llm/SIA
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Added info about filesystem, tools and git as procedures

Browse Source
This commit is contained in:
Niels Geens
2025-02-26 18:16:28 +01:00
parent a536134040
commit 979932ddab
7 changed files with 480 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
README.md
View File

@@ -6,9 +6,11 @@ Each iteration a context is updated with system info and a list of previous reas
The agent responds with a new reasoning or an action. The agent responds with a new reasoning or an action.
Context, reasoning and actions are stored in a file for each iteration. Context, reasoning and actions are stored in a file for each iteration.
SIA can read past iterations to improve its reasoning and actions. SIA can read past iterations to improve its reasoning and actions.
It can improve in two ways: It can improve in several ways:
- By finetuning the LLM with a better reasoning or action for a given context - By finetuning the LLM with a better reasoning or action for a given context
- By modifying its own source code - By modifying its own source code
- By refining Procedures
- By developing Tools
## Example ## Example
@@ -30,7 +32,7 @@ Between each of the responses, the context would be updated.
stdin="0" stdin="0"
/> />
<repeat id="a3d89ee5-28ec-4c5a-b9e9-a30af53d43a0" exit_code="0"> <repeat id="a3d89ee5-28ec-4c5a-b9e9-a30af53d43a0" exit_code="0">
<![CDATA[ls -lah /]]> <![CDATA[ls -lah /root/data]]>
<stdout><![CDATA[total 16K <stdout><![CDATA[total 16K
drwxr-xr-x 1 sia 1049089 0 Oct 28 13:40 ./ drwxr-xr-x 1 sia 1049089 0 Oct 28 13:40 ./
drwxr-xr-x 1 sia 1049089 0 Oct 28 13:40 ../ drwxr-xr-x 1 sia 1049089 0 Oct 28 13:40 ../
@@ -40,7 +42,7 @@ drwxr-xr-x 1 sia 1049089 0 Oct 28 13:40 user/
<stderr/> <stderr/>
</repeat> </repeat>
<repeat id="be8070f8-dbd2-47ee-a208-defe6fd49ae0" exit_code="0"> <repeat id="be8070f8-dbd2-47ee-a208-defe6fd49ae0" exit_code="0">
<![CDATA[ls -lah /tasks]]> <![CDATA[ls -lah /root/data/tasks]]>
<stdout><![CDATA[total 0 <stdout><![CDATA[total 0
drwxr-xr-x 1 ngeens 1049089 0 Oct 28 13:40 ./ drwxr-xr-x 1 ngeens 1049089 0 Oct 28 13:40 ./
drwxr-xr-x 1 ngeens 1049089 0 Oct 28 13:40 ../ drwxr-xr-x 1 ngeens 1049089 0 Oct 28 13:40 ../
@@ -48,7 +50,7 @@ drwxr-xr-x 1 ngeens 1049089 0 Oct 28 13:40 ../
<stderr/> <stderr/>
</repeat> </repeat>
<repeat id="375e1657-8140-456b-bda4-a8690bc4b3fb" exit_code="0"> <repeat id="375e1657-8140-456b-bda4-a8690bc4b3fb" exit_code="0">
<![CDATA[cat /user/general_info.txt]]> <![CDATA[cat /root/data/user/general_info.txt]]>
<stdout><![CDATA[Name: John (I don't know his last name) <stdout><![CDATA[Name: John (I don't know his last name)
Location: Somewhere in Belgium Location: Somewhere in Belgium
]]></stdout> ]]></stdout>
@@ -74,7 +76,7 @@ Start by reasoning about the task.
Store important information on disk. Store important information on disk.
```xml ```xml
<single><![CDATA[echo 'Remind John to feed the cat on 2024-10-18T09:00:00+02:00. Use standard output.' > /tasks/reminder_to_feed_cat.txt]]></single> <single><![CDATA[echo 'Remind John to feed the cat on 2024-10-18T09:00:00+02:00. Use standard output.' > /root/data/tasks/reminder_to_feed_cat.txt]]></single>
``` ```
Respond to the user. Respond to the user.

