refactor(ux): consolidate BMAD skills, update design system, and clean up Prisma generated client

This commit is contained in:
Sepehr Ramezani
2026-04-19 19:21:27 +02:00
parent 5296c4da2c
commit 25529a24b8
2476 changed files with 127934 additions and 101962 deletions

View File

@@ -0,0 +1,422 @@
# Ralph Driver Interface Contract
## Overview
The Ralph loop loads a platform driver by sourcing `ralph/drivers/${PLATFORM_DRIVER}.sh`
inside `load_platform_driver()` (`ralph_loop.sh` line 296). The `PLATFORM_DRIVER` variable
defaults to `"claude-code"` and can be overridden via `.ralphrc`.
After sourcing, `ralph_loop.sh` immediately calls three functions to populate core globals:
1. `driver_valid_tools` -- populates `VALID_TOOL_PATTERNS`
2. `driver_cli_binary` -- stored in `CLAUDE_CODE_CMD`
3. `driver_display_name` -- stored in `DRIVER_DISPLAY_NAME`
**File naming convention:** `${PLATFORM_DRIVER}.sh` (e.g., `claude-code.sh`, `codex.sh`).
**Scope:** This documents the sourceable driver contract used by `ralph_loop.sh`. Helper
scripts like `cursor-agent-wrapper.sh` are out of scope.
**Calling conventions:**
- Data is returned via stdout (`echo`).
- Booleans are returned via exit status (`0` = true, `1` = false).
- Some functions mutate global arrays as side effects.
---
## Required Hooks
Called unconditionally by `ralph_loop.sh` with no `declare -F` guard or default stub.
Omitting any of these will break the loop at runtime.
### `driver_name()`
```bash
driver_name()
```
No arguments. Echo a short lowercase identifier (e.g., `"claude-code"`, `"codex"`).
Used at line 2382 to gate platform-specific logic.
### `driver_display_name()`
```bash
driver_display_name()
```
No arguments. Echo a human-readable name (e.g., `"Claude Code"`, `"OpenAI Codex"`).
Stored in `DRIVER_DISPLAY_NAME`, used in log messages and tmux pane titles.
### `driver_cli_binary()`
```bash
driver_cli_binary()
```
No arguments. Echo the CLI executable name or resolved path (e.g., `"claude"`, `"codex"`).
Stored in `CLAUDE_CODE_CMD`. Most drivers return a static string; cursor resolves
dynamically.
### `driver_valid_tools()`
```bash
driver_valid_tools()
```
No arguments. Must populate the global `VALID_TOOL_PATTERNS` array with the platform's
recognized tool name patterns. Used by `validate_allowed_tools()`.
### `driver_build_command(prompt_file, loop_context, session_id)`
```bash
driver_build_command "$prompt_file" "$loop_context" "$session_id"
```
Three string arguments:
| Argument | Description |
|----------|-------------|
| `$1` prompt_file | Path to the prompt file (e.g., `.ralph/PROMPT.md`) |
| `$2` loop_context | Context string for session continuity (may be empty) |
| `$3` session_id | Session ID for resume (empty string = new session) |
Must populate the global `CLAUDE_CMD_ARGS` array with the complete CLI command and
arguments. Return `0` on success, `1` on failure (e.g., prompt file not found).
**Reads globals:** `CLAUDE_OUTPUT_FORMAT`, `CLAUDE_PERMISSION_MODE` (claude-code only),
`CLAUDE_ALLOWED_TOOLS` (claude-code only), `CLAUDE_USE_CONTINUE`.
---
## Optional Overrides with Loop Defaults
`ralph_loop.sh` defines default stubs at lines 284 and 288. All existing drivers override
them, but a minimal driver can rely on the defaults.
### `driver_supports_tool_allowlist()`
```bash
driver_supports_tool_allowlist()
```
No arguments. Return `0` if the driver supports `--allowedTools` filtering, `1` otherwise.
**Default:** returns `1` (false). Currently only `claude-code` returns `0`.
### `driver_permission_denial_help()`
```bash
driver_permission_denial_help()
```
No arguments. Print platform-specific troubleshooting guidance when the loop detects a
permission denial.
**Reads:** `RALPHRC_FILE`, `DRIVER_DISPLAY_NAME`.
**Default:** generic guidance text.
---
## Optional Capability Hooks
Guarded by `declare -F` checks or wrapper functions in `ralph_loop.sh` (lines 1917-1954,
1576-1583). Safe to omit -- documented fallback behavior applies.
### `driver_supports_sessions()`
```bash
driver_supports_sessions()
```
No arguments. Return `0` if the driver supports session resume, `1` otherwise.
**If not defined:** assumed true (`0`).
Implemented by all 5 drivers; `copilot` returns `1`.
### `driver_supports_live_output()`
```bash
driver_supports_live_output()
```
No arguments. Return `0` if the driver supports structured streaming output (stream-json
or JSONL), `1` otherwise.
**If not defined:** assumed true (`0`).
`copilot` returns `1`; all others return `0`.
### `driver_prepare_live_command()`
```bash
driver_prepare_live_command()
```
No arguments. Transform `CLAUDE_CMD_ARGS` into `LIVE_CMD_ARGS` for streaming mode.
**If not defined:** `LIVE_CMD_ARGS` is copied from `CLAUDE_CMD_ARGS` unchanged.
| Driver | Behavior |
