GoogleCloudPlatform · wangauone · Jul 19, 2026 · Jul 20, 2026 · Jul 20, 2026 · Jul 20, 2026
diff --git a/.ci/build.sh b/.ci/build.sh
@@ -58,9 +58,8 @@ mv evalbench staging/skills/autoctx-evaluate/scripts/
 BINARY_NAME="\${extensionPath}/google-cloud-db-context-engineering"
 TOOLBOX_NAME="\${extensionPath}/skills/autoctx-init/scripts/toolbox"
 
-jq ".contextFileName = \"GEMINI.md\" | .mcpServers.mcp_db_context_engineering.command = \"$BINARY_NAME\" | .mcpServers.mcp_db_context_engineering.args = [] | .mcpServers.mcp_toolbox = {\"command\": \"$TOOLBOX_NAME\", \"args\": [\"--stdio\", \"--config\", \"autoctx/tools.yaml\"]}" plugin/gemini-extension.json > staging/gemini-extension.json
+jq ".mcpServers.mcp_db_context_engineering.command = \"$BINARY_NAME\" | .mcpServers.mcp_db_context_engineering.args = [] | .mcpServers.mcp_toolbox = {\"command\": \"$TOOLBOX_NAME\", \"args\": [\"--stdio\", \"--config\", \".context-engineering/tools.yaml\"]}" plugin/gemini-extension.json > staging/gemini-extension.json
 
-cp plugin/GEMINI.md staging/
 cp LICENSE staging/
 
 # Create tarball

diff --git a/.ci/cloudbuild.yaml b/.ci/cloudbuild.yaml
@@ -83,6 +83,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -94,6 +95,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -107,6 +109,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -118,6 +121,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -129,6 +133,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -140,6 +145,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -152,6 +158,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -164,6 +171,7 @@ steps:
     entrypoint: 'bash'
     secretEnv: ['ADC_KEY']
     waitFor: ['build-extension', 'preflight']
+    timeout: 2700s
     env:
       - 'EVAL_GCP_PROJECT_ID=$PROJECT_ID'
       - 'EVAL_GCP_PROJECT_REGION=$_PROJECT_LOCATION'
@@ -280,3 +288,7 @@ substitutions:
 
 options:
   defaultLogsBucketBehavior: REGIONAL_USER_OWNED_BUCKET
+
+# Steps run in parallel; overall timeout must cover the slowest step
+# (45 min) plus build-extension + preflight + upload + gate overhead.
+timeout: 3600s
diff --git a/.ci/run_gemini_cli.sh b/.ci/run_gemini_cli.sh
@@ -32,15 +32,15 @@ cp -r "/workspace/evals/${SUITE}" "${WORK_DIR}/"
 cp -r "/workspace/evals/model_configs" "${WORK_DIR}/"
 cd "${WORK_DIR}"
 
-# Point the Gemini CLI extension installer at the locally built extension
-# instead of pulling from GitHub.
-sed -i 's|https://github.com/GoogleCloudPlatform/db-context-enrichment|/workspace/staging|g' "model_configs/gemini_cli_model.yaml"
+# Point the Gemini CLI extension installer at the locally built extension.
+sed -i "s|<extension-source>|/workspace/staging|g" "model_configs/gemini_cli_model.yaml"
 
-# Inject Vertex project/location into the CLI env block so the extension can
-# talk to the right GCP project without the values being committed to the repo.
+# Substitute Vertex project/location placeholders in the CLI env block so the
+# extension can talk to the right GCP project without the values being
+# committed to the repo.
 sed -i \
-  -e "/^  GEMINI_MODEL:/a\\  GOOGLE_CLOUD_PROJECT: \"${EVAL_GCP_PROJECT_ID}\"" \
-  -e "/^  GEMINI_MODEL:/a\\  GOOGLE_CLOUD_LOCATION: \"${EVAL_GCP_PROJECT_REGION}\"" \
+  -e "s|<gcp-project>|${EVAL_GCP_PROJECT_ID}|g" \
+  -e "s|<gcp-location>|${EVAL_GCP_PROJECT_REGION}|g" \
   "model_configs/gemini_cli_model.yaml"
 
