AutoR: Human-Centered AI Research Co-pilot

A terminal-first, file-based research workflow runner for long-form AI-assisted research.
It drives a fixed 8-stage research pipeline, requires human approval after every stage, and writes every prompt, log, summary, and artifact into an isolated run directory.

Why AutoR · Showcase · Quick Start · How It Works · Run Layout · Architecture · Roadmap

AutoR is not a chat demo, not a generic agent framework, and not a markdown-only research toy.

It is a research execution loop: goal -> literature -> hypothesis -> design -> implementation -> experiments -> analysis -> paper -> dissemination, with explicit human control at every stage and real artifacts on disk.

✨ Why AutoR

Most AI research demos stop at "the model wrote a plausible summary."

AutoR is built around a harder standard: the system should leave behind a run directory that another person can inspect, resume, audit, and critique.

What makes it different

AutoR does	Why it matters
Fixed 8-stage research workflow	The system behaves like a real research process instead of a free-form chat loop.
Mandatory human approval after every stage	AI executes; humans retain control at high-leverage decision points.
Full run isolation under runs/<run_id>/	Prompts, logs, stage outputs, code, figures, and papers are all auditable.
Draft -> validate -> promote for stage summaries	Half-finished summaries do not silently become official stage records.
Artifact-aware validation	Later stages must produce data, results, figures, LaTeX, PDF, and review assets, not just prose.
Resume and redo-stage support	Long runs are recoverable and partially repeatable.
Stage-local conversation continuation	Refinement improves the current stage instead of constantly resetting context.
Venue-aware writing stage	Stage 07 can target lightweight conference or journal-style paper packaging without pretending to be a full submission system.

Core guarantees

• A run is isolated under runs/<run_id>/.
• Claude never writes directly to the final stage summary file.
• Human approval is required before the workflow advances.
• Approved summaries are appended to memory.md; failed attempts are not.
• Stage 03+ must produce machine-readable data artifacts.
• Stage 05+ must produce machine-readable result artifacts.
• Stage 06+ must produce real figure files.
• Stage 07+ must produce a venue-aware manuscript package with a PDF.
• Stage 08+ must produce review and readiness materials.

🌟 Showcase

AutoR already has a full example run used throughout the repository: runs/20260330_101222.

That run produced:

• a compiled paper PDF: example_paper.pdf
• executable research code
• machine-readable datasets and result files
• real figures used in the paper
• review and dissemination materials

Highlighted outcomes from that run:

• AGSNv2 reached 36.21 ± 1.08 on Actor
• the system produced a full NeurIPS-style paper package
• the final run preserved the full human-in-the-loop approval trail

Terminal Experience

AutoR is designed for terminal-first execution, but the interaction layer is not limited to raw logs and plain prompts. The current UI supports banner-style startup, colored stage panels, parsed Claude event streams, wrapped markdown summaries, and a menu-driven approval loop suitable for demos and recordings.

Example Figures

Accuracy Comparison	Ablation + Actor Results
Two-Layer Narrative Figure

Paper Preview

Page 1 Title, abstract, framing

Page 5 Method and training algorithm

Page 7 Main tables and per-seed results

Human-in-the-Loop in Practice

The example run is interesting not because the AI was left alone, but because the human intervened at critical moments:

• Stage 02 narrowed the project to a single core claim.
• Stage 04 pushed the system to download real datasets and run actual pre-checks.
• Stage 05 forced experimentation to continue until real benchmark results were obtained.
• Stage 06 redirected the story away from leaderboard-only framing toward mechanism-driven analysis.

That is the intended shape of AutoR: AI handles execution load; humans steer the research when direction actually matters.

🚀 Quick Start

Prerequisites

• Python 3.10+
• Claude CLI available on PATH for real runs
• Local TeX tools are helpful for Stage 07, but not required for smoke tests

Start a new run

python main.py

Start with an explicit goal

python main.py --goal "Your research goal here"

Run a local smoke test without Claude

python main.py --fake-operator --goal "Smoke test"

Choose a Claude model

python main.py --model sonnet
python main.py --model opus

Choose a writing venue profile

python main.py --venue neurips_2025
python main.py --venue nature
python main.py --venue jmlr

If --venue is omitted, AutoR defaults to neurips_2025.

Resume or redo work inside the same run

python main.py --resume-run latest
python main.py --resume-run 20260329_210252 --redo-stage 03

Valid stage identifiers include 03, 3, and 03_study_design.