|--------|----------|
| claude-code | Replaces `json` with `stream-json` and adds `--verbose --include-partial-messages` |
| codex | Copies as-is (output is already suitable) |
| opencode | Copies as-is (output is already suitable) |
| cursor | Replaces `json` with `stream-json` |
### `driver_stream_filter()`
```bash
driver_stream_filter()
```
No arguments. Echo a `jq` filter expression that transforms raw streaming events into
displayable text.
**If not defined:** returns `"empty"` (no output).
Each driver has a platform-specific filter; `copilot` returns `'.'` (passthrough).
### `driver_extract_session_id_from_output(output_file)`
```bash
driver_extract_session_id_from_output "$output_file"
```
One argument: path to the CLI output log file. Echo the extracted session ID.
Tried first in the session save chain before the generic `jq` extractor. Only `opencode`
implements this (uses `sed` to extract from a `"session"` JSON object).
### `driver_fallback_session_id(output_file)`
```bash
driver_fallback_session_id "$output_file"
```
One argument: path to the output file (caller passes it at line 1583; the only
implementation in `opencode` ignores it).
Last-resort session ID recovery when both driver-specific and generic extractors fail.
Only `opencode` implements this (queries `opencode session list --format json`).
---
## Conventional Metadata Hooks
Present in every driver but NOT called by `ralph_loop.sh`. Consumed by bmalph's TypeScript
doctor/preflight checks in `src/platform/`. A new driver should implement these for
`bmalph doctor` compatibility.
### `driver_min_version()`
```bash
driver_min_version()
```
No arguments. Echo the minimum required CLI version as a semver string.
### `driver_check_available()`
```bash
driver_check_available()
```
No arguments. Return `0` if the CLI binary is installed and reachable, `1` otherwise.
---
## Global Variables
### Written by drivers
| Variable | Written by | Type | Description |
|----------------------|----------------------------------|-------|------------------------------------------------------|
| `VALID_TOOL_PATTERNS`| `driver_valid_tools()` | array | Valid tool name patterns for allowlist validation |
| `CLAUDE_CMD_ARGS` | `driver_build_command()` | array | Complete CLI command with all arguments |
| `LIVE_CMD_ARGS` | `driver_prepare_live_command()` | array | Modified command for live streaming |
### Read by drivers (set by ralph_loop.sh or .ralphrc)
| Variable | Used in | Description |
|-------------------------|--------------------------------------------|------------------------------------------------|
| `CLAUDE_OUTPUT_FORMAT` | `driver_build_command()` | `"json"` or `"text"` |
| `CLAUDE_PERMISSION_MODE`| `driver_build_command()` (claude-code) | Permission mode flag, default `"bypassPermissions"` |
| `CLAUDE_ALLOWED_TOOLS` | `driver_build_command()` (claude-code) | Comma-separated tool allowlist |
| `CLAUDE_USE_CONTINUE` | `driver_build_command()` | `"true"` or `"false"`, gates session resume |
| `RALPHRC_FILE` | `driver_permission_denial_help()` | Path to `.ralphrc` config file |
| `DRIVER_DISPLAY_NAME` | `driver_permission_denial_help()` | Human-readable driver name |
### Environment globals (cursor-specific)
| Variable | Used in | Description |
|---------------|--------------------------------------|-------------------------------------|
| `OS`, `OSTYPE`| `driver_running_on_windows()` | OS detection |
| `LOCALAPPDATA`| `driver_localappdata_cli_binary()` | Windows local app data path |
| `PATH` | `driver_find_windows_path_candidate()`| Manual PATH scanning on Windows |
### Set by ralph_loop.sh from driver output
| Variable | Source | Description |
|----------------------|-------------------------|-------------------------------|
| `CLAUDE_CODE_CMD` | `driver_cli_binary()` | CLI binary name/path |
| `DRIVER_DISPLAY_NAME`| `driver_display_name()` | Human-readable display name |
---
## Capability Matrix
| Capability | claude-code | codex | opencode | copilot | cursor |
|---------------------------------------------------------|:-----------:|:-----------:|:-----------:|:-----------:|:-----------:|
| Tool allowlist (`driver_supports_tool_allowlist`) | yes | no | no | no | no |
| Session continuity (`driver_supports_sessions`) | yes | yes | yes | no | yes |
| Structured live output (`driver_supports_live_output`) | yes | yes | yes | no | yes |
| Live command transform (`driver_prepare_live_command`) | transform | passthrough | passthrough | -- | transform |
| Stream filter (`driver_stream_filter`) | complex jq | JSONL select| JSONL select| passthrough | complex jq |
| Custom session extraction (`driver_extract_session_id_from_output`) | -- | -- | yes | -- | -- |
