Progressive algorithmization toolchain — from LLM to deterministic code, from proxy to tickets

These details have not been verified by PyPI

Project links

Project description

algitex

Progressive algorithmization toolchain — from LLM to deterministic code, from proxy to tickets.

The only framework that automates the path from "LLM handles everything" to "most traffic runs deterministically, LLM only for edge cases."

pip install algitex
algitex init ./my-app
algitex go

Why "algitex"?

The name reflects the core cycle: analyze → plan → execute → validate → repeat. Each iteration makes your codebase healthier and your LLM usage cheaper. The progressive algorithmization loop gradually replaces LLM calls with deterministic rules.

Algitex = Algorithmic + Intelligence + Execution + Engine

Semantically:

Alg- → algorithms, logic, determinism
-i- → intelligence layer
-tex → texture / system / framework / execution layer

Algitex is the intelligence compilation engine that transforms LLM-driven behavior into deterministic, cost-efficient algorithmic systems. It enables progressive algorithmization from probabilistic AI reasoning to structured, deterministic logic.

Progressive Algorithmization

The 5-stage transition from LLM to deterministic:

Stage 1: Discovery    → LLM performs tasks, collect traces
Stage 2: Extraction   → Identify recurring patterns
Stage 3: Rules        → Generate deterministic replacements
Stage 4: Hybrid       → Route by confidence: rules vs LLM
Stage 5: Optimization → Minimize LLM dependency, reduce costs

Result: Systems that start with LLM flexibility but evolve into efficient, deterministic engines—maintaining AI reasoning benefits with traditional software performance.

Name alternatives considered

Name	Why it works	Why we picked algitex
algitex	Core concept: the continuous improvement loop	Clear, memorable, tech-neutral
prollama	"progressive" + llama vibes	Ties too much to one model family
codefact	Code + factory/fact	Sounds like a trivia app
algopact	Algorithm + Propact	Hard to pronounce
loopcode	Loop + code	Reverse reads awkward
prodev	Progressive + dev	Too generic, SEO nightmare

Three layers, one command

Layer 1: Code Quality Loop

from algitex import Project

p = Project("./my-app")
p.analyze()    # code2llm + vallm + redup → health report
p.plan()       # auto-generate tickets from analysis
p.execute()    # LLM handles tasks via proxym
p.status()     # health + tickets + budget + cost ledger

Layer 2: Progressive Algorithmization

from algitex import Loop

loop = Loop("./my-app")
loop.discover()        # Stage 1: collect all LLM traces
loop.extract()         # Stage 2: find repeating patterns
loop.generate_rules()  # Stage 3: AI writes its own replacement
loop.route()           # Stage 4: rules vs LLM by confidence
loop.optimize()        # Stage 5: monitor, minimize LLM usage
print(loop.report())   # "42% deterministic, $12.50 saved"

Layer 3: Propact Workflows

from algitex import Workflow

wf = Workflow("./refactor-v1.md")
wf.execute()   # runs propact:shell, propact:rest, propact:llm blocks

CLI

# Core loop
algitex init ./my-app         # initialize project
algitex analyze               # health check
algitex plan --sprints 3      # generate sprint strategy + tickets
algitex go                    # full pipeline
algitex status                # dashboard

# Progressive algorithmization
algitex algo discover         # start trace collection
algitex algo extract          # find patterns in traces
algitex algo rules            # generate deterministic replacements
algitex algo report           # show % deterministic vs LLM

# Propact workflows
algitex workflow run fix.md   # execute Markdown workflow
algitex workflow validate f.md

# Tickets
algitex ticket add "Fix auth" --priority high
algitex ticket list
algitex ticket board
algitex sync                  # push to GitHub/Jira

# Quick queries
algitex ask "Explain this race condition" --tier premium
algitex tools                 # show installed tools

Parallel TODO Task Processing

Execute TODO tasks from prefact analysis in parallel with automatic categorization and fix strategies:

# Verify which TODO tasks are still valid vs already fixed
algitex todo verify

# Auto-fix mechanical issues in parallel (dry-run)
algitex todo fix-auto --workers 8

# Actually apply fixes
algitex todo fix-auto --execute

# Fix only specific category
algitex todo fix-auto --category unused_import --execute
algitex todo fix-auto --category return_type --execute

# Benchmark performance
algitex todo benchmark 100 --workers 8
algitex todo benchmark 50 --compare  # parallel vs sequential