⚙️ How It Works

AutoR uses a fixed 8-stage pipeline:

1. 01_literature_survey
2. 02_hypothesis_generation
3. 03_study_design
4. 04_implementation
5. 05_experimentation
6. 06_analysis
7. 07_writing
8. 08_dissemination

flowchart TD
    A[Start or resume run] --> S1[01 Literature Survey]
    S1 --> H1{Human approval}
    H1 -- Refine --> S1
    H1 -- Approve --> S2[02 Hypothesis Generation]
    H1 -- Abort --> X[Abort]

    S2 --> H2{Human approval}
    H2 -- Refine --> S2
    H2 -- Approve --> S3[03 Study Design]
    H2 -- Abort --> X

    S3 --> H3{Human approval}
    H3 -- Refine --> S3
    H3 -- Approve --> S4[04 Implementation]
    H3 -- Abort --> X

    S4 --> H4{Human approval}
    H4 -- Refine --> S4
    H4 -- Approve --> S5[05 Experimentation]
    H4 -- Abort --> X

    S5 --> H5{Human approval}
    H5 -- Refine --> S5
    H5 -- Approve --> S6[06 Analysis]
    H5 -- Abort --> X

    S6 --> H6{Human approval}
    H6 -- Refine --> S6
    H6 -- Approve --> S7[07 Writing]
    H6 -- Abort --> X

    S7 --> H7{Human approval}
    H7 -- Refine --> S7
    H7 -- Approve --> S8[08 Dissemination]
    H7 -- Abort --> X

    S8 --> H8{Human approval}
    H8 -- Refine --> S8
    H8 -- Approve --> Z[Run complete]
    H8 -- Abort --> X

Loading

Stage Attempt Loop

flowchart TD
    A[Build prompt from template + goal + memory + optional feedback] --> B[Start or resume stage session]
    B --> C[Claude writes draft stage summary]
    C --> D[Validate markdown and required artifacts]
    D --> E{Valid?}
    E -- No --> F[Repair, normalize, or rerun current stage]
    F --> A
    E -- Yes --> G[Promote draft to final stage summary]
    G --> H{Human choice}
    H -- 1 or 2 or 3 --> I[Continue current stage conversation with AI refinement]
    I --> A
    H -- 4 --> J[Continue current stage conversation with custom feedback]
    J --> A
    H -- 5 --> K[Append approved summary to memory.md]
    K --> L[Continue to next stage]
    H -- 6 --> X[Abort]

Loading

Approval semantics

• 1 / 2 / 3: continue the same stage conversation using one of the AI's refinement suggestions
• 4: continue the same stage conversation with custom user feedback
• 5: approve and continue to the next stage
• 6: abort the run

The stage loop is controlled by AutoR, not by Claude.

✅ Validation Bar

AutoR does not consider a run successful just because it generated a plausible markdown summary.

Stage	Required non-toy output
Stage 03+	Machine-readable data under workspace/data/
Stage 05+	Machine-readable results under workspace/results/
Stage 06+	Real figure files under workspace/figures/
Stage 07+	Venue-aware manuscript sources plus a compiled PDF
Stage 08+	Review and readiness assets under workspace/reviews/

Required stage summary shape:

# Stage X: <name>

## Objective
## Previously Approved Stage Summaries
## What I Did
## Key Results
## Files Produced
## Suggestions for Refinement
## Your Options

Additional rules:

• exactly 3 numbered refinement suggestions
• the fixed 6 user options
• no [In progress], [Pending], [TODO], [TBD], or similar placeholders
• concrete file paths in Files Produced

If a run only leaves behind markdown notes, it has not met AutoR's quality bar.