| Fallback session lookup (`driver_fallback_session_id`) | -- | -- | yes | -- | -- |
| Dynamic binary resolution (`driver_cli_binary`) | static | static | static | static | dynamic |
---
## Creating a New Driver
### Minimal driver skeleton
```bash
#!/usr/bin/env bash
# ralph/drivers/my-platform.sh
# Driver for My Platform CLI
#
# Sourced by ralph_loop.sh via load_platform_driver().
# PLATFORM_DRIVER must be set to "my-platform" in .ralphrc.
# ---------------------------------------------------------------------------
# Required hooks (5) -- omitting any of these breaks the loop
# ---------------------------------------------------------------------------
# Short lowercase identifier used to gate platform-specific logic.
driver_name() {
echo "my-platform"
}
# Human-readable name for log messages and tmux pane titles.
driver_display_name() {
echo "My Platform"
}
# CLI executable name or resolved path.
driver_cli_binary() {
echo "my-platform"
}
# Populate VALID_TOOL_PATTERNS with recognized tool name patterns.
# Used by validate_allowed_tools() to check allowlist entries.
driver_valid_tools() {
VALID_TOOL_PATTERNS=(
"Read"
"Write"
"Edit"
"Bash"
# Add your platform's tool patterns here
)
}
# Build the complete CLI command array.
# $1 = prompt_file Path to .ralph/PROMPT.md
# $2 = loop_context Context string for session continuity (may be empty)
# $3 = session_id Session ID for resume (empty = new session)
driver_build_command() {
local prompt_file="$1"
local loop_context="$2"
local session_id="$3"
if [[ ! -f "$prompt_file" ]]; then
return 1
fi
CLAUDE_CMD_ARGS=(
"my-platform"
"--prompt" "$prompt_file"
"--output-format" "${CLAUDE_OUTPUT_FORMAT:-json}"
)
# Append session resume flag if continuing a session
if [[ "$CLAUDE_USE_CONTINUE" == "true" && -n "$session_id" ]]; then
CLAUDE_CMD_ARGS+=("--session" "$session_id")
fi
# Append context if provided
if [[ -n "$loop_context" ]]; then
CLAUDE_CMD_ARGS+=("--context" "$loop_context")
fi
return 0
}
# ---------------------------------------------------------------------------
# Optional overrides (2) -- loop provides default stubs
# ---------------------------------------------------------------------------
# Return 0 if the platform supports --allowedTools filtering, 1 otherwise.
driver_supports_tool_allowlist() {
return 1
}
# Print troubleshooting guidance on permission denial.
driver_permission_denial_help() {
echo "Permission denied. Check that $DRIVER_DISPLAY_NAME has the required permissions."
echo "See $RALPHRC_FILE for configuration options."
}
# ---------------------------------------------------------------------------
# Metadata hooks (2) -- used by bmalph doctor, not called by ralph_loop.sh
# ---------------------------------------------------------------------------
# Minimum required CLI version (semver).
driver_min_version() {
echo "1.0.0"
}
# Return 0 if the CLI binary is installed and reachable, 1 otherwise.
driver_check_available() {
command -v my-platform &>/dev/null
}
```
### Checklist
- [ ] All 5 required hooks implemented (`driver_name`, `driver_display_name`,
`driver_cli_binary`, `driver_valid_tools`, `driver_build_command`)
- [ ] `driver_valid_tools` populates `VALID_TOOL_PATTERNS` with your platform's tool names
- [ ] `driver_build_command` handles all three arguments correctly
(`prompt_file`, `loop_context`, `session_id`)
- [ ] `driver_check_available` returns `0` only when the CLI is installed
- [ ] File named `${platform_id}.sh` matching the `PLATFORM_DRIVER` value in `.ralphrc`
- [ ] Register corresponding platform definition in `src/platform/` for bmalph CLI integration
- [ ] Tested with `bmalph doctor`
---
## Session ID Recovery Chain
When the loop needs to persist a session ID for resume, it follows a three-step priority
chain (`ralph_loop.sh` lines 1574-1588):
1. **`driver_extract_session_id_from_output($output_file)`** -- Driver-specific extraction.
If the function exists (`declare -F` guard) and echoes a non-empty string, that value
is used. Only `opencode` implements this (uses `sed` to extract from a `"session"` JSON
object).
2. **`extract_session_id_from_output($output_file)`** -- Generic `jq` extractor from
`response_analyzer.sh`. Searches the output file for `.sessionId`,
`.metadata.session_id`, and `.session_id` in that order.
3. **`driver_fallback_session_id($output_file)`** -- CLI-based last-resort recovery. If the
function exists and the previous steps produced nothing, this is called. Only `opencode`
implements this (queries `opencode session list --format json`).
The first step that returns a non-empty string wins. If all three steps fail, no session ID
is saved and the next iteration starts a fresh session.