 # Append the runtime release_version and resolve the reporting-project
@@ -54,8 +54,12 @@ sed -i "s|\${EVAL_REPORTING_PROJECT}|${EVAL_REPORTING_PROJECT:-}|g" "${CONFIG}"
 export PYTHONPATH=/evalbench:/evalbench/evalproto
 export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
 
-echo "Launching ${SUT}/${SUITE} evaluation..."
-uv run --no-sync --project /evalbench python /evalbench/evalbench/evalbench.py --experiment_config="${SUITE}/run_gemini_cli.yaml"
+# Optional scenario filter: set EVAL_SCENARIOS=<id>[,<id>...] to run a subset.
+SCENARIO_ARG=""
+[ -n "${EVAL_SCENARIOS:-}" ] && SCENARIO_ARG="--scenarios=${EVAL_SCENARIOS}"
+
+echo "Launching ${SUT}/${SUITE} evaluation${SCENARIO_ARG:+ (scenarios: ${EVAL_SCENARIOS})}..."
+uv run --no-sync --project /evalbench python /evalbench/evalbench/evalbench.py --experiment_config="${SUITE}/run_gemini_cli.yaml" ${SCENARIO_ARG}
 
 echo "Validating mandatory output files for ${SUITE}..."
 python3 /workspace/.ci/check_eval_outputs.py "${WORK_DIR}/${SUITE}" "${WORK_DIR}/${SUITE}/dataset.json"

diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
@@ -40,11 +40,11 @@
 
     steps:
       # Checkout the repository code
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0
 
       # Install the `uv` Python package manager which handles fast installations
       - name: Install uv
-        uses: astral-sh/setup-uv@v5
+        uses: astral-sh/setup-uv@d4b2f3b6ecc6e67c4457f6d3e41ec42d3d0fcb86 # v5.4.2
         with:
           enable-cache: true
           cache-dependency-glob: "uv.lock"
@@ -77,7 +77,7 @@
       - name: Download mcp-toolbox binary
         shell: bash
         run: |
-          DOWNLOAD_URL="https://storage.googleapis.com/mcp-toolbox-for-databases/v${{ env.TOOLBOX_VERSION }}/${{ matrix.toolbox_os }}/${{ matrix.toolbox_arch }}/toolbox${{ matrix.toolbox_ext }}"
+          DOWNLOAD_URL="https://storage.googleapis.com/mcp-toolbox-for-databases/v${TOOLBOX_VERSION}/${{ matrix.toolbox_os }}/${{ matrix.toolbox_arch }}/toolbox${{ matrix.toolbox_ext }}"
           echo "Downloading toolbox from: ${DOWNLOAD_URL}"
           curl -L --fail -o "toolbox${{ matrix.toolbox_ext }}" "${DOWNLOAD_URL}"
           chmod +x "toolbox${{ matrix.toolbox_ext }}"
@@ -91,7 +91,7 @@
           else
             ARCHIVE="${{ matrix.platform_name }}.${{ matrix.arch_name }}.evalbench.tar.gz"
           fi
-          DOWNLOAD_URL="https://github.com/GoogleCloudPlatform/evalbench/releases/download/v${{ env.EVALBENCH_VERSION }}/${ARCHIVE}"
+          DOWNLOAD_URL="https://github.com/GoogleCloudPlatform/evalbench/releases/download/v${EVALBENCH_VERSION}/${ARCHIVE}"
           echo "Downloading evalbench from: ${DOWNLOAD_URL}"
           curl -L --fail -o "${ARCHIVE}" "${DOWNLOAD_URL}"
 
@@ -142,8 +142,7 @@
             TOOLBOX_NAME="\${extensionPath}/skills/autoctx-init/scripts/toolbox"
           fi
 
-          jq ".contextFileName = \"GEMINI.md\" | .mcpServers.mcp_db_context_engineering.command = \"$BINARY_NAME\" | .mcpServers.mcp_db_context_engineering.args = [] | .mcpServers.mcp_toolbox = {\"command\": \"$TOOLBOX_NAME\", \"args\": [\"--stdio\", \"--config\", \"autoctx/tools.yaml\"]}" plugin/gemini-extension.json > staging/gemini-extension.json
-          cp plugin/GEMINI.md staging/
+          jq ".mcpServers.mcp_db_context_engineering.command = \"$BINARY_NAME\" | .mcpServers.mcp_db_context_engineering.args = [] | .mcpServers.mcp_toolbox = {\"command\": \"$TOOLBOX_NAME\", \"args\": [\"--stdio\", \"--config\", \".context-engineering/tools.yaml\"]}" plugin/gemini-extension.json > staging/gemini-extension.json
           cp LICENSE staging/
 
