feat(ci): two-phase update — haiku for deps, sonnet escalation for test fixes

Author: jdalton

.github/workflows/weekly-update.yml

@@ -94,14 +94,13 @@ jobs:
         with:
           gpg-private-key: ${{ secrets.BOT_GPG_PRIVATE_KEY }}
 
-      - name: Run updating skill with Claude Code
-        id: claude
-        timeout-minutes: 15
+      - name: Update dependencies (haiku — fast, cheap)
+        id: update
+        timeout-minutes: 10
         env:
           ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           GITHUB_ACTIONS: 'true'
         run: |
-          # Wrap pnpm through Socket firewall for all subprocesses (not just this shell).
           if [ -n "$SFW_BIN" ]; then
             mkdir -p /tmp/sfw-bin
             printf '#!/bin/bash\nexec "%s" pnpm "$@"\n' "$SFW_BIN" > /tmp/sfw-bin/pnpm
@@ -117,9 +116,9 @@ jobs:
 
           set +e
           pnpm exec claude --print \
-            --allowedTools "Bash(pnpm:*)" "Bash(git:*)" "Read" "Write" "Edit" "Glob" "Grep" \
-            --model sonnet \
-            --max-turns 25 \
+            --allowedTools "Bash(pnpm:*)" "Bash(git add:*)" "Bash(git commit:*)" "Bash(git status:*)" "Bash(git diff:*)" "Bash(git log:*)" "Bash(git rev-parse:*)" "Read" "Write" "Edit" "Glob" "Grep" \
+            --model haiku \
+            --max-turns 15 \
             "$(cat <<'PROMPT'
           /updating
 
@@ -132,7 +131,7 @@ jobs:
           Update all dependencies to their latest versions.
           Create one atomic commit per dependency update with a conventional commit message.
           Leave all changes local — the workflow handles pushing and PR creation.
-          Skip running builds, tests, and type checks — CI runs those separately.
+          Do not run builds or tests — the next step handles that.
           </instructions>
 
           <success_criteria>
@@ -142,7 +141,7 @@ jobs:
           </success_criteria>
           PROMPT
           )" \
-            2>&1 | tee claude-output.log
+            2>&1 | tee claude-update.log
           CLAUDE_EXIT=${PIPESTATUS[0]}
           set -e
 
@@ -152,15 +151,112 @@ jobs:
             echo "success=false" >> $GITHUB_OUTPUT
           fi
 