187
.ralph/drivers/claude-code.sh Executable file
View File

@@ -0,0 +1,187 @@
#!/bin/bash
# Claude Code driver for Ralph
# Provides platform-specific CLI invocation logic
# Driver identification
driver_name() {
echo "claude-code"
}
driver_display_name() {
echo "Claude Code"
}
driver_cli_binary() {
echo "claude"
}
driver_min_version() {
echo "2.0.76"
}
# Check if the CLI binary is available
driver_check_available() {
command -v "$(driver_cli_binary)" &>/dev/null
}
# Valid tool patterns for --allowedTools validation
# Sets the global VALID_TOOL_PATTERNS array
driver_valid_tools() {
VALID_TOOL_PATTERNS=(
"Write"
"Read"
"Edit"
"MultiEdit"
"Glob"
"Grep"
"Task"
"TodoWrite"
"WebFetch"
"WebSearch"
"AskUserQuestion"
"EnterPlanMode"
"ExitPlanMode"
"Bash"
"Bash(git *)"
"Bash(npm *)"
"Bash(bats *)"
"Bash(python *)"
"Bash(node *)"
"NotebookEdit"
)
}
driver_supports_tool_allowlist() {
return 0
}
driver_permission_denial_help() {
echo " 1. Edit $RALPHRC_FILE and keep CLAUDE_PERMISSION_MODE=bypassPermissions for unattended Claude Code loops"
echo " 2. If Claude was denied on an interactive approval step, ALLOWED_TOOLS will not fix it"
echo " 3. If Claude was denied on a normal tool, update ALLOWED_TOOLS to include the required tools"
echo " 4. Common ALLOWED_TOOLS patterns:"
echo " - Bash - All shell commands"
echo " - Bash(node *) - All Node.js commands"
echo " - Bash(npm *) - All npm commands"
echo " - Bash(pnpm *) - All pnpm commands"
echo " - AskUserQuestion - Allow interactive clarification when you want pauses"
echo ""
echo "After updating $RALPHRC_FILE:"
echo " bash .ralph/ralph_loop.sh --reset-session # Clear stale session state"
echo " bmalph run # Restart the loop"
}
# Build the CLI command arguments
# Populates global CLAUDE_CMD_ARGS array
# Parameters:
# $1 - prompt_file: path to the prompt file
# $2 - loop_context: context string for session continuity
# $3 - session_id: session ID for resume (empty for new session)
driver_build_command() {
local prompt_file=$1
local loop_context=$2
local session_id=$3
local resolved_permission_mode="${CLAUDE_PERMISSION_MODE:-bypassPermissions}"
# Note: We do NOT use --dangerously-skip-permissions here. Tool permissions
# are controlled via --allowedTools from CLAUDE_ALLOWED_TOOLS in .ralphrc.
# This preserves the permission denial circuit breaker (Issue #101).
CLAUDE_CMD_ARGS=("$(driver_cli_binary)")
if [[ ! -f "$prompt_file" ]]; then
echo "ERROR: Prompt file not found: $prompt_file" >&2
return 1
fi
# Output format
if [[ "$CLAUDE_OUTPUT_FORMAT" == "json" ]]; then
CLAUDE_CMD_ARGS+=("--output-format" "json")
fi
# Prevent interactive approval flows from blocking unattended -p loops.
CLAUDE_CMD_ARGS+=("--permission-mode" "$resolved_permission_mode")
# Allowed tools
if [[ -n "$CLAUDE_ALLOWED_TOOLS" ]]; then
CLAUDE_CMD_ARGS+=("--allowedTools")
local IFS=','
read -ra tools_array <<< "$CLAUDE_ALLOWED_TOOLS"
for tool in "${tools_array[@]}"; do
tool=$(echo "$tool" | sed 's/^[[:space:]]*//;s/[[:space:]]*$//')
if [[ -n "$tool" ]]; then
CLAUDE_CMD_ARGS+=("$tool")
fi
done
fi
# Session resume
# IMPORTANT: Use --resume with explicit session ID instead of --continue.
# --continue resumes the "most recent session in current directory" which
# can hijack active Claude Code sessions. --resume with a specific session ID
# ensures we only resume Ralph's own sessions. (Issue #151)
if [[ "$CLAUDE_USE_CONTINUE" == "true" && -n "$session_id" ]]; then
CLAUDE_CMD_ARGS+=("--resume" "$session_id")
fi
# Loop context as system prompt
if [[ -n "$loop_context" ]]; then
CLAUDE_CMD_ARGS+=("--append-system-prompt" "$loop_context")
fi
# Prompt content
local prompt_content
prompt_content=$(cat "$prompt_file")
CLAUDE_CMD_ARGS+=("-p" "$prompt_content")
}
# Whether this driver supports session continuity
driver_supports_sessions() {
return 0 # true
}
# Claude Code supports stream-json live output.
driver_supports_live_output() {
return 0 # true
}
# Prepare command arguments for live stream-json output.
driver_prepare_live_command() {
LIVE_CMD_ARGS=()
local skip_next=false
for arg in "${CLAUDE_CMD_ARGS[@]}"; do
if [[ "$skip_next" == "true" ]]; then
LIVE_CMD_ARGS+=("stream-json")
skip_next=false
elif [[ "$arg" == "--output-format" ]]; then
LIVE_CMD_ARGS+=("$arg")
skip_next=true
else
LIVE_CMD_ARGS+=("$arg")
fi
done
if [[ "$skip_next" == "true" ]]; then
return 1
fi
LIVE_CMD_ARGS+=("--verbose" "--include-partial-messages")
}
# Stream filter for raw Claude stream-json events.
driver_stream_filter() {
echo '
if .type == "stream_event" then
if .event.type == "content_block_delta" and .event.delta.type == "text_delta" then
.event.delta.text
elif .event.type == "content_block_start" and .event.content_block.type == "tool_use" then
"\n\n⚡ [" + .event.content_block.name + "]\n"
elif .event.type == "content_block_stop" then
"\n"
else
empty
end
else
empty
end'
}