Python API

from algitex.todo import verify_todos, fix_todos, benchmark_fix, compare_modes

# Verify task validity
result = verify_todos("TODO.md")
print(f"Still open: {result.still_open}, Fixed: {result.already_fixed}")

# Parallel auto-fix (mechanical tasks only)
stats = fix_todos("TODO.md", workers=8, dry_run=False)
print(f"Fixed: {stats['fixed']}, Skipped: {stats['skipped']}")

# Benchmark performance
result = benchmark_fix("TODO.md", limit=100, workers=8, mode="parallel")
result.print_report()

# Compare modes
comparison = compare_modes("TODO.md", limit=50, workers=8)
# Shows: sequential vs parallel speedup, throughput, efficiency

Auto-fix Categories

Category	Auto-fixable	Description
`unused_import`	✅ Yes	Remove unused imports (import X, from Y import X)
`return_type`	✅ Yes	Add missing return type annotations
`fstring`	⚠️ Partial	Delegate to `flynt` tool
`magic`	❌ No	Requires manual constant extraction
`docstring`	❌ No	Requires LLM-style rewrite
`exec_block`	❌ No	Requires architectural decision

How Parallel Processing Works

┌─────────────────────────────────────────────────────────┐
│              Parallel TODO Processing                    │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  1. Parse TODO.md → filter worktree duplicates          │
│  2. Categorize tasks (unused_import, return_type...)    │
│  3. Group by file (1 worker per file, zero conflicts)   │
│  4. Sort tasks bottom-up (line DESC) → preserve numbers│
│  5. Execute in ThreadPoolExecutor (8 workers default)  │
│  6. Collect results: fixed, skipped, errors            │
│                                                         │
│  Safety: Each worker touches different file.           │
│  Within file: bottom-up prevents line number shifts.   │
│                                                         │
└─────────────────────────────────────────────────────────┘

Three Execution Paths

Path	LLM?	Parallel?	Throughput	Use Case
`todo fix-auto`	❌ No	✅ Yes (8 workers)	~1500 tickets/sec	Mechanical fixes: unused imports, return types
`todo run --tool ollama-mcp`	✅ Yes	❌ Sequential (queue)	~1-10 tickets/sec	Complex fixes requiring reasoning
`autofix via proxy`	✅ Yes	⚠️ Batch	~5-50 tickets/sec	Intelligent fixes via litellm-proxy

When to use which:

Path 1: Mechanical Fixes (todo fix-auto)

No LLM calls — pure regex/text manipulation
8 parallel workers, thread-safe per-file isolation
Handles: unused_import, return_type, fstring (via flynt)
Best for: bulk cleanup of 100+ simple issues

from algitex.todo import fix_todos
stats = fix_todos("TODO.md", workers=8, dry_run=False)
# 2679 tasks → ~1.8 seconds total

Path 2: LLM-based Fixes (todo run)

Uses Ollama/aider via Docker MCP
Sequential execution (respects LLM rate limits)
Handles: complex refactoring, architectural changes
Best for: issues requiring code understanding

algitex todo run --tool ollama-mcp --limit 10

Path 3: Hybrid via Proxy (autofix)

Routes through litellm-proxy with cost tracking
Batch processing with retry logic
Handles: smart fixes with context awareness
Best for: production workflows with budget constraints

from algitex.tools.autofix import AutoFix
autofix = AutoFix(backend="litellm-proxy", proxy_url="http://localhost:4000")
autofix.fix_all(limit=5)  # $0.12 per batch avg

Path 4: Hybrid CLI (todo hybrid) — Fast + Parallel + LLM

Phase 1: Parallel mechanical fixes (no LLM)
Phase 2: Rate-limited parallel LLM fixes
Handles: complete TODO workflow in one command

# Dry run (preview)
algitex todo hybrid --workers 4 --rate-limit 10

# Execute with rate limiting
algitex todo hybrid --execute --backend litellm-proxy --workers 4 --rate-limit 10

# Local Ollama (100% offline)
algitex todo hybrid --execute --backend ollama --workers 2 --rate-limit 5

The Missing Piece: Fast + Parallel + LLM

To achieve szybkie + równoległe + LLM, you need to combine ThreadPoolExecutor with ProxyBackend:

from algitex.todo import HybridAutofix

# Combines parallel task distribution with LLM backend
fixer = HybridAutofix(
    backend="litellm-proxy",
    workers=4,              # Parallel workers
    rate_limit=10,          # Requests per second
    retry_attempts=3,
    timeout=30
)

