[LifeSci Skill] Add bounded target-evidence research

plugins/life-science-research/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "life-science-research",
-  "version": "1.0.3",
+  "version": "1.1.0",
   "description": "General life-sciences research workflows with query routing, evidence synthesis, and optional parallel subagent analysis across genetics, omics, biology, chemistry, structure, clinical evidence, and public dataset discovery.",
   "author": {
     "name": "OpenAI"

plugins/life-science-research/README.md

@@ -45,7 +45,7 @@ This plugin is meant to support workflows like:
 
 ## Skill Families
 
-The plugin currently bundles 50 skills. The most useful way to think about them is by research area rather than as a flat list.
+The plugin currently bundles 51 skills. The most useful way to think about them is by research area rather than as a flat list.
 
 ### Human Genetics And Variant Evidence
 
@@ -94,6 +94,7 @@ The plugin currently bundles 50 skills. The most useful way to think about them
 
 ### Clinical, Translational, And Disease Evidence
 
+- `research-target-evidence-skill`
 - `clinicaltrials-skill`
 - `cbioportal-skill`
 - `civic-skill`
@@ -165,6 +166,7 @@ Each subagent should receive a bounded objective and return concise findings, ca
 - `Map the most plausible causal genes at this inflammatory bowel disease locus and explain why.`
 - `Summarize known structure, ligand, and pathway information for EGFR.`
 - `Pull ClinicalTrials.gov, ChEMBL, and PharmGKB context for JAK inhibitors in alopecia areata.`
+- `Use $research-target-evidence-skill to separate human and preclinical evidence for ROR1 biology, therapeutic programs, and safety.`
 - `Find metabolomics and proteomics resources relevant to MASLD and PPARG.`
 - `Interpret this variant using ClinVar, gnomAD, Ensembl, and cohort association evidence.`
 

plugins/life-science-research/skills/research-router-skill/SKILL.md

@@ -68,6 +68,8 @@ Choose the smallest set of skills that can answer the question well.
 
 Examples:
 
+- bounded target biology, program, and safety review:
+  `research-target-evidence-skill`
 - target or disease evidence review:
   `opentargets-skill`, `gwas-catalog-skill`, `gtex-eqtl-skill`, `human-protein-atlas-skill`
 - variant interpretation:
@@ -79,7 +81,7 @@ Examples:
 - chemistry and pharmacology:
   `chembl-skill`, `bindingdb-skill`, `pubchem-pug-skill`, `pharmgkb-skill`
 - clinical and translational:
-  `clinicaltrials-skill`, `cbioportal-skill`, `civic-skill`
+  `research-target-evidence-skill`, `clinicaltrials-skill`, `cbioportal-skill`, `civic-skill`
 - literature and dataset discovery:
   `ncbi-entrez-skill`, `ncbi-pmc-skill`, `biorxiv-skill`, `biostudies-arrayexpress-skill`, `ncbi-datasets-skill`
 

plugins/life-science-research/skills/research-target-evidence-skill/SKILL.md

@@ -0,0 +1,62 @@
+---
+name: research-target-evidence-skill
+description: Produce bounded, source-backed evidence briefs for one or more biological targets, covering biology, therapeutic programs, human safety, preclinical evidence, and cross-target comparison. Use when a user asks for a target assessment, target comparison, translational evidence, program history, modality comparison, or human-versus-preclinical safety review and wants a fast primary-source research pass.
+---
+
+## Research Target Evidence
+
+Use the bundled script exactly once for all requested targets. Let it plan,
+deduplicate, globally pace, and batch requests across PubMed and
+ClinicalTrials.gov.
+
+Do not decompose the request into additional source calls unless the script
+returns `ok=false` or the user explicitly asks for a deeper follow-up.
+
+## Execution
+
+For one target, run:
+
+```bash
+python scripts/research_target_evidence.py \
+  --target "<target>" \
+  --questions biology programs safety \
+  --separate-human-preclinical
+```
+
+For a comparison, repeat `--target` in the same command:
+
+```bash
+python scripts/research_target_evidence.py \
+  --target "<target 1>" \
+  --target "<target 2>" \
+  --target "<target 3>" \
+  --mode compare \
+  --questions biology programs safety \
+  --separate-human-preclinical
+```
+
+Do not run one process per target. The shared process enforces global source
+pacing, preserves landmark-program and evidence-class quotas, and emits a
+single size-bounded JSON result.
+
+## Synthesis
+
+- Lead with the target-level conclusion and the largest uncertainty.
+- Separate human evidence from preclinical evidence.
+- For comparisons, use the same evidence axes for every target and finish with
+  a compact comparison of validation, selectivity, safety, modality maturity,
+  and uncertainty.
+- Cover only modalities supported by the returned papers or trial records.
+- Link each PMID as `[PMID <id>](https://pubmed.ncbi.nlm.nih.gov/<id>/)`.
+- Link each trial as `[<NCT id>](https://clinicaltrials.gov/study/<NCT id>)`.
+- Distinguish observed human toxicity from preclinical or theoretical risk.
+- Treat registry adverse-event counts as non-attributed unless the record says
+  otherwise.
+- Preserve the returned limitations and current registry statuses.
+- Report per-target errors or omitted evidence counts rather than silently
+  treating a partial result as complete.
+- Keep the answer concise enough that the evidence hierarchy remains visible.
+
+The retrieval is bounded and relevance-ranked, not a systematic review. Check
+the script's heuristic human/preclinical classification during synthesis. Do
+not include retrieval telemetry in the user-facing brief unless requested.

plugins/life-science-research/skills/research-target-evidence-skill/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Research Target Evidence"
+  short_description: "Bounded target evidence and comparison"
+  default_prompt: "Use $research-target-evidence-skill to assess or compare target biology, therapeutic programs, and safety evidence."