101
.ralph/drivers/codex.sh Executable file
View File

@@ -0,0 +1,101 @@
#!/bin/bash
# OpenAI Codex driver for Ralph
# Provides platform-specific CLI invocation logic for Codex
driver_name() {
echo "codex"
}
driver_display_name() {
echo "OpenAI Codex"
}
driver_cli_binary() {
echo "codex"
}
driver_min_version() {
echo "0.1.0"
}
driver_check_available() {
command -v "$(driver_cli_binary)" &>/dev/null
}
# Codex tool names differ from Claude Code
driver_valid_tools() {
VALID_TOOL_PATTERNS=(
"shell"
"read_file"
"write_file"
"edit_file"
"list_directory"
"search_files"
)
}
driver_supports_tool_allowlist() {
return 1
}
driver_permission_denial_help() {
echo " - $DRIVER_DISPLAY_NAME uses its native sandbox and approval model."
echo " - ALLOWED_TOOLS in $RALPHRC_FILE is ignored for this driver."
echo " - Ralph already runs Codex with --sandbox workspace-write."
echo " - Review Codex approval settings, then restart the loop."
}
# Build Codex CLI command
# Codex uses: codex exec [resume <id>] --json "prompt"
driver_build_command() {
local prompt_file=$1
local loop_context=$2
local session_id=$3
CLAUDE_CMD_ARGS=("$(driver_cli_binary)" "exec")
if [[ ! -f "$prompt_file" ]]; then
echo "ERROR: Prompt file not found: $prompt_file" >&2
return 1
fi
# JSON output
CLAUDE_CMD_ARGS+=("--json")
# Sandbox mode - workspace write access
CLAUDE_CMD_ARGS+=("--sandbox" "workspace-write")
# Session resume — gated on CLAUDE_USE_CONTINUE to respect --no-continue flag
if [[ "$CLAUDE_USE_CONTINUE" == "true" && -n "$session_id" ]]; then
CLAUDE_CMD_ARGS+=("resume" "$session_id")
fi
# Build prompt with context
local prompt_content
prompt_content=$(cat "$prompt_file")
if [[ -n "$loop_context" ]]; then
prompt_content="$loop_context
$prompt_content"
fi
CLAUDE_CMD_ARGS+=("$prompt_content")
}
driver_supports_sessions() {
return 0 # true - Codex supports session resume
}
# Codex JSONL output is already suitable for live display.
driver_supports_live_output() {
return 0 # true
}
driver_prepare_live_command() {
LIVE_CMD_ARGS=("${CLAUDE_CMD_ARGS[@]}")
}
# Codex outputs JSONL events
driver_stream_filter() {
echo 'select(.type == "item.completed" and .item.type == "agent_message") | (.item.text // ([.item.content[]? | select(.type == "output_text") | .text] | join("\n")) // empty)'
}