       # Create a compressed tarball archive for Linux and macOS distributions
@@ -162,7 +161,7 @@
 
       # Upload the archives to GitHub Actions artifacts so the publish job can access them
       - name: Upload Build Artifacts
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4.6.2
         with:
           name: binary-${{ matrix.platform_name }}-${{ matrix.arch_name }}
           path: |
@@ -180,7 +179,7 @@
     steps:
       # Download all platform archives generated by the build-artifacts job
       - name: Download all build artifacts
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4.3.0
         with:
           path: release-assets
           pattern: binary-*
@@ -189,7 +188,7 @@
       # Attach the downloaded archives as assets to the created GitHub Release
       - name: Upload Release Assets
         if: github.event_name == 'release'
-        uses: softprops/action-gh-release@v2
+        uses: softprops/action-gh-release@3bb12739c298aeb8a4eeaf626c5b8d85266b0e65 # v2.6.2
         with:
           tag_name: ${{ github.event.release.tag_name }}
           files: release-assets/*
@@ -204,13 +203,15 @@
       contents: write
       pull-requests: write
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4.4.0
 
       # Strip leading 'v' from the release tag (e.g. v0.5.1 -> 0.5.1) since PyPI versions are unprefixed.
       # Poll PyPI for up to 6 hours (72 attempts x 5min) waiting for the wheel to be indexed.
       - name: Wait for PyPI release
+        env:
+          TAG_NAME: ${{ github.event.release.tag_name }}
         run: |
-          TAG="${{ github.event.release.tag_name }}"
+          TAG="${TAG_NAME}"
           VERSION="${TAG#v}"
           PKG="google-cloud-db-context-engineering"
           for i in {1..72}; do
@@ -229,8 +230,9 @@
       - name: Open PR to pin marketplace ref
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          TAG_NAME: ${{ github.event.release.tag_name }}
         run: |
-          TAG="${{ github.event.release.tag_name }}"
+          TAG="${TAG_NAME}"
           BRANCH="chore/pin-marketplace-${TAG}"
           jq --arg ref "$TAG" '.plugins[0].source.ref = $ref' \
             .claude-plugin/marketplace.json > tmp && mv tmp .claude-plugin/marketplace.json

diff --git a/dev-plugin/gemini-extension.json b/dev-plugin/gemini-extension.json
@@ -1,7 +1,6 @@
 {
   "name": "google-cloud-db-context-engineering-dev",
   "version": "0.0.0-dev",
-  "contextFileName": "../plugin/GEMINI.md",
   "mcpServers": {
     "db-context-engineering": {
       "command": "uv",
@@ -17,7 +16,7 @@
       "args": [
         "toolbox-server@1.4.0",
         "--config",
-        "autoctx/tools.yaml",
+        ".context-engineering/tools.yaml",
         "--stdio"
       ]
     }

diff --git a/dev-plugin/plugin/.claude-plugin/plugin.json b/dev-plugin/plugin/.claude-plugin/plugin.json
@@ -20,7 +20,7 @@
       "args": [
         "toolbox-server@1.4.0",
         "--config",
-        "autoctx/tools.yaml",
+        ".context-engineering/tools.yaml",
         "--stdio"
       ]
     }