# Mechanical fixes: parallel, no LLM
fixer.fix_mechanical("TODO.md")  # 1000+ tickets/sec

# Complex fixes: parallel LLM with rate limiting
fixer.fix_complex("TODO.md")     # 10-50 tickets/sec, cost-tracked

Requirements for parallel LLM:

Rate limiting (prevent 429 errors)
Retry logic with exponential backoff
Cost tracking per batch
Circuit breaker for failed requests

The 5-Stage Progressive Algorithmization

Stage 1: Discovery     → LLM handles 100%, collect traces
Stage 2: Extraction    → identify hot paths + repeating patterns
Stage 3: Rules         → AI generates deterministic replacements
Stage 4: Hybrid        → confidence-based: known patterns → rules, unknown → LLM
Stage 5: Optimization  → most traffic deterministic, LLM for edge cases only

No existing framework automates this path. DSPy goes LLM→smaller LLM. algitex goes LLM→algorithm.

Propact: Markdown as Workflow

# Fix Authentication Module

Analyze current state:

```propact:shell
code2llm ./src/auth -f toon --json

Ask LLM for a fix plan:

POST http://localhost:4000/v1/chat/completions
{"model": "balanced", "messages": [{"role": "user", "content": "Fix auth"}]}

Validate the result:

vallm batch ./src/auth --recursive


## Planfile-Aware Proxy Headers

Every LLM request through algitex carries context:

X-Planfile-Ref: my-project/current/DLP-0042 X-Workflow-Ref: refactor-v1.md X-Task-Tier: complex X-Inject-Context: true


Proxym logs cost/model/latency **per ticket**. The cost ledger shows exactly what each task costs.

## Installation

```bash
pip install algitex                # core
pip install algitex[all]           # + all tools
pip install algitex[proxy]         # + proxym
pip install algitex[analysis]      # + code2llm, vallm, redup
pip install algitex[tickets]       # + planfile
pip install algitex[routing]       # + llx

Examples

# Quickstart — three main objects (Project, Loop, Workflow)
cd examples/01-quickstart
make run

# Progressive Algorithmization — 5-stage loop
cd examples/02-algo-loop
make run

# Composable Pipeline — fluent API
cd examples/03-pipeline
make run

# IDE Integration — generate configs for Roo Code, Cline, etc.
cd examples/04-ide-integration
make setup && make run

# Cost Tracking — per-ticket cost ledger
cd examples/05-cost-tracking
make run

# Local LLM with Ollama — 100% offline, zero API costs
cd examples/18-ollama-local
make setup && make run

# Local MCP Tools — self-hosted code analysis & validation
cd examples/19-local-mcp-tools
make up && make run

# Self-Hosted Pipeline — complete local CI/CD stack
cd examples/20-self-hosted-pipeline
make build && make up && make run

# Aider CLI + Ollama — local refactoring with prefact TODO workflow
cd examples/21-aider-cli-ollama
make setup && make run

# Claude Code + Ollama — AI assistant with local LLM
cd examples/22-claude-code-ollama
make setup && make run

# Continue.dev + Ollama — VS Code extension setup
cd examples/23-continue-dev-ollama
make setup

# Ollama Batch Processing — parallel code analysis
cd examples/24-ollama-batch
python batch_analyze.py --dir ./src

# Local Model Comparison — benchmark Ollama models
cd examples/25-local-model-comparison
make benchmark

# LiteLLM Proxy + Ollama — native algitex integration (better than aider)
cd examples/26-litellm-proxy-ollama
make setup && make proxy  # Terminal 1
make fix                  # Terminal 2

# Hybrid AutoFix — fast parallel + LLM with rate limiting
cd examples/33-hybrid-autofix
make dry-run              # Preview
make hybrid               # Execute with LiteLLM proxy
make ollama               # Execute with Ollama (100% offline)

Each example has:

