file (looper.toml) workflow = "parallel" ``` ### Workflow Configuration Each workflow can be configured via `[workflows.<name>]` sections in `looper.toml`. This allows you to customize workflow behavior without modifying code. #### Configuration Pattern ```toml # Select workflow workflow = "parallel" # Configure parallel workflow [workflows.parallel] max_concurrent = 5 # Maximum concurrent tasks fail_fast = true # Stop on first error # Configure code-review workflow [workflows.code-review] diff_path = "." # Path to review (default: ".") review_stages = ["analyze", "security", "style", "performance"] require_approval = true approval_file = ".looper/approval.txt" # Configure incident-triage workflow [workflows.incident-triage] severity_levels = ["critical", "high", "medium", "low"] auto_assign = true notify_slack = false slack_webhook = "https://hooks.slack.com/services/..." ``` #### Workflow-Specific Options **Parallel Workflow** (`parallel`) The parallel workflow executes multiple tasks concurrently with bounded concurrency. | Option | Type | Default | Description | |--------|------|---------|-------------| | `max_concurrent` | int | `3` | Maximum number of tasks to run simultaneously | | `fail_fast` | bool | `false` | Stop execution immediately if any task fails | Example: ```toml [workflows.parallel] max_concurrent = 5 # Run up to 5 tasks at once fail_fast = true # Stop on first error ``` Use cases: - Large backlogs of independent tasks - Projects where task isolation is guaranteed - Faster completion when tasks don't depend on each other **Code Review Workflow** (`code-review`) Multi-stage code review with git diff analysis and optional manual approval. | Option | Type | Default | Description | |--------|------|---------|-------------| | `diff_path` | string | `"."` | Path to run git diff from | | `review_stages` | array | `["analyze", "security", "style"]` | Review stages to execute | | `require_approval` | bool | `true` | Require manual approval after review | | `approval_file` | string | `".looper/approval.txt"` | Path to approval marker file | Built-in stages: - `analyze` - Code quality, structure, bugs, performance, testing - `security` - SQL injection, XSS, auth issues, sensitive data - `style` - Formatting, naming, comments, idiomatic patterns Example: ```toml [workflows.code-review] diff_path = "src/" # Review only src/ directory review_stages = ["analyze", "security", "style", "performance"] require_approval = true approval_file = ".looper/approval.txt" ``` **Custom Stages**: Add custom review stages by defining them in `review_stages` and configuring a stage-specific agent: ```toml # Add a custom stage agent [agents.agent_performance] binary = "claude" model = "" [workflows.code-review] review_stages = ["analyze", "security", "style", "performance"] ``` **Incident Triage Workflow** (`incident-triage`) Automated incident classification, assignment, and notification. | Option | Type | Default | Description | |--------|------|---------|-------------| | `severity_levels` | array | `["critical", "high", "medium", "low"]` | Available severity levels | | `auto_assign` | bool | `true` | Automatically assign based on severity | | `notify_slack` | bool | `false` | Send Slack notifications | | `slack_webhook` | string | `""` | Slack webhook URL | Default assignment rules: - `critical` → `oncall-senior` - `high` → `oncall` - `medium` → `team-backend` - `low` → `backlog` Example: ```toml [workflows.incident-triage] severity_levels = ["critical", "high", "medium", "low"] auto_assign = true notify_slack = true slack_webhook = "https://hooks.slack.com/services/YOUR/WEBHOOK/URL" ``` **Traditional Workflow** (`traditional`) The default looper workflow with sequential task execution. No specific configuration options - uses standard looper settings like `max_iterations`, `schedule`, `roles`, etc. #### Discovering Configuration Options Use `looper workflow describe` to see configuration options for a specific workflow: ```bash looper workflow describe parallel looper workflow describe code-review looper workflow describe incident-triage ``` ### Workflow Descriptions #### Traditional The default looper workflow with sequential task execution: - Processes one task per iteration - Runs review pass when tasks are exhausted - Supports task repair via agent - Configurable iteration schedules (agent selection, odd-even, round-robin) ```toml workflow = "traditional" ``` #### Parallel Execute tasks concurrently with bounded concurrency: ```toml workflow = "parallel" [workflows.parallel] max_concurrent = 3 # Max concurrent tasks (default: 3) fail_fast = false # Stop on first error (default: false) ``` Use parallel workflow for: - Independent tasks that can run simultaneously - Faster completion of large task backlogs - Projects where task isolation is guaranteed #### Code Review Multi-stage code review with git diff analysis: ```toml workflow = "code-review" [workflows.code-review] diff_path = "." # Path to review (default: ".") review_stages = ["analyze", "security", "style"] require_approval = true # Require manual approval approval_file = ".looper/approval.txt" ``` Stages: - **analyze** - Code quality, structure, bugs, performance, testing - **security** - Security vulnerabilities (SQL injection, XSS, auth issues) - **style** - Code formatting, naming, comments, idiomatic patterns Custom stages can be added to `review_stages` with stage-specific agent configuration: ```toml [agents.agent_performance] binary = "claude" model = "" [workflows.code-review] review_stages = ["analyze", "security", "style", "performance"] ``` #### Incident Triage Automated incident classification and assignment: ```toml workflow = "incident-triage" [workflows.incident-triage] severity_levels = ["critical", "high", "medium", "low"] auto_assign = true notify_slack = false slack_webhook = "https://hooks.slack.com/services/..." ``` The workflow: 1. Finds tasks tagged with "incident" 2. Classifies severity using an AI agent 3. Auto-assigns based on severity (oncall-senior, oncall, team-backend, backlog) 4. Optionally sends Slack notifications Assignment rules: - **critical** → oncall-senior - **high** → oncall - **medium** → team-backend - **low** → backlog ### Custom Workflow Plugins Workflows can be implemented as external plugins. Create a workflow plugin: ```bash looper plugin create my-workflow --type workflow ``` The plugin will be discovered automatically when placed in: - `.looper/plugins/` (project-level) - `~/.looper/plugins/` (user-level) Workflow plugin manifest (`looper-plugin.toml`): ```toml name = "my-workflow" version = "1.0.0" category = "workflow" [plugin] binary = "./bin/my-workflow" [workflow] type = "my-workflow" supports_approval = false ``` See the Plugins section for more details on plugin development. ### Workflow System Architecture The workflow system uses a registry pattern similar to agents: ```go // Workflow interface type Workflow interface { Run(context.Context) error Description() string } // Register workflows in init() func init() { workflows.Register(workflows.WorkflowType("my-workflow"), MyWorkflowFactory) } ``` Built-in workflows are in `internal/workflows/`: - `types.go` - Core interfaces and types - `registry.go` - Workflow registration and discovery - `traditional.go` - Default looper loop - `parallel.go` - Concurrent execution - `code_review.go` - Multi-stage code review - `incident_triage.go` - Incident response - `plugin_workflow.go` - Plugin workflow wrapper ## Release Workflow The `looper push` command runs a release workflow via an AI agent: ```bash looper push [options] ``` The push command will: 1. Verify git is available and you're in a git repository 2. Check for `gh` (GitHub CLI) and verify authentication 3. Offer to create a GitHub repository if no remote exists and `gh` is available 4. Run an agent with instructions to follow the release-runbook skill at `skills/release-runbook/SKILL.md` 5. Execute the release workflow: run tests, bump version, tag, push, create GitHub release, update Homebrew formula **Options:** - `--agent <agent>` - Agent to use for release workflow (default: codex) - `-y` - Skip confirmation prompts **Example:** ```bash # Run the release workflow with codex looper push # Run with claude and skip confirmation looper push --agent claude -y ``` The agent will follow the workflow steps defined in `skills/release-runbook/SKILL.md`, which includes: - Checking repo state and verifying tools - Determining the appropriate version bump - Running tests - Updating version references in files - Committing release changes - Creating and pushing git tags - Creating GitHub releases - Updating Homebrew formulas (if present) The output is logged to the log directory with label `push` for traceability. ## Log Cleanup The `looper clean` command removes old log runs by age or count, with dry-run support and clear summaries. ```bash # Dry run - see what would be deleted looper clean --dry-run # Keep only the N most recent runs looper clean --keep 10 # Delete logs older than a duration looper clean --age 7d # 7 days looper clean --age 24h # 24 hours looper clean --age 30m # 30 minutes # Combine filters looper clean --keep 5 --age 30d ``` **Options:** - `--dry-run` - Show what would be deleted without actually deleting - `--keep N` - Keep N most recent runs (0 = delete all) - `--age DURATION` - Delete logs older than duration (e.g., `7d`, `24h`, `30m`) The command will: 1. Find all log runs in the log directory 2. Group files by run ID (timestamp-pid) 3. Filter based on your criteria 4. Show a summary of what will be deleted 5. Ask for confirmation before deleting (unless using filters or `--dry-run`) ## Git Behavior Looper can automatically initialize a git repository when `git_init` is enabled: ```bash looper run --git-init # Or via config: git_init = true # Or via env: LOOPER_GIT_INIT=1 ``` When `git init` is enabled and git is available, Looper runs `git init` before the bootstrap phase if the current directory is not already a git repository. When git is available, Looper uses it to resolve the project root for log grouping. ## Dev Mode (Prompt Development) For prompt development and debugging, set `LOOPER_PROMPT_MODE=dev` to enable: - `--prompt-dir` - Override the prompt directory - `--print-prompt` - Print rendered prompts before running Example: ```bash LOOPER_PROMPT_MODE=dev looper run --print-prompt --prompt-dir ./my-prompts ``` **Note:** Dev mode is intentionally hidden from normal usage and not shown in help output unless enabled. You can also set `LOOPER_PROMPT_DIR` and `LOOPER_PRINT_PROMPT=1` instead of flags. ## Building ```bash make build # Build for current platform make test # Run tests make smoke # Run smoke test ``` ## Architecture Looper is written in Go with a clean, modular architecture: - `cmd/looper` - CLI entrypoint - `internal/config` - Multi-source configuration loading (TOML, env, flags) - `internal/prompts` - Prompt store and rendering - `internal/todo` - Task file types and validation - `internal/loop` - Orchestration state machine - `internal/agents` - Modular agent system with registry pattern - `internal/coreplugins` - Bundled core plugins (claude, codex, traditional) with auto-extraction - `internal/parsers` - Plugin-based parser system for agent output - `internal/workflows` - Pluggable workflow system (traditional, parallel, code-review, incident-triage) - `internal/plugin` - Plugin management and discovery - `internal/logging` - JSONL logging with charmbracelet/log console output - `internal/hooks` - Hook invocation - `internal/utils` - Shared utilities (platform, scheduling) See [ARCHITECTURE.md](ARCHITECTURE.md) for detailed design documentation. ### Agent Registry Looper uses a registry-based agent system that allows custom agent types to be registered and used throughout the system. Built-in agents (`codex`, `claude`) are registered at initialization, but additional agent types can be registered programmatically. To register a custom agent type, use the `agents.RegisterAgent()` function: ```go import "github.com/nibzard/looper-go/internal/agents" // Define an agent type const AgentTypeOpenCode agents.AgentType = "opencode" // Register the agent factory agents.RegisterAgent(AgentTypeOpenCode, func(cfg agents.Config) (agents.Agent, error) { return &myOpenCodeAgent{cfg: cfg}, nil }) ``` Once registered, the agent type can be used in schedules, step agents (repair/review/bootstrap), and configuration files. The `doctor` command will automatically validate and show registered agent types. ### Parser Configuration Agents can use custom parsers for extracting summaries from their output. Parsers are configured in `looper.toml`: ```toml [agents.claude] binary = "claude" parser = "builtin:claude" # Use built-in Go parser [agents.codex] binary = "codex" parser = "codex_parser.py" # Use bundled Python parser [agents.custom] binary = "custom-agent" parser = "~/.looper/parsers/custom.py" # Use custom parser ``` Parser search paths: 1. Absolute path (if starts with `/` or `~/`) 2. `./looper-parsers/` (project-level) 3. `~/.looper/parsers/` (user-level) 4. Bundled parsers (`claude_parser.py`, `codex_parser.py`, `opencode_parser.py`) ## Migration from Shell Version See [MIGRATION.md](MIGRATION.md) for details on migrating from the previous shell script implementation.