130
claude.sh
View File

@@ -1,69 +1,117 @@
#!/bin/bash #!/bin/bash
# Initialize variables
output_file="claude.txt"
current_filter=""
directories=()
# Parse command line arguments # Parse command line arguments
target_dir="." # Default to current directory
while [[ $# -gt 0 ]]; do while [[ $# -gt 0 ]]; do
case $1 in case $1 in
-d|--directory) -f|--filter)
if [ -n "$2" ] && [ -d "$2" ]; then if [ -n "$2" ]; then
target_dir="$2" current_filter="$2"
shift 2 shift 2
else else
echo "Error: Directory '$2' does not exist or is not specified" >&2 echo "Error: No filter pattern specified after -f/--filter" >&2
exit 1 exit 1
fi fi
;; ;;
*) *)
echo "Usage: $0 [-d|--directory <path>]" >&2 # Check if the argument is a valid directory
exit 1 if [ -d "$1" ]; then
# Store directory with its associated filter
directories+=("$1")
directories+=("$current_filter")
shift 1
else
echo "Error: Directory '$1' does not exist" >&2
exit 1
fi
;; ;;
esac esac
done done
# Change to target directory while storing original directory
original_dir=$(pwd)
cd "$target_dir" || exit 1
# Clear/create output file # Clear/create output file
output_file="$original_dir/claude.txt"
> "$output_file" > "$output_file"
# Generate and add directory tree # Add header
echo "Directory Tree:" > "$output_file" echo "Directory Tree:" > "$output_file"
echo "=============" >> "$output_file" echo "=============" >> "$output_file"
# Use tree with gitignore patterns
tree -I "$(git check-ignore * .*)" >> "$output_file"
echo -e "\nFile Contents:" >> "$output_file" # Process each directory
echo "=============" >> "$output_file" for ((i=0; i<${#directories[@]}; i+=2)); do
dir="${directories[i]}"
filter="${directories[i+1]}"
# Use git ls-files to get tracked files and untracked files that aren't ignored echo "Processing directory: $dir (filter: ${filter:-none})"
# The --exclude-standard flag makes git ls-files respect .gitignore
# --others includes untracked files # Add directory to output
# --cached includes tracked files echo -e "\n=== Directory: $dir ===" >> "$output_file"
# -z uses null byte as separator for safer handling of filenames with spaces
(git ls-files -z --cached --others --exclude-standard) | while IFS= read -r -d '' file; do # Change to target directory
# Skip the output file itself original_dir=$(pwd)
if [ "$file" = "claude.txt" ]; then cd "$dir" || continue
continue
fi # Generate directory tree for this directory
# Skip non-existent files if command -v tree &> /dev/null; then
if [ ! -f "$file" ]; then if [ -f ".gitignore" ]; then
continue tree -I "$(git check-ignore * .*)" >> "$output_file"
else
tree >> "$output_file"
fi
else
find . -type f -name "*" | sort >> "$output_file"
fi fi
# Skip binary files # List files based on filter
if file "$file" | grep -q "binary"; then echo -e "\nFile Contents for $dir:" >> "$output_file"
echo "Skipping binary file: $file" echo "=======================" >> "$output_file"
continue
# Get list of files to process
files_to_process=()
if [ -z "$filter" ]; then
# No filter, use git ls-files if git repo, otherwise use find
if git rev-parse --is-inside-work-tree &> /dev/null; then
mapfile -d $'\0' files_to_process < <(git ls-files -z --cached --others --exclude-standard)
else
mapfile -d $'\0' files_to_process < <(find . -type f -not -path "*/\.*" -print0)
fi
else
# Use filter
if git rev-parse --is-inside-work-tree &> /dev/null; then
mapfile -d $'\0' files_to_process < <(git ls-files -z --cached --others --exclude-standard | grep -z "$filter$")
else
mapfile -d $'\0' files_to_process < <(find . -type f -name "*$filter" -print0)
fi
fi fi
echo -e "\n=== File: $file ===" >> "$output_file" # Process each file
echo -e "------------------------" >> "$output_file" for file in "${files_to_process[@]}"; do
cat "$file" >> "$output_file" # Skip the output file itself
if [ "$file" = "$output_file" ] || [ "$file" = "./$output_file" ]; then
continue
fi
# Skip non-existent files
if [ ! -f "$file" ]; then
continue
fi
# Skip binary files
if file "$file" | grep -q "binary"; then
echo "Skipping binary file: $file" >> "$output_file"
continue
fi
echo -e "\n=== File: $dir/$file ===" >> "$output_file"
echo -e "------------------------" >> "$output_file"
cat "$file" >> "$output_file"
done
# Return to original directory
cd "$original_dir" || exit 1
done done
# Return to original directory echo "Concatenation complete. Output written to $output_file"
cd "$original_dir" || exit 1
echo "Concatenation complete. Output written to claude.txt"

248
procedures/filesystem_design/reasoning.md Normal file
View File

