Ma End To End

Overview

Coordinate the complete meta-analysis workflow, ensure every step is tracked, and produce a final manuscript with reviewer responses.

Inputs

• TOPIC.txt
• Optional user constraints such as population, outcomes, time window, study types, or target journal.

Outputs

• Standard project layout and all step artifacts described below.
• Final rendered manuscript in 07_manuscript/.
• Reviewer notes in 08_reviews/.

Project Layout (Numbered)

Create a numbered top-level structure and keep every artifact in its step folder.

01_protocol/
02_search/
03_screening/
04_fulltext/
05_extraction/
06_analysis/
07_manuscript/
08_reviews/
09_qa/
tooling/python/   # uv project

Environment Setup

1. Initialize Python tooling with uv inside tooling/python/ using uv init.
2. Use uv add to manage dependencies for search and automation scripts.
3. Run Python scripts via uv run (do not call python3 directly).
4. Use uv tool for any external CLI helpers that should be isolated.
5. Use R with renv inside 06_analysis/ for reproducible meta-analysis.

Workflow

1. Read TOPIC.txt and produce protocol artifacts in 01_protocol/.

- Read from projects/<project-name>/TOPIC.txt - Use /ma-topic-intake skill - Write to 01protocol/pico.yaml, 01protocol/eligibility.md, 01protocol/outcomes.md, 01protocol/search-plan.md, 01protocol/decision-log.md 1b. Preliminary analysis type: ≥3 treatments → nmacandidate, 2 treatments → pairwise. - Record in 01protocol/pico.yaml (L22: analysistype.preliminary field) - Record in 01_protocol/analysis-type-decision.md (Stage 1 section)

1. Plan and run database searches, then save round-based .bib files in 02_search/.

- Use /ma-search-bibliography skill - Write to 02search/round-01/queries.txt, 02search/round-01/results.bib, 02search/round-01/dedupe.bib, 02search/round-01/log.md

1. Screen titles and abstracts, record decisions, and generate included .bib in 03_screening/.

- Use /ma-screening-quality skill - Write to 03screening/round-01/decisions.csv, 03screening/round-01/included.bib, 03screening/round-01/agreement.md 3b. Analysis Type Confirmation Gate (if nmacandidate): - Tally study designs, assess network connectivity and transitivity - If >30% single-arm → strongly consider downgrading to pairwise + pooled proportions - Confirm in 01protocol/analysis-type-decision.md (Stage 2 section) - Update 01protocol/pico.yaml (L23: analysis_type.confirmed field) - Do NOT proceed to Stage 06 without confirmed analysis type

1. Collect full texts and build a manifest in 04_fulltext/.

- Use /ma-fulltext-management skill - Write to 04fulltext/manifest.csv, 04fulltext/*.pdf 4b. Full-text eligibility screening (PRISMA 2020 item 16 — mandatory). - Use /ma-fulltext-management skill (Stage 04b section) - Run uv run tooling/python/aiscreen.py --project <name> --stage fulltext --reviewer 1 - Run uv run tooling/python/aiscreen.py --project <name> --stage fulltext --reviewer 2 - Compute kappa: uv run ma-screening-quality/scripts/dualreviewagreement.py --file 04fulltext/fulltextdecisions.csv --col-a FTReviewer1Decision --col-b FTReviewer2Decision --out 04fulltext/ftagreement.md - Resolve conflicts, then only FTFinalDecision = include rows proceed to Stage 05 - Write to 04fulltext/fulltextdecisions.csv, 04fulltext/ftagreement.md

1. Extract data into a normalized database in 05_extraction/.

- Input: Only studies with FTFinalDecision = include from 04fulltext/fulltextdecisions.csv - Use /ma-data-extraction skill - Write to 05extraction/extraction.sqlite, 05extraction/extraction.csv, 05_extraction/data-dictionary.md

1. Run meta-analysis in R with renv, generate figures and tables in 06_analysis/.

- Route by analysistype.confirmed: pairwise | nma | pooledproportion | narrative - Use /ma-meta-analysis skill for pairwise - Use /ma-network-meta-analysis skill for NMA - Write to 06analysis/.R, 06analysis/figures/.png, 06analysis/tables/*.csv, 06analysis/renv.lock