正在加载仓库 README…

这个仓库中的 skills

软件开发工程师软件质量保证分析师与测试员

2 个职业分类 · 已分类 57%

skill

职业分类

描述

更新

looper-go-dev

软件开发工程师

Looper-Go development workflows: build, test, run, debug, and manage the looper-go project. Use when working on looper-go codebase, running tests, building binaries, or debugging issues.

原文语言：英语

2026-03-09

looper-live-test

软件质量保证分析师与测试员

Run a live Looper-Go smoke test by creating a temporary project (PROJECT.md + to-do.json), executing looper for one iteration, observing output, and reporting results. Use when verifying looper-go behavior or demonstrating a live run.

原文语言：英语

2026-02-13

looper-go-dev

软件开发工程师

Looper-Go development workflows: build, test, run, debug, and manage the looper-go project. Use when working on looper-go codebase, running tests, building binaries, or debugging issues.

原文语言：英语

2026-01-23

looper-live-test

软件质量保证分析师与测试员

原文语言：英语

2026-01-23

release-runbook

未分类

Release preparation and publish workflow: run tests, bump version, tag, push, and create a GitHub release (and update Homebrew formula if present). Use when asked to cut a release, bump version, create tags, or publish a release.

原文语言：英语

2026-01-23

git-conventional-commit

未分类

Create clean git commits using the Conventional Commits format (type(scope): summary). Use when Codex needs to stage changes, craft commit messages, and record completed work in a Git repository.

原文语言：英语

2026-01-21

todo-json-manager

未分类

Manage project task lists stored in to-do.json and to-do.schema.json, including bootstrapping a backlog, selecting the highest priority task with jq, and updating task statuses. Use when Codex needs to read or update a JSON task list.

原文语言：英语

2026-01-21

已展示 7 / 7 个已收集 Skill。

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/nibzard/looper-go.