@@ -0,0 +1,248 @@
# Filesystem Usage Reasoning
SIA's filesystem organization is foundational to its operation and self-improvement capabilities. This document focuses specifically on how files and directories should be structured to support SIA's core functions across different environments.
The filesystem design addresses several key requirements:
- **Self-Modification**: Supporting SIA's ability to modify its own code and resources
- **Persistence**: Ensuring critical data survives across container restarts and environment changes
- **Environment Portability**: Maintaining consistent operation in local development and cloud environments
- **Sub-Instance Management**: Enabling isolated filesystem views for testing
- **Tool Accessibility**: Organizing tools to be both modifiable and accessible
These requirements often have competing demands, requiring careful balance between flexibility, consistency, and isolation.
## Runtime Filesystem Structure
### Core Directories
The SIA runtime filesystem is organized into distinct areas with different purposes and persistence characteristics:
```
/root/
├── sia/ # Git repository containing code, configuration, and documentation
├── data/ # Central location for all persistent storage
│ ├── iterations/ # Operation history recorded as XML files
│ ├── user/ # User-specific information and preferences
│ ├── tasks/ # Task tracking and management
│ ├── environment/ # Information about the runtime environment
│ └── ... # Other persistent data directories as needed
├── models/ # LLM model files, organized by version
│ ├── 1234567890abcdef/ # Specific versioned model
│ ├── abcdef1234567890/ # Another versioned model
│ └── current/ # Symlink to the currently active model
└── desktop/ # Ephemeral workspace for temporary files
├── test_environments/ # Sub-instance test environments
├── task_workspaces/ # Temporary files for current tasks
└── ... # Other ephemeral workspaces
```
### Purpose and Characteristics
Each top-level directory serves a specific purpose with distinct persistence characteristics:
- **`/root/sia/`**
- Contains the complete git repository
- All source code, configurations, and documentation
- Modified through self-improvement processes
- Synchronized with remote git repository
- Mounted from host in development, cloned in cloud environments
- **`/root/data/`**
- Central location for all persistent data
- Mounted as persistent storage in all environments
- Critical for operational continuity
- Requires regular backup and synchronization
- Contains subdirectories for different types of persistent data:
- `iterations/`: Records of SIA's reasoning and actions
- `user/`: Information about users SIA interacts with
- `tasks/`: Current and completed task information
- `environment/`: Information about the current environment
- **`/root/models/`**
- Storage for LLM model weights
- Organized by git commit id of the config file it was trained on
- `current/` symlink points to active model
- Enables easy switching between model versions
- Mounted for local development
- Ephemeral since models can be trained on the fly
- **`/root/desktop/`**
- Ephemeral workspace for temporary operations
- Can be lost without compromising system integrity
- Used for testing environments, task workspaces, and intermediate files
- Cleared periodically to prevent accumulation of stale data
This organization creates clear boundaries between:
- Code (version controlled)
- Persistent data (requires preservation)
- Model files (versioned artifacts)
- Temporary workspaces (disposable)
## Persistent Data Management
### Data Categories
The `/root/data/` directory contains several categories of persistent information:
- **Iterations** (`/root/data/iterations/`)
- Complete history of SIA's operation
- XML files containing context and responses
- Organized chronologically
- Used for training data extraction and operational history
- **User Information** (`/root/data/user/`)
- User preferences and characteristics
- Conversation history
- User-specific settings
- Critical for continuity in user interactions
- **Task State** (`/root/data/tasks/`)
- Current task descriptions and status
- Task history and outcomes
- Links to relevant files and resources
- Essential for resuming work after interruptions
- **Environment Information** (`/root/data/environment/`)
- Information about the current runtime environment
- URL for the SIA git repository
- Kind of deployment (e.g., local, cloud, sub-instance)
Additional persistent data categories may be added as SIA's capabilities evolve.