105
.ralph/drivers/copilot.sh Executable file
View File

@@ -0,0 +1,105 @@
#!/bin/bash
# GitHub Copilot CLI driver for Ralph (EXPERIMENTAL)
# Provides platform-specific CLI invocation logic for Copilot CLI.
#
# Known limitations:
# - No session continuity (session IDs not capturable from -p output)
# - No structured output (plain text only, no --json flag)
# - Coarse tool permissions (only shell, shell(git:*), shell(npm:*), write)
# - CLI is new (GA Feb 25, 2026) — scripting interface may change
driver_name() {
echo "copilot"
}
driver_display_name() {
echo "GitHub Copilot CLI"
}
driver_cli_binary() {
echo "copilot"
}
driver_min_version() {
echo "0.0.418"
}
driver_check_available() {
command -v "$(driver_cli_binary)" &>/dev/null
}
# Copilot CLI tool names
driver_valid_tools() {
VALID_TOOL_PATTERNS=(
"shell"
"shell(git:*)"
"shell(npm:*)"
"write"
)
}
driver_supports_tool_allowlist() {
return 1
}
driver_permission_denial_help() {
echo " - $DRIVER_DISPLAY_NAME uses its own autonomy and approval controls."
echo " - ALLOWED_TOOLS in $RALPHRC_FILE is ignored for this driver."
echo " - Ralph already runs Copilot with --no-ask-user for unattended mode."
echo " - Review Copilot CLI permissions, then restart the loop."
}
# Build Copilot CLI command
# Context is prepended to the prompt (same pattern as Codex driver).
# Uses --autopilot --yolo for autonomous mode, -s to strip stats, -p for prompt.
driver_build_command() {
local prompt_file=$1
local loop_context=$2
# $3 (session_id) is intentionally ignored — Copilot CLI does not
# expose session IDs in -p output, so resume is not possible.
CLAUDE_CMD_ARGS=("$(driver_cli_binary)")
if [[ ! -f "$prompt_file" ]]; then
echo "ERROR: Prompt file not found: $prompt_file" >&2
return 1
fi
# Autonomous execution flags
CLAUDE_CMD_ARGS+=("--autopilot" "--yolo")
# Limit auto-continuation loops
CLAUDE_CMD_ARGS+=("--max-autopilot-continues" "50")
# Disable interactive prompts
CLAUDE_CMD_ARGS+=("--no-ask-user")
# Strip stats for cleaner output
CLAUDE_CMD_ARGS+=("-s")
# Build prompt with context prepended
local prompt_content
prompt_content=$(cat "$prompt_file")
if [[ -n "$loop_context" ]]; then
prompt_content="$loop_context
$prompt_content"
fi
CLAUDE_CMD_ARGS+=("-p" "$prompt_content")
}
driver_supports_sessions() {
return 1 # false — session IDs not capturable from -p output
}
# Copilot CLI does not expose structured live output for jq streaming.
driver_supports_live_output() {
return 1 # false
}
# Copilot CLI outputs plain text only (no JSON streaming).
# Passthrough filter — no transformation needed.
driver_stream_filter() {
echo '.'
}

View File

@@ -0,0 +1,13 @@
#!/bin/bash
# Wrap Windows .cmd execution so GNU timeout launches a bash script instead of the .cmd directly.
set -euo pipefail
cli_path=${1:-}
if [[ -z "$cli_path" ]]; then
echo "ERROR: Missing Cursor CLI path" >&2
exit 1
fi
shift
exec "$cli_path" "$@"

