Theta (θ)

A science co-researcher agent for the pi coding-agent harness.

Theta is the fundamental parameter of the coalescent — the standard model of how genetic diversity arises — and the universal symbol for the parameters a model fits. One letter, both halves of "computational biology + AI."

Status

Very early development (0.3.0). The architecture, skill-selection plan, and milestones are being designed in the open. M1 through M3 are built: the theta launcher plus the Tier-1 research-loop skill library — K-Dense skills vendored verbatim plus the self-authored theta-research-loop orchestrator (see Vendored skills) — the first connector skill, theta-lit-search (PubMed, arXiv, Semantic Scholar; connectors are skills per ADR 0002), and the project notebook, theta-notebook (dated provenance entries carrying content-addressed, graded evidence snapshots, recorded inside the research loop and after standalone searches per ADR 0003). The milestones below continue from there.

What Theta will be

Describe a scientific task in plain language; Theta plans, searches literature, writes and runs code, executes analyses, and reports results with a visible trace — built as a pi package rather than a monolithic fork:

• Skills: curated science skills (literature review, hypothesis generation, experimental design, statistics, domain libraries) following the open Agent Skills specification
• Connectors: literature and database tools (PubMed, arXiv, UniProt, …) as Theta-owned connector skills
• Compute: remote Slurm-cluster jobs via an MCP compute gateway, with a light local sandbox for small tasks
• Models: any of pi's available providers — bring your own key

Theta works today as a pi package from a checkout. The staged plan (M7–M8) grows it into a standalone theta application with pi embedded as the core engine — macOS and Linux first, Windows later.

Roadmap

• M1 — package skeleton, theta launcher (bin), Tier-1 research-loop skills
• M2 — literature connectors (PubMed, arXiv, Semantic Scholar)
• M3 — project notebook: provenance entries with graded evidence
• M4 — compute: Slurm MCP gateway + typed AlphaFold3 tool
• M5 — safety gates + containerized execution
• M6 — eval suite (MCP client is built into pi)
• M7 — standalone theta shell on pi's headless core (macOS/Linux)
• M8 — installers & release channel: Homebrew, deb/AUR, Windows

Install

pi install npm:theta-agent@0.3.0

Requires theta-agent ≥ 0.3.0 on the registry — 0.2.0 and earlier predate the project notebook, and a version-pinned install fails loudly until 0.3.0 is visible. If the registry is not there yet, run Theta from a checkout instead — see Try it.

Try it

theta rides on the pi coding agent, and pi is a peer dependency — install it first. The zero-registry path is a checkout:

git clone https://github.com/hugolld/theta-agent
cd theta-agent
npm i -g @earendil-works/pi-coding-agent
node bin/theta.mjs --dev

node bin/theta.mjs --dev launches pi with this checkout preloaded; from a checkout you can also name another directory with $THETA_DEV_ROOT (theta never loads an arbitrary current directory this way). Theta targets pi ≥ 0.99; if the launcher reports pi not found, install pi with the command above and retry.

Once theta-agent ≥ 0.3.0 is on the registry (0.2.0 and earlier predate the project notebook — until 0.3.0 is visible, stay on the checkout path above), the installed form is:

npm i -g theta-agent@0.3.0
theta

That launches pi with the Theta package preloaded.

Vendored skills

The research-loop skill library starts from the K-Dense scientific-agent-skills library, vendored verbatim per ADR 0001: upstream names, directory shapes, and content unchanged except a provenance header in each vendored SKILL.md. Re-syncs happen at milestone boundaries.

Currently vendored (upstream release v2.72.0, commit 526ebce) — the Tier-1 research-loop set:

• literature-review — systematic, scoping, and narrative literature reviews with reproducible searches and citation checks
• paper-lookup — searches 18 scholarly APIs for papers, citations, abstracts, and open-access full text with reproducible provenance
• paperzilla — reads Paperzilla projects, searches project feeds, and retrieves recommendations and canonical papers
• bgpt-paper-search — searches BGPT scientific papers by topic or DOI and retrieves claim-level evidence from full text
• research-lookup — compiles current scholarly evidence for manuscripts and research briefs via Parallel Search, Extract, and Research
• citation-management — citation search, metadata validation, and BibTeX generation across OpenAlex, PubMed, and Google Scholar
• pyzotero — manages Zotero reference libraries through the pyzotero Python client and the Zotero Web API
• peer-review — evidence-bounded, constructive peer-review drafts and structured manuscript assessments
• scholar-evaluation — qualitative-first, evidence-traceable developmental review of scholarly works and research-assessment rubrics
• scientific-writing — drafts, revises, and audits manuscripts with explicit evidence provenance and reporting-guideline coverage
• scientific-critical-thinking — evaluates scientific claims and evidence quality: validity, bias, confounders, and evidence grading
• scientific-brainstorming — evidence-aware scientific ideation with structured discussion, adversarial review, and decision logs
• hypothesis-generation — evidence-bounded candidate hypotheses, rival explanations, and preregistration-ready analysis plans
• hypogenic — plans and audits LLM-assisted hypothesis generation from labeled text datasets with HypoGeniC/HypoRefine
• experimental-design — designs experiments before data collection: randomization, blocking, factorial designs, and treatment layouts
• statistical-analysis — guided test selection, assumption checking, effect sizes, power analysis, and APA-formatted reporting
• statistical-power — sample-size and power calculation, closed-form and simulation-based, for study planning
• uncertainty-and-units — unit tracking and measurement-uncertainty propagation in scientific calculations
• exploratory-data-analysis — bounded, local first-pass analysis profiles of supported scientific file formats

The theta-research-loop orchestrator that routes the research loop into the library is Theta's own, self-authored work.

One deliberate omission: pi-agent, a Tier-1 candidate that documents the pi harness itself. It would duplicate pi's own documentation, drift against installed pi releases, and add nothing to the research loop. The skip is recorded on issue #5.

K-Dense's skills are MIT-licensed, © K-Dense Inc. If you use them in research, please cite their paper:

Kassis, T., Agarwal, V., He, Y., Patel, D., & Brueckner, A. M. (2026). Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents. arXiv:2609.00065. https://doi.org/10.48550/arXiv.2609.00065

License

MIT