Parallax

Scientific augmentation tool encoding best practices for reproducible, hypothesis-driven science into agentic AI workflows.

True depth requires two lines of sight.

Philosophy

Accelerate as fast as is safe, quantifiable, and verifiable — but no faster. Parallax encodes scientific best practices into agentic AI workflows so that speed gains never come at the cost of rigor, reproducibility, or correctness. AI supplements scientific work; it never drives it.

See CONSTITUTION.md for core values.

Architecture

3-layer model:

1. Convention System (MVP-alpha) — CLI interview (parallax init) generates project config (CLAUDE.md, PARALLAX.md, templates). Claude Code skills and hooks enforce scientific best practices. CI as hard enforcement layer.

2. State + Workflow Engine (MVP-beta) — SQLite-backed hypothesis lifecycle tracking, git worktree parallel exploration, auto-documentation, regression tracking, agent handoff summaries.

3. Full Orchestrator (v1+) — Dashboard, multi-agent coordination, provenance chains, literature integration, JupyterLab hub.

See VISION.md for details.

Repo Structure

src/parallax/           # Main package
  cli/                  # Typer CLI (init, refine, config)
  core/                 # Config, interview, renderer, refiner
  db/                   # SQLite models (Layer 2)
  templates/            # string.Template files for init output
    agents/             # Agent definition templates
    skills/             # Skill templates
    hooks/              # Hook script templates
tests/                  # pytest (mirrors src structure)
docs/                   # VISION.md, ROADMAP.md, plans/
.claude/                # Skills (skill-name/SKILL.md) and hooks for development

Prerequisites

• pixi -- package/environment management
• Claude Code -- required for auto-refinement during parallax init

Installation

Install pixi:

# macOS / Linux
curl -fsSL https://pixi.sh/install.sh | bash

# macOS (Homebrew)
brew install pixi

# Windows
powershell -c "irm https://pixi.sh/install.ps1 | iex"

Then:

pixi install

Development

pixi run test        # pytest
pixi run lint        # ruff check
pixi run format      # ruff format
pixi run typecheck   # mypy --strict
pixi run check       # all of the above

Usage

# Initialize a new Parallax-managed project
parallax init

# With options
parallax init -t /path/to/project   # target directory
parallax init -y                     # accept defaults, skip optional
parallax init -f                     # overwrite existing files
parallax init --token-tier 5x        # set model tier for agents
parallax init --skip-refine          # skip auto-refinement
parallax init -b                     # run refinement in background (headless)
parallax init -k                     # keep interview cache after init

# Post-init refinement
parallax refine                      # launch interactive refinement session
parallax refine -t /path/to/project  # target directory
parallax refine --done               # strip refinement comment blocks

# Post-init config changes
parallax config set token-tier 5x    # update agent model selection

parallax init runs a structured interview generating:

• CLAUDE.md -- project-specific AI agent guide
• PARALLAX.md -- scientific workflow rules
• CONSTITUTION.md -- core scientific principles
• .claude/skills/ -- hypothesis, handoff, audit, experiment, session-start, manuscript-review, latex-guide skills
• .claude/agents/ -- hypothesis-explorer, experiment-runner, literature-reviewer, result-validator, paper-writer, presentation-writer, manuscript-reviewer agents
• .claude/hooks/ -- test guard, lint check, stop check enforcement scripts
• .claude/settings.json -- hook configuration referencing scripts above

Token tiers control agent model selection:

• pro (default) -- conservative: haiku exploration, sonnet validation
• 5x -- balanced: opus exploration, sonnet runner
• 20x -- generous: opus for most tasks
• api -- unconstrained: opus everywhere

Current Status

Layer 1 (Convention System) functional. parallax init, parallax refine, hook enforcement, and skills all implemented.

What exists:

• parallax init: structured interview + template rendering + auto-refinement
• Merge mode: parallax init into repos with existing .claude/ files -- suffixes conflicts, never overwrites, writes merge guide
• parallax refine: interactive refinement session (auto-detects merge guide for merge assistance)
• parallax refine --done: strip refinement comment blocks
• parallax config: post-init configuration changes (token tier)
• Hook enforcement: test guard (blocks test weakening), lint check (ruff feedback), stop check (uncommitted work reminder)
• Full skill definitions: /hypothesis, /handoff, /audit, /experiment, /session-start, /manuscript-review, /latex-guide
• Agent definitions: hypothesis-explorer, experiment-runner, literature-reviewer, result-validator, paper-writer, presentation-writer, manuscript-reviewer
• Token tier system: model selection per agent based on usage tier (pro/5x/20x/api)
• CI pipeline (ruff, mypy --strict, pytest)
• Integration test suite validating generated output

What's next:

• Layer 2: SQLite hypothesis lifecycle, git worktrees
• Template versioning / migration
• Semantic version validation in CI

See ROADMAP.md for the full backlog.

Contributing

This is a personal project in early development. If you're interested, open an issue.

License

All rights reserved. License to be determined.