Autonomous research you can verify.

Pi is the model runtime, Git holds reviewable research content, and SQLite holds operational state. You drive it all from a terminal UI built with OpenTUI.

• ProjectThe unit of research, roughly one paper

  One Git repository plus an approved research contract. The contract sets scope, datasets and licenses, and the test and evaluation commands that workers may run. Each Project has its own database, branches, worktrees, and budgets, and no context leaks between Projects. Change the contract later with /amend-contract.

  • Research chatWhere you work with the Project agent

    Discovery happens here. It asks one question at a time and proposes the contract. Afterwards, this is where you ask questions, plan, stage work, and start Missions and Directions.

  • MissionA goal with a finish line

    An approved objective with success criteria, budgets, and a final review rubric. A Mission Director plans the work as a graph of Nodes and repairs the plan as results come in. Accepted work merges onto the Mission's own branch, nosh/mission-<id>, so you merge one branch at the end.

    • NodeOne step of the plan

      Each Node has a type, and it runs only after the Nodes it depends on are accepted. The type decides who does the work: a librarian-researcher for literature, a reviewer for reviews and approvals, and a general worker for everything else.

      literature_review · reproduction · implementation · dataset · analysis · research_direction · autoresearch · integration · synthesis · paper_section · claim_review · approval · final_review
      • TaskOne worker doing one bounded job

        A single agent session in its own Git worktree, with a token budget. It finishes with one typed JSON envelope. The daemon checks the work, and then a separate reviewer decides whether the Node is accepted.

  • Research DirectionOne question, answered rigorously

    A falsifiable question with a scope, stopping rules, and a frozen evaluation contract, so every result inside it is comparable. A Direction can stand on its own or serve as one Node of a Mission.

    • BaselineThe starting point everything is measured against

      The daemon runs the evaluation on the unchanged code, and a reviewer accepts the numbers. The baseline is the root of the experiment tree.

    • AutoresearchA search loop under fixed rules

      Runs rounds of experiments against one primary metric. It enforces guardrails, a minimum meaningful effect, and limits on iterations and compute. It can also run on its own, outside a Direction.

      • ExperimentOne measured variant

        A change on its own branch with exactly one parent, compared against that parent and the baseline. Once an experiment is evaluated, its code is frozen, and any further change becomes a child experiment. See how an experiment is kept.

    • ClosureThe Direction's answer

      The Direction closes as supported, refuted, or inconclusive, with a reviewed closure packet. A negative answer still counts as an answer, and it stays in the record.

  • EvidenceHow results become a paper

    Missions and Directions both write here. Reviewers read these records for themselves instead of trusting a worker's summary.

    ArtifactA content-addressed output: a diff, run result, source, or report.

    EvidenceAn Artifact with a meaning, such as a baseline run, an ablation, a citation, or a review verdict.

    ClaimA statement linked to the evidence that supports, contradicts, or qualifies it.

    PaperBuilt from reviewed claims, and exported to LaTeX or PDF.

1. Proposal

   An agent proposes a change and the effect it expects.

2. Implementation

   A worker builds it in its own Git worktree.

3. Evaluation Job

   The daemon runs the contract's evaluation command as a supervised Job.

4. Result

   Metrics and guardrails are checked against the frozen contract.

5. Review

   A separate session reads the diff and the run records, not the worker's claims.

6. Decision

   Only reviewed winners are promoted.

   promoteholdreject

keep = run.valid
     ∧ every guardrail passes
     ∧ improvement ≥ contract.min_effect
     ∧ independent review agrees

spent = input + output
      + cache_writes
      + 0.1 × cache_reads  // per task and step

Turn a repository into a Project

Discovery proposes a research contract with your test and evaluation commands. On approval, NOSH commits just the contract files.

Run Missions

Pause, steer, retry, or stop at any time. Every result is reviewed before it counts.

Answer research questions

A Direction measures a baseline, runs Autoresearch, and closes as supported, refuted, or inconclusive.

Search for improvements

Autoresearch evaluates each experiment under the same contract, enforces guardrails, and promotes only reviewed winners.

Keep long calls off the critical path

Tool calls in one turn run in parallel. A slow test run returns a placeholder, and the result wakes the agent when it lands.

Work through many files in one step

Agents write a short program that reads files together and keeps only what matters. In our tests, a 13-file task used about a fifth of the tokens.

Run your commands safely

Workers never get a shell. They ask the daemon to run commands declared in the contract, each as a supervised Job.

See what is happening

Press Ctrl+G in the terminal UI for a live map of Missions, Directions, experiments, workers, and Jobs.

Use any model account

Sign in with a ChatGPT or Claude subscription, or use an API key from OpenAI, Anthropic, Gemini, OpenRouter, DeepSeek, Groq, Mistral, xAI, and others.

Survive restarts

Closing the terminal UI does not stop the daemon or its Jobs. Orchestration resumes from its last checkpoint.