+      - name: Run tests
+        id: tests
+        if: steps.update.outputs.success == 'true'
+        run: |
+          if [ -n "$SFW_BIN" ]; then
+            mkdir -p /tmp/sfw-bin
+            printf '#!/bin/bash\nexec "%s" pnpm "$@"\n' "$SFW_BIN" > /tmp/sfw-bin/pnpm
+            chmod +x /tmp/sfw-bin/pnpm
+            export PATH="/tmp/sfw-bin:$PATH"
+          fi
+
+          set +e
+          pnpm build 2>&1 | tee build.log
+          BUILD_EXIT=${PIPESTATUS[0]}
+
+          pnpm test 2>&1 | tee test.log
+          TEST_EXIT=${PIPESTATUS[0]}
+          set -e
+
+          if [ "$BUILD_EXIT" -eq 0 ] && [ "$TEST_EXIT" -eq 0 ]; then
+            echo "tests-passed=true" >> $GITHUB_OUTPUT
+          else
+            echo "tests-passed=false" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Fix test failures (sonnet — smarter, escalated)
+        id: claude
+        if: steps.update.outputs.success == 'true' && steps.tests.outputs.tests-passed == 'false'
+        timeout-minutes: 15
+        env:
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          GITHUB_ACTIONS: 'true'
+        run: |
+          if [ -n "$SFW_BIN" ]; then
+            mkdir -p /tmp/sfw-bin
+            printf '#!/bin/bash\nexec "%s" pnpm "$@"\n' "$SFW_BIN" > /tmp/sfw-bin/pnpm
+            chmod +x /tmp/sfw-bin/pnpm
+            export PATH="/tmp/sfw-bin:$PATH"
+          fi
+
+          FAILURE_LOG="$(cat build.log test.log 2>/dev/null)"
+
+          set +e
+          pnpm exec claude --print \
+            --allowedTools "Bash(pnpm:*)" "Bash(git add:*)" "Bash(git commit:*)" "Bash(git status:*)" "Bash(git diff:*)" "Bash(git log:*)" "Bash(git rev-parse:*)" "Read" "Write" "Edit" "Glob" "Grep" \
+            --model sonnet \
+            --max-turns 25 \
+            "$(cat <<PROMPT
+          <context>
+          You are an automated CI agent in a weekly dependency update workflow.
+          Git is configured with GPG signing. A branch has been created for you.
+          Dependencies were updated in the previous step but build/tests failed.
+          </context>
+
+          <failure_log>
+          $FAILURE_LOG
+          </failure_log>
+
+          <instructions>
+          The dependency updates above caused build or test failures.
+          Diagnose the failures from the logs and fix the code so it builds and tests pass.
+          Create one atomic commit per fix with a conventional commit message.
+          Run pnpm build && pnpm test to verify your fixes.
+          Leave all changes local — the workflow handles pushing and PR creation.
+          </instructions>
+
+          <success_criteria>
+          pnpm build succeeds.
+          pnpm test succeeds.
+          Each fix has its own commit.
+          No uncommitted changes remain in the working tree.
+          </success_criteria>
+          PROMPT
+          )" \
+            2>&1 | tee claude-fix.log
+          CLAUDE_EXIT=${PIPESTATUS[0]}
+          set -e
+
+          if [ "$CLAUDE_EXIT" -eq 0 ]; then
+            echo "success=true" >> $GITHUB_OUTPUT
+          else
+            echo "success=false" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Set final status
+        id: final
+        if: always()
+        run: |
+          if [ "${{ steps.update.outputs.success }}" = "true" ] && [ "${{ steps.tests.outputs.tests-passed }}" = "true" ]; then
+            echo "success=true" >> $GITHUB_OUTPUT
+          elif [ "${{ steps.update.outputs.success }}" = "true" ] && [ "${{ steps.claude.outputs.success }}" = "true" ]; then
+            echo "success=true" >> $GITHUB_OUTPUT
+          else
+            echo "success=false" >> $GITHUB_OUTPUT
+          fi
+
       - name: Validate changes
         id: validate
-        if: steps.claude.outputs.success == 'true'
+        if: steps.final.outputs.success == 'true'
         run: |
-          # Only allow changes to dependency-related files.
           UNEXPECTED=""
           for file in $(git diff --name-only origin/main..HEAD); do
             case "$file" in
               package.json|*/package.json|pnpm-lock.yaml|*/pnpm-lock.yaml|.npmrc|pnpm-workspace.yaml) ;;
+              src/*|test/*) ;;
+              *.ts|*.mts|*.js|*.mjs) ;;
               *) UNEXPECTED="$UNEXPECTED $file" ;;
             esac
           done
@@ -181,13 +277,13 @@ jobs:
           fi
 
       - name: Push branch
-        if: steps.claude.outputs.success == 'true' && steps.validate.outputs.valid == 'true' && steps.changes.outputs.has-changes == 'true'
+        if: steps.final.outputs.success == 'true' && steps.validate.outputs.valid == 'true' && steps.changes.outputs.has-changes == 'true'
         env:
           BRANCH_NAME: ${{ steps.branch.outputs.branch }}
         run: git push origin "$BRANCH_NAME"
 
       - name: Create Pull Request
-        if: steps.claude.outputs.success == 'true' && steps.validate.outputs.valid == 'true' && steps.changes.outputs.has-changes == 'true'
+        if: steps.final.outputs.success == 'true' && steps.validate.outputs.valid == 'true' && steps.changes.outputs.has-changes == 'true'
         env:
           GH_TOKEN: ${{ github.token }}
           BRANCH_NAME: ${{ steps.branch.outputs.branch }}
@@ -216,7 +312,7 @@ jobs:
             --base main
 
       - name: Add job summary
-        if: steps.claude.outputs.success == 'true' && steps.validate.outputs.valid == 'true' && steps.changes.outputs.has-changes == 'true'
+        if: steps.final.outputs.success == 'true' && steps.validate.outputs.valid == 'true' && steps.changes.outputs.has-changes == 'true'
         env:
           BRANCH_NAME: ${{ steps.branch.outputs.branch }}
         run: |
@@ -231,7 +327,11 @@ jobs:
         uses: actions/upload-artifact@bbbca2ddaa5d8feaa63e36b76fdaad77386f024f # v7.0.0
         with:
           name: claude-output-${{ github.run_id }}
-          path: claude-output.log
+          path: |
+            claude-update.log
+            claude-fix.log
+            build.log
+            test.log
           retention-days: 7
 
       - uses: SocketDev/socket-registry/.github/actions/cleanup-git-signing@6096b06b1790f411714c89c40f72aade2eeaab7c # main