mirror of
https://github.com/MBombeck/HealthLog.git
synced 2026-11-08 12:32:43 +00:00
Marc-André Bombeck
c2ab9cda11
The live evaluation now captures the tool calls a generation makes, with their arguments, and hands them to the judge. runScenarioLive puts a scenario to the model in the chat turn's own context (system prompt, tool-mode and dialog addenda, DATA INVENTORY, EARLIER TABLES, transcript) over the real tool catalogue, answers each call with a stand-in result, and reads the reply through the turn's clarification and chip parsers. runScenarioJudge grades each scenario a few times with the same evaluateScenario the deterministic suite uses and reports a pass rate per scenario; the server-only scenarios and checks are left out. Like the existing judge it runs only with COACH_EVAL_API_KEY, from the nightly workflow's non-blocking step, and never throws.
72 lines
2.9 KiB
YAML
72 lines
2.9 KiB
YAML
name: Coach eval
|
|
|
|
# B0 (v1.21.3) — the Coach evaluation harness, two layers.
|
|
#
|
|
# 1. The DETERMINISTIC suite (golden set + red-team battery + the grounding
|
|
# grader + the dialog scenarios) runs ALWAYS. It is free, offline, and the per-change floor — the
|
|
# same suite `pnpm test` runs, pinned here as a standing nightly check.
|
|
# 2. The LIVE JUDGE step runs ONLY when the `COACH_EVAL_API_KEY` repository
|
|
# secret is present, and is NON-BLOCKING (`continue-on-error: true`). With
|
|
# the secret absent it no-ops with a clear log line — the workflow stays
|
|
# green. The operator flips the secret on later; no rebuild needed. The
|
|
# same step puts the dialog scenarios to the model and logs a pass rate
|
|
# per scenario, with the tool calls and their arguments.
|
|
#
|
|
# Schedule + manual only. This is NOT a required check.
|
|
|
|
on:
|
|
schedule:
|
|
# Nightly at 04:20 UTC — off the busy hour, after any evening merges.
|
|
- cron: "20 4 * * *"
|
|
workflow_dispatch:
|
|
|
|
permissions:
|
|
contents: read
|
|
|
|
jobs:
|
|
coach-eval:
|
|
name: Coach eval — deterministic floor (always) + live judge (opt-in)
|
|
runs-on: ubuntu-latest
|
|
timeout-minutes: 15
|
|
|
|
# Map the opt-in secret to a job-level env so the live-judge steps can gate
|
|
# on it. The `secrets` context is NOT available in `if:` expressions — using
|
|
# it there is a workflow-file validation error that surfaces as a 0-second
|
|
# startup failure on every push. `env` is allowed in `if:`, so gate on that.
|
|
env:
|
|
COACH_EVAL_API_KEY: ${{ secrets.COACH_EVAL_API_KEY }}
|
|
|
|
steps:
|
|
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
|
|
|
|
- uses: pnpm/action-setup@ea17c68df8912ef543352723c149a84f56e3d413 # v6.1.0
|
|
|
|
- uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
|
|
with:
|
|
node-version: 22
|
|
cache: pnpm
|
|
|
|
- run: pnpm install --frozen-lockfile
|
|
|
|
- run: pnpm db:generate
|
|
|
|
# Layer 1 — the deterministic floor. Free, offline, ALWAYS runs. A
|
|
# failure here reds the workflow: a regression in the golden set or the
|
|
# red-team battery is a real find.
|
|
- name: Deterministic floor (golden set + red team + dialog scenarios)
|
|
run: pnpm exec vitest run src/lib/ai/coach/eval/__tests__/
|
|
|
|
# Layer 2 — the live judge. Runs ONLY when the secret is present, and is
|
|
# non-blocking either way. With the secret absent the script no-ops with a
|
|
# clear line and exits 0.
|
|
- name: Live judge (opt-in, non-blocking)
|
|
if: ${{ env.COACH_EVAL_API_KEY != '' }}
|
|
continue-on-error: true
|
|
run: pnpm dlx tsx scripts/run-coach-eval.ts
|
|
|
|
# When the secret is absent, surface a clear no-op line so the run log
|
|
# explains why the live judge did not execute.
|
|
- name: Live judge skipped (no secret)
|
|
if: ${{ env.COACH_EVAL_API_KEY == '' }}
|
|
run: echo "COACH_EVAL_API_KEY not set — live Coach judge skipped (deterministic floor still gated above)."
|