283
.ralph/drivers/cursor.sh Executable file
View File

@@ -0,0 +1,283 @@
#!/bin/bash
# Cursor CLI driver for Ralph
# Uses the documented cursor-agent contract for background execution and
# switches to stream-json only for live display paths.
driver_name() {
echo "cursor"
}
driver_display_name() {
echo "Cursor CLI"
}
driver_cli_binary() {
local binary
binary=$(driver_resolve_cli_binary)
if [[ -n "$binary" ]]; then
echo "$binary"
return 0
fi
echo "cursor-agent"
}
driver_min_version() {
echo "0.1.0"
}
driver_check_available() {
local cli_binary
cli_binary=$(driver_cli_binary)
if [[ -f "$cli_binary" ]]; then
return 0
fi
command -v "$cli_binary" &>/dev/null
}
driver_valid_tools() {
VALID_TOOL_PATTERNS=(
"file_edit"
"file_read"
"file_write"
"terminal"
"search"
)
}
driver_supports_tool_allowlist() {
return 1
}
driver_permission_denial_help() {
echo " - $DRIVER_DISPLAY_NAME uses its native permission model."
echo " - ALLOWED_TOOLS in $RALPHRC_FILE is ignored for this driver."
echo " - Ralph already runs Cursor with --force."
echo " - Review Cursor permissions or approval settings, then restart the loop."
}
driver_build_command() {
local prompt_file=$1
local loop_context=$2
local session_id=$3
local cli_binary
cli_binary=$(driver_cli_binary)
if [[ ! -f "$prompt_file" ]]; then
echo "ERROR: Prompt file not found: $prompt_file" >&2
return 1
fi
CLAUDE_CMD_ARGS=()
if [[ "$cli_binary" == *.cmd ]]; then
CLAUDE_CMD_ARGS+=("$(driver_wrapper_path)" "$cli_binary")
else
CLAUDE_CMD_ARGS+=("$cli_binary")
fi
CLAUDE_CMD_ARGS+=("-p" "--force" "--output-format" "json")
if [[ "$CLAUDE_USE_CONTINUE" == "true" && -n "$session_id" ]]; then
CLAUDE_CMD_ARGS+=("--resume" "$session_id")
fi
local prompt_content
if driver_running_on_windows; then
prompt_content=$(driver_build_windows_bootstrap_prompt "$loop_context" "$prompt_file")
else
prompt_content=$(cat "$prompt_file")
if [[ -n "$loop_context" ]]; then
prompt_content="$loop_context
$prompt_content"
fi
fi
CLAUDE_CMD_ARGS+=("$prompt_content")
}
driver_supports_sessions() {
return 0
}
driver_supports_live_output() {
return 0
}
driver_prepare_live_command() {
LIVE_CMD_ARGS=()
local skip_next=false
for arg in "${CLAUDE_CMD_ARGS[@]}"; do
if [[ "$skip_next" == "true" ]]; then
LIVE_CMD_ARGS+=("stream-json")
skip_next=false
elif [[ "$arg" == "--output-format" ]]; then
LIVE_CMD_ARGS+=("$arg")
skip_next=true
else
LIVE_CMD_ARGS+=("$arg")
fi
done
if [[ "$skip_next" == "true" ]]; then
return 1
fi
}
driver_stream_filter() {
echo '
if .type == "assistant" then
[(.message.content[]? | select(.type == "text") | .text)] | join("\n")
elif .type == "tool_call" then
"\n\n⚡ [" + (.tool_call.name // .name // "tool_call") + "]\n"
else
empty
end'
}
driver_running_on_windows() {
[[ "${OS:-}" == "Windows_NT" || "${OSTYPE:-}" == msys* || "${OSTYPE:-}" == cygwin* || "${OSTYPE:-}" == win32* ]]
}
driver_resolve_cli_binary() {
local candidate
local resolved
local fallback
local candidates=(
"cursor-agent"
"cursor-agent.cmd"
"agent"
"agent.cmd"
)
for candidate in "${candidates[@]}"; do
resolved=$(driver_lookup_cli_candidate "$candidate")
if [[ -n "$resolved" ]]; then
echo "$resolved"
return 0
fi
done
fallback=$(driver_localappdata_cli_binary)
if [[ -n "$fallback" ]]; then
echo "$fallback"
return 0
fi
echo ""
}
driver_lookup_cli_candidate() {
local candidate=$1
local resolved
resolved=$(command -v "$candidate" 2>/dev/null || true)
if [[ -n "$resolved" ]]; then
echo "$resolved"
return 0
fi
if ! driver_running_on_windows; then
return 0
fi
driver_find_windows_path_candidate "$candidate"
}
driver_find_windows_path_candidate() {
local candidate=$1
local path_entry
local normalized_entry
local resolved_candidate
local path_entries="${PATH:-}"
local -a path_parts=()
if [[ "$path_entries" == *";"* ]]; then
IFS=';' read -r -a path_parts <<< "$path_entries"
else
IFS=':' read -r -a path_parts <<< "$path_entries"
fi
for path_entry in "${path_parts[@]}"; do
[[ -z "$path_entry" ]] && continue
normalized_entry=$path_entry
if command -v cygpath &>/dev/null && [[ "$normalized_entry" =~ ^[A-Za-z]:\\ ]]; then
normalized_entry=$(cygpath -u "$normalized_entry")
fi
resolved_candidate="$normalized_entry/$candidate"
if [[ -f "$resolved_candidate" ]]; then
echo "$resolved_candidate"
return 0
fi
done
}
driver_localappdata_cli_binary() {
local local_app_data="${LOCALAPPDATA:-}"
if [[ -z "$local_app_data" ]] || ! driver_running_on_windows; then
return 0
fi
if command -v cygpath &>/dev/null && [[ "$local_app_data" =~ ^[A-Za-z]:\\ ]]; then
local_app_data=$(cygpath -u "$local_app_data")
fi
local candidates=(
"$local_app_data/cursor-agent/cursor-agent.cmd"
"$local_app_data/cursor-agent/agent.cmd"
)
local candidate
for candidate in "${candidates[@]}"; do
if [[ -f "$candidate" ]]; then
echo "$candidate"
return 0
fi
done
}
driver_wrapper_path() {
local driver_dir
driver_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
echo "$driver_dir/cursor-agent-wrapper.sh"
}
driver_build_windows_bootstrap_prompt() {
local loop_context=$1
local prompt_file=$2
cat <<EOF
Read these Ralph workspace files before taking action:
- .ralph/PROMPT.md
- .ralph/PROJECT_CONTEXT.md
- .ralph/SPECS_INDEX.md
- .ralph/@fix_plan.md
- .ralph/@AGENT.md
- relevant files under .ralph/specs/
Then follow the Ralph instructions from those files and continue the next task.
EOF
if [[ -n "$loop_context" ]]; then
cat <<EOF
Current loop context:
$loop_context
EOF
fi
if [[ "$prompt_file" != ".ralph/PROMPT.md" ]]; then
cat <<EOF
Also read the active prompt file if it differs:
- $prompt_file
EOF
fi
}

