Install scientific skills. Manage a filesystem-backed research wiki. Build on the full Oh My Pi agent platform while the higher-level scientific orchestration layers come online.
Built on @mariozechner's Pi, adding scientific workflow orchestration, a structured wiki, self-evolving skills, and model-tier safety — everything you need to run AI-assisted research from the terminal.
One harness, three modes. Choose the level of coordination your experiment needs.
Loop mode — single agent, tight optimization cycle: propose → test → measure → repeat. Inspired by Karpathy's autoresearch. Good for hyperparameter sweeps, ablation studies.
Pipeline mode — predefined multi-step workflows with specialized agents. Fixed roles, fixed handoffs. Good for literature review, RNA-seq analysis, docking pipelines.
Team mode — self-organizing agent teams with discussion, critique, and queue-based work distribution. The AutoScientists pattern. Good for open-ended exploration.
Five entity types track every part of the scientific process. Filesystem-backed — zero dependencies, portable anywhere.
Paper — literature references with structured citations, DOI, and extracted findings.
Hypothesis — testable statements with predictions, rationale, and evidence links. Status flows: draft → proposed → testing → confirmed | rejected.
Experiment — base-delta protocol from ELF. Reference a base protocol, describe only what changed. Minimal documentation overhead.
Evidence — findings with strength assessment (weak/moderate/strong). Linked to source hypothesis and experiment.
Insight — freeform notes. Ideas sparked by reading, discussion, or unexpected results.
Scidekick consumes the existing Agent Skills ecosystem instead of creating a parallel one. Today the product slice supports listing and installing science-focused skills from a Git repository into the standard skill discovery paths.
Local or remote — install from a GitHub repo, shorthand, or local fixture repository.
Discovery-compatible — installed skills land in the same directories the agent already scans.
Tracked — every install is recorded in SQLite so validation metadata can be layered on next.
AutoScientists (MIMS Harvard, 2026) proved that Haiku-class models fail catastrophically at scientific reasoning — they hallucinate API unavailability and “describe instead of do.”
Scidekick already ships the model-tier classification library: haiku, sonnet, and opus tiers are inferred from model ID heuristics.
The next wiring step is product enforcement: warn or block only when a scientific workflow is active, without penalizing ordinary coding sessions.
SkillLens (MSR 2026) found that 25% of skill–target pairs cause negative transfer — the skill actively harms performance. The safe first step is durable metadata: know what was installed, where it came from, and what it was validated against.
Today Scidekick records installed skills in a SQLite registry with validation metadata fields: tested models, performance delta, and rubric score.
The next product slice wires that registry into scientific workflows so unvalidated skills warn or block before use.
Scidekick inherits the full Pi/omp platform. LSP wired into every write. Native Rust ripgrep + glob. Hashline content-addressed edits. Subagent isolation with 10+ CoW backends. Every feature ships in the same binary.
14 LSP operations, 28 debug adapter ops. Renames go through willRenameFiles. Debuggers attach via lldb, dlv, debugpy.
Persistent Python + Bun Worker. Either kernel calls back into agent tools. %%bash, %pip, %time. 5-min idle, max 4 sessions.
Content-hash anchors. Stale files rejected before corruption. Grok 4 Fast: −61% output tokens on the same work.
Schema-validated yields. 10+ filesystem isolation backends. IRC channels for peer coordination.
ripgrep, glob, find: in-process. brush shell with persistent sessions. Same binary on macOS, Linux, Windows.
pr://, issue://, agent://, skill://, rule:// resolve inside read/write/edit. No separate tools to learn.
Internal schemes resolve transparently in every FS-shaped tool. read pr://1428 returns the same shape as read src/foo.ts. One tool surface — the model already knows how to use it.
Default install path uses the latest GitHub release binary. No Bun required. Bun ≥ 1.3.14 is only needed for --source installs.