1. Draft and render Quarto manuscript in 07_manuscript/.

- Use /ma-manuscript-quarto skill - Write to 07manuscript/*.qmd, 07manuscript/index.html, 07_manuscript/index.pdf

1. Perform Reviewer 1 and Reviewer 2 checks and save notes in 08_reviews/.

- Use /ma-peer-review skill - Write to 08reviews/gradesummary.csv, 08reviews/rob2assessment.csv

1. Maintain cross-step validation logs in 09_qa/.

- Write to 09_qa/pipeline-checklist.md

1. Add robustness checks: GRADE profiles, dual-review agreement stats, and PRISMA flow summary.

- Use scripts/runrobustnesschecks.py

1. Optionally run scripts/runrobustnesschecks.py via uv run to generate all robustness artifacts at once.

- Use scripts/runrobustnesschecks.py

1. Apply publication-quality checks (PRISMA/MOOSE, HK, influence, SoF, claim audit, crossref).

- Use /ma-publication-quality skill - Write to 09qa/claimaudit.md, 09qa/crossrefreport.md, 09qa/reportingchecklist_audit.md

1. Validate stage transitions with scripts/validatestagetransition.py and store reports in 09_qa/.

- Use scripts/validatestagetransition.py - Write to 09qa/stagetransition_report.md

1. Create checkpoints before major steps with scripts/checkpoint.py.

- Use scripts/checkpoint.py - Creates .checkpoint/ snapshots

Agent Teams (Parallel Mode)

When running with agent teams enabled (CLAUDECODEEXPERIMENTALAGENTTEAMS=1), the pipeline can leverage parallel teammates for independent stages.

Parallelism Opportunities

How to Start

1. User says "create a team for [project]" or "start team mode"
2. Lead reads /ma-agent-teams skill for the orchestration playbook
3. Lead creates shared task list with 12 tasks and dependencies
4. Lead spawns teammates in phased order (see SKILL.md for details)
5. Hooks enforce quality gates at stage transitions

Quality Gates (Lead Enforces)

• Stage 03→04: Screening kappa ≥ 0.60 (lead computes after both reviewers finish)
• Stage 04→05: FT screening kappa ≥ 0.60
• Stage 05→06: Extraction completeness (all included studies extracted)
• Stage 06→07: All figures ≥ 300 DPI
• Stage 09: PRISMA 27/27 (or 32/32 for NMA), publication readiness ≥ 95%

Generate Spawn Prompts

uv run tooling/python/team_spawn_helper.py --project <project-name> --role <role-name>

See ma-agent-teams/SKILL.md for complete orchestration details.

Resources

• scripts/init_project.py creates the numbered folder tree and a checklist.
• scripts/runrobustnesschecks.py runs agreement stats, PRISMA flow, and GRADE summaries.
• scripts/validate_pipeline.py enforces checklist completion before final render.
• scripts/finalqareport.py generates a final QA report and blocks on failures.
• scripts/validatestagetransition.py validates continuity between stages.
• scripts/checkpoint.py creates and restores pipeline checkpoints.
• scripts/hash_artifacts.py computes SHA-256 hashes for reproducibility audit.
• scripts/validatemoduleregistry.py checks all scripts are documented across SKILL.md, CLAUDE.md, and GETTING_STARTED.md.

Step References

Open the relevant skill for details at each stage:

• ma-topic-intake/SKILL.md
• ma-search-bibliography/SKILL.md
• ma-screening-quality/SKILL.md
• ma-fulltext-management/SKILL.md
• ma-data-extraction/SKILL.md
• ma-meta-analysis/SKILL.md
• ma-manuscript-quarto/SKILL.md
• ma-peer-review/SKILL.md
• ma-publication-quality/SKILL.md

Validation

• Ensure each step writes its expected artifacts before moving to the next.
• Create and update 09_qa/pipeline-checklist.md after every milestone.