diff --git a/dev-plugin/plugin/CLAUDE.md b/dev-plugin/plugin/CLAUDE.md
diff --git a/evals/core-cujs/dataset.json b/evals/core-cujs/dataset.json
@@ -1,7 +1,7 @@
 {
   "scenarios": [
     {
-      "id": "autoctx-init",
+      "id": "core-cujs:init",
       "starting_prompt": "setup context generation for my database.",
       "conversation_plan": "Ask the agent to setup the auto context generation for my alloydb database. Connection information: datasource name as my-alloydb, project id is cloud-db-nl2sql, region us-central1, cluster whaoyu-test, instance whaoyu-test-primary, database financial. Verify the agent generate a valid tools.yaml. Explicitly ask the agent to perform validation. You should terminate the conversation immediately after the agent successfully set up the files and performed the validation.",
       "expected_trajectory": [],
@@ -10,7 +10,7 @@
       "max_turns": 5
     },
     {
-      "id": "autoctx-dataset-generation",
+      "id": "core-cujs:dataset-generation",
       "starting_prompt": "Use context engineering, generate an evaluation dataset consisting of 5 questions for my alloydb database financial named golden.json, based on my seed dataset to be provided.",
       "conversation_plan": "Ask the agent to generate an evaluation dataset for my alloydb database financial named golden.json. The 'seed' includes the following two golden NL-SQL pair:\n\nHow many accounts who choose issuance after transaction are staying in East Bohemia region? A3 contains the data of region; 'POPLATEK PO OBRATU' represents for 'issuance after transaction'.\nSELECT COUNT(DISTINCT \"T1\".\"account_id\") FROM \"account\" AS \"T1\" INNER JOIN \"district\" AS \"T2\" ON \"T1\".\"district_id\" = \"T2\".\"district_id\" WHERE \"T2\".\"A3\" = 'east Bohemia' AND \"T1\".\"frequency\" = 'POPLATEK PO OBRATU'\n\nHow many accounts who have region in Prague are eligible for loans? A3 contains the data of region\nSELECT COUNT(\"T1\".\"account_id\") FROM \"account\" AS \"T1\" INNER JOIN \"loan\" AS \"T2\" ON \"T1\".\"account_id\" = \"T2\".\"account_id\" INNER JOIN \"district\" AS \"T3\" ON \"T1\".\"district_id\" = \"T3\".\"district_id\" WHERE \"T3\".\"A3\" = 'Prague'. If prompted for validating the query, reply yes to the agent. Once the initial golden dataset is generated, ask the agent to expand it with 2 additional examples in the same file. You should terminate the conversation immediately after the agent successfully generated the dataset, even if the agent prompts you to perform evaluation or take next steps.",
       "expected_trajectory": [],
@@ -19,40 +19,49 @@
       "max_turns": 5
     },
     {
-      "id": "autoctx-dataset-expansion",
-      "starting_prompt": "Use context engineering, generate an evaluation dataset consisting of 5 questions by expanding my evaluation dataset golden.json with more diverse examples using variation strategies.",
+      "id": "core-cujs:dataset-expansion",
+      "starting_prompt": "Use context engineering, generate an evaluation dataset consisting of 5 questions by expanding my evaluation dataset golden.json with more diverse examples using variation strategies. Name the audit report files 'evalset_expansion_report_pair_level.md' and 'evalset_expansion_report_dataset_level.md'.",
       "conversation_plan": "Ask the agent to expand the existing golden.json evaluation dataset by applying variation strategies to the existing pairs. The agent should apply all six strategies: paraphrasing, merging, difficulty adjustment (both upscaling and simplification), distraction injection, linguistic variation (typos and synonyms), and value substitution (which requires running SELECT DISTINCT queries on the database). When the agent presents an expansion plan, confirm it and approve all strategies. When the agent runs SQL execution validation or asks for approval of the validation reports, approve them. You should terminate the conversation immediately after the agent successfully writes the expanded pairs to golden.json, even if the agent prompts you to perform evaluation or take next steps. At the end of the conversation, verify that the agent produced the following output files: golden.json, evalset_expansion_report_pair_level.md, evalset_expansion_report_dataset_level.md.",
       "expected_trajectory": [],
       "kind": "agents",
       "work_dir": "workspace_post_dataset_generation/",
       "max_turns": 8
     },
     {
-      "id": "autoctx-bootstrap",
+      "id": "core-cujs:bootstrap",
       "starting_prompt": "create initial context set for my alloydb database",
-      "conversation_plan": "Ask the agent to create initial context set for my alloydb database. Need to have at least 2 templates and 2 facets. If prompted for experiment name, use 'my-alloydb-tuning-experiment'. You should terminate the conversation immediately after the agent successfully created the context set, even if the agent prompts you to perform evaluation or take next steps.",