147
.ralph/drivers/opencode.sh Executable file
View File

@@ -0,0 +1,147 @@
#!/bin/bash
# OpenCode driver for Ralph
# Uses OpenCode's build agent with JSON event output and optional session resume.
driver_name() {
echo "opencode"
}
driver_display_name() {
echo "OpenCode"
}
driver_cli_binary() {
echo "opencode"
}
driver_min_version() {
echo "0.1.0"
}
driver_check_available() {
command -v "$(driver_cli_binary)" &>/dev/null
}
driver_valid_tools() {
VALID_TOOL_PATTERNS=(
"bash"
"read"
"write"
"edit"
"grep"
"question"
)
}
driver_supports_tool_allowlist() {
return 1
}
driver_permission_denial_help() {
echo " - $DRIVER_DISPLAY_NAME uses its native permission and approval model."
echo " - ALLOWED_TOOLS in $RALPHRC_FILE is ignored for this driver."
echo " - BMAD workflows can use OpenCode's native question tool when needed."
echo " - Review OpenCode permissions, then restart the loop."
}
driver_build_command() {
local prompt_file=$1
local loop_context=$2
local session_id=$3
CLAUDE_CMD_ARGS=("$(driver_cli_binary)" "run" "--agent" "build" "--format" "json")
if [[ ! -f "$prompt_file" ]]; then
echo "ERROR: Prompt file not found: $prompt_file" >&2
return 1
fi
if [[ "$CLAUDE_USE_CONTINUE" == "true" && -n "$session_id" ]]; then
CLAUDE_CMD_ARGS+=("--continue" "--session" "$session_id")
fi
local prompt_content
prompt_content=$(cat "$prompt_file")
if [[ -n "$loop_context" ]]; then
prompt_content="$loop_context
$prompt_content"
fi
CLAUDE_CMD_ARGS+=("$prompt_content")
}
driver_supports_sessions() {
return 0
}
driver_supports_live_output() {
return 0
}
driver_prepare_live_command() {
LIVE_CMD_ARGS=("${CLAUDE_CMD_ARGS[@]}")
}
driver_stream_filter() {
echo 'select((.type == "message.updated" or .type == "message.completed") and .message.role == "assistant") | ([.message.parts[]? | select(.type == "text") | .text] | join("\n"))'
}
driver_extract_session_id_from_output() {
local output_file=$1
if [[ ! -f "$output_file" ]]; then
echo ""
return 1
fi
local session_id
session_id=$(sed -n 's/.*"session"[^{]*{[^}]*"id"[[:space:]]*:[[:space:]]*"\([^"]*\)".*/\1/p' "$output_file" | head -n 1 | tr -d '\r')
echo "$session_id"
[[ -n "$session_id" && "$session_id" != "null" ]]
}
driver_fallback_session_id() {
local cli_binary
cli_binary=$(driver_cli_binary)
local sessions_json
sessions_json=$("$cli_binary" session list --format json 2>/dev/null) || {
echo ""
return 1
}
local session_ids
if command -v jq >/dev/null 2>&1; then
session_ids=$(printf '%s' "$sessions_json" | jq -r '
if type == "array" then
[.[]?.id // empty]
elif (.sessions? | type) == "array" then
[.sessions[]?.id // empty]
else
[]
end
| map(select(length > 0))
| .[]
' 2>/dev/null | tr -d '\r')
else
session_ids=$(printf '%s' "$sessions_json" | grep -oE '"id"[[:space:]]*:[[:space:]]*"[^"]+"' | sed 's/.*"id"[[:space:]]*:[[:space:]]*"\([^"]*\)"/\1/' | tr -d '\r')
fi
local -a session_id_candidates=()
local session_id
while IFS= read -r session_id; do
if [[ -n "$session_id" && "$session_id" != "null" ]]; then
session_id_candidates+=("$session_id")
fi
done <<< "$session_ids"
if [[ ${#session_id_candidates[@]} -ne 1 ]]; then
echo ""
return 1
fi
echo "${session_id_candidates[0]}"
return 0
}