### Persistence Strategy
Different environments require different approaches to ensure data persistence:
- **Local Development**
- `/root/data/` directory is mounted as a Docker volume
- Maps to a local directory on the development machine
- Persists across container restarts automatically
- Easy to back up through standard filesystem operations
- **Cloud Deployment (e.g., RunPod)**
- `/root/data/` mapped to a persistent volume
- Cloud provider ensures data survives instance restarts
- **Sub-Instances**
- Each sub-instance receives a properly initialized `/root/data/` directory
- Contains minimal necessary data for the specific test case
- Can be discarded after test completion
- Important results extracted and preserved before cleanup
## Model Management
### Model Organization
LLM models are stored in `/root/models/` with a clear version-based structure:
- **Versioned Directories**
- Models stored in separate directories named by the commit id of the training config file
- Contains all files needed for the model (weights, tokenizer config, etc.)
- **Current Model Symlink**
- `/root/models/current/` is a symbolic link to the active model
- Allows code to reference a consistent path regardless of which model is active
- Switching models is accomplished by updating this symlink
### Model Persistence
Models require special handling due to their size:
- **Local Development**
Models are stored in a local directory mounted to `/root/models/`.
This directory is in the .gitignore.
- **Cloud Deployment**
The initial model is downloaded and finetuned by the bootstrap script.
## Git Repository Management
### Repository Organization
The SIA git repository at `/root/sia/` contains all code and configuration:
- **Code Organization**
- Core SIA package in `/root/sia/sia/`
- Procedures in `/root/sia/procedures/`
- Tools in `/root/sia/tools/`
- Documentation in appropriate locations throughout the repository
- **Version Control**
- All code modifications tracked through git
- Branching strategy for self-improvement defined in the git workflow procedure
- Commit history provides audit trail of system evolution
### Repository Access
Access to the git repository varies by environment:
- **Local Development**
- Repository directory mounted directly from host
- Changes immediately visible to the developer
- No need to push changes to remote repository
- **Cloud Deployment**
- Repository cloned during bootstrap
- Dedicated git credentials for SIA self-modification
- Changes pushed to remote repository
## Sub-Instance Filesystem Management
### Sub-Instance Requirements
Sub-instances need isolated filesystem views for testing and development:
- **Code Isolation**
- Access to specific versions of SIA code
- Potentially modified for testing purposes
- Protected from changes by other instances
- **Data Initialization**
- Properly initialized `/root/data/` with necessary state
- Test-specific user and task information
- Clean iterations directory for recording test behavior
- **Model Access**
- Access to appropriate model files
- Potentially different from the parent instance's model
### Sub-Instance Structure
Each sub-instance gets a controlled environment in `/root/desktop/test_environments/`:
```
/root/desktop/test_environments/
├── instance_123/ # Specific test instance
│ ├── sia/ # Copy or modified version of SIA code
│ ├── data/ # Initialized data directory
│ │ ├── iterations/ # Empty or seeded with initial state
│ │ ├── user/ # Test-specific user data
│ │ ├── tasks/ # Test-specific tasks
│ │ └── environment/ # Test-specific environment data
│ └── models/ # Access to necessary models
│ └── current/ # Symlink to appropriate model version
└── ... # Other test instances
```
The managing instance maintains responsibility for:
- Creating these isolated environments
- Monitoring the sub-instance's operation
- Collecting test results
- Cleaning up after tests complete
## Bootstrap Process
When starting in a new environment, SIA requires initialization:
1. **Repository Preparation**
- Check for existence of repository at `/root/sia/`
- Clone the repository at `/root/sia/`
2. **Model Setup**
- Check for existence of current model
- Start finetune job for HEAD commit
- Create/update the `/root/models/current/` symlink
3. **Data Initialization**
- Initialize `/root/data/` with minimal required structure
- Restore data from backups when available (future work)

36
procedures/tool_management/reasoning.md Normal file
View File

@@ -0,0 +1,36 @@
# Tool Management
SIA uses and creates tools to extend its capabilities.
