feat(e2e): run long-horizon Bub workloads

.github/workflows/e2e-harness.yml

@@ -94,92 +94,4 @@ jobs:
         if: always()
         env:
           POWERCONTEXT_E2E_DATABASE: ${{ matrix.database }}
-        run: e2e/bub/run.sh down
-
-  # live-replay:
-  #   needs: acceptance
-  #   if: github.event_name != 'pull_request'
-  #   runs-on: ubuntu-latest
-  #   timeout-minutes: 50
-  #   strategy:
-  #     fail-fast: false
-  #     matrix:
-  #       database: [sqlite, oceanbase]
-  #   name: Live replay (${{ matrix.database }})
-  #   steps:
-  #     - name: Check out
-  #       uses: actions/checkout@v7
-
-  #     - name: Check provider configuration
-  #       id: provider
-  #       env:
-  #         API_KEY: ${{ secrets.POWERCONTEXT_E2E_API_KEY }}
-  #         MODEL: ${{ vars.POWERCONTEXT_E2E_MODEL }}
-  #       run: |
-  #         if [[ -n "${API_KEY}" && -n "${MODEL}" ]]; then
-  #           echo "configured=true" >> "${GITHUB_OUTPUT}"
-  #         else
-  #           echo "configured=false" >> "${GITHUB_OUTPUT}"
-  #           echo "Live replay skipped because provider credentials are not configured." >> "${GITHUB_STEP_SUMMARY}"
-  #         fi
-
-  #     - name: Run live replay
-  #       if: steps.provider.outputs.configured == 'true'
-  #       env:
-  #         BUB_API_BASE: ${{ vars.POWERCONTEXT_E2E_API_BASE }}
-  #         BUB_API_KEY: ${{ secrets.POWERCONTEXT_E2E_API_KEY }}
-  #         BUB_MODEL: ${{ vars.POWERCONTEXT_E2E_MODEL }}
-  #         POWERCONTEXT_E2E_DATABASE: ${{ matrix.database }}
-  #         POWERCONTEXT_E2E_OUTPUT: ${{ github.workspace }}/.powercontext-e2e/${{ matrix.database }}/live
-  #       run: make harness-compose-live
-
-  #     - name: Scan replay evidence
-  #       id: evidence_scan
-  #       if: always() && steps.provider.outputs.configured == 'true'
-  #       continue-on-error: true
-  #       env:
-  #         EVIDENCE_PATH: ${{ github.workspace }}/.powercontext-e2e/${{ matrix.database }}/live
-  #       run: |
-  #         test -d "${EVIDENCE_PATH}"
-  #         docker run --rm \
-  #           --volume "${EVIDENCE_PATH}:/evidence:ro" \
-  #           "${TRUFFLEHOG_IMAGE}" \
-  #           filesystem /evidence \
-  #           --no-verification \
-  #           --results=verified,unknown,unverified \
-  #           --fail \
-  #           --fail-on-scan-errors \
-  #           --no-update \
-  #           --json > "${RUNNER_TEMP}/trufflehog-live-${{ matrix.database }}.jsonl"
-
-  #     - name: Report suppressed replay evidence
-  #       if: >-
-  #         always() &&
-  #         steps.provider.outputs.configured == 'true' &&
-  #         steps.evidence_scan.outcome != 'success'
-  #       run: echo "::warning::Replay evidence was not published because secret scanning did not complete cleanly."
-
-  #     - name: Publish replay summary
-  #       if: >-
-  #         always() &&
-  #         steps.provider.outputs.configured == 'true' &&
-  #         steps.evidence_scan.outcome == 'success'
-  #       run: cat ".powercontext-e2e/${{ matrix.database }}/live/report.md" >> "${GITHUB_STEP_SUMMARY}"
-
-  #     - name: Upload replay evidence
-  #       if: >-
-  #         always() &&
-  #         steps.provider.outputs.configured == 'true' &&
-  #         steps.evidence_scan.outcome == 'success'
-  #       uses: actions/upload-artifact@v7
-  #       with:
-  #         name: e2e-live-${{ matrix.database }}-${{ github.sha }}
-  #         path: .powercontext-e2e/${{ matrix.database }}/live
-  #         if-no-files-found: error
-  #         retention-days: 14
-
-  #     - name: Stop harness environment
-  #       if: always()
-  #       env:
-  #         POWERCONTEXT_E2E_DATABASE: ${{ matrix.database }}
-  #       run: e2e/bub/run.sh down
+        run: make harness-compose-down

Makefile

@@ -52,15 +52,9 @@ harness-check: ## Validate the Bub replay harness and committed scenarios.
 	@uv run --project e2e/bub powercontext-e2e --help >/dev/null
 
