肿瘤与正常样本长读长研究流程

references/agent-playbook.md

1. Establish one true matched PacBio HiFi tumor/normal pair from acquisition metadata. Require patient and sample IDs; do not invent clinical identifiers.
2. Select a reference with exact assembly, contig naming and resource provenance. FASTA alone does not supply VEP, SV annotation, clonality or signature resources.
3. Prepare artifacts with the helper's --dry-run and --use-current-path. Report each warning and separate file/name checks from unperformed BAM, schema, network, container and scheduler checks.
4. Review samplesheet.csv: tumor status 1, normal status 0, same patient, distinct samples/files. Confirm that optional .pbi indexes correspond to their BAM. Confirm that MM/ML modification tags survived input processing before expecting methylation results.
5. Review commit pin, params, profiles and generated shell script. Read local infrastructure configuration before launch. --executor slurm by itself does not set the task executor to Slurm. Default config.yaml is descriptive and is not loaded by Python.
6. Execute only when requested. A new run uses --run; reusing generated artifacts needs --overwrite. If using a local checkout, record git HEAD and any uncommitted diff separately and omit --pipeline-version.
7. Return the exact artifact/script paths, run type and checked scope, requested pipeline revision, submission job ID if parsed, and one next QC/triage action. Do not label a queued job complete or helper dry-run as pipeline validation.
8. On failure, preserve the driver log, .nextflow.log, work task logs, params and samplesheet. Do not delete work/cache before deciding whether to resume. Keep secrets and identifiable sequencing metadata out of shared logs/reports.

All example commands in the main skill use repository-root paths. When the skill is installed elsewhere, resolve scripts/run_pacsomatic.py relative to its skill directory; .github/skills/pacsomatic is not this repository's layout.