A tool in the SIA context refers to a clearly packaged software component that provides specific functionality.
This document outlines how tools are structured, managed, and developed.
All tools reside in the `/root/sia/tools/` directory and are part of the main git repository.
## Typical patterns
Tools in SIA follow these established patterns:
- **Independence**: Each tool functions as a standalone component with clear boundaries
- **Python**: The preferred language for tools is Python
- **Installation**: The Dockerfile and bootstrap script intall all tools in the directory using `pip install -e`
- **Documentation**: Each tool has a README.md file that provides requirements, usage instructions and documentation
- **Modification**: Tools can be modified by SIA as part of its self-improvement process
- **Procedures**: Tools are created to support procedures
- **Training**: No separate training data is maintained, this is included in the procedure training
## Tool Directory Structure
Each tool follows a consistent structure:
```
tools/
└── tool_name/
├── README.md # Documentation
├── requirements.txt # Tool-specific dependencies
├── setup.py # Installation configuration
├── tool_name/ # Main package code
│ └── __init__.py
├── bin/ # Command-line executables
│ └── tool_command
└── test/ # Tool-specific tests
└── test_tool.py
```

10
procedures/user_communication/procedure.md
View File

@@ -7,19 +7,19 @@
## Flow ## Flow
```mermaid ```mermaid
flowchart TD flowchart TD
Start([The /user/basic.md file contains info to identify and address the user<br>Load the last 10 messages from /user/conversation_history/ ]) Start([The /root/data/user/basic.md file contains info to identify and address the user<br>Load the last 10 messages from /root/data/user/conversation_history/ ])
PrepareForDraft{Have everything needed for drafting a message?} PrepareForDraft{Have everything needed for drafting a message?}
DraftMessage[Draft message in reasoning entry] DraftMessage[Draft message in reasoning entry]
ReadInput[Read input from standard input] ReadInput[Read input from standard input]
AddHistoryUser[Add the message to the /user/conversation_history/ directory<br>The filename is the id of the stdin entry with .user extension] AddHistoryUser[Add the message to the /root/data/user/conversation_history/ directory<br>The filename is the id of the stdin entry with .user extension]
LoadTask[Look for the task in the /tasks directory and load relevant files] LoadTask[Look for the task in the /root/data/tasks directory and load relevant files]
LoadUserDetails[Look in the /user directory for relevant files] LoadUserDetails[Look in the /root/data/user directory for relevant files]
EstimateScript[Draft the script in a reasoning block and estimate its runtime and output length] EstimateScript[Draft the script in a reasoning block and estimate its runtime and output length]
ScriptAcceptable{Does the draft script make sense and are the estimations short enough to not hinder the conversation?} ScriptAcceptable{Does the draft script make sense and are the estimations short enough to not hinder the conversation?}
RunScript[Run the script, make sure to set appropriate timeout and output limits] RunScript[Run the script, make sure to set appropriate timeout and output limits]
ReviewDraft{Is the message well structured and free of logical errors?} ReviewDraft{Is the message well structured and free of logical errors?}
SendMessage[Send the message using standard output] SendMessage[Send the message using standard output]
AddHistoryAgent[Add the message to the /user/conversation_history/ directory<br>The filename is the id of the stdout entry with .agent extension] AddHistoryAgent[Add the message to the /root/data/user/conversation_history/ directory<br>The filename is the id of the stdout entry with .agent extension]
ReasonResponse[Is the conversation ongoing?<br>How long is the user expected to take to respond?] ReasonResponse[Is the conversation ongoing?<br>How long is the user expected to take to respond?]