 .PHONY: harness-acceptance
-harness-acceptance: ## Run all deterministic Bub session replay scenarios against a Server.
-	@uv run --project e2e/bub powercontext-e2e acceptance e2e/bub/scenarios/*.yaml \
-		--output "$${POWERCONTEXT_E2E_OUTPUT:-e2e/bub/results}"
-
-.PHONY: harness-live
-harness-live: ## Run one real-model Bub session replay scenario against a Server.
-	@uv run --project e2e/bub powercontext-e2e live \
-		"$${POWERCONTEXT_E2E_SCENARIO:-e2e/bub/scenarios/project-database-decision.yaml}" \
-		--output "$${POWERCONTEXT_E2E_OUTPUT:-e2e/bub/results/live}"
+harness-acceptance: ## Evaluate workloads by ID or category against an existing Server.
+	@uv run --project e2e/bub powercontext-e2e acceptance \
+		--output "$${POWERCONTEXT_E2E_OUTPUT:-e2e/bub/results}" $(ARGS)
 
 .PHONY: harness-rescore
 harness-rescore: ## Rescore REPLAY without rerunning Bub or PowerContext.
@@ -74,17 +68,12 @@ harness-compose-check: ## Validate the SQLite and OceanBase Compose environments
 	@POWERCONTEXT_E2E_DATABASE=oceanbase e2e/bub/run.sh check
 
 .PHONY: harness-compose-acceptance
-harness-compose-acceptance: ## Build and run deterministic replay scenarios in containers.
-	@e2e/bub/run.sh acceptance
-
-.PHONY: harness-compose-live
-harness-compose-live: ## Build and run one real-model replay in containers.
-	@e2e/bub/run.sh live
+harness-compose-acceptance: ## Build and evaluate workloads by ID or category in the fixed harness.
+	@e2e/bub/run.sh acceptance $(ARGS)
 
 .PHONY: harness-compose-down
-harness-compose-down: ## Stop both isolated harness environments and remove their volumes.
-	@POWERCONTEXT_E2E_DATABASE=sqlite e2e/bub/run.sh down
-	@POWERCONTEXT_E2E_DATABASE=oceanbase e2e/bub/run.sh down
+harness-compose-down: ## Stop the selected isolated harness environment and remove its volumes.
+	@e2e/bub/run.sh down
 
 .PHONY: contract-test
 contract-test: api-generate-check ## Verify generated API code and contract bindings.

docker/Dockerfile

@@ -1,6 +1,6 @@
 # syntax=docker/dockerfile:1.7
 
-FROM ghcr.io/astral-sh/uv:0.10.12 AS uv
+FROM docker.io/astral/uv:0.10.12 AS uv
 
 
 FROM python:3.11-slim-bookworm AS builder

e2e/bub/Dockerfile

@@ -1,14 +1,24 @@
 # syntax=docker/dockerfile:1.7
 
-FROM ghcr.io/astral-sh/uv:0.10.12 AS uv
+FROM docker.io/astral/uv:0.10.12 AS uv
+
+FROM docker.io/library/docker:27.5.1-dind AS docker
 
 FROM python:3.12-slim-bookworm
 
 COPY --from=uv /uv /uvx /bin/
+COPY --from=docker /usr/local/bin/ /usr/local/bin/
+COPY --from=docker /usr/local/libexec/docker/cli-plugins/ /usr/local/libexec/docker/cli-plugins/
+
+RUN apt-get update \
+    && apt-get install --yes --no-install-recommends ca-certificates git iproute2 iptables pigz socat \
+    && rm -rf /var/lib/apt/lists/* \
+    && docker compose version
 
 ENV PATH="/workspace/powercontext/e2e/bub/.venv/bin:$PATH" \
     PYTHONUNBUFFERED=1 \
     PYTHONDONTWRITEBYTECODE=1 \
+    POWERCONTEXT_E2E_REPOSITORY=/workspace/powercontext \
     UV_COMPILE_BYTECODE=1 \
     UV_LINK_MODE=copy
 
@@ -19,4 +29,6 @@ COPY . .
 RUN --mount=type=cache,id=powercontext-bub-e2e-uv,target=/root/.cache/uv \
     uv sync --project e2e/bub --frozen --no-dev --no-editable
 
-ENTRYPOINT ["powercontext-e2e"]
+COPY --chmod=0755 e2e/bub/container-entrypoint.sh /usr/local/bin/powercontext-e2e-container
+
+ENTRYPOINT ["powercontext-e2e-container"]