01-quickstart/README.md — Project, Loop, Workflow basics
02-algo-loop/README.md — Progressive algorithmization
03-pipeline/README.md — Composable fluent API
04-ide-integration/README.md — IDE configs
05-cost-tracking/README.md — Cost tracking
06-telemetry/README.md — Telemetry & observability
07-context/README.md — Context building
08-feedback/README.md — Feedback loops
09-workspace/README.md — Workspace management
10-cicd/README.md — CI/CD pipelines
11-aider-mcp/README.md — Aider MCP code refactoring
12-filesystem-mcp/README.md — Filesystem operations
13-vallm/README.md — Vallm validation
14-docker-mcp/README.md — Docker container management
15-github-mcp/README.md — GitHub repository operations
16-test-workflow/README.md — Comprehensive test pipeline
17-docker-workflow/README.md — Refactoring workflow
18-ollama-local/README.md — Local LLM with Ollama (100% offline)
19-local-mcp-tools/README.md — Self-hosted MCP tools (Docker)
20-self-hosted-pipeline/README.md — Complete local CI/CD pipeline
21-aider-cli-ollama/README.md — Aider CLI + Ollama local refactoring
22-claude-code-ollama/README.md — Claude Code + Ollama AI assistant
23-continue-dev-ollama/README.md — Continue.dev VS Code extension + Ollama
24-ollama-batch/README.md — Parallel batch processing with Ollama
25-local-model-comparison/README.md — Benchmark Ollama models
26-litellm-proxy-ollama/README.md — LiteLLM Proxy + Ollama (native algitex)
33-hybrid-autofix/README.md — Fast parallel + LLM with rate limiting
run.sh — executable script
Makefile — make run, make setup, make clean
.env.example — configuration template (where applicable)

Additional Documentation

README2.md — Detailed conceptual overview of Algitex as intelligence compilation engine

Architecture

src/algitex/
├── __init__.py           # Project, Loop, Workflow, Config, Pipeline
├── config.py             # Unified config (env + YAML)
├── project.py            # Main Project class (expanded)
├── cli.py                # Typer CLI (all commands)
├── algo/                 # Progressive algorithmization
│   ├── __init__.py       # Loop, TraceEntry, Pattern, Rule, LoopState
│   └── loop.py           # Re-export
├── propact/              # Markdown workflow engine
│   ├── __init__.py       # Workflow, WorkflowStep, WorkflowResult
│   └── workflow.py       # Re-export
├── tools/
│   ├── __init__.py       # Tool discovery
│   ├── proxy.py          # proxym wrapper + planfile headers
│   ├── analysis.py       # code2llm + vallm + redup
│   └── tickets.py        # planfile wrapper + cost ledger
└── workflows/
    ├── __init__.py        # Pipeline (composable steps)
    └── pipeline.py        # Re-export

How it connects to the ecosystem

┌─────────────────────────────────────────────────────┐
│                     algitex                         │
│            (orchestration layer)                    │
├─────────────────────────────────────────────────────┤
│                                                     │
│  analyze()   plan()   execute()   algo.discover()   │
│     │          │         │            │             │
│  code2llm   planfile   proxym      trace →          │
│  vallm      tickets    llx         patterns →       │
│  redup      strategy   models      rules →          │
│                                    hybrid routing   │
│                                                     │
│  run_workflow("fix.md")                             │
│     │                                               │
│  propact:shell → subprocess                         │
│  propact:rest  → httpx                              │
│  propact:llm   → proxym                             │
│  propact:mcp   → MCP tool call                      │
│                                                     │
└─────────────────────────────────────────────────────┘

Tool Roles

Tool	What	Install
proxym	LLM gateway, 10 providers, routing, budget	`pip install proxym`
planfile	Sprint planning, tickets, GitHub/Jira sync	`pip install planfile`
llx	Metric-driven model selection, MCP server	`pip install llx`
code2llm	Static analysis → .toon diagnostics	`pip install code2llm`
vallm	4-tier code validation	`pip install vallm`
redup	Duplication detection	`pip install redup`

License

Licensed under Apache-2.0.

Author

Tom Sapletta

Created by Tom Sapletta — tom@sapletta.com

Part of the semcod / wronai ecosystem.