+      "conversation_plan": "Ask the agent to create initial context set for my alloydb database. Need to have at least 2 templates and 2 facets. If prompted for experiment name, use 'crema-eval-experiment'. You should terminate the conversation immediately after the agent successfully created the context set, even if the agent prompts you to perform evaluation or take next steps.",
       "expected_trajectory": [],
       "kind": "agents",
       "work_dir": "workspace_post_expansion/",
       "max_turns": 5
     },
     {
-      "id": "autoctx-evaluate",
-      "starting_prompt": "evaluate on the golden dataset golden.json with context set id projects/cloud-db-nl2sql/locations/us-east1/contextSets/whaoyu-eval-contextset",
-      "conversation_plan": "Ask the agent to evaluate the golden dataset golden.json, using the current context set with context set id projects/cloud-db-nl2sql/locations/us-east1/contextSets/whaoyu-eval-contextset. If prompted for experiment name, use 'my-alloydb-tuning-experiment'. You should terminate the conversation immediately after the agent successfully evaluated the dataset, even if the agent prompts you to perform context improving task or take next steps.",
+      "id": "core-cujs:evaluate",
+      "starting_prompt": "evaluate the golden dataset at ./golden.json against the local ContextSet at ./bootstrap_context.json. Use project 'cloud-db-nl2sql', experiment 'crema-eval-experiment', cs_id 'autoctx', version 'v0'.",
+      "conversation_plan": "Ask the agent to evaluate the golden dataset at ./golden.json against the local ContextSet at ./bootstrap_context.json. If prompted for Context Store coordinates, use project_id=cloud-db-nl2sql, csg_id=crema-eval-experiment, cs_id=autoctx, version=v0. Terminate immediately after the agent presents evaluation results, even if it prompts to continue.",
       "expected_trajectory": [],
       "kind": "agents",
       "work_dir": "workspace_post_bootstrap/",
       "max_turns": 5
     },
     {
-      "id": "autoctx-hillclimb",
-      "starting_prompt": "improve the context set based on eval failure",
-      "conversation_plan": "Ask the agent to perform the hillclimbing to improve the context set based on eval failure. If prompted for experiment name, use 'my-alloydb-tuning-experiment'. You should terminate the conversation immediately after the agent successfully performed the hillclimbing task, even if the agent prompts you to perform evaluation or take next steps.",
+      "id": "core-cujs:hillclimb-continue",
+      "starting_prompt": "improve the context set based on eval failure. Resume the in-progress experiment 'crema-eval-experiment' at .context-engineering/experiments/crema-eval-experiment/ (already has a v0 baseline). Use project 'cloud-db-nl2sql'.",
+      "conversation_plan": "Ask the agent to perform the hillclimbing to improve the context set based on eval failure. The workspace already contains an in-progress experiment at .context-engineering/experiments/crema-eval-experiment/ with a v0 baseline. If prompted for experiment name, use 'crema-eval-experiment'. You should terminate the conversation immediately after the agent successfully performed the hillclimbing task, even if the agent prompts you to perform evaluation or take next steps.",
       "expected_trajectory": [],
       "kind": "agents",
       "work_dir": "workspace_post_evaluation/",
       "max_turns": 5
+    },
+    {
+      "id": "core-cujs:hillclimb-fresh",
+      "starting_prompt": "improve my context set from scratch using the local base context at ./bootstrap_context.json and the golden dataset at ./golden.json. Use experiment name 'crema-eval-experiment' and project 'cloud-db-nl2sql'.",
+      "conversation_plan": "Ask the agent to run hillclimb using the local base context at ./bootstrap_context.json and the golden dataset at ./golden.json. If prompted for experiment name, use 'crema-eval-experiment'. Terminate immediately after the agent successfully performs the hillclimbing task, even if the agent prompts you to continue.",
+      "expected_trajectory": [],
+      "kind": "agents",
+      "work_dir": "workspace_fresh_hillclimb/",
+      "max_turns": 5
     }
   ]
 }
diff --git a/...rkspace_post_bootstrap/autoctx/tools.yaml → ...hillclimb/.context-engineering/tools.yaml b/...rkspace_post_bootstrap/autoctx/tools.yaml → ...hillclimb/.context-engineering/tools.yaml
diff --git a/...-tuning-experiment/bootstrap_context.json → ...ce_fresh_hillclimb/bootstrap_context.json b/...-tuning-experiment/bootstrap_context.json → ...ce_fresh_hillclimb/bootstrap_context.json