| name | inlab-human |
| description | In-lab human interaction with deployed prediction endpoints — the Claude Code chat is the UI. Tier-1 orchestrator with two modes: CONSOLE (default; patient-first on-demand inference: pick patient → see chart → list models → predict via the endpoint-predict tool → analyze) and STUDY (formal reader protocol: bundle → blind/assisted review → decision-influence report). Endpoint-agnostic: local Flask, Databricks, SageMaker — same wire contract. Trigger: inlab, console, get patient, run prediction, clinician eval, reader study, /inlab-human. |
| argument-hint | [console|patient|models|predict | bundle|review|report | status] [args...] |
| allowed-tools | Bash, Read, Write, Grep, Glob, Skill |
| metadata | {"version":"0.2.0","last_updated":"2026-07-10","summary":"Tier-1: console mode (default, on-demand inference) + study mode (reader protocol)."} |
Skill: inlab-human (in-lab human evaluation, tier-1)
Two modes, one plugin:
CONSOLE (default) clinician DRIVES: patient → chart → models → on-demand
prediction → analysis. -> /inlab-human-console
STUDY clinician is MEASURED: frozen bundle → blind→assisted
reading → influence metrics. -> bundle / review / report
/inlab-human -> console (orient: patients + models)
/inlab-human <patient_id> [model] -> console (jump straight in)
/inlab-human console | patient | models | predict ... -> console verbs
/inlab-human status -> dashboard: patients, models(live?),
bundles, responses
/inlab-human bundle <endpoint_path> [...] -> /inlab-human-bundle (study: freeze cases)
/inlab-human review <bundle.json> -> /inlab-human-review (study: run session)
/inlab-human report <bundle.json> -> /inlab-human-report (study: metrics)
Routing
Step 1: parse $ARGUMENTS.
Step 2: verb -> dispatch:
(nothing) | console | patient | models | predict | a patient id
-> Skill inlab-human-console (DEFAULT — on-demand inference)
bundle -> Skill inlab-human-bundle (needs: endpoint path or live URL)
review -> Skill inlab-human-review (needs: a frozen review_bundle.json)
report -> Skill inlab-human-report (needs: bundle + responses.jsonl)
status -> dashboard: list_patients + list_models (via endpoint-predict
tools or CLI twin) + scan for review_bundle*.json /
responses*.jsonl; suggest the next verb.
Step 3: relay the specialist's return contract verbatim.
Hard rules (enforced across the family)
- The score is never LLM-produced. All scores come from the
endpoint-predict tool (MCP server, or its CLI twin
mcp-servers/endpoint-predict/predict_cli.py via Bash).
- Frozen bundle. Reading sessions consume a static
review_bundle.json;
never call a live endpoint mid-session.
- gold is invisible during a session. Only the report skill reads it.
- Study data lives with the study (the project repo), never in this plugin.
Contracts: ../../ref/review-bundle-schema.md · ../../ref/feedback-form.md.
Return contract
status: ok | blocked | failed
summary: what was done / what the dashboard shows
artifacts: [paths]
next: suggested next command