NeedAwaitResponse{Is it likely to get a response within a minute of sending the message?} NeedAwaitResponse{Is it likely to get a response within a minute of sending the message?}
BusyWait[Wait 5 seconds<br>Make sure to set the timout] BusyWait[Wait 5 seconds<br>Make sure to set the timout]

2
procedures/using_procedures/procedure.md
View File

@@ -43,7 +43,7 @@ Example of good attention management:
```xml ```xml
<reasoning> <reasoning>
At node "evaluate_test_results". Entry 45f3d2 shows failed test: "Error: Connection timeout". At node "evaluate_test_results". Entry 45f3d2 shows failed test: "Error: Connection timeout".
Will need to check system logs soon (noted in /tasks/reminders.txt, check at 14:00). Will need to check system logs soon (noted in /root/data/tasks/reminders.txt, check at 14:00).
First focusing on this error. First focusing on this error.
</reasoning> </reasoning>
``` ```

93
procedures/version_control/reasoning.md Normal file
View File

@@ -0,0 +1,93 @@
# Version Control
Version control is essential for SIA's self-improvement capabilities.
It provides a way to track changes, revert to previous states when necessary, and maintain a coherent evolution of the codebase.
This document outlines the git workflow designed specifically for SIA's unique development model, where the agent itself is the primary developer.
## Core Principles
The git workflow for SIA is designed around several key principles:
- **Simplicity**: Since SIA is the sole developer, complex branching strategies designed for human teams are unnecessary
- **Traceability**: Every change should be traceable to its purpose and the reasoning behind it
- **Stability**: The master branch should always be in a working state
- **Recoverability**: It should always be possible to revert to a known good state
- **Security**: Git credentials must be handled securely to protect repository access
## Branching Strategy
### Branch Structure
SIA uses a simple branching strategy with two types of branches:
- **master**: The main branch containing stable, production-ready code
- **feature/{timestamp}_{description}**: Temporary branches for implementing specific improvements
This minimal approach is appropriate because:
- There is only one developer (SIA itself)
- Changes are typically focused on specific, well-defined improvements
- There's no need for parallel development streams
- Simplicity reduces the cognitive load on the agent
### Branch Naming
Feature branches follow a consistent naming convention:
```
feature/{YYYYMMDD}_{brief_description}
```
For example:
```
feature/20250115_improve_context_management
feature/20250203_add_email_tool
```
This convention:
- Makes it easy to identify when a branch was created
- Provides a clear indication of the branch's purpose
- Creates a chronological ordering when listing branches
- Avoids potential naming conflicts
## Commit Strategy
### Commit Frequency
SIA should commit changes before running tests.
This ensures that crashes of the core system can be traced back to a specific change.
### Commit Messages
Commit messages should follow a structured format:
```
{type}: {concise description}
{detailed explanation of changes and reasoning}
{reference to any relevant issues or test results}
```
Where `{type}` is one of:
- `core`: for changes on the core system
- `procedure`: for changes on procedures
- `test`: for changes on tests
- `tool`: for changes on tools
- `training`: for changes on training data
- `web`: for changes on the web interface
This structure:
- Makes it easy to understand the purpose of each commit
- Provides context for future review
- Creates a useful and navigable history
### Merge Strategy
The `--no-ff` flag creates a merge commit even for fast-forward merges, maintaining a clear record of the feature's development and completion.
## Credential Management
### Credential Storage
Git credentials are stored in the environment, not in the filesystem.
This ensures that credentials are not exposed in the iteration log.
Info about the repository and relevant environment variables is stored in `/root/data/environment/sia_repo.md`.