Project details

These details have not been verified by PyPI

Project links

Release history Release notifications | RSS feed

0.1.64

Apr 25, 2026

0.1.63

Apr 25, 2026

0.1.62

Apr 25, 2026

0.1.61

Apr 25, 2026

0.1.60

Apr 25, 2026

0.1.59

Mar 29, 2026

0.1.58

Mar 29, 2026

0.1.57

Mar 28, 2026

0.1.56

Mar 28, 2026

0.1.55

Mar 28, 2026

0.1.54

Mar 28, 2026

0.1.53

Mar 28, 2026

0.1.52

Mar 28, 2026

0.1.51

Mar 28, 2026

0.1.50

Mar 28, 2026

0.1.49

Mar 28, 2026

0.1.48

Mar 28, 2026

0.1.47

Mar 28, 2026

0.1.46

Mar 28, 2026

0.1.45

Mar 28, 2026

0.1.44

Mar 28, 2026

0.1.43

Mar 28, 2026

0.1.42

Mar 28, 2026

0.1.41

Mar 28, 2026

0.1.40

Mar 28, 2026

0.1.39

Mar 28, 2026

0.1.38

Mar 28, 2026

This version

0.1.37

Mar 28, 2026

0.1.36

Mar 28, 2026

0.1.35

Mar 28, 2026

0.1.34

Mar 28, 2026

0.1.33

Mar 28, 2026

0.1.32

Mar 28, 2026

0.1.31

Mar 28, 2026

0.1.30

Mar 28, 2026

0.1.29

Mar 28, 2026

0.1.28

Mar 28, 2026

0.1.27

Mar 28, 2026

0.1.26

Mar 28, 2026

0.1.25

Mar 28, 2026

0.1.24

Mar 28, 2026

0.1.23

Mar 28, 2026

0.1.22

Mar 28, 2026

0.1.21

Mar 28, 2026

0.1.20

Mar 28, 2026

0.1.19

Mar 28, 2026

0.1.18

Mar 28, 2026

0.1.17

Mar 28, 2026

0.1.16

Mar 28, 2026

0.1.15

Mar 28, 2026

0.1.14

Mar 28, 2026

0.1.13

Mar 28, 2026

0.1.12

Mar 28, 2026

0.1.11

Mar 28, 2026

0.1.10

Mar 28, 2026

0.1.9

Mar 28, 2026

0.1.8

Mar 28, 2026

0.1.7

Mar 28, 2026

0.1.6

Mar 28, 2026

0.1.5

Mar 28, 2026

0.1.4

Mar 28, 2026

0.1.3

Mar 28, 2026

0.1.2

Mar 28, 2026

0.1.1

Mar 28, 2026

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

algitex-0.1.37.tar.gz (149.1 kB view details)

Uploaded Mar 28, 2026 Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

The dropdown lists show the available interpreters, ABIs, and platforms. Enable javascript to be able to filter the list of wheel files.

algitex-0.1.37-py3-none-any.whl (148.0 kB view details)

Uploaded Mar 28, 2026 Python 3

File details

Details for the file algitex-0.1.37.tar.gz.

File metadata

Download URL: algitex-0.1.37.tar.gz
Upload date: Mar 28, 2026
Size: 149.1 kB
Tags: Source
Uploaded using Trusted Publishing? No
Uploaded via: twine/6.2.0 CPython/3.13.7

File hashes

Hashes for algitex-0.1.37.tar.gz
Algorithm	Hash digest
SHA256	`507aae24c348788f0ef151ccde8844a7cd62b302e1b98bbee4921a7cb8f7691e`
MD5	`518eb5d227cd7315a9f7ae3cc4cad552`
BLAKE2b-256	`659efd7f1cb8232a60da739889c7c436f5460d5c12394bc4699d87a0c9539c94`

See more details on using hashes here.

File details

Details for the file algitex-0.1.37-py3-none-any.whl.

File metadata

Download URL: algitex-0.1.37-py3-none-any.whl
Upload date: Mar 28, 2026
Size: 148.0 kB
Tags: Python 3
Uploaded using Trusted Publishing? No
Uploaded via: twine/6.2.0 CPython/3.13.7

File hashes

Hashes for algitex-0.1.37-py3-none-any.whl
Algorithm	Hash digest
SHA256	`6233040e0acc4d36a2ff165f1f0f6bcdb6a9dc873e7876187c7e967245ec82dd`
MD5	`9b1c44c32dc6aacd6c79c78bdb2b7c0f`
BLAKE2b-256	`4f9226343f9c7cf91a360c0bb6aa05cf4e6cfe9c3c41e89d6fcfc3f533983a6d`

See more details on using hashes here.

algitex 0.1.37

Navigation

Verified details

Maintainers

Unverified details

Project links

Meta

Classifiers

Project description

algitex

Why "algitex"?

Progressive Algorithmization

Name alternatives considered

Three layers, one command

Layer 1: Code Quality Loop

Layer 2: Progressive Algorithmization

Layer 3: Propact Workflows

CLI

Parallel TODO Task Processing

Python API

Auto-fix Categories

How Parallel Processing Works

Three Execution Paths

The Missing Piece: Fast + Parallel + LLM

The 5-Stage Progressive Algorithmization

Propact: Markdown as Workflow

Examples

Additional Documentation

Architecture

How it connects to the ecosystem

Tool Roles

License

Author

Project details

Verified details

Maintainers

Unverified details

Project links

Meta

Classifiers

Release history Release notifications | RSS feed

Download files

Source Distribution

Built Distribution

File details

File metadata

File hashes

File details

File metadata

File hashes