Skip to content

ROADMAP.md

Latest commit

 

History

History
88 lines (67 loc) · 4.74 KB

ROADMAP.md

File metadata and controls

88 lines (67 loc) · 4.74 KB

Parallax Roadmap

Living document tracking decisions, open questions, and future work.

Current Status

Layer 1 (Convention System) — functional. parallax init interview + template rendering implemented. parallax refine implemented. Hook enforcement scripts implemented. Skills fully specified.

Known Gaps (Post-Bootstrap Design Work)

1. PARALLAX.md template content (resolved)

Implemented in template with hypothesis protocol, experiment manifest, handoff format.

2. Structured interview questions (resolved)

parallax init interview implemented: 2-phase (core + detailed), editor-based multi-line input, defaults, validation.

3. Template versioning/migration

When Parallax ships new templates, how do existing projects upgrade? Merge? Override? Side-by-side? Need a migration strategy before users depend on generated configs.

4. Data management

DVC/git-lfs/provenance unspecified for data-heavy science. Layer 3 mentions it but no concrete design. Critical for any project with non-trivial datasets.

5. Dogfooding plan

Test Parallax on a real scientific project ASAP. Identifies gaps theory can't. Candidate: use Parallax to manage its own development (meta-bootstrap).

6. Agent failure recovery

Partial state from midway failures. What happens when an agent dies mid-hypothesis? Recovery strategy, state rollback, idempotent operations.

7. PARALLAX.md + existing CLAUDE.md interaction

If a user already has CLAUDE.md, how does Parallax-generated content interact? Merge/append/@import/conflict resolution. Must not clobber user's existing rules.

8. Exploratory vs hypothesis-driven modes

Current design assumes hypothesis-driven work. Many scientific tasks are exploratory (data exploration, literature review, feasibility studies). Needs a "discovery mode" alongside hypothesis mode.

9. Claude Code API stability

Hooks/skills format could change upstream. No compatibility strategy. Should we pin Claude Code versions? Adapter layer? Accept breakage and fix forward?

Layer 1 Remaining Work

  • parallax init interview design + implementation
  • Template files: PARALLAX.md, CLAUDE.md, CONSTITUTION.md, settings.json, skills
  • Claude Code skills: /hypothesis, /handoff, /audit, /experiment, /session-start
  • Hook scripts: test protection, lint check, stop check
  • Custom agent definitions: hypothesis-explorer, experiment-runner, literature-reviewer, result-validator
  • Token tier system: model selection per agent (pro/5x/20x/api)
  • Auto-refinement via Claude CLI (interactive plan-mode session; -b for headless background)
  • parallax config set token-tier for post-init changes
  • memory: project on hypothesis + experiment skills
  • Writing agents: paper-writer (LaTeX), presentation-writer (Beamer)
  • Writing skills: /manuscript-review (auto-detects paper/Beamer), /latex-guide (troubleshooting, auto-invocable)
  • manuscript-reviewer agent (read-only critique, Bash for compilation checks)
  • Layer 2 design doc (005_layer2_design.md)
  • Doc staleness check (scripts/check_doc_staleness.py + CI step)
  • Semantic version validation in CI

Layer 2 Features (MVP-beta)

  • SQLite schema for hypothesis lifecycle + test results (skill memory: project as short-term proxy)
  • Git worktree integration for parallel hypotheses (Claude Code handles plumbing natively; Parallax defines workflows)
  • Agent handoff summary system
  • Semantic versioning automation
  • Conversation/session logging
  • Auto-documentation generation
  • Negative result logging
  • Token usage tracking (basic)

Layer 3 Features (v1+)

  • Multi-agent orchestration with token budgets
  • TUI/web dashboard
  • Regression metric visualization
  • JupyterLab hub
  • Literature/DOI integration
  • Full provenance chains
  • Data versioning (DVC/git-lfs)
  • Multi-language support
  • Voice interface
  • Fun scientist on startup
  • cmux / terminal multiplexer integration (plan)

Deferred Decisions

  • License: All rights reserved for now. Evaluate open-source options when project matures.
  • Team features: Personal use only for MVP. Multi-user/org support is v2+.
  • Container strategy: pixi for now. Evaluate Docker/Apptainer for HPC reproducibility in Layer 2/3.
  • Notebook format: Jupytext (markdown-based, diffable). Evaluate when Layer 2 auto-docs land.
  • Terminal multiplexer docs: Add tmux/cmux usage tutorial to docs/ when multiplexer integration lands.
  • Vision/roadmap as generated convention: Parallax-dev uses VISION.md + ROADMAP.md for strategic context in /session-start, but these aren't part of the generated template yet. Scope/format TBD after dogfooding.