📂 Run Layout

Every run lives entirely inside its own directory.

runs/<run_id>/
├── user_input.txt
├── memory.md
├── run_config.json
├── logs.txt
├── logs_raw.jsonl
├── prompt_cache/
├── operator_state/
├── stages/
└── workspace/
    ├── literature/
    ├── code/
    ├── data/
    ├── results/
    ├── writing/
    ├── figures/
    ├── artifacts/
    ├── notes/
    └── reviews/

Directory semantics

• literature/: reading notes, survey tables, benchmark notes
• code/: runnable code, scripts, configs, implementations
• data/: machine-readable data and manifests
• results/: machine-readable experiment outputs
• writing/: LaTeX sources, sections, bibliography, tables
• figures/: real plots and paper figures
• artifacts/: compiled PDFs and packaged deliverables
• notes/: temporary or supporting research notes
• reviews/: readiness, critique, and dissemination materials

🧠 Execution Model

For each stage attempt, AutoR assembles a prompt from:

1. the stage template from src/prompts/
2. the required stage summary contract
3. execution-discipline constraints
4. user_input.txt
5. approved memory.md
6. optional refinement feedback
7. for continuation attempts, the current draft/final stage files and workspace context

The assembled prompt is written to runs/<run_id>/prompt_cache/, per-stage session IDs are stored in runs/<run_id>/operator_state/, and Claude is invoked in live streaming mode.

Exact Claude CLI pattern

First attempt for a stage:

claude --model <model> \
  --permission-mode bypassPermissions \
  --dangerously-skip-permissions \
  --session-id <stage_session_id> \
  -p @runs/<run_id>/prompt_cache/<stage>_attempt_<nn>.prompt.md \
  --output-format stream-json \
  --verbose

Continuation attempt for the same stage:

claude --model <model> \
  --permission-mode bypassPermissions \
  --dangerously-skip-permissions \
  --resume <stage_session_id> \
  -p @runs/<run_id>/prompt_cache/<stage>_attempt_<nn>.prompt.md \
  --output-format stream-json \
  --verbose

Important behavior:

• refinement attempts reuse the same stage conversation whenever possible
• streamed Claude output is shown live in the terminal
• raw stream-json output is captured in logs_raw.jsonl
• if resume fails, AutoR can fall back to a fresh session
• if stage markdown is incomplete, AutoR can repair or normalize it locally

🏗️ Architecture

The main code lives in:

• main.py
• src/manager.py
• src/operator.py
• src/utils.py
• src/writing_manifest.py
• src/prompts/

flowchart LR
    A[main.py] --> B[src/manager.py]
    B --> C[src/operator.py]
    B --> D[src/utils.py]
    B --> E[src/writing_manifest.py]
    B --> F[src/prompts/*]
    C --> D

Loading

File boundaries

• main.py: CLI entry point; starts new runs or resumes old ones
• src/manager.py: owns the 8-stage loop, approval flow, repair flow, and stage continuation policy
• src/operator.py: invokes Claude CLI, streams output, persists session IDs, resumes stage conversations, and falls back on resume failure
• src/utils.py: stage metadata, run paths, prompt assembly, markdown validation, artifact validation, and venue resolution
• src/writing_manifest.py: scans figures, results, data files, and stage summaries to generate Stage 07 writing context
• src/prompts/: one prompt template per stage

📌 Scope

Included in the current mainline

• fixed 8-stage workflow
• mandatory human approval after every stage
• one primary Claude invocation per stage attempt
• stage-local continuation within the same Claude session
• prompt caching via @file
• live streaming terminal output
• repair passes and local fallback normalization
• draft-to-final stage promotion
• artifact-aware validation
• resume and --redo-stage
• lightweight venue profiles for Stage 07 writing

Intentionally out of scope

• generic multi-agent orchestration
• database-backed runtime state
• concurrent stage execution
• heavyweight platform abstractions
• dashboard-first productization

🛣️ Roadmap

The most valuable next steps are the ones that make AutoR more like a real research workflow, not more like a demo framework.

• Cross-stage rollback and invalidation Later-stage failures should be able to mark downstream work as stale.
• Machine-readable run manifest Add a lightweight source of truth for stage status, stale dependencies, and artifact pointers.
• Continuation handoff compression Make long stage refinement more stable without bloating context.
• Stronger automated tests Cover repair flow, resume fallback, artifact validation, and approval-loop correctness.
• Artifact indexing Add lightweight metadata around data/, results/, figures/, and writing/.
• Frontend run browser A lightweight UI for browsing runs, stages, logs, and artifacts, driven by the run directory itself.

📝 Notes

• runs/ is gitignored.
• AutoR controls workflow orchestration, not scientific truth.
• Submission-grade output still depends on the environment, model quality, local tools, and available datasets.
• Stage 07 venue support is intentionally lightweight metadata-driven packaging, not a promise of full official template compliance for every venue.

🌍 Community

Join the project community channels:

Discord	WeChat	WhatsApp

⭐ Star History